P1: JZP

0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

This page intentionally left blank

P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

INDUCTIVE REASONING

Without inductive reasoning, we couldn’t generalize from one instance to an-

other, derive scientific hypotheses, or predict that the sun will rise again tomor-
row morning. Despite the prevalence of inductive reasoning, books on this topic
are rare. Indeed, this is the first book on the psychology of inductive reasoning
in twenty years. The chapters survey recent advances in the study of inductive
reasoning and address questions about how it develops, the role of knowledge in
induction, how best to model people’s reasoning, and how induction relates to
other forms of thinking. Written by experts in philosophy, developmental science,
cognitive psychology, and computational modeling, the contributions here will
be of interest to a general cognitive science audience as well as to those with a
more specialized interest in the study of thinking.

Aidan Feeney is Senior Lecturer in Psychology at Durham University. He received

his B.A. in psychology from Trinity College, Dublin, in 1992 and completed his
Ph.D. at the Center for Thinking and Language at the University of Plymouth. He
was appointed Lecturer in the Department of Psychology at Durham University
in 1998 and became Senior Lecturer in 2005. Dr. Feeney’s research has been sup-
ported by a number of grants from the Economic and Social Research Council
(UK). He has published approximately thirty journal articles, book chapters, and
papers on the psychology of hypothesis testing and reasoning.

Evan Heit is currently Professor of Psychology and Cognitive Science, and Found-
ing Faculty, at the University of California, Merced. Previously, Dr. Heit was on
the faculty in the Psychology Department of the University of Warwick, UK. He
has undergraduate degrees in computer science and psychology from the Univer-
sity of Pennsylvania and a Ph.D. from Stanford University. He also carried out
postdoctoral research at the University of Michigan and Northwestern University.
Professor Heit has published more than fifty papers on the psychology of reason-
ing, memory, and categorization. His research has been funded by the National
Science Foundation, the National Institutes of Health, the Economic and Social
Research Council (UK), and the Biotechnology and Biological Sciences Research
Council (UK). He is currently on the editorial board of Memory and Cognition
and the Journal of Experimental Psychology: Learning, Memory, and Cognition and
is Associate Editor of the Journal of Memory and Language.
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

Inductive Reasoning
Experimental, Developmental, and
Computational Approaches

Edited by
AIDAN FEENEY
Durham University

EVAN HEIT
University of California, Merced
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521856485

© Cambridge University Press 2007

This publication is in copyright. Subject to statutory exception and to the provision of

relevant collective licensing agreements, no reproduction of any part may take place
without the written permission of Cambridge University Press.
First published in print format 2007

ISBN-13 978-0-511-34179-3 eBook (EBL)

ISBN-10 0-511-34179-2 eBook (EBL)

ISBN-13 978-0-521-85648-5 hardback

ISBN-10 0-521-85648-5 hardback

ISBN-13 978-0-521-67244-3 paperback

ISBN-10 0-521-67244-9 paperback

Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

Contents

List of Figures page vii

List of Tables ix
List of Contributors xi
Preface xiii

1 What Is Induction and Why Study It? 1

Evan Heit
2 The Development of Inductive Reasoning 25
Brett K. Hayes
3 Interpreting Asymmetries of Projection in Children’s
Inductive Reasoning 55
Douglas L. Medin and Sandra Waxman
4 Property Generalization as Causal Reasoning 81
Bob Rehder
5 Availability in Category-Based Induction 114
Patrick Shafto, John D. Coley, and Anna Vitkin
6 From Similarity to Chance 137
Sergey Blok, Daniel Osherson, and Douglas L. Medin
7 Theory-Based Bayesian Models of Inductive Reasoning 167
Joshua B. Tenenbaum, Charles Kemp, and Patrick Shafto
8 Use of Single or Multiple Categories in Category-Based
Induction 205
Gregory L. Murphy and Brian H. Ross

v
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

vi Contents

9 Abductive Inference: From Philosophical Analysis to Neural

Mechanisms 226
Paul Thagard
10 Mathematical Induction and Induction in Mathematics 248
Lance J. Rips and Jennifer Asmuth
11 Induction, Deduction, and Argument Strength in Human
Reasoning and Argumentation 269
Mike Oaksford and Ulrike Hahn
12 Individual Differences, Dual Processes, and Induction 302
Aidan Feeney
13 Taxonomizing Induction 328
Steven A. Sloman

Index 345
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

List of Figures

1.1 Criterion-shift account of deduction and induction. page 9

2.1 Schematic stimulus example from Thompson and Hayes
(2005). 39
2.2 Mean proportion of causal choices on the Thompson and
Hayes (2005) causal induction task. 40
4.1 Results from Hadjichristidis et al. (2004), Experiment 2. 88
4.2 Results from Rehder (in press), Experiment 3. 90
4.3 Results from Nair (2005). 93
4.4 Results from Rehder (in press), Experiment 1. 99
4.5 Results from Rehder (in press), Experiment 3. 101
4.6 Network topologies tested in Rehder and Hastie (2004),
Experiment 2. 104
4.7 Generalization ratings from Rehder and Hastie (2004),
Experiment 1. 106
6.1 Predicted versus observed probabilities for the 200 arguments. 149
6.2 Predicted versus observed probabilities for thirty-two
arguments. 150
6.3 Predicted versus observed probabilities for the four sets of
stimuli. 162
7.1 (a) A folk taxonomy of mammal species. (b–e) Examples of
mutation histories. 180
7.2 One simulated sample from the causal-transmission model,
for the foodweb shown in Figure 7.4a. 185
7.3 Comparing models of property induction with human
judgments, for reasoning in a default biological context with
generic anatomical or physiological properties. 190
7.4 Multiple relational structures over the same domains of
species. 195
vii
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

viii List of Figures

7.5 Comparing models of induction with human judgments, for

two kinds of properties: disease properties and genetic
properties. 196
8.1 An example similar to the categories used in Murphy and Ross
(1994), illustrating the main comparison of that set of studies. 210
9.1 The process of abductive inference. 229
9.2 Abduction as a neural process. 234
10.1 Stimulus problems from a study of geometry problem solving
(adapted from Koedinger & Anderson, 1991, Figure 2). 262
10.2 Stimulus problem for a study of non-Euclidean geometry. 263
12.1 Interaction between Ability and Conclusion Type from
Feeney, Shafto, and Dunning (in press). 321
12.2 Means involved in the interaction between Ability and Validity
from Feeney, Dunning, and Over (in submission). 323
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

List of Tables

2.1 Summary of major category-based induction phenomena in

infants and children. page 42
3.1 Ross et al. (2003) induction probabilities, indicating
asymmetries. 70
4.1 Features and causal relationship for Romanian Rogos, an
artifical category. 92
4.2 Category structure from Nair (2005), Experiment 2, consisting
of eight Rogos (R1–8) and eight non-Rogos (NR1–8). 94
5.1 Sample items from adult open-ended projection study. 122
6.1 Distances (in millimeters) in Figure 1 of Rips, Shoben, &
Smith (1973). 158
6.2 The 36 complex statements figuring in the experiment. 159
6.3 Average probabilities assigned to variables. 160
6.4 Correlations between predicted and observed probabilities. 161
9.1 Abductive inference in five domains, specifying what needs to
be explained and the kinds of hypotheses that provide
explanations. 227
12.1 Examples of the materials used by Rips (2001) to demonstrate
a dissociation between induction and deduction. 322

ix
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

List of Contributors

Jennifer Asmuth Department of Psychology, Northwestern University, 2029

Sheridan Road, Evanston, IL 60208-2710, USA
Sergey Blok Department of Psychology, 1 University Station A8000, Univer-
sity of Texas at Austin, Austin, TX 78712, USA
John D. Coley Department of Psychology, MS 0125 NI, Northeastern Uni-
versity, 360 Huntington Avenue, Boston, MA 02115-5000, USA
Aidan Feeney Department of Psychology, Durham University, Queen’s Cam-
pus, Thornaby, Stockton-on-Tees, TS17 6BH, UK
Ulrike Hahn School of Psychology, Cardiff University, Tower Building, Park
Place, Cardiff, CF10 3AT, UK
Brett K. Hayes School of Psychology, University of New South Wales, Sydney,
NSW, Australia, 2052
Evan Heit School of Social Sciences, Humanities, and Arts, University of
California, Merced, PO Box 2039, Merced, CA 95344, USA
Charles Kemp Department of Brain and Behavioral Sciences, 46-4053, Mas-
sachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge,
MA 02139, USA
Douglas L. Medin Department of Psychology, Northwestern University, Swift
Hall, 2029 Sherdian Road, Evanston, IL 60208-2710, USA
Gregory L. Murphy Department of Psychology, New York University, 6
Washington Place, New York, NY 10003, USA
Mike Oaksford School of Psychology, Birkbeck College London, Malet Street,
London, WC1E 7HX, UK
xi
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

xii List of Contributors

Daniel Osherson Department of Psychology, Green Hall 3-S-12, Princeton

University, Princeton, NJ 08544, USA
Bob Rehder Department of Psychology, New York University, 6 Washington
Place, New York, NY 10003, USA
Lance J. Rips Psychology Department, Northwestern University, 2029 Sheri-
dan Road, Evanston, IL 60208, USA
Brian H. Ross Beckman Institute, University of Illinois, 405 N. Mathews Ave.,
Urbana, IL 61801, USA
Patrick Shafto Department of Brain and Behavioral Sciences, 46-4053, Mas-
sachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA
02139, USA
Steven Sloman Cognitive & Linguistic Sciences, Brown University, Box 1978,
Providence, RI 02912, USA
Joshua B. Tenenbaum Department of Brain and Behavioral Sciences, 46-4015,
Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge,
MA 02139-4307, USA
Paul Thagard Philosophy Department, University of Waterloo, Waterloo,
Ontario, Canada N2L 3G1
Anna Vitkin Department of Psychology, MS 0125 NI, Northeastern Univer-
sity, 360 Huntington Avenue, Boston, MA 02115-5000, USA
Sandra Waxman Department of Psychology, Northwestern University, Swift
Hall, 2029 Sherdian Road, Evanston, IL 60208-2710, USA
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

Preface

Books on induction are rare; in fact, you are holding in your hands the first
book devoted solely to the psychology of inductive reasoning in twenty years.
And yet induction is a central topic in cognitive science, fundamental to
learning, discovery, and prediction. We make inductive inferences every time
we use what we already know to deal with novel situations. For example,
wherever you found this book – in your university library, online, or in an
academic bookshop – before picking it from the shelf or clicking on a link,
you will have made some inductive inferences. Amongst other things, these
inferences will have concerned the book’s likely content given its title, its
editors, its publisher, or its cover. Common to all of these inferences will have
been the use of what you already know – about us, or the topic suggested
by our title, or Cambridge University Press – to make predictions about the
likely attributes of this book.
It is not only in publishers’ catalogues that attention to induction has
been scant. Despite its obvious centrality to an understanding of human
cognition and behaviour, much less work has been carried out by psychologists
on induction than on deduction, or logical reasoning. As a consequence,
although there have been several edited collections of work on deduction
and on decision making (Connolly, Arkes, & Hammond, 2000; Leighton &
Sternberg, 2004; Gilovich, Griffin, & Kahneman, 2002; Manktelow & Chung,
2004), to the best of our knowledge there has never previously been an edited
collection of papers on the psychology of inductive inference.
To attempt to redress the balance somewhat, in 2004 we organised a sympo-
sium on induction at the Fifth International Conference on Thinking, which
was held in the historical setting of the University of Leuven in Belgium.
The series of international conferences on thinking, of which the meeting
in Leuven was part, has been primarily concerned with deductive inference,
problem solving, and decision making. Our symposium was intended to raise
xiii
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

xiv Preface

the profile, in Europe particularly, of work on induction. Many, but not all, of
the chapter authors for this book talked at the symposium, which took place
on Saturday 24 July. Doug Medin, Pat Shafto, Paul Thagard, Brett Hayes, Josh
Tenenbaum, Evan Heit, and Aidan Feeney all talked about their work on in-
duction, and Steven Sloman was the discussant. John Coley, Mike Oaksford,
and Bob Rehder were also present and contributed greatly on the day, and
Lance Rips had also been vesting Leuven. The symposium went so well that
we decided to press on with plans for a book. But what kind of book should
it be?
The last book devoted wholly to the psychology of induction was Hol-
land, Holyoak, Nisbett, and Thagard’s landmark book on induction, which
appeared in 1986. Broadly speaking, Holland and colleagues’ book attempted
to apply just one general explanatory framework to a variety of inductive
phenomena including analogy, scientific discovery, generalisation, and rule
learning. Since 1986 the field has changed and developed enormously. There
have been great accomplishments in the study of how inductive reasoning
develops, as well as a focus on precise mathematical modelling of induc-
tion throughout the 1990s, culminating in very powerful Bayesian models
of inductive thinking that have emerged over the last few years. Other de-
velopments include a recent focus on the relationship between induction
and deduction, widespread consideration of the role of categories and back-
ground knowledge in induction, and significant progress on questions about
individual and cultural differences in inductive generalisation.
To emphasise how the field has changed since the mid-1980s, and because
of the range of work on induction that has emerged in the intervening time,
this book reverses the tactic employed by Holland and colleagues; instead of
applying one approach to a variety of inductive phenomena, for the most part
this book will focus on how a wide range of methodological and modelling
approaches have increased our understanding of induction.
There is often a concern that edited books can lack coherence. Given that
we wished to collect contributions on a wide range of approaches, this could
have been a problem here. However, as most chapter authors presented at,
or attended, the symposium in Leuven, we were all aware while writing our
chapters of the concerns of other chapter authors. In addition, every chapter
was read by two other chapter authors and every chapter was redrafted in the
light of other authors’ comments. Our goal throughout has been to achieve
integration and overlap between chapters, and we hope that we have achieved
coherence.
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

Preface xv

    

This book is intended to be read by anyone interested in thinking. Our goal
was to capture the current state of the art in our field. Thus all readers – be they
psychologists, mathematical modellers, cognitive scientists, or philosophers –
with even a passing interest in induction should find something to hold
their attention here. Although we think that postgraduate students might be
interested in most chapters, we also hope that undergraduates will find useful
material. Whilst some of the chapters in this book focus in great detail on
a very specific issue, others give very accessible overviews of big swathes of
the literature. Although editors should probably refrain from favouring some
contributions over others, because we think that all of the chapters in this
book are excellent, we are happy to particularly recommend Chapters 1, 8,
and 9 to undergraduates exploring the psychology of induction for the first
time.
We also hope that many of our colleagues researching other forms of think-
ing such as deduction, decision making, analogy, and problem solving will
be interested in this book. Whereas some of the issues, particularly those
concerning development, culture, and individual differences, are likely to be
similar across many forms of thinking, others are not. The particular issue that
is distinctive to induction is that of knowledge; induction involves reasoning
from all relevant belief. Whereas in some forms of reasoning it is an error to
consider one’s own beliefs, for inductive reasoning it is a necessity. Accord-
ingly, to explain induction one has to consider how knowledge is structured,
which type of knowledge is most relevant when, and how that knowledge is
brought to bear on the premises of inductive arguments. Although these are
not issues the literature on deduction, for example, has always highlighted,
nevertheless we think that consideration of the role of different types of
knowledge in reasoning might be very beneficial for researchers interested in
deduction.
Finally, we think that Holland and colleagues were correct in suggesting that
questions to do with induction are central to the study of cognitive science.
How agents reuse previous experience stored in the form of knowledge is a key
question for anyone interested in cognition. The diverse perspectives offered
in this book should therefore be of interest to a wide range of researchers.
Furthermore, we have tried very hard in this book to ensure coverage of
molehills and mountains: molehills for the specialists and mountains for the
general cognitive science reader.
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

xvi Preface

   

When we drew up our list of invitees, the most important consideration was
that we include a coherent set of contributions reflecting a range of approaches
and methods important in contemporary cognitive science. These approaches
are summarised in Chapter 1 by Evan Heit. Following a discussion of the dif-
ferences between induction and deduction, the chapter uses the phenomenon
of evidential diversity to explore approaches from philosophy, experimental
psychology, mathematical modelling, and developmental science. After these
introductory chapters, the book contains a further four sections, which we
will briefly summarise.
The centrality of induction to cognitive development, and the centrality
of the developmental approach to research on induction, can be seen in the
chapters by Hayes and by Medin and Waxman. In Chapter 2 Brett Hayes
provides a comprehensive overview of our current understanding of how in-
ductive reasoning develops. He discusses the information that children use
for inductive reasoning and presents the development of induction as be-
ing about the development of knowledge. Hayes also describes the uses to
which developmental data may be put when evaluating theories of induction.
Chapter 3, written by Doug Medin and Sandra Waxman, nicely illustrates how
developmental data can constrain theory. Medin and Waxman concern them-
selves with asymmetries (preferences for arguments from certain categories
over arguments from other categories) in children’s reasoning. In contrast to
Chapters 1 and 2, the chapter by Medin and Waxman is tightly focused on
a single issue. However, by considering a range of possible explanations for
asymmetries in the light of novel developmental evidence, some of which is
cross-cultural, the authors are able to come down in favour of one account
over all others.
The next three chapters all concern knowledge and induction. In Chapter 4,
Bob Rehder makes a strong case for the importance of causal knowledge in
induction. This chapter describes evidence that causal knowledge about a to-
be-projected property affects its projectability, and it discusses the relationship
between knowledge about similarity or taxonomic relations and knowledge
about causal relations. Whereas the chapter by Rehder considers cases in which
the two forms of knowledge are in competition, Chapter 5, contributed by
Pat Shafto, John Coley, and Anna Vitkin, considers the role of availability
in determining whether taxonomic or ecological knowledge dominates in
induction. Shafto and colleagues consider two factors that might affect which
type of knowledge is, or becomes, most available: context and experience.
This framework proves a neat way of organising recent findings concerning
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

Preface xvii

the effects of content, property, expertise, and culture on inductive reasoning.

Together these chapters demonstrate how far the field has moved beyond
original attempts to model inductive inference using notions such as typicality
and similarity.
Of course, similarity remains an extremely important construct in the liter-
ature on induction, and many inductive inferences will be based on similarity.
One of the problems for models that have exploited similarity as an explana-
tory mechanism is the property that people are asked to project from members
of one category to another. Typically, these models do not take account of the
meaning of properties. In Chapter 6, Blok, Osherson, and Medin describe and
experimentally verify a mathematical model of induction that takes as input
(a) the similarity between the categories in the argument and (b) premise and
conclusion probabilities. As the nature of the property determines the prob-
ability of the statements in an argument, this model can capture results for
meaningful properties. Thus, it places similarity-based accounts on stronger
footing.
The chapter by Blok and colleagues comes out of the very rich tradition of
mathematical modelling in the literature on induction. So too does Chapter 7
on Bayesian models by Josh Tenenbaum, Charles Kemp, and Pat Shafto. The
main contribution of this chapter is to show how certain assumptions about
how knowledge is represented allow us to generate priors for a Bayesian belief-
updating process. Thus, inductive reasoning can be given a comprehensive
Bayesian treatment, which captures many of the knowledge effects discussed
in previous chapters.
Chapter 8, by Greg Murphy and Brian Ross, offers an interesting perspective
on Bayesian approaches to inductive reasoning. When generating a rating of
argument strength, the models described by Tenenbaum and colleagues work
through all of the hypotheses that are consistent with a particular inductive
argument. Murphy and Ross show that, in many situations, people tend
not to consider alternative hypotheses when evaluating inductive arguments.
Murphy and Ross’s chapter is problematic for the Bayesian view and asks
questions about the link between normative and psychological models (as
do Steven Sloman’s comments in his integrative chapter), which researchers
interested in Bayes would do well to address.
The volume closes with a set of chapters concerned with relations between
induction and other forms of thinking. Paul Thagard outlines a theory of
abductive reasoning, a form of induction in which one reasons to an expla-
nation. This chapter touches on several themes in contemporary cognitive
science as it posits an important role for the emotions and attempts to give a
neuropsychologically plausible account of cognitive processes. Thagard was
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

xviii Preface

one of the authors of the 1986 book on induction, and his chapter brings
an especially broad perspective to the study of thinking. As is apparent from
this chapter, there are many types of thinking, and Thagard’s comments on
the likely frequency of deductive and inductive inference in everyday life are
sobering.
Chapter 10, by Lance Rips and Jennifer Asmuth, although explicitly about
a rarefied form of reasoning in mathematics, is ultimately concerned with the
relationship between deduction and induction. Previous work (Rips, 2001)
suggests that inductive and deductive reasoning are dissociated. Rips and
Asmuth consider the case of mathematical induction, which they view as a
form of deductive thinking. Interestingly, even quite advanced students of
mathematics have problems with this form of reasoning.
Mike Oaksford and Ulrike Hahn in Chapter 11 argue for a probabilistic
approach to argument evaluation. They claim that induction and deduction
may be treated as belief-updating problems, although they do not rule out
the possibility that some small amount of deduction takes place in the world.
Because of its Bayesian character, this chapter is complementary to the chapter
by Tenenbaum and colleagues. Also notable in this chapter is an attempt
to get to grips with, from a Bayesian point of view, well-known informal
reasoning fallacies, such as the argument from ignorance. Oaksford and Hahn
demonstrate that some of these so-called fallacies may be justified from a
Bayesian point of view.
In Chapter 12, Feeney takes a different perspective on the relationship
between induction and other forms of thinking. Whereas Rips and Asmuth
suggest that induction and deduction are different types of thinking, and
Oaksford and Hahn suggest that, generally speaking, there is only one type
of thinking, Feeney makes the argument that induction is no different from
deduction in that both generally require the operation of two different types
of thinking process. One of these is fast and associative whilst the other is
slow and symbol manipulating (see Evans & Over, 1996; Sloman, 1996). The
dual-process argument is not new, but its application here leads to a novel
conception of inductive reasoning that is often thought of as involving only
associative processes. Interestingly, however, causal knowledge is somewhat
problematic for this dual-process view.
Finally, in Chapter 13, Steven Sloman comments on the other chapters in
the book and draws together their common themes and implications in an
interesting and provocative way. We do not want to spoil his punchlines here,
but he has many interesting things to say about the relationship between types
of thinking and about Bayesian approaches to induction. In many ways, this
is the most important chapter in this book, as its role is to explicitly comment
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19

Preface xix

on the state of the field rather than merely describing and interpreting what
is to be found in the literature.


We wish to thank the organisers of the Fifth International Conference on
Thinking for providing the forum for the meeting that led to this book. Darren
Dunning helped us prepare the manuscript for submission to the publisher.
Finally, we would like to thank all of the contributors to the symposium on
induction, including Dinos Hadjichristidis, Fenna Poletiek, and Vinod Goel,
and all of the contributors to this book.

References
Connolly, T., Arkes, H. R., & Hammond, K. R. (2000). Judgement and decision making:
An interdisciplinary reader. Cambridge, UK: Cambridge University Press.
Evans, J. St. B. T., & Over, D. E. (1996). Rationality and reasoning. Hove, UK: Psychology
Press.
Leighton, J. P., & Sternberg, R. J. (2004). The nature of reasoning. Cambridge, UK:
Cambridge University Press.
Manktelow, K., & Chung, M. C. (2004). Psychology of reasoning: Theoretical and historical
perspectives. Hove, UK: Psychology Press.
Rips, L. J. (2001). Two kinds of reasoning. Psychological Science, 12, 129–134.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psychological
Bulletin, 119, 3–22.

Editors
Aidan Feeney
Durham University
Evan Heit
University of California, Merced
P1: JZP
0521856485pre CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:19
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It?

Evan Heit

Why study induction, and indeed, why should there be a whole book devoted
to the study of induction? The first reason is that inductive reasoning cor-
responds to probabilistic, uncertain, approximate reasoning, and as such, it
corresponds to everyday reasoning. On a daily basis we draw inferences such
as how a person will probably act, what the weather will probably be like, and
how a meal will probably taste, and these are typical inductive inferences. So
if researchers want to study a form of reasoning that is actually a pervasive
cognitive activity, then induction is of appropriate interest.
The second reason to study induction is that it is a multifaceted cognitive
activity. It can be studied by asking young children simple questions involving
cartoon pictures, or it can be studied by giving adults a variety of complex
verbal arguments and asking them to make probability judgments. Although
induction itself is uncertain by nature, there is still a rich, and interesting, set
of regularities associated with induction, and researchers are still discovering
new phenomena.
Third, induction is related to, and it could be argued is central to, a number
of other cognitive activities, including categorization, similarity judgment,
probability judgment, and decision making. For example, much of the study
of induction has been concerned with category-based induction, such as
inferring that your next door neighbor sleeps on the basis that your neighbor
is a human animal, even if you have never seen your neighbor sleeping.
And as will be seen, similarity and induction are very closely related, many
accounts of induction using similarity as their main currency (Heit & Hayes,
2005).
Finally, the study of induction has the potential to be theoretically revealing.
Because so much of people’s reasoning is actually inductive reasoning, and
because there is such a rich data set associated with induction, and because
induction is related to other central cognitive activities, it is possible to find
1
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

2 Evan Heit

out a lot about not only reasoning but cognition more generally by studying
induction.
Induction is traditionally contrasted with deduction, which is concerned
with drawing logically valid conclusions that must follow from a set of
premises. The following section will consider possible definitions of induction
by describing possible relations between induction and deduction. But first it
is useful to briefly mention that the reasons for studying induction to some
extent are linked to the differences between induction and deduction. That is,
it could be argued that induction, in comparison to deduction, characterizes
more of everyday reasoning, has the potential to be studied with a broader
range of tasks and materials, and is closely related to other cognitive activities
that help people manage uncertainty.

     

Although it might be natural to ask “how are induction and deduction differ-
ent?” that would presuppose the conclusion that they are actually different.
Although induction and deduction are traditionally considered alternatives
to each other, as will be seen under some conceptions the similarities are
much greater than the differences. Before assessing to what extent induction
and deduction are similar or different, it is first important to consider just
what kind of entities induction and deduction are. Although not always made
explicit by researchers, there are two views on this issue, namely, the “problem
view” and the “process view.” According to the problem view, induction and
deduction refer to particular types of reasoning problems. So from looking
at a particular problem, say a question on a piece of paper in a psychologi-
cal experiment on reasoning, it should be possible to say whether this is an
induction problem or a deduction problem (or possibly it could be deemed
debatable whether it is one or the other). In contrast, according to the process
view, the locus of the question is not on the paper but in the head. That is,
induction and deduction refer to psychological processes. For a given prob-
lem, it may be possible to answer it using induction processes or deduction
processes. Likewise, we can investigate what is the relation between the two
kinds of processing.
The problem view and the process view have to a large extent been con-
founded in the literature. That is, researchers who have studied problems
that are traditionally thought of as induction have typically been interested
in different psychological theories than researchers who have studied tradi-
tional deduction problems. However, for the sake of clarity it is better to treat
the two views separately, namely, how problems of induction may differ from
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 3

problems of deduction, and how inductive processes may differ (or not differ)
from deductive processes. These two views will now be addressed in turn.

The Problem View

general and specific. It is sometimes said that induction goes from the
specific to the general, and deduction goes from the general to the specific.
For example, after observing that many individual dogs bark, one might
induce a more general belief that all dogs bark. Alternately, having the general
belief that all dogs bark, one might deduce that some particular dog will bark.
However, there are difficulties with this version of the problem view. Consider
the following arguments. (The statement above the line is a premise that is
assumed to be true, and the task is to consider the strength of the conclusion,
below the line.)

Dogs have hearts (1)

---------------------------------
All mammals have hearts

All mammals have hearts (2)

---------------------------------
Dogs have hearts

Dogs have hearts (3)

-------------------------
Wolves have hearts

Dogs have hearts (4)

--------------------------
At least some mammals have hearts

Argument (1) is a good example of an inductive argument going from

specific to general, and likewise argument (2) is a good example of a deductive
argument going from general to specific. Yet arguments (3) and (4) do not fit
neatly into this scheme. Argument (3) is somewhat plausible but surely not
deductively valid, so it is better thought of as an inductive argument. Yet it goes
from specific to specific rather than specific to general. Finally, argument (4)
seems to be deductively valid, yet it starts with a specific statement. Still, it is
possible to disagree about these last two arguments. For argument (3), it could
be said that there is an intervening general conclusion, such as “All mammals
have hearts.” For argument (4), it could be said that there is a hidden general
premise, such as “All dogs are mammals.” The key point is that one can’t just
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

4 Evan Heit

look at the written form of an argument, in terms of whether it goes from

specific to general or vice versa, and easily state whether it is inductive or
deductive in nature.

defining validity. Hence, it would seem profitable to take a more subtle

approach to the problem view. Perhaps the most defensible version of the
problem view is to define deductively valid arguments, and relate other kinds
of arguments to those that are deductively valid. One standard definition of
deductively valid arguments is that these are arguments following the rules of a
well-specified logic. Assuming that one can specify the rules of one’s preferred
logic, say in terms of truth tables for various symbolic combinations, then
it should be possible (if not easy) to determine whether any given argument
is deductively valid or not. It might be seen as a small disadvantage of this
approach that deductive validity is not defined in absolute terms but only
relative to a logic. Different people might endorse different logics and hence
disagree about which arguments are deductively valid. On the other hand,
defining deductive validity relative to a logic could be seen as an advantage
in terms of giving flexibility and in terms of appreciating that there is not a
single true logic that is universally agreed.
A more serious problem with this version of the problem view is that it does
not say much about inductive problems. Once the deductively valid arguments
are defined, what remains are the deductively invalid ones. Presumably some
of these are stronger than others, in terms of induction. For example, compare
argument (1) above to argument (5) below.
Dogs have hearts (5)
---------------------------------
All living things have hearts

It should be clear that neither (1) nor (5) is deductively valid, yet some-
how (1) seems more plausible in terms of being a good inductive argument.
Whatever rules of logic are used to define deductive arguments may not be
too useful in determining that (1) is stronger than (5).

levels of certainty. Another approach to the problem view is to describe

arguments in terms of the certainty of their conclusions (Skyrms, 2000).
Consider argument (6).
Dogs have hearts (6)
-------------------------
Dogs have hearts

In this case, it seems absolutely certain that if the premise is taken to be

true, then the conclusion must necessarily follow. This must be a perfectly
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 5

valid argument. On the other hand, an argument such as (2) above might
seem to have a very certain conclusion, perhaps 99.5% certain. This level of
certainty could still be well over the threshold that is required for saying that
an argument is deductively valid. Let’s say, hypothetically, that arguments
with conclusions below the 99% level of certainty will be called deductively
invalid. Even among these arguments, this version of the problem view allows
a great deal of differentiation. For example, argument (1) might be associated
with 80% certainty and argument (5) might be associated with 10% cer-
tainty. Hence (1) would be considered a much stronger inductive argument
in comparison to (5).
Perhaps the greatest appeal of this version of the problem view is that
it allows for deduction and induction to be placed on a common scale of
argument strength. In principle, any argument would have a place on this
scale, and whether it is deductively valid, inductively strong, or inductively
weak would be determined by the value on the scale. The most obvious
problem, though, is that there is still a need for assessing the place of each
argument on the scale. One nice idea might be an inductive logic, that is,
some set of rules or operations that for a set of premises can assign a certainty
value for a conclusion.
A subtler problem is that “certainty” itself would need to be defined better.
For example, in argument (1), either the conclusion that all mammals have
hearts is true or it is not, so the conversion from probability to certainty
may not be obvious. For example, it would seem a little funny to assign a
certainty level from 0% to 100% to a statement that is either true or false.
(Perhaps certainty could be related to the proportion of mammals with hearts,
rather than to whether it is true that all mammals have hearts.) Another issue
to clarify is the distinction between argument strength and certainty of the
conclusion. Argument (1) may seem strong simply because people believe
the conclusion that all mammals have hearts. Now compare that argument to
argument (7), below.

Lemons have seeds (7)

---------------------------------
All mammals have hearts

Here is a situation where the two arguments have the same conclusion,
which is equally certain in each case, but (1) seems much stronger than
(7). It could be valuable to consider other ways of representing argument
strength here, such as the conditional probability of the conclusion given
the premise, or the difference between the unconditional probability of
the conclusion and the conditional probability of the conclusion, given the
premise.
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

6 Evan Heit

matters of convention. A final perspective on the problem view is simply

to be descriptive, that is, to enumerate what kinds of arguments are studied by
researchers under the topics of induction and deduction. Induction is poten-
tially a very broad topic and a variety of cognitive activities have been referred
to as induction, including categorization, reasoning by analogy, and proba-
bility judgment. However, many of the chapters in this book focus on a more
particular kind of induction, namely, category-based induction, involving ar-
guments about categories and their properties. (Most of the examples in this
chapter represent typical examples of category-based induction.) Research on
adults’ reasoning usually involves presenting arguments like these in written
form; however, for children it is possible to present similar information with
pictures. Studies of induction typically ask people to make judgments of ar-
gument strength, such as to judge which of two arguments is stronger, or with
a single argument to make a continuous judgment of strength or probability.
In comparison to induction, research in deduction has used a narrower
range of problems. One typical area of research within deduction is condi-
tional reasoning – arguments involving ifs and thens, examining reasoning
involving classic rules such as modus ponens and modus tollens. Another
area of research within deduction is syllogistic reasoning – reasoning with
arguments with statements like “All artists are beekeepers.” Indeed, for syl-
logisms involving two premises, there are only sixty-four classical forms of
syllogism. Research on deduction tends to ask people to assess logical validity
of conclusions (a yes or no question) rather than make continuous judgments.
Overall, the conventional approach is like other approaches to the problem
view in that there is a relatively narrow range of arguments corresponding to
deduction and a wider, somewhat ill-defined, range corresponding to induc-
tion. Yet even within the area of deduction research, there are lively debates
about what exactly is a problem of deduction. For example, Wason’s selection
task, involving selecting cards to test a rule such as “If a card has a vowel on
one side then it has an even number on the other side,” has been variously
argued to be a problem of deduction or induction (e.g., Feeney & Handley,
2000; Oaksford & Chater, 1994; Poletiek, 2001).

evaluation of the problem view. Perhaps the most appealing aspect

of the problem view is that it offers the possibility of defining induction
and deduction in an objective way, in terms of the problem being solved or
the question being asked. (The problem view is more impressive in terms
of defining deduction in comparison to defining induction, though.) From
the point of view of psychologists, this strength would also be the greatest
weakness, namely, that the problem view does not itself refer to psychological
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 7

processes. Just because one problem is defined as an induction problem and

another is defined as a deduction problem does not guarantee that people will
engage in inductive reasoning processes for one task and deductive reasoning
processes for the other task. The same processes could be used for both, or the
delimitation between types of psychological processes might not correspond
at all to the agreed definition of problems, or any problem might engage a
mixture of processes. In the terminology of memory research, there are no
process-pure tasks. Of course, for computer scientists or logicians, reference
to psychological processes may not be a priority. Still, it does seem desirable to
consider the alternative of treating induction and deduction as possible kinds
of psychological process. Hence, this chapter will next turn to the process
view.

The Process View

According to the process view, comparing induction and deduction is a mat-
ter of specifying the underlying psychological processes. According to one-
process accounts, the same kind of processing underlies both induction and
deduction. Another way to describe this idea is that there is essentially one
kind of reasoning, which may be applied to a variety of problems that could
be considered either inductive or deductive in nature (Harman, 1999). In
contrast, according to two-process accounts, there are two distinct kinds of
reasoning. It is possible that these two kinds of reasoning directly correspond
to induction and deduction. Alternately, the two kinds of reasoning might
correspond to some other distinction, such as intuitive reasoning versus delib-
erative reasoning, that could be related to the distinction between induction
and deduction. It should be acknowledged at the start that one-process and
two-process accounts are somewhat poorly named. That is, at some level, rea-
soning surely involves many different psychological processes. The question,
though, is whether the same processing account is applied to both induction
and deduction, or whether two different processing accounts are applied.
Some examples of one-process and two-process accounts will now be de-
scribed, followed by the presentation of some experimental evidence aimed
at assessing these accounts.

one-process accounts. One of the most widely known theories of reason-

ing is mental model theory, which proposes that people solve reasoning prob-
lems extensionally by constructing models of possible states of the world and
performing operations and manipulation on them (Johnson-Laird, 1983).
Mental model theory is usually thought of as an account of deduction, and
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

8 Evan Heit

indeed it has been extensively applied to conditional-reasoning and syllogistic-

reasoning problems. However, it has also been argued that mental model
theory can be applied to problems of induction, namely, probabilistic reason-
ing tasks (Johnson-Laird, 1994; Johnson-Laird, Legrenzi, Girotto, Legrenzi,
& Caverni, 1999). Hence, mental model theory is a one-process account, in
the sense that it is aimed at giving a singular account for problems both of
induction and deduction.
A newer alternative to mental model theory is the probabilistic account,
which aims to account for a variety of reasoning phenomena, particularly
traditional deduction problems in terms of probabilistic formulas, such as
from Bayesian statistics (Chater & Oaksford, 2000; Oaksford & Chater, 1994).
Essentially, the probabilistic account is saying that people solve deduction
problems by means of induction processes. This account does not propose
different kinds of processing for performing deduction, and hence the prob-
abilistic account is also a one-process account.
The previous accounts are aimed mainly at problems of deduction. In con-
trast, other reasoning accounts have focused on problems of induction, such
as category-based induction (Heit, 1998; Osherson, Smith, Wilkie, Lopez,
& Shafir, 1990; Sloman, 1993). These accounts are aimed at predicting the
judged strength of various inductive arguments, for example, that (1) above
seems stronger or more plausible than (5). Typically, these accounts of in-
duction are based on some measure of similarity or overlap between premise
and conclusion categories, in terms of existing knowledge. In this example,
there is more known overlap between dogs and mammals than between dogs
and living things; hence the argument relating dogs and mammals seems
stronger than the argument relating dogs and living things. Now refer back
to argument (6). Here, there is perfect overlap between the premise cate-
gory and the conclusion category – in this case the categories are both dog.
Hence, there is perfect overlap between premise and conclusion categories,
and these accounts of induction should predict that (6) is perfectly strong.
In other words, accounts of induction can treat some deductively valid argu-
ments as a special case rather than as being wholly different than inductively
weak or strong arguments. The same processing mechanisms – for example,
for assessing overlap – would be applied to problems of induction and de-
duction. In this way, these accounts of induction are one-process accounts.
However, it should be made clear that these accounts of induction do not
give complete accounts of deductive phenomena. For example, many deduc-
tively valid arguments in conditional and syllogistic reasoning could not be
assessed simply in terms of feature overlap between premise and conclusion
categories.
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 9

Criterion 2

DEDUCTIVELY DEDUCTIVELY
INVALID VALID
Minimum Maximum
Argument Argument
Strength Strength
INDUCTIVELY INDUCTIVELY
WEAK STRONG

Criterion 1
figure 1.1. Criterion-shift account of deduction and induction.

In addition to these one-process accounts that specialize mainly in either

deduction or induction problems, there is an alternative that does not give a
detailed account of either deduction or induction but does offer a balanced
view of how deduction and induction are related to each other. The criterion-
shift account (described by Rips, 2001) is closely related to the levels-of-
certainty version of the problem view and is illustrated in Figure 1.1. Under
this account, assessing the strength of an argument involves finding its place
on a one-dimensional scale ranging from minimum argument strength (the
most unconvincing argument possible) to maximum strength (an utterly and
completely compelling argument). To assess whether an argument should be
considered inductively strong, its strength is compared to a criterion, such
as criterion 1 in the figure. To assess whether an argument is deductively
valid, the criterion is shifted to the right, to criterion 2. By this criterion,
an argument would have to be judged extremely strong before it could be
called deductively valid. The same reasoning mechanisms would be used for
different argument types. The only difference between performing induction
or deduction would be represented as a shift of the criterion.

two-process accounts. In contrast to one-process accounts, other re-

searchers have emphasized a distinction between two kinds of reasoning (e.g.,
Evans & Over, 1996; Sloman, 1996; Stanovich, 1999). In these two-process
accounts there is one system that is relatively fast but heavily influenced by
context and associations, and another system that is more deliberative and an-
alytic or rule based. These two systems do not necessarily correspond directly
to induction and deduction. That is, the traditional distinction between these
two forms of reasoning may not be the best way to divide things in psycho-
logical terms. Still, it is plausible that induction would depend more on the
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

10 Evan Heit

first system, whereas deduction would depend more on the second system.
These two-process accounts have been used to explain a variety of findings
in reasoning, concerning individual differences, developmental patterns, and
relations between reasoning and processing time. For example, in Stanovich’s
work there is a rich account of how reasoning in the more deliberative system
is correlated with IQ, accounting for patterns of individual differences in a
variety of problems that would rely more on one system or the other.

evaluating one-process and two-process accounts. How would it be

possible to decide in favor of either one-process or two-process accounts?
There is some neuropsychological evidence, based on brain imaging, for two
anatomically separate systems of reasoning (Goel, Gold, Kapur, & Houle,
1997; Osherson et al., 1998). In the studies, subjects were given a set of
arguments to evaluate. Half the subjects were asked to judge deductive validity
and the other half were asked to judge inductive plausibility. Within each study,
there were distinct brain areas implicated for deduction versus induction.
What is particularly relevant for present purposes is that the problems were
the same for the two conditions, but subjects were asked to perform either
deduction or induction. Hence, this is a case of unconfounding the process
view from the problem view – presumably all that varied between conditions
was processes, unlike the situation in most previous studies of deduction,
induction, or both, which used very different problems for one task or the
other.
One does not require expensive brain imaging equipment to compare
deduction versus induction instructions for a common set of problems. Rips
(2001) used the same logic in a much cheaper pen-and-paper study, in which
subjects were instructed to judge either deductive correctness or inductive
strength for two types of arguments. One type of argument was deductively
correct but causally inconsistent, such as “Jill rolls in the mud and Jill gets
clean, therefore Jill rolls in the mud,” and the other type was deductively
incorrect but causally consistent, such as “Jill rolls in the mud, therefore Jill
rolls in the mud and Jill gets dirty.” In terms of the criterion-shift version of
the one-process account, if one type of argument is stronger than another for
deduction, then the same type of argument should be stronger for induction.
In Figure 1.1, let the dots on the scale represent different types of argument.
If one type is stronger, that is, further to the right end of the scale, then it
should be stronger regardless of whether the induction or deduction criterion
is used. Yet the results were that subjects in the deduction condition gave more
positive judgments to the correct but inconsistent arguments, whereas subjects
in the induction condition gave more positive judgments to the incorrect
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 11

but consistent arguments. Rips concluded that these results contradicted the
criterion-shift account, which predicted a monotonic ordering of arguments
in the two conditions. (See Heit & Rotello, 2005, for further examinations of
this kind, leading to the same conclusion.)
In sum, there is some evidence already that giving people deduction versus
induction instructions leads to qualitatively different results. It would seem
difficult to explain these results by assuming that deduction and induction
processes are essentially the same, except that deduction involves a stricter
criterion for giving a positive response. Yet at the same time, it seems too early
to abandon the one-process accounts, which do provide detailed and accurate
predictions about a range of phenomena, usually either concerning deductive
or inductive problems. In contrast, the two-process view does not seem to
be as well developed in terms of providing detailed process accounts and
predictions. More experimentation, directly aimed at comparing the one-
and two-process views and at further developing the two-process view, is
clearly needed.
At a more general level, the process view itself seems to be a rich and
worthwhile approach to studying induction. Certainly for psychologists, the
problem view does not seem viable. It is a mistake to assume that people
are performing deduction processes on designated deduction problems and
induction processes on designated induction problems. Likewise, even for
psychologists who are developing process level accounts of reasoning, it is
important to keep in mind the wide range of possible reasoning problems.
Assuming there is at least considerable overlap between deduction and in-
duction processes, an ideal theory of reasoning would not be limited to either
traditional problems of deduction or induction but would encompass both
types of problem.

   :    

 
Now that the position of induction with respect to deduction has been ex-
plored, the next step will be to introduce the study of induction in more
depth. Of course, this whole book is devoted to presenting research on in-
duction, so the following material is intended to serve as a microcosm of
the book rather than a complete review of induction research. The focus of
the remainder of this chapter will be the diversity principle, namely, the idea
that more diverse evidence seems to be stronger evidence than less diverse
evidence. This principle seems to be ubiquitous in various forms of reasoning.
For example, in social reasoning, seeing a person act in an extroverted way, in
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

12 Evan Heit

a variety of contexts, suggests that the person is truly extroverted by nature.

In comparison, seeing a person repeatedly act in an extroverted way, but only
in one context, does not give as good support for inferences that the person
is truly extroverted. In legal reasoning, having two independent witnesses to
a crime seems to be stronger evidence than two witnesses who may share a
similar perspective, such as two members of the same family. Note that none
of this represents anything like a deductively valid inference; however, there
is still a sense of stronger evidence coming from more diverse observations.
The diversity principle has been an object of interest in terms of various
approaches to induction, including the historical approach, the philosophical
approach, the experimental approach, the developmental approach, and the
model-based approach. These approaches will be reviewed in relation to how
they have addressed the diversity principle.

The Historical Approach

The diversity principle has been historically important in science. Scien-
tists have tended to favor testing a theory with a diverse set of experiments
rather than repeatedly conducting the same experiment or close replications.
Imagine two scientists who are testing their respective theories. One scientist
essentially conducts the same experiment ten times, whereas the other con-
ducts a variety of experiments assessing different aspects of the theory. Which
theory will seem to have stronger support?
An early example of the diversity principle in action was in Bacon’s Novum
Organum (1620/1898), which cautioned scientists of the day against inferences
drawn from narrow samples. Bacon illustrated this point with the concept of
heat, listing twenty-eight different kinds of heat and hot things that would
need to be observed in a study of heat. For example, Bacon listed the rays
of the sun, fiery meteors, natural warm baths, hot vapor in furnaces, sparks,
stacks of hay, and the insides of animals.
In a somewhat more modern example, Salmon (1984) described how early
in the twentieth century, chemists and physicists had developed a wide variety
of experimental methods for deriving Avogadro’s number (6.02 x 1023 ), the
number of particles in a mole of any substance. These methods included Brow-
nian movement, alpha-particle decay, X-ray diffraction, black-body radiation,
and electrochemistry. Together, these techniques led to strong support for the
existence of atoms and molecules because the existence of these particles was
the basis of a unified explanation for a set of highly diverse results. Salmon
argued that any one of these techniques taken alone, no matter how carefully
applied, would not have been sufficient to convince scientists of that period
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 13

to accept the atomic theory over its principal rival, known as energeticism,
which conceived of matter as being continuous rather than being composed
of particles.
More generally, it should be possible to study inductive reasoning by study-
ing historical examples of reasoning, whether by scientists or others. One
advantage of studying scientific reasoning is that the evidence and theories
are usually explicit, in comparison to just studying everyday examples of rea-
soning. (See the previous book on induction, by Holland, Holyoak, Nisbett,
and Thagard, 1986, as well as a more recent collection by Gorman, Tweney,
Gooding, and Kincannon, 2005, for further examples.) There is an interest-
ing parallel between the historical approach to the study of induction and
the historical approach to the study of creativity (e.g., Weisberg, 1986). In
each case, it seems that much can be learned about cognition by looking at
paragon cases of thinking and reasoning, even outside the bounds of tightly
controlled psychological studies.

The Philosophical/Normative Approach

This historical approach is complemented by attempts of philosophers and
statisticians to argue that certain patterns of induction are normative or
correct. The philosophical approach follows in the footsteps of Hume’s (1777)
problem of induction, namely, the problem of whether an inductive inference
can ever be justified or considered valid. Without solving Hume’s problem in
absolute terms, it still might be possible to argue that performing induction
in one way is better than another. As such, there have been various attempts to
argue for or even prove the greater strength of diverse evidence in comparison
to less diverse evidence.
For example, Nagel (1939) argued that a scientific theory should be derived
from diverse observations to obtain more reliable estimates. He gave the
example of inspecting the quality of coffee beans delivered on a ship. It would
be better to inspect small samples of beans from various parts of the ship
than to inspect a large number of beans from just one location. Carnap
(1950) linked the collection of diverse evidence to the principle that scientific
theories should make novel predictions rather than merely redescribe old
data. Similarly, Hempel (1966) related the collection of diverse evidence to
a falsifying research strategy, namely, it is better to test theories with a wide
variety of challenging experiments.
These intuitive arguments did not lead directly to more formal proofs of
the diversity principle. For example, Carnap (1950) promised a proof of the
diversity principle in a future edition of his book that never did appear. More
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

14 Evan Heit

recently, however, there have been several attempts to formalize the advan-
tage for following the diversity principle. As reviewed by Wayne (1995), there
have been two approaches. The first approach compares correlated sources
of evidence to independent sources of evidence, in statistical terms. For for-
mal treatments of this correlation approach, linking similarity to probability
theory, see Earman (1992) and Howson and Urbach (1993). The second ap-
proach is the eliminative approach. The idea behind the eliminative approach
is that diverse data sets will be particularly useful for eliminating plausible
but incorrect hypotheses, allowing stronger inferences to be drawn based on
the remaining, contending hypotheses. In contrast, non-diverse data sets will
likely be consistent with too many hypotheses to allow any strong inferences.
For a formal treatment of this approach, including a geometric proof, see
Horwich (1982), and see Heit (1998) and Tenenbaum and Griffiths (2001)
for some psychological applications.
Moreover, there have been arguments that following the diversity prin-
ciple is not normative. For example, using Earman’s (1992) derivations of
the diversity principle, Wayne (1995) showed that there can be exceptions,
namely, that non-diverse observations can lead to strong inferences if this
evidence is nonetheless very surprising. Wayne pointed to the case of the
near-simultaneous discovery in 1974 of a previously unknown subatomic
particle in two laboratories being a case of non-diverse evidence that still
had strong implications for revision of theories in physics. Lo, Sides, Rozelle,
and Osherson (2002) raised a related criticism of the normative status of the
diversity principle. They too argued that what is crucial is not diversity of ob-
servations but rather surprisingness of observations. Lo et al. also suggested
a set of exceptions, such as the following:

Squirrels can scratch through Bortex fabric in less than 10 seconds. (8)
Bears can scratch through Bortex fabric in less than 10 seconds.
---------------------------------------------------------------------
All forest mammals can scratch through Bortex fabric in less than 10 seconds.

Squirrels can scratch through Bortex fabric in less than 10 seconds. (9)
Mice can scratch through Bortex fabric in less than 10 seconds.
---------------------------------------------------------------------
All forest mammals can scratch through Bortex fabric in less than 10 seconds.

It seems intuitive that squirrels and bears are a more diverse pair than
squirrels and mice. Yet Lo et al. argued that (9) is stronger than (8), because
the evidence about squirrels and mice is more surprising than the evidence
about squirrels and bears. That is, the knowledge that small animals are less
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 15

capable of feats of strength than are large animals makes the evidence
about squirrels and mice more surprising than evidence about squirrels and
bears.
Heit, Hahn, and Feeney (2005) argued that these exceptions to the diver-
sity principle, suggested by Wayne (1995) and Lo et al. (2002), are indeed
exceptions, but they do not undermine the normative status of the diversity
principle itself. In the example of the discovery of a new subatomic particle
in 1974, physicists were influenced not only by diversity but also by many
other sources of knowledge in particle physics. In the example of scratching
through Bortex fabric, people would be influenced not only by diversity but
also by other knowledge about animals and their strength. In other words,
these exceptions as stated do not contain all the premises upon which the
arguments are based. Reasoning about these arguments is also influenced by
other hidden premises or background knowledge, so that diversity is not being
assessed in isolation. Therefore, these counterexamples do not invalidate the
diversity principle, because they are not pure tests of diversity. Rather, they
show that people will use other knowledge when possible. Indeed, philoso-
phers of science have not claimed that the diversity principle is the sole
principle for assessing evidence. For example, Popper (1963, p. 232) listed
diversity of supporting evidence as one of six criteria for assessing a scientific
theory.
In more general terms, it should be possible to consider a variety of patterns
in inductive reasoning in the light of the normative question, namely, what is
good reasoning. For example, one of the most pervasive findings in psycho-
logical research on induction is the similarity effect, namely, that arguments
such as (3) above concerning dogs and wolves are considered stronger than
arguments such as (10).
Dogs have hearts. (10)
-------------------------
Bees have hearts.

Argument (10) seems much weaker, and this seems to be a consequence of

the low similarity between dogs and bees. However, providing a proof of why
it is normative to reason on the basis of similarity has been an elusive task for
philosophers and psychologists (Goodman, 1972, but see Shepard, 1987).

The Experimental Approach

In addition to the normative perspective on the diversity principle, there has
been a sustained effort by psychologists to document how well the diversity
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

16 Evan Heit

principle serves as a descriptive account of how people carry out informal,

inductive reasoning. Osherson et al. (1990) documented diversity effects in
adults by using written arguments like the following:

Hippos require Vitamin K for the liver to function. (11)

Rhinos require Vitamin K for the liver to function.
---------------------------------------------------------
All mammals require Vitamin K for the liver to function.

Hippos require Vitamin K for the liver to function. (12)

Hamsters require Vitamin K for the liver to function.
---------------------------------------------------------
All mammals require Vitamin K for the liver to function.

The subjects judged arguments like (12) to be stronger than arguments like
(11), in response to the greater diversity of hippos and hamsters compared to
hippos and rhinos. Indeed, there is a great deal of evidence that adults, mainly
Western university students, follow the diversity principle when evaluating
written arguments (see Heit, 2000, for a review).
However, when looking to other subject populations, and to evidence col-
lected at a greater distance from the psychology lab, there seem to be exceptions
to the diversity principle as a descriptive account. In their study of Itzaj-
Mayan adults from the rainforests of Guatemala, Lopez, Atran, Coley, Medin,
and Smith (1997) did not find evidence for diversity-based reasoning, using
arguments with various living things and questions about disease transmis-
sion. Indeed, sometimes the Itzaj reliably chose arguments with non-diverse
premise categories over arguments with diverse categories. It appears that
they were using other knowledge about disease transmission that conflicted
with diversity-based reasoning. For example, given a non-diverse argument
that two similar kinds of tall palm trees get some disease, one person claimed
it would be easy for shorter trees, located below, to get the disease as well.
Giving further support to this idea that other strategies and knowledge can
overrule diversity, Proffitt, Coley, and Medin (2000) reported that American
adults who are tree experts (such as landscapers) did not show strong diversity
effects when reasoning about trees and their diseases. The tree experts seemed
to be relying on the knowledge that tree diseases tend to spread readily within
tree families such as elms and maples.
Medin, Coley, Storms, and Hayes (2003) reported further exceptions to the
diversity principle. One exception, referred to as non-diversity by property
reinforcement, potentially makes a direct challenge to the diversity principle
that is not as easily explained in terms of the use of other knowledge. The idea
behind non-diversity by property reinforcement is that two diverse categories
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 17

may nonetheless have some characteristic in common and tend to generalize

only to other categories with this same characteristic. In the non-diversity by
property reinforcement effect, “if an otherwise diverse set of premises shares
a salient property not shared by the conclusion category, the reinforcement
of the property might weaken that argument relative to a related argument
with less diverse premises” (p. 523). This phenomenon is illustrated by the
following example:

Polar bears have property X. (13)

Antelopes have property X.
---------------------------------------
All animals have property X.

Polar bears have property X. (14)

Penguins have property X.
---------------------------------------
All animals have property X.

When given a forced choice between polar bears and antelopes versus polar
bears and penguins, subjects judged the two animals from the same biological
class, polar bears and antelopes, to be more similar than the two animals from
different biological classes, polar bears and penguins. However, when asked
to assess the inductive strength of each argument, argument (14) was judged
to be less convincing than argument (13). That is, argument (13) had less
diverse evidence, yet it was the stronger argument. Intuitively, although polar
bears and penguins are from different biological classes, they still share the
characteristic of living in a cold climate. It might seem than property X does
not apply to all animals but only to animals living in cold climates.
Heit and Feeney (2005) investigated the non-diversity by property rein-
forcement effect further and came to a somewhat different conclusion. Es-
sentially, their subjects judged polar bears and penguins to be more similar
than polar bears and antelopes. Hence, when argument (13) was judged to be
stronger than argument (14), Heit and Feeney’s subjects were showing a diver-
sity effect rather than a non-diversity effect. Heit and Feeney concluded that
the diversity effect was indeed robust, and their results suggest that exceptions
may be hard to show consistently.

The Developmental Approach

The first experiment on diversity-based reasoning in induction was actually
a developmental one by Carey (1985), comparing six-year-olds and adults.
Carey looked at patterns of inductive projection given the premises that two
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

18 Evan Heit

diverse animals, dogs and bees, have some biological property. The purpose
of this study was to see whether subjects reason that “if two such disparate
animals as dogs and bees” have this property then “all complex animals must”
(p. 141). Indeed, adults made broad inferences to all animals, extending the
property not only to things that were close to the premises (other mammals
and insects) but also to other members of the animal category (such as birds
and worms). In contrast, the children seemed to treat each premise separately;
they drew inferences to close matches such as other mammals and insects, but
they did not use the diversity information to draw a more general conclusion
about animals. Therefore in this first attempt there was evidence for effects
of diversity in adults but not children. In a follow-up study, Carey looked
at diversity effects based on the concept living thing rather than animal.
The key comparison was that children were taught a biological fact either
about dogs and bees or about dogs and flowers, with the latter being even
more diverse than the former. Given a fact about dogs and flowers, children
did tend to generalize fairly broadly, suggesting that children may have some
sensitivity to diversity of premise categories. However, if anything they tended
to overgeneralize, extending the property not only to other living things but
often to inanimate objects as well. Hence, there was suggestive evidence for
the impact of diversity of premise categories in this study, although children
did not show the same pattern as adults.
Continuing along this line of research that looks for diversity effects in
children, Lopez et al. (1992) found limited evidence for nine-year-olds and no
evidence for five-year-olds. For the five-year-olds, choices in a picture-based
task did not show any sensitivity to diversity of premise categories, even when
the diversity was explicitly mentioned by the experimenter. However, nine-
year-olds did show sensitivity to diversity of premises, but only for arguments
with a general conclusion category such as animal rather than a specific
conclusion category such as kangaroo. Gutheil and Gelman (1997) attempted
to find evidence of diversity-based reasoning for specific conclusions in nine-
year-olds, using category members at lower taxonomic levels which would
presumably enhance reasoning. However, like Lopez et al. (1992), Gutheil
and Gelman did not find diversity effects in nine-year-olds, although in a
control condition with adults, there was robust evidence for diversity effects.
More recently, however, Heit and Hahn (2001) reported diversity effects in
children younger than nine years in experiments that used pictures of people
and everyday objects as stimuli rather than animals with hidden properties.
For example, children were shown a diverse set of dolls (a china doll, a stuffed
doll, and a Cabbage Patch doll), all being played with by a girl named Jane.
Also children were shown a non-diverse set, three pictures of Barbie dolls,
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 19

being played with by Danielle. The critical test item was another kind of doll,
a baby doll, and the question was who would like to play with this doll. In
another stimulus condition, there was a diverse set of hats worn by one person,
and a non-diverse set worn by another person, and again, the critical question
was whether another hat would belong to the person with diverse hats or the
person with non-diverse hats. For 74% of these critical test items, children
age five to eight years made the diverse choice rather than the non-diverse
choice. It seems from the Heit and Hahn experiments that children can follow
the diversity principle at some level. However, it will take further work to
establish the critical differences that led the past studies to not find diversity
effects in children. (See Lo et al., 2002, for additional results, and Gelman,
2003, for further discussion.)

The Model-Based Approach

The psychological study of inductive reasoning has benefited from the de-
velopment of computational models. Perhaps what is most impressive about
these models is that they systematize what is inherently unsystematic, namely,
probabilistic inference. Furthermore, these models capture important empir-
ical phenomena in inductive reasoning, such as the diversity effect. In general,
modeling work and experimental work on induction have been closely linked.
In the following section, there are brief, conceptual descriptions of some of
the earlier models and how they address the diversity effect. Several chapters
in this book present newer modeling work in more detail (see Heit & Hayes,
2005, for another review of models of induction).

osherson et al. (1990). The most influential model of category-based in-

duction was developed by Osherson et al. (1990). This model has two main
components. The first component assesses the similarity between the premise
category or categories and the conclusion category. For example, argument (3)
seems fairly strong because of the high level of similarity between dogs and
wolves. In comparison, argument (15) below seems weaker, due to the lower
similarity between dogs and parrots.
Dogs have hearts. (15)
-------------------------
Parrots have hearts.

The second component of the model measures how well the premise cate-
gories cover the superordinate category that includes all the categories men-
tioned in an argument. For example, in arguments (11) and (12), the relevant
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

20 Evan Heit

superordinate category is mammal. Intuitively, hippos and rhinos are some-

what similar mammals that do not span the whole superordinate category. In
comparison, hippos and hamsters cover a broader range of mammals. The
Osherson et al. model formalizes the calculation of coverage based on the
structure of the mammal category. In effect, different instantiations of mam-
mal are considered, from mice to cows to elephants, with the model looking
for matches between the premise categories and these instantiations. Hippos
and rhinos are a fairly good match to elephants and other large mammals, but
they will be a poor match to smaller mammals. In comparison, the diverse
set, hippos and hamsters, gives good matches to both large mammals such as
elephants and small mammals such as mice. The diverse set is considered to
have greater coverage because it yields a wider range of good matches with
members of the superordinate category. More generally speaking, the simi-
larity and coverage components of this model can be used to explain not only
the diversity effect but also a dozen other phenomena in induction.

sloman (1993). This model was implemented as a connectionist network,

and perhaps the most important difference from the Osherson et al. (1990)
model is that it does not have a separate component for assessing coverage
of a superordinate category. Indeed, the Sloman model is valuable because
it shows how much can be accomplished without this second mechanism,
indeed addressing many of the same phenomena as the Osherson et al. model
but without coverage. In brief, the way this model works is that premises of
an argument are encoded by training the connectionist network to learn asso-
ciations between input nodes representing the features of premise categories
and an output node for the property to be considered. Then the model is
tested by presenting the features of the conclusion category and measuring
the activation of the same output node. For example, after the network has
been trained that dogs have hearts, it will give a strong response if the features
of dog are presented again as input. In addition, the network would give a
strong output signal to wolves, indicating that wolves have hearts, because the
featural representations of dogs and wolves would have a lot of overlap. The
model accounts for diversity effects because training on a diverse set of cate-
gories will tend to strengthen a greater number of connections than training
on a narrow range of categories. For example, training the network that both
hippos and hamsters have a certain property would activate a broad range of
features that apply to various mammals, leading to a strong conclusion that
all mammals have that property. That is, hippos and hamsters would activate
different features and different connections. In contrast, training the network
that hippos and rhinos have a property would activate only a narrow range of
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 21

features and connections. Although this model does have a notion of breadth
of features, there is no distinct component for assessing coverage of a super-
ordinate category, and indeed the model does not even rely on knowledge
about superordinate categories. Nonetheless, the Sloman model can account
for not only diversity effects but a variety of other phenomena.

heit (1998). The final model to be discussed is the Bayesian model by Heit
(1998). This model is linked to eliminative accounts of hypothesis testing and
as such is a normative model of how to reason with a hypothesis space. In
addition, this account is fairly successful as a descriptive account in the sense
that it predicts most of the same psychological phenomena as the Osherson
et al. (1990) and Sloman (1993) models. According to the Bayesian model,
evaluating an inductive argument is conceived of as learning about a property,
in particular learning for which categories the property is true or false. For
example, upon learning that dogs have some novel property X, the goal
would be to infer which other animals have property X and which do not. For
example, do wolves have the property and do parrots have the property? The
key assumption is that for a novel property such as in this example, people
would rely on prior knowledge about familiar properties to derive a set of
hypotheses about what the novel property may be like. People already know
a relatively large number of properties that are true of both dogs and wolves,
so if property X applies to dogs, then it probably applies to wolves too. On the
other hand, people know a relatively small number of properties that are true
of both dogs and parrots. Hence property X is relatively unlikely to extend to
parrots.
How does the Bayesian model explain diversity effects? In brief, diverse
categories bring to mind a very different set of hypotheses than non-diverse
categories. If hippos and hamsters have some novel property X in common,
one considers familiar properties that are shared by all mammals, such as
warm-bloodedness. Hence, property X too seems to extend broadly to all
mammals, assuming that it is distributed in a similar way as other properties
that are brought to mind. In contrast, if hippos and rhinos have property X in
common, it is easier to think of familiar properties that are shared by hippos
and rhinos but not most other mammals, such as being large and having a
tough skin. Property X, too, may be distributed in the same way, namely,
only to large, thick-skinned mammals, and seems less likely to extend to all
mammals.
In sum, this section has illustrated different modeling approaches to in-
duction, which have subsequently developed in later work. Interestingly,
these three models address a similar set of phenomena with different
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

22 Evan Heit

theoretical frameworks, namely, in terms of hierarchically structured cate-

gories (Osherson et al., 1990), features in a connectionist network (Sloman,
1993), and beliefs within a hypothesis space (Heit, 1998). Not only is it pos-
sible to systematize what is unsystematic by nature, namely, probabilistic
inference, but there is more than one way of doing so.


This chapter should at the very least illustrate the richness of research on
induction. Research on this topic might seem to face a lot of challenges. Af-
ter all, the degree of overlap with deduction has not yet been determined,
and some accounts of reasoning simply define induction in terms of not be-
ing deduction. By their very nature, inductive inferences do not have a “right
answer” in the same way as deductive inferences. Yet there are still regularities,
such as the diversity principle, which can be studied from a variety of per-
spectives, including historical, philosophical, experimental, developmental,
and computational. By no means is this regularity completely deterministic;
indeed, there are well-documented exceptions to the diversity principle that
are themselves of interest.
The material in this chapter should be seen as an invitation to consider
different approaches to induction and different phenomena in induction,
including those presented in the subsequent chapters of this book. All of
the chapters refer to the experimental approach to at least some extent. The
chapters by Hayes and by Medin and Waxman focus on the developmental
approach. The chapters by Tenenbaum and Blok, Osherson, and Medin largely
take the modeling approach. The chapters by Rips and Asmuth, and by
Thagard, involve the historical approach, and the chapters by Oaksford and
Hahn, and by Thagard, involve the philosophical approach. Finally, several
of the chapters (by Rehder, Rips & Asmuth, Oaksford & Hahn, Thagard, and
Feeney) directly address the question of what induction is.

References
Bacon, F. (1620–1898). Novum organum. London: George Bell and Sons.
Carey, S. (1985). Conceptual change in childhood. Cambridge, MA: Bradford Books.
Carnap, R. (1950). Logical foundations of probability. Chicago: University of Chicago
Press.
Chater, N., & Oaksford, M. (2000). The rational analysis of mind and behavior. Synthese,
122, 93–131.
Earman, J. (1992). Bayes or bust? A critical examination of Bayesian confirmation theory.
Cambridge, MA: MIT Press.
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

What Is Induction and Why Study It? 23

Evans, J. St. B. T., & Over, D. E. (1996). Rationality and reasoning. Hove, UK: Psychology
Press.
Feeney, A., & Handley, S. J. (2000). The suppression of q-card selections: Evidence
for deductive inference in Wason’s selection task. Quarterly Journal of Experimental
Psychology, 53A, 1224–1242.
Gelman, S. A. (2003). The essential child: Origins of essentialism in everyday thought. New
York: Oxford University Press.
Goel, V., Gold, B., Kapur, S., & Houle, S. (1997). The seats of reason: A localization
study of deductive and inductive reasoning using PET (O15) blood flow technique.
NeuroReport, 8, 1305–1310.
Goodman, N. (1972). Problems and projects. Indianapolis: Bobbs-Merrill.
Gorman, M. E., Tweney, R. D., Gooding, D. C., & Kincannon, A. P. (Eds.), (2005).
Scientific and technological thinking. Mahwah, NJ: Lawrence Erlbaum Associates.
Gutheil, G., & Gelman, S. A. (1997). Children’s use of sample size and diversity in-
formation within basic-level categories. Journal of Experimental Child Psychology, 64,
159–174.
Harman, G. (1999). Reasoning, meaning, and mind. Oxford: Oxford University Press.
Heit, E. (1998). A Bayesian analysis of some forms of inductive reasoning. In M. Oaksford
& N. Chater (Eds.), Rational models of cognition, 248–274. Oxford: Oxford University
Press.
Heit, E. (2000). Properties of inductive reasoning. Psychonomic Bulletin & Review, 7,
569–592.
Heit, E., & Feeney, A. (2005). Relations between premise similarity and inductive
strength. Psychonomic Bulletin & Review, 12, 340–344.
Heit, E., & Hahn, U. (2001). Diversity-based reasoning in children. Cognitive Psychology,
43, 243–273.
Heit, E., Hahn, U., & Feeney, A. (2005). Defending diversity. In W. Ahn, R. Goldstone,
B. Love, A. Markman, & P. Wolff (Eds.), Categorization inside and outside of the
laboratory: Essays in honor of Douglas L. Medin, 87–99. Washington, DC: American
Psychological Association.
Heit, E., & Hayes, B. K. (2005). Relations among categorization, induction, recognition,
and similarity. Journal of Experimental Psychology: General, 134, 596–605.
Heit, E., & Rotello, C. M. (2005). Are there two kinds of reasoning? In Proceedings of
the Twenty-Seventh Annual Conference of the Cognitive Science Society. Hillsdale, NJ:
Lawrence Erlbaum Associates.
Heit, E., & Rubinstein, J. (1994). Similarity and property effects in inductive reason-
ing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 411–
422.
Hempel, C. G. (1966). Philosophy of natural science. Englewood Cliffs, NJ: Prentice Hall.
Holland, J. H., Holyoak, K. J., Nisbett, R. E., & Thagard, P. (1986). Induction: Processes
of inference, learning, and discovery. Cambridge, MA: MIT Press.
Horwich, P. (1982). Probability and evidence. Cambridge, UK: Cambridge University
Press.
Howson, C., & Urbach, P. (1993). Scientific reasoning: The Bayesian approach. Chicago:
Open Court.
Hume, D. (1777). An enquiry concerning human understanding. Oxford: Clarendon Press.
Johnson-Laird, P. (1983). Mental models. Cambridge, MA: Harvard University Press.
P1: JZP
0521672443c01 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:24

24 Evan Heit

Johnson-Laird, P. N. (1994). Mental models and probabilistic thinking. Cognition, 50,

189–209.
Johnson-Laird, P. N., Legrenzi, P., Girotto, V., Legrenzi, M. A., & Caverni, J. P. (1999).
Naive probability: A mental model theory of extensional reasoning. Psychological
Review, 106, 62–88.
Lo, Y., Sides, A., Rozelle, J., & Osherson, D. (2002). Evidential diversity and premise
probability in young children’s inductive judgment. Cognitive Science, 16, 181–206.
Lopez, A., Gelman, S. A., Gutheil, G., & Smith, E. E. (1992). The development of
category-based induction. Child Development, 63, 1070–1090.
Lopez, A., Atran, S., Coley, J. D., Medin, D. L., & Smith, E. E. (1997). The tree of life:
Universal and cultural features of folkbiological taxonomies and inductions. Cognitive
Psychology, 32, 251–295.
Medin, D. L., Coley, J. D., Storms, G., & Hayes, B. K. (2003). A relevance theory of
induction. Psychonomic Bulletin & Review, 10, 517–532.
Nagel, E. (1939). Principles of the theory of probability. Chicago: University of Chicago
Press.
Oaksford, M., & Chater, N. (1994). A rational analysis of the selection task as optimal
data selection. Psychological Review, 101, 608–631.
Osherson, D. N., Smith, E. E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Osherson, D. N., Perani, D., Cappa, S., Schnur, T., Grassi, F., & Fazio, F. (1998). Distinct
brain loci in deductive versus probabilistic reasoning. Neuropsychologia, 36, 369–376.
Poletiek, F. (2001). Hypothesis testing behaviour. Hove, UK: Psychology Press.
Popper, K. R. (1963). Conjectures and refutations: The growth of scientific knowledge.
London: Routledge.
Proffitt, J. B., Coley, J. D., & Medin, D. L. (2000). Expertise and category-based induction.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 811–828.
Rips, L. J. (2001). Two kinds of reasoning. Psychological Science, 12, 129–134.
Salmon, W. C. (1984). Scientific explanation and the causal structure of the world. Prince-
ton, NJ: Princeton University Press.
Shepard, R. N. (1987). Towards a universal law of generalization for psychological
science. Science, 237, 1317–1323.
Skyrms, B. (2000). Choice and chance: An introduction to inductive logic. (Fourth edition).
Belmont, CA: Wadsworth.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psychological
Bulletin, 119, 3–22.
Stanovich, K. E. (1999). Who is rational? Studies of individual differences in reasoning.
Mahwah, NJ: Lawrence Erlbaum Associates.
Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalization, similarity, and Bayesian
inference. Behavioral and Brain Sciences, 24, 629–641.
Wayne, A. (1995). Bayesianism and diverse evidence. Philosophy of Science, 62, 111–121.
Weisberg, R. W. (1986). Creativity: Genius and other myths. New York: Freeman.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning

Brett K. Hayes

From one perspective virtually every cognitive act carried out by young chil-
dren involves induction. Compared to older children and adults, young chil-
dren have little background knowledge about the world and have only a shaky
grasp of the rules that govern propositional reasoning. They are continually
faced with the task of making inferences based on their past observations
and experience. This is true whether the child is trying to determine which
kinds of animals have red blood cells, whether a specific tool can be used to
assemble a bike, or whether the new kid who has moved next door is likely
to be friendly. This chapter focuses on a particularly important aspect of the
development of inductive reasoning – the way that children use their under-
standing of categories and category structure to generalize attributes from
familiar to novel instances.
Category-based induction typically involves three components. First, ob-
serving that an inductive base or “premise” has the property P (e.g., that a
shark has fins); second, deciding that X and a target or “conclusion” item Y
are related in some way (e.g., that a shark and a trout are both fish); and third,
inferring whether Y shares property P.1 Investigating children’s inductive rea-
soning allows us to determine when this reasoning ability first emerges and
to chart important age-related changes in induction. It is also a valuable tool
for understanding the development of children’s category representations. In
fact, some important theories of conceptual development (e.g., Carey, 1985)
have been based, for the most part, on the results of studies of age changes in
children’s inductive generalizations.
This chapter will review and discuss work on the development of category-
based induction from infancy through early and later childhood. The first part

1 Throughout this chapter the terms “base” and “premise,” and “target” and “conclusions” are
considered synonymous and will be used interchangeably.

25
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

26 Brett K. Hayes

of the chapter will consider just what kinds of information infants and children
use to decide whether the properties of one category generalize to other
categories. To preface the main conclusions it will be argued that, by five years
of age, children seem aware of a number of different ways in which two or more
categories can be related, and they can use these different kinds of relations
as a basis for inductive inference. In addition to relatively obvious relations
such as the perceptual similarity of the inductive base and target, young
children use taxonomic, hierarchical, and causal principles to guide property
inferences. They are also sensitive to the fact that inferences often depend on
the kind of property being considered. More complex heuristics that require
the integration of information across multiple instances or premises (e.g., the
premise diversity principle described in Chapter 1) may emerge somewhat
later in development. However, even here recent evidence suggests that school-
age children are better at using these principles than was once thought.
The second part of the chapter will examine the implications of this de-
velopmental research for general theories of induction, highlighting the ways
that the developmental data can be used to evaluate and constrain competing
theories. The third and final part will consider the best way of explaining de-
velopmental change in induction. Again, to preface the main argument, the
current evidence suggests that many of the fundamental processes involved
in inductive reasoning are developmentally stable from the preschool years
and that the most significant age-related changes in induction may be due to
the growth of more sophisticated knowledge about objects, object properties,
and category structures.

.     



Perceptual Similarity
Classic theories of cognitive development emphasized the role played by
perceptual similarity in determining early categorization and induction
(e.g., Bruner, Olver, & Greenfield, 1966; Inhelder & Piaget, 1958; Vygotsky,
1934/1986). Subsequent work has confirmed that the influence of perceptual
similarity on property inferences emerges at an early point in development.
Baldwin, Markman, and Melartin (1993), for example, allowed infants aged
between nine and sixteen months to inspect an object (the inductive base) that
had an unusual and unexpected property (e.g., it wailed when it was shaken
or tilted). Infants were then shown other toys that varied in their perceptual
similarity to the base. In the critical test condition these were modified so
that they could not generate the unexpected sound. Nevertheless, even the
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 27

youngest infants persistently tried to reproduce the unexpected property with

the toys that resembled the base. They rarely generalized the property to ob-
jects with a different colour or surface pattern (also see Graham, Kilbreath, &
Welder, 2004; Rakison & Hahn, 2004).
This influence of perceptual similarity on induction continues throughout
infancy and childhood (Farrar, Raney, & Boyer, 1992). Indeed, even most
major theories of adult induction (e.g., Osherson et al., 1990; Sloman, 1993)
acknowledge that similarity, in terms of the overlap in structural features
between an inductive base and target, plays a major role in determining the
strength of inductive inferences.
A more contentious issue though is whether perceptual similarity is the
primary principle that governs early induction. Many early developmental
theories assumed that young children’s reasoning was “perceptually bound” in
that perceptual similarity overshadowed consideration of logical or taxonomic
relations (Inhelder & Piaget, 1958; Springer, 2001). Much of the work reviewed
in the following sections challenges this view.

Taxonomic Relations
One of the most consistent themes in this book is that people are sensitive
to the fact that the members of closely related categories are likely to share
novel, non-obvious properties. It is important therefore to establish just when
and how children first make use of shared taxonomic relations as a basis for
inductive inference. The most common approach to studying this issue has
been to examine whether infants and children are more likely to generalize
a novel property from an inductive base to target when these are given the
same noun label.
studies with infants. Welder and Graham (2001) examined the effects of
verbal labels and perceptual similarity on induction by sixteen- to twenty-one-
month-old infants. Infants were shown a base object with a novel property
(e.g., a cloth-covered toy that squeaked when squeezed). This was presented
together with either a novel noun label (e.g., “Look at this blint”) or without a
label (e.g., “Look at this one”). Infants were then shown test objects that varied
in their similarity to the shape of the base, with the label present or absent.
When there was no label infants relied on shape similarity in generalizing the
property. In the shared label conditions generalization of the novel property
depended on both shared labels and shape. This same pattern of inference
has been shown with thirteen-month-olds who are just beginning to acquire
productive language (Graham et al., 2004).
Many have interpreted such findings as evidence that infants understand
that noun labels supply information about category membership and that
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

28 Brett K. Hayes

members of the same categories are likely to share non-obvious properties

(e.g., Gelman, 2003; Graham et al., 2004). An alternative view is that infants
have simply learned that objects with the same label are usually similar in
many other ways (Jones & Smith, 1993; Quinn & Eimas, 2000). It may also
be that category labels simply count as additional perceptual features, albeit
highly salient ones, when children compare base and target instances (Sloutsky
& Fisher, 2004). In the latter cases the argument is that the use of shared labels
in induction does not necessarily imply an understanding of the categorical
implications of the label. This issue will be examined in more depth after
reviewing relevant work with older children.

studies with preschoolers and older children. One of the first sug-
gestions that children look beyond surface similarity when doing induction
came from the work of Carey (1985). Children aged from four to ten years,
and adults, were shown a picture of a base item such as a person or other ani-
mal and were told that it had an unobservable novel property (e.g., “a spleen
inside it”). Carey then presented a range of other objects, including dogs,
bees, and flowers, as well as inanimate kinds like rocks, and asked people to
decide which ones shared the novel property. Although children’s inferences
were often influenced by the similarity between the base and target items,
Carey also found some important exceptions. For example, young children
were more likely to generalize a biological property of humans to “bugs” than
from bees to bugs. Carey explained this finding by arguing that humans were
regarded by young children as a kind of ideal or prototypical animal, so that
their properties were highly likely to be shared by other animals (see Chap-
ter 3 by Medin and Waxman for further discussion).
In a series of studies spanning two decades, Susan Gelman and her col-
leagues have tried to systematically separate out the effects of taxonomic and
perceptual factors on young children’s induction (see Gelman, 2003, for a
review). Gelman and Markman (1986), for example, used a triad task in
which four-year-olds were taught a novel property about a base picture (e.g.,
a tropical fish) with a familiar category label (“fish”) and shown two target
pictures. One of these was perceptually similar to the base but had a different
label (e.g., dolphin). The other looked less like the base but had the same
label (e.g., a picture of a shark labelled as a “fish”). Children were asked to
choose which of the targets was more likely to have the novel property. The
majority of four-year-olds chose the taxonomic rather than the perceptual
match. Subsequent work has found the same pattern of inference in children
as young as two years of age (Gelman & Coley, 1990). Preschoolers’ preference
for inferences based on category labels over perceptual appearances has been
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 29

found with realistic photographs as well as line drawings of animals (Deák &
Bauer, 1996; Florian, 1994), and extends to artifact (Gelman, 1988) and social
categories (e.g., Diesendruck, 2003; Heyman & Gelman, 2000a, b).
As we saw earlier there are at least two ways to interpret such findings. First,
labels can have an indirect effect on inductive inferences by serving as a cue
that the base and target instances belong to the same category. Alternatively,
common labels could have a direct effect by serving as identical auditory
features making the premise and conclusion items more perceptually similar.
A series of control studies support the argument that young children use
shared category membership rather than just the perceptual aspects of shared
labels as the basis for their inferences. Preschool children reliably gener-
alize novel properties from premise to conclusion items with perceptually
dissimilar but synonymous labels (e.g., “bunny” and “rabbit”) (Gelman &
Markman, 1986). Critically, however, preschool children do not use shared
labels as a basis for induction when these labels lack information about cat-
egory membership. Gelman and Coley (1990), for example, found that the
presence of a shared label referring to a transient state like “is sleepy” did not
promote the generalization of biological properties. Similarly, young children
do not use shared proper names like “Anna” as a basis for generalizing be-
havioural traits, presumably because they recognize that such names refer to
individuals rather than members of a social category (Heyman & Gelman,
2000a). Young children do not generalize properties on the basis of shared
labels when these conflict with existing category knowledge (e.g., when catlike
animals are labelled as “dogs”) (Jaswal, 2004), or when they cross superordi-
nate category boundaries (e.g., when the same label is given to animals and
artifacts) (Davidson & Gelman, 1990).
In some cases labels may not be necessary for preschool children to gen-
eralize properties along taxonomic lines. Deák and Bauer (1996) presented
young children with realistic photographs of premise objects (e.g., a panther)
and conclusion alternatives from the same basic-level categories (e.g., a Tabby
house cat) or a different category (e.g., a black horse). The different-category
photo was always more similar in appearance to the premise than the same-
category item. Even when no labels were given, children were more likely
to generalize novel properties based on shared category membership than
on overall perceptual appearance. Children used the information contained
in the photographs to infer that the panther and cat belonged to the same
biological category, and they used this relationship rather than appearance as
a basis for property induction.
A final line of evidence suggesting that young children use taxonomic re-
lations in induction is their understanding that typical category members
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

30 Brett K. Hayes

support stronger inferences than less typical instances. López, Gelman,

Gutheil, and Smith (1992) showed that kindergarten children were more
likely to generalize a biological property from “dogs” to “animals” than from
“bats” to “animals” (also see Carey, 1985, Lo et al., 2002). López et al. presented
the “animal” category only verbally, so it is hard to see how this result could
be explained in terms of the perceptual similarity of base and target items.2
Taken together, this is strong evidence that linguistic or structural
cues suggesting the shared category membership of base and target items
strengthens children’s inductive inferences. Common noun labels can serve
as a powerful cue to promote inference by promoting a belief in shared
category membership.
Given the strong evidence that preschool children use labels as markers
of category membership during induction, is there evidence that infants use
labels in the same way? The kinds of controlled comparisons reviewed above
that reveal the contexts in which children do or do not use shared labels
for induction have yet to be carried out with infants. At least two kinds of
evidence, however, suggest that by two years of age children have some grasp of
the categorical implications of shared labels. First, two-year-olds understand
that when adults use novel noun labels they are usually referring to classes of
objects and that labels serve a different linguistic function from other kinds
of utterances such as adjectives or expressions of personal preference (e.g.,
Henderson & Graham, 2005; Waxman & Booth, 2001). Second, studies of
name extension suggest that, by eighteen months of age, infants understand
that labels convey information about the kind of thing an object is. Booth,
Waxman, and Huang (2005), for example, presented eighteen- and twenty-
four-month-olds with a novel object called a “dax” together with vignettes that
implied that daxes were either animals (e.g., “has a mommy who loves it very
much”) or artifacts (e.g., “was bought at a store”). Infants exposed to different
vignettes attended to different kinds of perceptual features when generalizing
the label to other objects. In other words infants were interpreting the label
as implying membership of a particular global category and generalized the
label accordingly.3

2 It should be noted that López et al. (1992) offer an alternative explanation of typicality effects,
arguing that they can arise through an assessment of the similarity of the features of premise
and conclusion items, without the activation of categorical information.
3 A little caution is required in drawing implications about the development of property inference
from studies using a name extension paradigm. Adults and children usually distinguish names
from other object properties and may generalize each in different ways (Markman & Ross, 2003).
Nevertheless, the cited studies establish that infants understand the categorical implications of
shared labels.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 31

Hierarchical Relations
Taxonomic knowledge also entails an understanding that objects are embed-
ded in hierarchies based on class inclusion relations. Daffodils are understood
to be a kind of flower which is, in turn, a living thing. Like older children
and adults, preschool children are most likely to generalize a novel property
from a category to other items at the same hierarchical level, and they show
decreasing generalization to test items located at more abstract levels (Carey,
1985; Gelman & O’Reilly, 1988; Inagaki & Hatano, 1996).
An important finding in other conceptual tasks like categorization and
naming is that certain hierarchical levels are psychologically basic or “privi-
leged” (Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976). For example,
children usually acquire the names of objects at intermediate levels of a hi-
erarchy (e.g., apple) before they learn the names for things located at a more
subordinate (e.g., Delicious apple) or superordinate level (e.g., fruit). In the
case of induction the privileged level can be defined as the highest level in a
hierarchy where there is a strong belief that the properties of objects at that
level can be generalized to an immediate superordinate.
Coley, Hayes, Lawson, and Moloney (2004) examined the location of the
privileged level within a biological hierarchy in five-year-olds, eight-year-olds,
and adults. All groups were given two kinds of conceptual tasks; feature listing
and inductive inference. In the first task participants listed the features they
thought were shared by the members of each of four hierarchical levels (e.g.,
animals, birds, sparrows, house sparrows). In the induction task participants
were taught a novel property about an instance from a given hierarchical level
and asked whether this generalized to categories at the immediate superor-
dinate level (e.g., from house sparrows to sparrows, from sparrows to birds,
from birds to animals). The key finding was that, for eight-year-olds and
adults, the location of the privileged level often differed across the two tasks.
In feature listing the highest level in the hierarchy where objects were believed
to share many features was the life-form (e.g., bird, fish, tree). For adults and,
in many cases, for eight-year-olds, the privileged level for induction was the
more specific folk-generic level (e.g., sparrow, maple). Adults and eight-year-
olds generalized novel properties from very specific biological kinds such as
house sparrow and sugar maple to sparrows and maples, respectively, but
they were less likely to generalize from sparrows to birds or maples to trees.
Five-year-olds showed a different pattern with the life-form level privileged
for both feature listing and induction in the majority of biological categories.
The task discrepancies in the location of the basic level show that by eight
years of age children often rely on different sorts of information for feature
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

32 Brett K. Hayes

listing and induction. Feature listing reflects knowledge about category mem-
bers. Adults and children knew more about the kinds of features shared by the
members of life-form categories than they did about more specific categories.
But their inferences reflected an expectation that features would cluster at the
folk-generic level. One source of such expectations is the nomenclature for
members of subordinate categories. In English (and most other languages)
levels below the folk-generic are marked as subtypes by the use of compound
names (e.g., desert oak). Coley et al. (2004) show that by eight years of age
children recognize the implications of such subordination for inductive in-
ference. An understanding of linguistic hierarchies alone, however, is unlikely
to explain the emergence of a privileged level for induction. In a related study
with adults, Coley, Medin, and Atran (1997) found that the privileged level
for induction remained at the folk-generic level even when category labels
contained a reference to the life-form level (e.g., goldfish, catfish). Expecta-
tions about feature clusters at the folk-generic level therefore may also reflect a
more abstract belief that category members at this level share some underlying
biological property that gives rise to those features.
More research is needed to clarify the factors that lead certain hierarchical
levels to be privileged in children’s induction. Nevertheless, the findings of
Coley et al. (2004) highlight the importance of children’s beliefs about category
hierarchies in induction and indicate that these beliefs involve more than
just an appreciation of the common and distinctive features of instances at
different hierarchical levels.

Relations Involving Multiple Premises: Diversity and Monotonicity

When drawing inductive inferences adults are able to integrate information
across more than one premise. Adult reasoners, for example, are sensitive to
the effects of sample size or “premise monotonicity.” In general, the more
premise categories that share a particular property, the more likely it is that
the property will be generalized to other members of the same superordinate
category (Osherson et al., 1990). Adults are also sensitive to premise diversity.
Cats and buffalos sharing some property, for example, is seen as a stronger
basis for generalization to other animals than if the property was shared by
cows and buffalos. The conventional explanation for this effect is that the
more diverse set samples a greater range of the possible exemplars of the
relevant superordinate and therefore provides a stronger base for induction
(Osherson et al., 1990).
Intuitively, combining information across multiple premises seems to
involve more complex processing than many of the inductive arguments
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 33

discussed so far. Perhaps not surprisingly, therefore, early investigations of

monotonicity- and diversity-based reasoning in children (e.g., Carey, 1985;
Gutheil & Gelman, 1997; López et al., 1992) suggested that an understanding
of these principles emerged rather late in development. López et al. (1992),
for example, examined the generalizations made by five- and eight-year-olds
presented with a variety of single- and multiple-premise arguments patterned
after those used by Osherson et al. (1990). They found that young children
were able to draw appropriate conclusions from certain kinds of arguments
that included multiple premises. Five-year-olds, for example, judged that a
property shared by horses and zebras was more likely to generalize to donkeys
than to squirrels. Young children also exhibited “nonmonotonic” reasoning
for some pairs of arguments. That is, adding a premise to an argument could
weaken inductive strength if the added premise was not a member of the con-
clusion category (e.g., children were more likely to project a property from
lions to animals than from lions and apples to animals). An appreciation of
the role of premise monotonicity and premise diversity, however, took con-
siderably longer to develop. Eight-year-olds, but not younger children, used
monotonicity and premise diversity when the conclusion category was a more
general one like “animals” that included the premise categories, but not when
the conclusion was at the same level as the premises (e.g. generalizing from
cats and buffalos to kangaroos) (also see Gutheil & Gelman, 1997; Lo, Sides,
Rozelle, & Osherson, 2002).
Lopez et al. (1992) argued that the developmental progression from no
understanding of monotonicity and diversity at age five, to an understand-
ing of general-conclusion forms at around eight years, to an understanding
of both general and specific forms in adulthood, fits well with Osherson
et al.’s (1990) Similarity-Coverage model of induction. According to Osher-
son et al. understanding arguments involving premise-conclusion similarity
involves a relatively straightforward computation of the similarity between
the features of the premise and conclusion categories. Monotonicity and di-
versity, however, should be more difficult for young children because they
require the additional process of “coverage” which involves computing the
average maximum similarity of the premise categories to samples of the
lowest-level category which includes both the premises and conclusions. In
the general forms of these arguments the conclusion category is given in the
argument. In the specific forms it has to be generated, a process that is likely
to be particularly demanding for younger children (Markman & Callanan,
1984).
It is possible, however, that the negative conclusions regarding children’s
understanding of these more complex forms of induction may reflect their
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

34 Brett K. Hayes

difficulties with the performance aspects of the tasks used by Lopez et al.
(1992) and Gutheil and Gelman (1997). Heit and Hahn (2001) suggest that
young children may have difficulty in applying principles like diversity and
monotonicity to hidden or invisible properties (e.g., “has leucocytes inside”).
Moreover, Lopez et al. (1992) required children to respond to a variety of
inductive arguments (e.g., typicality, similarity, monotonicity, and diversity)
administered in the same testing session. To perform well children would have
to not only understand each kind of argument, but also switch between induc-
tive strategies. Heit and Hahn (2001) tried to overcome these performance
barriers by examining five- and nine-year-olds’ generalization of concrete,
observable properties from more or less diverse premise sets. For example,
children were shown photographs of a diverse set of three dolls and told that
they belonged to a girl named Jane. They were also shown a set of three very
similar-looking dolls and told that they belonged to a girl named Danielle.
They were then shown a new doll and asked who was the most likely owner.
For observable properties, even five-year-olds reliably based their inferences
on the diverse rather than the non-diverse premises. When hidden properties
such as “has nuts inside” were used, only the nine-year-olds were sensitive to
premise diversity.
In a similar vein Hayes and Kahn (2005) re-examined sensitivity to premise
monotonicity in five- and nine-year-olds using items that always involved a
choice between inferences based on a small or a large premise set. For example,
children were presented with a small set of two grey turtles which shared a
property such as “eats snails” or a larger set of six brown turtles that shared a
different property such as “eats worms.” The premises were presented with the
cover story that they were samples from a larger inclusive category (i.e., both
kinds of turtles came from the same zoo). Children were then asked to decide
which property was most likely to be found in either another animal at the
same hierarchical level (e.g., a turtle whose colour was unknown) or in most
of the other animals of the same type at the zoo. Nine-year-olds were more
likely to generalize the property of the larger set to both specific and general
types of conclusions. Younger children did so only for the general conclusions
(see Lo et al., 2002, for additional positive findings regarding children’s use of
monotonicity).
These data suggest that, by five years of age, children are able to use inductive
heuristics like premise monotonicity and diversity. Exactly how they do this
remains to be explained. One interpretation is that an ability to compute
coverage develops by around five years of age. On the other hand it could
be that Osherson et al.’s assumptions about how people process multiple
premises in diversity and monotonicity tasks need to be revised. If young
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 35

children have some understanding of these principles, as well as effects like

premise-conclusion similarity and premise typicality, this may imply that
similar kinds of computational processes subserve each of these phenomena.
This possibility will be considered in more detail in Part II.

Property Knowledge
In all of the cases considered so far the focus has been on how children use
various kinds of relations between premise and conclusion categories to draw
inferences about properties. Unlike most of the items used in experimental
induction tasks, however, most cases of everyday induction involve some
knowledge of the properties to be inferred, as well as the relevant categories.
Whether a property is generalized from one bird to another depends on
whether the property is an internal organ, nesting pattern, or type of birdsong.
There is plenty of evidence that adult inductive inference varies as a function of
both knowledge about the categories and the predicate in question (e.g., Heit
& Rubinstein, 1994; Nisbett, Krantz, Jepson, & Kunda, 1983). The evidence
reviewed below suggests that preschool children, and possibly younger infants,
are also sensitive to category-property interactions.

infant studies. Infants’ understanding of how certain properties might

generalize across some categories but not others was examined by Mandler and
McDonough (1996, 1998). Infants aged between nine and fourteen months
were presented with instances of different superordinate categories (e.g., a dog
and a car). Experimenters then modelled different kinds of actions for each
instance (e.g. giving a drink to the dog or riding on the car). Infants as young
as nine months of age generalized the actions observed with the base animal to
other animals, and the actions observed with the base vehicle to other vehicles.
However, they seemed to respect the boundaries between global categories
and did not imitate “animal” actions in the presence of vehicles or vice versa.
Furthermore, Mandler and McDonough (1998) found that fourteen-month-
olds distinguished between actions that were appropriate to only animate or
only inanimate kinds from those that are appropriate for both domains (e.g.,
being washed). They concluded that infants form global concepts of animals
and vehicles and have some understanding of the characteristic properties
within each domain.
These conclusions, however, have been challenged on both empirical and
theoretical grounds (see McClelland & Rogers, 2003; Rakison & Hahn, 2004
for critiques). One problem is that Mandler and her colleagues may have
underestimated the role played by perceptual similarity in driving induction.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

36 Brett K. Hayes

Although the members of her global categories were perceptually diverse, it

could be argued that the members of the animal category were still more
similar to one another than they were to vehicles, and vice versa. Hence, the
generalization to items from the same superordinate may in part be driven by
perceptual similarity. Some aspects of Mandler’s data are consistent with this
argument. Although Mandler and McDonough (1998) found that domain-
general properties like “being washed” were generalized to both animals and
vehicles, the test object which infants imitated first was usually the one that
was most similar in perceptual terms to the model.
Rakison and Hahn (2004) suggest that infants’ inductive inferences are
driven by their knowledge and expectations about the functional properties
of object features rather than an understanding of the properties associated
with different superordinate categories. In a variant on the Mandler and Mc-
Donough (1998) procedure, eighteen-month-olds were shown models that
moved along a non-linear path (thought to be characteristic of animate mo-
tion), or a linear path (thought to be characteristic of inanimate motion).
Eighteen-month-olds generalized these actions based on a functional rela-
tionship with specific object parts rather than along taxonomic lines. For
example, when the premise item was a cat who walked along a winding path,
infants were more likely to imitate this action with an inanimate object that
had a perceptual feature associated with walking (e.g., a bed with four legs)
than with animals that lacked these features (e.g., a dolphin).
Such findings were seen to point to the primacy of perceptual features
(i.e., object parts) over superordinate relations in infant induction (Rakison
& Hahn, 2004). However, it needs to be noted that the infants in these studies
were using perceptual information in a relatively sophisticated way. They
understood that the ability to move in a particular way is causally dependent
on having certain kinds of structural features like legs. Moreover, they used
this knowledge as a basis for inductive inference. So by eighteen months of age
infants are aware of differences in the generalizability of functional properties,
although this awareness seems to be based on their knowledge of object parts
rather than category membership.

studies with preschoolers and older children. By three and a half to

four years of age children are aware that certain properties are more generaliz-
able than others. Children do not generalize transient properties like “is cold”
or accidental properties like “fell on the floor this morning” even when the
base and target have the same name and similar appearance (Gelman, 1988;
Gelman & Markman, 1986, 1987; Springer, 1992). Young children believe that
characters labelled with the same personality traits share novel behavioural
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 37

properties such as “likes to play tibbits” but not transient properties such as
“feeling thirsty” (Heyman & Gelman, 2000a).
Children also appear sensitive to certain kinds of specific property-category
relations. In an ingenious study, Kalish and Gelman (1992) presented three-
and four-year-olds with novel objects that had two labels, one that referred to
object kind and one that described the material that the object was made from.
In the critical experimental condition these labels supported different kinds
of inferences. For example, whether “glass scissors” are judged to be fragile
depends on which of the two labels is seen as more relevant to the judgment.
Kalish and Gelman found that young children were sensitive to the kinds of
labels that were relevant to generalizing different sorts of properties. When
asked whether objects would share a novel dispositional property (e.g., “will
fracture in water”), young children based their judgments on the material
label. When asked whether they shared a novel functional property (e.g.,
“used for partitioning”), children used the object label.
Kelemen, Widdowson, Posner, Brown, and Casler (2003) have also shown
that the direction of children’s inferences within biological categories depends
on the kind of property that is being generalized. Three- and four-year-
olds based inferences about which animals had similar behavioural features
(e.g., “fights off dangerous animals”) on shared body parts relevant to this
behaviour (e.g., having horns) rather than on overall perceptual similarity.
This pattern reversed when children were generalizing a novel category label
(cf. McCarrell & Callanan, 1995).
With age there appears to be an increasing appreciation of the role of
category-property interactions in induction. Gelman (1988) presented four-
and eight-year-olds with natural kinds such as flowers and artifacts such
as chairs. Children were asked to decide whether the properties of these
objects generalized to other natural kinds and artifacts. Some children were
asked about properties which (from an adult perspective) were thought to
be most generalizable within the domain of natural kinds (e.g., “needs CO2
to grow”), while others were asked about properties thought to be more
generalizable for artifacts (e.g., “you can loll with it”). Control conditions
also examined properties thought to be generalizable to neither domain or
to be equally generalizable to both. The older children were sensitive to the
relations between categories and properties, being more likely to generalize
“natural kind properties” to natural kind exemplars than to artifacts, and vice
versa. Four-year-olds, however, did not discriminate between the properties
typically associated with the two domains.
One way of interpreting the developmental trends described by Gelman
(1988) is that sensitivity to category-property relations in induction increases
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

38 Brett K. Hayes

as children learn more about the causal principles that operate in different
conceptual domains and the specific properties associated with exemplars
from each domain. Although a direct test of this developmental hypothesis
has yet to be carried out, indirect support comes from studies of the effects
of expertise on adult reasoning. Shafto and Coley (2003) found that fish
experts used causal and ecological relations in generalizing familiar properties
like “has disease X.” When generalizing properties about which they knew
little (e.g., “has sarca”), both experts and novices relied on similarity-based
strategies. Hence, increasing domain knowledge is associated with greater use
of inductive reasoning based on the interactions between specific categories
and properties.

Causal Relations
The previous section showed that children’s background knowledge of both
categories and properties plays an important role in their inductive inferences.
Exactly how does such knowledge interact with more “domain-general” prin-
ciples such as premise-conclusion similarity, monotonicity, diversity, and so
on? Recent studies with adults suggest that when people have some knowl-
edge about the relations between premises this can override more general
inductive principles. Medin, Coley, Storms, and Hayes (2003) showed that
diversity-based reasoning is undermined when people perceive a salient or
distinctive relation between the premises that is unlikely to be shared by a
conclusion category. Adults, for example, judged that a novel property shared
by fleas and butterflies was more likely to generalize to sparrows than a prop-
erty shared by dogs and fleas, even though they saw the second set of premises
to be more diverse. In other words, they were aware of the ecological relations
between dogs and fleas and judged that this special relation was unlikely to
generalize to other animals (but see Heit and Feeney, 2005, for a different
interpretation).
An important goal for studies of the development of inductive reasoning
is to specify how children integrate their causal knowledge with an under-
standing of more general inductive principles, and to examine whether this
process changes as children develop more knowledge of the world. As a first
step in this direction Thompson and Hayes (2005) devised a task in which
five-year-olds, nine-year-olds, and adults were forced to choose between in-
ductive inferences based on shared causal relations or featural similarity (cf.
Lassaline, 1996). The basic procedure is illustrated in Figure 2.1. Attribute base
items were fictional animals made up of three features that were not linked via
explicit causal relations but which could plausibly co-exist. Causal base items
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 39

Attribute Base Causal Base

The following are true of Animal A: The following are true of Animal B:

• Has long fur • Has a large protective shell

• Has claw-like feet • Can stay in the sun without getting
• Eats leaves from trees burned
• The protective shell allows the animal
to stay in the sun without getting
burned

Target
The following facts are true of Animal C

• Has long fur

• Has claw-like feet
• Has a large protective shell

Induction Test

Animal C has one other feature. Do you think it:

Eats leaves from trees?
OR
Can stay in the sun without getting burned?

figure 2.1. Schematic stimulus example from Thompson and Hayes (2005).

contained two features that could be linked by causal relations, together with
a statement of the direction of causation. All features were illustrated with
line drawings. After presentation of the bases, a target instance was described.
As the table shows, this always had a higher featural similarity to the attribute
base but contained the antecedent feature from the causal base. Participants
had to predict which of two other properties was most likely to be found in
the target, the third feature of the attribute base or the consequent feature
from the causal base.
The main results are summarized in Figure 2.2. The proportion of “causal”
choices at test exceeded chance for each age group. There was also a clear
increase in the proportion of causal choices with age. When children’s justifi-
cations for their choices were considered, there was even more robust evidence
for the early development of an awareness of the role of causal relations in
induction. Five-year-olds mentioned the causal relation in over forty percent
of justifications. By comparison, explicit appeals to feature similarity as a basis
for induction were relatively rare for all age groups. This shows that causal
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

40 Brett K. Hayes

0.9

0.8
Proportion of causal choices

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
5 year olds 9 year olds Adults
Age

figure 2.2. Mean proportion of causal choices on the Thompson and Hayes (2005)
causal induction task.

relations between the features of premise and conclusion items can override
featural similarity as a basis for induction, even in young children. It remains
to be shown whether these results generalize to other category domains like
artifacts and social kinds. Further work is also needed to establish just which
aspects of causal relationships lead children to make stronger inductions.

Summary and Discussion

This review shows that, by the preschool years, children make inductive infer-
ences based on a variety of premise-conclusion relations. As well as perceptual
and taxonomic relations, children are sensitive to the role of hierarchical and
causal relations, as well as the ways that these interact with certain kinds of
property knowledge. There is some evidence that by five years children may
use general heuristics like diversity and monotonicity. Even preverbal infants
appear capable of going beyond perceptual similarity in their generalization
of novel properties.
While these demonstrations of precocious induction are impressive, they
need to be viewed within a broader perspective. An important agenda driving
much of the research in this field has been to show that young children are not
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 41

perceptually bound in their reasoning. To achieve this researchers often place

perceptual cues in conflict with other cues like category labels and force
children to choose one cue as a basis for inference. This approach can reveal
the dominant cue for induction only in a given situation. It tells us little about
how perceptual and conceptual cues might interact. Such interactions are
likely to be important since in everyday induction it is not always easy to draw a
clear distinction between “perceptual” and “conceptual” relations (Goldstone,
1994; Murphy, 2004). Moreover, the more perceptually obvious similarities
between base and target items often provide useful clues to underlying causal
or taxonomic relations (Gentner, 1988). As Madole and Oakes (1999) put it,
“Objects that look like dogs also tend to sound like dogs, move like dogs, and
smell like dogs. Moreover, they usually are dogs” (p. 270 [original emphasis]).
One implication is that although the various relations that influence chil-
dren’s induction have been treated separately in this review, this is not the
way that they always appear in the child’s (or even the adult’s) world. Many
of the more abstract or “conceptual” aspects of categorization involve under-
standing and explaining the links between observable features, such as why an
object that has wings and feathers is likely to fly. Given this interdependence
of perceptual and conceptual features, it makes little sense to ask whether
children rely on “perceptual,” “taxonomic,” or “causal” relations when they
do induction. In fact, adult studies that have manipulated both the number of
features shared by base and target items and their causal relations have found
that induction is influenced by both kinds of factors (e.g., Lassaline, 1996).
Perhaps the best way of interpreting the studies reviewed here, therefore, is
that they tell us about the kinds of cues that draw children’s attention to-
wards or away from the more observable or more abstract relations between
bases/premises and targets/conclusions during induction.

.     

In addition to revealing the origins and early developmental course of induc-
tion, developmental data may also be useful for evaluating competing process
models of induction. The major inductive phenomena in infants and children
reviewed so far are summarized in Table 2.1. How well do process models
of induction, originally developed to explain adult data, account for these
phenomena? The key issues here are whether existing models can explain the
range of relations that children are capable of using as a basis for inductive
inferences, and whether the models can predict the kinds of factors that direct
children’s attention to particular kinds of relations.
 P1: JZP
0521672443c02

table 2.1. Summary of major category-based induction phenomena in infants and children

Induction Phenomenon By 18 months By 5 years

Premise-conclusion similarity Yes Yes
(Baldwin et al., 1993; Graham et al., 2004; (Carey, 1985; Gelman & Markman, 1986, 1987;
Welder & Graham, 2001) Gelman, 1988; Sloutsky & Fisher, 2004)
Shared labels override perceptual Yes Yes
CUFX144-Feeney 0 521 85648 5

similarity (Graham et al., 2004; Welder & Graham, (Deak & Bauer, 1996; Florian, 1994; Gelman, 1988;
2001) Gelman & Coley, 1990; Gelman & Markman, 1986,
1987; Heyman & Gelman, 2000a, b; Sloutsky &
Fisher, 2004)
Similarity overrides shared labels

42
r When labels conflict with existing —– Yes
knowledge (Davidson & Gelman, 1990; Jaswal, 2004)
r When adjective labels are —– Yes
used (Gelman & Coley, 1990; Gelman & Heyman, 1999)
r When labels are proper names —– Yes
(Heyman & Gelman, 2000a)
Conclusion homogeneity —– Yes
(Carey, 1985; López et al., 1992)
Premise Typicality —– Yes
(Lo et al., 2002; López et al., 1992)
Premise diversity: Specific conclusions —– Yes
(Heit & Hahn, 2001)
July 19, 2007
14:27
 P1: JZP
0521672443c02

Premise diversity: General conclusions —– No

(Lo et al., 2002; López et al., 1992)
Premise Monotonicity (Sample size): —– No
Specific conclusions (Hayes & Kahn, 2005; López et al., 1992)
Premise Monotonicity: General conclusions —– Yes
(Hayes & Kahn, 2005; Lo et al., 2002)
Premise Nonmonotonicity: Specific —– (Yes)
conclusions (López et al., 1992)
CUFX144-Feeney 0 521 85648 5

Premise Nonmonotonicity: General —– Yes

conclusions (López et al., 1992)
Category-property interactions

43
r Transient vs. entrenched — Yes
properties (Gelman, 1988; Gelman & Coley, 1990)
r Property generalization depends on the (Yes) Yes
category domain (Mandler & McDonough, 1996, 1998) (Carey, 1985; Gelman, 1988; Heyman & Gelman,
2000b; Kalish & Gelman, 1992)
Shared functional features promote Yes Yes
inference (McCarrell & Callanan, 1995; Rakison & (Kelemen et al., 2003)
Hahn, 2004)
Shared causal relations promote inference —– Yes
(Thompson & Hayes, 2005)

Note: Yes = Robust evidence for the phenomenon; (Yes) = Some positive evidence but some controversy still exists; No = Robust evidence against the phenomenon;
— = No studies with the relevant age group.
July 19, 2007
14:27
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

44 Brett K. Hayes

The Similarity-Coverage Model (SCM)

As noted above, this model suggests that induction is driven by two principles,
similarity and coverage. Arguments are perceived as strong or convincing
to the extent that 1) the premise categories are similar to the conclusions
and 2) the premises “cover” the lowest-level category that includes both
premise and conclusion categories (Osherson et al., 1990). Because many of
the induction phenomena that have been studied in children were derived
from adult studies of the SCM (e.g., premise-conclusion similarity, typicality,
nonmonotonicity), it is hardly surprising that these can be readily explained
by that model.
Because the SCM model does not discriminate between “perceptual” and
other forms of similarity, it is unclear whether it can explain competition
between shared category labels and perceptual similarity in determining in-
ductive generalization, although this may be possible if additional assump-
tions regarding similarity computations are added (Heit & Hayes, 2005). A
more serious problem for the model is children’s use of knowledge about the
entrenchment or distribution of feature properties to promote or limit gener-
alization (e.g., Gelman, 1988; Kelemen et al., 2003). The SCM model does not
consider knowledge about specific properties and is unable to explain these
effects. Further, because SCM focuses on the global similarity between cate-
gories, it is unable to explain children’s induction based on causal relations
shared by the premise and conclusion.

Feature-Based Induction Model (FBI)

Like SCM, Sloman’s (1993) FBI model emphasizes the similarity of premise
and conclusion categories as a basis for induction. However, it does so without
the retrieval of superordinate categories. Instead, novel properties are gener-
alized on the basis of the overlap between the known features of the premises
and conclusions. Property generalization is proportional to the number of
conclusion category features present in the premises. In terms of explaining
induction in children, FBI shares many of the strengths and weaknesses of
the Similarity-Coverage approach. It gives ready explanations of phenomena
such as premise-conclusion similarity and typicality. By adding assumptions
that shared causal features should be given more weight in similarity com-
putations, it may also be able to account for the causal induction data of
Thompson and Hayes (2005).
Generalization based on knowledge about the distribution of features
within a category, however, lies outside of the scope of the model. Without
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 45

additional assumptions, FBI also has difficulty in explaining children’s sensi-

tivity to nonmonotonicity effects. Moreover, the finding that children’s feat-
ural knowledge does not always predict which hierarchical level will be priv-
ileged for induction (Coley et al., 2004) runs contrary to one of the model’s
key assumptions.

Similarity, Induction, Naming, and Categorization (SINC)

Sloutsky and Fisher (2004) proposed this model to explain developmental
changes in induction, categorization, and naming. In the case of induction, the
probability that a property will be generalized from a base to a target instance
is a function of the similarity of these instances. Similarity is computed both
in terms of the level of overlap in (visual) perceptual features and similarity in
the labels attached to the premise and conclusion items. The relative salience
of these two kinds of features is assumed to change with age, with similar
labels being more salient for younger children. This model differs from other
“similarity-based” models such as SCM and FBI in the different roles accorded
to visual features and labels. It is also different in that the model focuses on
induction over specific instances rather than categories.
So far SINC has been applied only to data from triad tasks like those
developed by Gelman and Markman (1986). Moreover, it does not explain
cases where young children fail to use shared labels as a basis for induction.
It is also not clear how it would explain effects readily accounted for by other
models, such as typicality, diversity, and monotonicity. Like other similarity-
based models, it also offers no principled account of how children’s knowledge
about predicates influences inductive generalization. More generally, despite
the claim that it differs from induction models like SCM that allow for the
operation of taxonomic as well as perceptual relations, SINC makes many
assumptions that parallel closely those made by other models. In particular,
the constraint that shared labels should be given greater weight than shared
visual features is common to many other models of inductive reasoning (Heit
& Hayes, 2005).

Bayesian Models
Heit (1998) proposed a Bayesian model of induction such that generalization
of a novel property from a premise to a conclusion category depends on
prior knowledge about known properties shared by these categories. This
knowledge is used to generate hypotheses about the projection of the novel
property (e.g., whether it is shared by both the premises and conclusion
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

46 Brett K. Hayes

categories or whether it is shared by the premises only). The information in the

premise(s) is then used as evidence to evaluate these hypotheses using Bayesian
rules (cf. McDonald, Samuels, & Rispoli, 1996). Heit (1998) showed that,
with appropriate assumptions about knowledge of feature distribution, the
Bayesian model could explain phenomena like premise-conclusion similarity,
typicality, diversity, and monotonicity. An important finding supporting the
Bayesian approach is that children’s generalization of novel properties follows
the same patterns as their generalization of familiar properties (Gelman,
1988). Because it accords a central role to property knowledge, the model
can also explain children’s sensitivity to differences in the generalizability of
various features within and between categories. Moreover, the model posits
the same basic processes to explain how people evaluate arguments involving
premise-conclusion similarity, diversity, and monotonicity. It can therefore
explain data suggesting that some understanding of premise diversity and
monotonicity is present by age five.
The Bayesian model, however, has difficulty in explaining children’s sensi-
tivity to nonmonotonicity effects. Another limitation is that the validity of the
model rests on its assumptions about children’s and adult’s prior knowledge
about category features. So far, however, there have been no tests of these as-
sumptions in real categories (see Chapter 7 by Tenenbaum, Kemp, and Shafto
for a related approach).

Relevance Theory
With a view to explaining inductive phenomena in which people have varying
levels of knowledge about inductive predicates, Medin et al. (2003) proposed
a relevance framework such that people assume that the set of premises
provided in inductive arguments are informative and will search for distinc-
tive relations between them. For example, if told that koalas, kangaroos, and
wallabies all shared some property, you might infer that the property had
something to do with being an Australian animal (or more specifically, an
Australian marsupial). Moreover, the activation of this distinctive relation
can take place before considering the similarity between the premises and
conclusions. If the conclusion category shared the distinctive relation (e.g.,
if it was another Australian animal), then the novel property is likely to be
generalized. According to this approach, causal relations between premises
are highly distinctive and therefore likely to influence induction.
A detailed process-model of relevance has yet to be described, so it is
difficult to compare this approach with more explicit models like SCM and
FBI. In general terms, however, this approach can explain why features that are
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 47

believed to be causally linked to category membership (e.g., biological features

in animal categories) are more likely to be generalized to other animals than
transient behavioural features. It may also account for the fact that causal
relations can override perceptual similarity in children’s induction. In this
approach an assessment of the featural similarity and taxonomic relations
between premises and conclusions becomes the default if no more distinctive
relations exist, so the approach can also account for the more “standard”
induction phenomena such as premise-conclusion similarity, typicality, and
so on.

Summary and Discussion

It appears that no existing process model gives an entirely satisfactory ac-
count of children’s induction. Models like SCM and FBI provide detailed
explanations of the cognitive processes that lead to effects that arise early
in development, like premise-conclusion similarity and typicality, but fail to
predict when children will or will not use labels as a basis for induction and
to account for differences in inductive projection due to property knowledge.
The Bayesian approach and Relevance theory on the other hand hold some
promise of explaining children’s induction with both blank and non-blank
predicates within a single framework. Both of these approaches, however,
currently lack detailed assumptions and processing constraints. Having said
that, it seems that the broader perspective offered by the Bayesian and Rele-
vance frameworks is likely to be necessary because children are able to flexibly
employ a number of different premise-conclusion relations as a basis for prop-
erty inference. The main goal for future theories of inductive development is
to provide a more detailed process account of these multifaceted inferences
and to specify the factors that lead certain kinds of relations to become more
salient in specific induction contexts.

.     

Looking over Table 2.1, one might be struck by the degree of developmental
continuity found in induction phenomena. By thirteen months of age children
are sensitive to multiple relations between inductive bases and targets such
as perceptual similarity and shared labels. From around five years of age
children’s inductive inferences seem to be influenced by the same broad
range of cues and relations that influence adult judgments. Where there are
age changes these seem to be mainly gradual and quantitative, with certain
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

48 Brett K. Hayes

kinds of information (e.g., causal relations) being used more frequently and
across a wider range of domains.
This final part of the chapter addresses two issues relating to this develop-
mental pattern. First, how do children first acquire an understanding of the
relations that can be used as a basis for induction? Second, where there are
age-related changes in induction, can these simply be explained by the gradual
expansion of children’s background knowledge about categories, exemplars,
and exemplar properties?

Early Acquisition of Inductive Potential

Two different kinds of answers have been proposed to explain the early emer-
gence of induction and the early shifts in the kinds of relations used to
make property inferences. Some have argued that cognitive architectures
which track correlations between object parts, features, and labels can learn
to group instances according to higher-order properties such as animacy
(Loose & Mareschal, 1999; McClelland & Rogers, 2003; Rakison & Hahn,
2004). According to this “bottom-up” approach, learning about the inductive
potential of categories requires only that infants perceive and interact with
(lots of) objects over time. Hence, children shift from inferences based on
appearances to inferences based on labels, not because they appreciate the
conceptual significance of the labels, but because labels provide a more re-
liable basis of predicting the presence or absence of category features across
exemplars (Loose & Mareschal, 1999).
A different kind of answer is that from a very early age children are biased or
constrained to pay special attention to categorical cues like noun labels which
imply a commonality of observed and hidden features (Keil, 1991; Gelman,
2003). Gelman (2003), for example, argues that from a very young age children
hold essentialist beliefs about many categories; that is, they believe that there
is some unobservable property or essence that causes things to be the way
they are and that everyday categories reflect this essentialist structure.4
It is difficult to adjudicate between these competing approaches on the basis
of empirical evidence alone. This is because much of the existing evidence
on children’s induction can be interpreted as in line with either account.
Sloutsky and Fisher (2004), for example, showed that when taught a novel
property about exemplars from an animal category like “cat,” five-year-olds
are less likely than eleven-year-olds and adults to spontaneously generalize this

4 This of course begs the question as to the origins of essentialist biases or beliefs. See Gelman
(2003) for a review of possible answers.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 49

property to other category members. They also show that, with appropriate
training, four-year-olds can readily learn to generalize in the same way as
older children. Sloutsky and Fisher argue that their results show that young
children do not have an a priori assumption that category members share
hidden properties. However, the rapidity with which children learned to
generalize along categorical lines could also be seen as consistent with a bias
which favours category-based induction.
So far “bottom-up” models of category acquisition and induction have
been applied to only a small part of the developmental induction data in
Table 2.1. They can account for children’s increasing reliance on labels as a
basis for induction over time. However, it is unclear how they would explain
why different kinds of labels have different effects on inductive inference, and
why structural and functional features sometimes override labels as a basis for
induction. Such models also offer no principled account of property-category
interactions. It is too early to tell, therefore, whether these approaches will
succeed without the incorporation of additional constraints like an essentialist
bias. Nevertheless, the tension between these approaches has had a positive
effect on the field by stimulating the development of more explicit models of
the mechanisms involved in learning which object features are most important
in promoting induction.

Are There Developmental Changes in Inductive Processes?

There seems little doubt that, in just about all conceptual domains, chil-
dren’s knowledge of perceptual, taxonomic, and causal relations, as well as
the distribution of object properties, expands with age (Farrar et al., 1992).
An important question is whether this expanding knowledge base is suffi-
cient to explain the development of induction. Over and above these changes
in knowledge, is there evidence of significant changes in the fundamental
processes that children use to compare and integrate inductive premises and
conclusions? So far two detailed accounts of such a change have been exam-
ined: Sloutsky and Fisher’s (2004) proposal that there is a shift from treating
labels as perceptual features to treating them as markers of category member-
ship, and Lopez et al.’s (1992) argument that older children are more capable
of computing the “coverage” components of the Similarity-Coverage model.
Neither approach has been shown to give an adequate account of children’s
induction.
This should not, however, discourage us from the goal of devising more
detailed process models of induction and applying these to the study of in-
ductive development. In part, this will require some consideration of how
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

50 Brett K. Hayes

the development of induction fits in with other kinds of age-related changes

in children’s reasoning and cognition. Developmental changes in children’s
ability to complete conditional or deductive reasoning tasks have been cited
in support of more general models of developmental change in children’s
capacity to process more complex relations (e.g., Halford & Andrews, 2004).
By contrast, there have been few attempts to explore the relationship between
the processing mechanisms involved in children’s category-based induction
and those involved in the development of other kinds of reasoning. This is
surprising given that other forms of reasoning are often subject to constraints
that are similar to those that operate in induction. Like category-based induc-
tion, children’s analogical reasoning is influenced by the similarity between
the inductive base and target items and by children’s level of knowledge of the
causal relations involved in the analogy (Goswami, 2002). Some theories
of analogical reasoning also posit developmental change in the factors that
children attend to when drawing analogies. Gentner (1988), for example,
suggests that young children are strongly influenced by the similarity between
the surface features of base and target items but become more sensitive to the
importance of shared relations as they grow older, a developmental change
referred to as a relational shift. It would be interesting to examine the gener-
ality of this shift and, in particular, whether children’s sensitivity to relational
information in induction follows a similar developmental trajectory.

.  

This review of the development of category-based induction highlights the
range and complexity of factors influencing children’s judgements about
whether and how far a property generalizes. Children draw inductive in-
ferences based on a range of possible relations between premises and conclu-
sions, as well as beliefs about the to-be-generalized property. Children derive
information about each of these components from their perceptual analysis
of object features, domain knowledge, and linguistic conventions. Describing
the interaction between these components, and the way that this interaction
changes with age, is a challenging but necessary goal for future models of
induction.


I would like to thank Susan Thompson and Tamara Cavenett for their valuable
comments and assistance in the preparation of this chapter. The research
described in the chapter was supported by Australian Research Council Grant
DP0344436.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 51

References
Baldwin, D. A., Markman, E. M., & Melartin, R. L. (1993). Infants’ ability to draw
inferences about nonobvious object properties: Evidence from exploratory play. Child
Development, 64, 711–728.
Booth, A. E., Waxman, S. R., & Huang, Y. T. (2005). Conceptual information permeates
word learning in infancy. Developmental Psychology, 41, 491–505.
Bruner, J. S., Olver, R. R., & Greenfield, P. M. (1966). Studies in cognitive growth. New
York: Wiley.
Carey, S. (1985). Conceptual change in childhood. Cambridge, MA: MIT Press, Bradford
Books.
Coley, J. D., Hayes, B., Lawson, C., & Moloney, M. (2004). Knowledge, expectations, and
inductive reasoning within conceptual hierarchies. Cognition, 90, 217–253.
Coley, J. D., Medin, D. L., & Atran, S. (1997). Does rank have its privilege? Inductive
inferences within folkbiological taxonomies. Cognition, 64, 72–112.
Davidson, N. S., & Gelman, S. A. (1990). Inductions from novel categories: The role of
language and conceptual structure. Cognitive Development, 5, 151–176.
Deák, G. O., & Bauer, P. (1996). The dynamics of preschoolers’ categorization choices.
Child Development, 67, 740–767.
Diesendruck, G. (2003). Categories for names or names for categories? The interplay
between domain-specific conceptual structure and language. Language and Cognitive
Processes, 18, 759–787.
Farrar, M. J., Raney, G. E., & Boyer, M. E. (1992). Knowledge, concepts, and inferences
in childhood. Child Development, 63, 673–691.
Florian, J. E. (1994). Stripes do not a zebra make, or do they? Conceptual and perceptual
information in inductive inference. Developmental Psychology, 30, 88–101.
Gelman, S. A. (1988). The development of induction within natural kind and artifact
categories. Cognitive Psychology, 20, 65–95.
Gelman, S. A. (2003). The essential child: Origins of essentialism in everyday thought.
Oxford: Oxford University Press.
Gelman, S. A., & Coley, J. D. (1990). The importance of knowing a dodo is a bird:
Categories and inferences in 2-year-old children. Developmental Psychology, 26, 796–
804.
Gelman, S. A., & Markman, E. M. (1986). Categories and induction in young children.
Cognition, 23, 183–209.
Gelman, S.A., & Markman, E. M. (1987). Young children’s inductions from natural
kinds: The role of categories and appearances. Child Development, 8, 157–167.
Gelman, S. A., & O’Reilly, A. W. (1988). Children’s inductive inferences within super-
ordinate categories: The role of language and category structure. Child Development,
59, 876–887.
Gentner, D. (1988). Metaphor as structure mapping: The relational shift. Child Devel-
opment, 59, 47–59.
Goldstone, R. L. (1994). The role of similarity in categorization: Providing a groundwork.
Cognition, 52, 125–157.
Goswami, U. (2002). Inductive and deductive reasoning. In U. Goswami (Ed.), Blackwell
handbook of cognitive development (pp. 282–302). Oxford, UK: Blackwell.
Graham, S. A., Kilbreath, C. S., & Welder, A. N. (2004). 13-month-olds rely on shared
labels and shape similarity for inductive inferences. Child Development, 75, 409–427.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

52 Brett K. Hayes

Gutheil, G., & Gelman, S. A. (1997). Children’s use of sample size and diversity in-
formation within basic-level categories. Journal of Experimental Child Psychology, 64,
159–174.
Halford, G., & Andrews, G. (2004). The development of deductive reasoning: How
important is complexity? Thinking and Reasoning, 23, 123–145.
Hayes, B. K., & Kahn, T. (2005). Children’s sensitivity to sample size in inductive rea-
soning. Paper presented at the 14th Biennial Conference of the Australasian Human
Development Association, Perth, July.
Heit, E. (1998). A Bayesian analysis of some forms of inductive reasoning. In M. Oaksford
& N. Chater (Eds.), Rational models of cognition (pp. 248–274). Oxford: Oxford
University Press.
Heit, E., & Feeney, A. (2005). Relations between premise similarity and inductive
strength. Psychonomic Bulletin & Review, 12, 340–344.
Heit, E., & Hahn, U. (2001). Diversity-based reasoning in children. Cognitive Psychology,
47, 243–273.
Heit, E., & Hayes, B. K. (2005). Relations between categorization, induction, recognition
and similarity. Journal of Experimental Psychology: General, 134, 596–605.
Heit, E., & Rubinstein, J. (1994). Similarity and property effects in inductive reasoning.
Journal of Experimental Psychology: Learning, Memory, & Cognition, 20, 411–422.
Henderson, A. M., & Graham, S. A. (2005). Two-year-olds’ appreciation of the shared
nature of novel object labels. Journal of Cognition and Development, 6, 381–402.
Heyman, G. D., & Gelman, S. A. (2000a). Preschool children’s use of trait labels to make
inductive inferences. Journal of Experimental Child Psychology, 77, 1–19.
Heyman, G. D., & Gelman, S. A. (2000b). Preschool children’s use of novel predicates to
make inductive inferences about people. Cognitive Development, 15, 263–280.
Inagaki, K., & Hatano, G. (1996). Young children’s recognition of the commonalities
between animals and plants. Child Development, 67, 2823–2840.
Inhelder, B., & Piaget, J. (1958). The growth of logical thinking from childhood to adoles-
cence. New York: Basic Books.
Jaswal, V. (2004). Don’t believe everything you hear: Preschoolers’ sensitivity to speaker
intent in category induction. Child Development, 75, 1871–1885.
Jones, S. S., & Smith, L. S. (1993). The place of perception in children’s concepts. Cognitive
Development, 8, 113–139.
Kalish, C. W., & Gelman, S. A. (1992). On wooden pillows: Multiple classifications and
children’s category-based inductions. Child Development, 63, 1536–1557.
Keil, F. (1991). The emergence of theoretical beliefs as constraints on concepts. In S. Carey
& R. Gelman (Eds.), The epigenesis of mind (pp. 133–169). Hillsdale, NJ: Erlbaum.
Kelemen, D., Widdowson, D., Posner, T., Brown, A. L., & Casler, K. (2003). Teleo-
functional constraints on preschool children’s reasoning about living things. Devel-
opmental Science, 6, 329–345.
Lassaline, M. E. (1996). Structural alignment in induction and similarity. Journal of
Experimental Psychology: General, 22, 754–770.
Lo, Y., Sides, A., Rozelle, J., & Osherson, D. (2002). Evidential diversity and premise
probability in young children’s inductive judgment. Cognitive Science, 26, 181–206.
Loose, J. J., & Mareschal, D. (1999). Inductive reasoning revisited: Children’s reliance on
category labels and appearances. In Proceedings of the Twenty-First Annual Conference
of the Cognitive Science Society (pp. 320–325). Mahwah, NJ: Erlbaum.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

The Development of Inductive Reasoning 53

López, A., Gelman, S. A., Gutheil, G., & Smith, E. E. (1992). The development of
category-based induction. Child Development, 63, 1070–1090.
Madole, K. L., & Oakes, L. (1999). Making sense of infant categorization: Stable processes
and changing representations. Developmental Review, 19, 263–296.
Mandler, J. M., & McDonough, L. (1996). Drinking and driving don’t mix: Inductive
generalization in infancy. Cognition, 59, 307–335.
Mandler, J. M., & McDonough, L. (1998). Studies in inductive inference in infancy.
Cognitive Psychology, 37, 60–96.
Markman, A. B., & Ross, B. (2003). Category use and category learning. Psychological
Bulletin, 129, 592–613.
Markman, E. M., & Callanan, M. A. (1984). An analysis of hierarchical classification.
In R. Sternberg (Ed.), Advances in the psychology of human intelligence, Vol. 2 (pp.
325–366). Hillsdale, NJ: Erlbaum.
McCarrell, N. S., & Callanan, M. A. (1995). Form-function correspondences in children’s
inference. Child Development, 66, 532–546.
McClelland, J. L., & Rogers, T. T. (2003). The parallel distributed processing approach
to semantic cognition. Nature Reviews Neuroscience, 4(4), 310–322.
McDonald, J., Samuels, M., & Rispoli, J. (1996). A hypothesis assessment model of
categorical argument strength. Cognition, 59, 199–217.
Medin, D. L., Coley, J., Storms, G., & Hayes, B. K. (2003). A relevance theory of induction.
Psychonomic Bulletin & Review, 10, 517–532.
Murphy, G. L. (2004). On the conceptual-perceptual divide in early concepts. Develop-
mental Science, 7, 513–515.
Nisbett, R. E., Krantz, D. H., Jepson, C., & Kunda, Z. (1983). The use of sta-
tistical heuristics in everyday inductive reasoning. Psychological Review, 90, 339–
363.
Osherson, D. N., Smith, E. E., Wilkie, O., López, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Quinn, P. C., & Eimas, P. D. (2000). The emergence of category representations dur-
ing infancy: Are separate perceptual and conceptual processes required? Journal of
Cognition and Development, 1, 55–61.
Rakison, D. H., & Hahn, E. (2004). The mechanisms of early categorization and in-
duction: Smart or dumb infants? In R. Kail (Ed.), Advances in child development and
behavior, Vol. 32 (pp. 281–322). New York: Academic Press.
Rosch, E., Mervis, C. G., Gray, W. D., Johnson, D. M., & Boyes-Braem, P. (1976). Basic
objects in natural categories. Cognitive Psychology, 8, 382–439.
Shafto, P., & Coley, J. D. (2003). Development of categorization and reasoning in the
natural world: Novices to experts, naive similarity to ecological knowledge. Journal of
Experimental Psychology: Learning, Memory, & Cognition, 29, 641–649.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Sloutsky, V. M., & Fisher, A. V. (2004). Induction and categorization in young children:
A similarity-based model. Journal of Experimental Psychology: General, 133, 166–
188.
Springer, K. (1992). Children’s awareness of the biological implications of kinship. Child
Development, 63, 950–959.
Springer, K. (2001). Perceptual boundedness and perceptual support in conceptual
development. Psychological Review, 108, 691–708.
P1: JZP
0521672443c02 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:27

54 Brett K. Hayes

Thompson, S., & Hayes, B. K. (2005). Causal induction in adults and children. Paper
presented at 32nd Australasian Experimental Psychology Conference, Melbourne
University, April.
Vygotsky, L. S. (1986). Thought and language. Cambridge, MA: MIT Press (original work
published 1934).
Waxman, S. R., & Booth, A. E. (2001). Seeing pink elephants: Fourteen-month-olds’
interpretations of novel nouns and adjectives. Cognitive Psychology, 43, 217–242.
Welder, A. N., & Graham, S. A. (2001). The influence of shape similarity and shared
labels on infants’ inductive inferences about nonobvious object properties. Child
Development, 72, 1653–1673.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection in Children’s

Inductive Reasoning
Douglas L. Medin and Sandra Waxman

. 
Like adults, children use categories as a basis for inductive inference. Having
learned that some property is true of some individual (e.g., “My dog, Magic,
likes marshmallows”), a child might assume both that other members of the
same category (dogs) share this property and that members of other, simi-
lar categories (e.g., cats) might also share this property. In short, inductive
inference may be guided by and reflect categorical relationships, hence the
term “category-based induction” or CBI. Cognitive and developmental re-
searchers have used a category-based induction paradigm not only to study
the use of categories in reasoning, but also to draw inferences from patterns of
reasoning about the nature of conceptual structures themselves (Atran et al.,
2001; Carey, 1985; Gelman & Markman, 1986; Gutheil, Bloom, Valderrama,
& Freedman, 2004; Gutheil, Vera, & Keil, 1998; Hatano & Inagaki, 1994; Ina-
gaki, 2002; Johnson & Carey, 1998; Medin & Smith, 1981; Ross, Medin, Coley,
& Atran, 2003; Waxman, Lynch, Casey, & Baer, 1997).
One of the most influential examples of the use of inductive projections
to draw inferences about conceptual organization comes from research in
the domain of folkbiology involving categories of living things, including
humans, nonhuman animals, and plants. Developmental evidence has re-
vealed, in particular, certain systematic asymmetries in inductive strength
among these categories. For example, researchers have found that children
are more willing to project properties from humans to dogs than from dogs
to humans. As we will see, one interpretation of this result is that children’s
biological knowledge is organized around humans as the prototype and that it
is more natural to generalize or project from the prototype to its variants than
from variants to the prototype (more about this later). There is even some
evidence that young, urban children may violate the principle of similarity

55
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

56 Douglas L. Medin and Sandra Waxman

by generalizing more from humans to bugs than from bees to bugs (Carey,
1985).
In this chapter, our focus will be on asymmetries in children’s inductive pro-
jections. We begin by describing Carey’s studies and outlining the framework
within which she interpreted her findings. Subsequent work has questioned
the generality of Carey’s findings across populations and procedures, but rel-
atively little attention has been directed toward the theoretical basis for the
asymmetries that she observed. We point out that although Carey’s theoret-
ical position is consistent with the observed asymmetries, it is not unique
in this regard. We describe several other alternative interpretations and their
associated claims about conceptual organization and processing principles.
The upshot is that human-animal asymmetries underdetermine theoretical
accounts, and that further empirical constraints are needed.
We next continue to examine asymmetries but broaden their range to
include asymmetries involving humans, nonhuman mammals, plants, and
insects. We also expand the empirical base by drawing on category-based
induction in children from a range of cultures. The data show certain asym-
metries that appear to hold across populations. These additional sources of
evidence offer considerable leverage for evaluating alternative interpretations
of asymmetries. To foreshadow, from this broader view, we find that the
asymmetries in inductive inference are not well characterized as knowledge
or typicality effects favoring humans, as Carey has suggested. Instead, these
asymmetries seem to reflect the manner by which distinctive categories and
features become activated by the comparison processes associated with induc-
tive inferences. On this account, a major source of human-animal asymme-
tries is the ambiguous status of humans as members of the category “animal”
and as members of a category “human” that is contrastive with animal. The
distinctive features/categories account leads to further predictions that are
supported by children’s open-ended justifications of their responses. We close
with a discussion of implications of this newer interpretation of asymmetries
for the understanding of category structure, conceptual organization, and
conceptual development.

.   - 

An early source of evidence for asymmetries in young children’s inductive
reasoning comes from Carey’s (1985) CBI task. In her task, she described a
novel property (e.g., “has an omentum”) that was attributed to a base (e.g.,
“human”) and then tested children’s willingness to project the property to
various targets. She used “human,” “dog,” and “bee” as bases and multiple
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 57

living and nonliving kinds as targets. The now-classic finding was that children
from four to six years of age willingly project novel properties from humans
to nonhuman animals but are reluctant to make the converse generalization
from nonhuman animals to humans. More specifically, young children who
were told that people have a little green thing inside them, called an omentum,
were quite willing to assert that dogs also have an omentum. In contrast, when
children were told that dogs have an omentum, they were reluctant to infer
that people also have an omentum. In general, the youngest children were very
reluctant to generalize from any base other than humans to other kinds. One
striking consequence of this tendency is that the youngest children were more
likely to generalize from humans to bugs than from bees to bugs, despite the
stronger perceptual similarity and taxonomic relation between the latter than
the former pair. Carey took this asymmetry as a reflection of the underlying
conceptual structure of the young child’s mind, in which biological knowledge
is organized around humans as the prototype humans. Let’s take a closer look.
An important assumption underlying Carey’s interpretation of her results
is that children’s development can be understood in terms of domain-specific
competences. At a minimum, one can distinguish between three domains
in which children form naı̈ve theories: physics, which is concerned with the
physical principles that govern the actions of objects in the world (e.g., Spelke,
1990); psychology, which deals with beliefs, desires, and intentions (Leslie,
1984; Wellman & Gelman, 1992); and biology, which revolves around knowl-
edge of plants and animals (Carey, 1985). One issue is whether these domains
are present at birth and become elaborated with development, or whether
they are wholly acquired. Carey and her colleagues argued that the pattern
of findings on biological induction supports the view that young children do
not have a distinct naı̈ve biology, but rather their reasoning reflects a naı̈ve
psychology where humans are the prototypical psychological entity. Only
later, when children are roughly seven to ten years of age, do they acquire a
distinct biology, in which humans are seen as one animal among many. Older
children do generalize from bases like “dog” and “bee,” and at this devel-
opmental point the asymmetries in reasoning are no longer present. Notice
that on this view, there is radical conceptual change, as children’s biological
reasoning shifts from an intuitive theory of psychology to a naı̈ve biology.
There are, however, lingering questions concerning Carey’s interpretation.
One set of questions concerns the conditions under which these asymmetries
in inductive reasoning arise. We touch on this only briefly in the current
chapter (see Ross et al., 2003; Gutheil et al., 1998; and Atran et al., 2001,
for a broader analysis of the generality of human animal asymmetries). More
central to this chapter are the questions concerning the underlying basis of the
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

58 Douglas L. Medin and Sandra Waxman

asymmetries, whenever they do arise. In the next section, we outline a broad

range of possibilities, focusing initially on the observed human–nonhuman
mammal asymmetries. As will be seen, most of these alternative interpre-
tations can straightforwardly account for the human–nonhuman mammal
asymmetries. Ultimately, then, it will be necessary to broaden our focus
to consider asymmetries among other categories of living things, including
mammals, insects, and plants, and to use these additional sets of asymmetries
to select among the alternatives.

.    

We will consider three broad classes of interpretations of the observed asym-
metries in inductive inference. Within each class there are important varia-
tions. Our analysis begins with Carey’s original interpretation and moves on
to consider two others.

A. Typicality Effects
The explanations included in this section are consistent with the view es-
poused by Carey. In essence, the argument is that the asymmetries in induc-
tive inference reflect an underlying conceptual structure in which humans are
considered the prototypical entity. There are two variants of this general view,
as described below, one relying on a central-tendency notion of typicality and
the other on typicality as based on ideals.

1. Inductive Confidence Displays a Central-Tendency-Based Typicality

Effect. Asymmetries in inductive confidence, which have been documented
in adults, have often been attributed to differences in typicality (Osherson
et al., 1990). On this view, inferences from typical to atypical category ex-
amples are stronger than the reverse direction of comparison. This typicality
view was endorsed by Carey (1985) to account for the human–nonhuman
animal asymmetries displayed by young children. Specifically, Carey suggests
that young children treat humans as the prototypic animal, and that the
asymmetries are just typicality effects.
To bolster this view, one would need independent, converging evidence
that young children treat humans as the prototypical animal. It would also
be beneficial to provide an account of how differences in typicality lead to
asymmetries. First let’s consider typicality. The most common reading of
typicality is in terms of central tendency – being similar to other members
of the same category and dissimilar to members of contrasting categories.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 59

On the surface it does not appear that humans look more like other animals
than, say, a dog does. But if we allow frequency of exposure to examples of
the concept to bias typicality (see Barsalou, 1985, for evidence that frequency
of instantiation as a member of the category is the more relevant variable),
then humans should be the prototype, at least for urban children who may
have little experience with other animals.
How might typicality effect induction? Let’s examine how the similarity-
coverage model (SCM) of Osherson et al. accounts for typicality effects. True
to its name, the model has two components. The first is similarity. The greater
the similarity between the base and the target, the greater the confidence that
some novel property true of the base will be true of the target. The SCM
assumes that similarity is symmetrical (the similarity of A to B is the same
as the similarity of B to A), and therefore asymmetries do not arise from
the similarity component. Instead, it is the coverage component that leads to
typicality effects.
The coverage component works as follows: In addition to considering the
similarity of the base and target categories, the model assumes that partici-
pants generate examples of the lowest-level category that encompasses both
the base and target and then compute their similarity to the base. Because
participants tend to generate typical examples (even if sampling is random,
because a body of other work suggests that natural object categories have
a typicality structure; see Rosch & Mervis, 1975; Smith, Shoben, & Rips,
1974, for evidence and Smith & Medin, 1981, for a review), on average they
will be more similar to a typical than to an atypical base. Because inductive
confidence is assumed to increase with similarity, typical bases will support
stronger inductive inferences since they have greater similarity than atypi-
cal bases to the examples generated by the coverage component process. For
example, consider an inference from penguins to cardinals compared to an
inference from cardinals to penguins. Although the SCM treats similarity as
symmetrical, the latter inference is predicted to be stronger than the former.
The lowest-level superordinate in each case is “bird,” and participants should
think of specific (and typical) birds like robin, blue jay, and sparrow. Because
these are more similar to cardinals than to penguins, cardinal will be stronger
than penguin as an inductive base. In short, the SCM model predicts that
inference will be stronger for typical bases than for atypical bases and that the
coverage component of the model can generate asymmetries.
Although this coverage component accounts for asymmetries at a theoret-
ical level, it is hard to see how it would account for the human asymmetries
observed in young children. More specifically, it is difficult to see why “hu-
mans” would have better coverage than other mammals (e.g., “dog,” “wolf ”)
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

60 Douglas L. Medin and Sandra Waxman

since humans are not especially similar to other mammals. We might need to
amend the SCM model to allow frequency to strongly affect the category ex-
amples generated and then argue that young children tend to think of humans
when generating examples of the superordinate category “animal.” (We note,
however, that recent evidence from our ongoing developmental studies sug-
gests that this is not the case.) A different strategy is to conceptualize humans
as a prototype in the ideal, rather than similarity, sense and to use something
other than the similarity-coverage model to account for the asymmetries. We
now turn to this option.

2. Asymmetries May Represent Ideal-Based Prototypicality Effects. An alter-

native view is that asymmetries derive from typicality effects but that these
typicality effects are based on ideals rather than central tendency (Medin &
Atran, 2004). To get on with developing this position, two related assump-
tions are needed, and both of them are drawn from Carey (1985). One is
the assumption that because young children do not yet have a distinct naı̈ve
biology, their biological induction draws on an undifferentiated module (in
which naı̈ve psychology and naı̈ve biology are fused) that is organized pri-
marily by naı̈ve psychology. The second and uncontroversial assumption is
that humans are the prototypic psychological entity. If we adopt these as-
sumptions, we can appeal to other research (Gleitman, Gleitman, Miller, &
Ostrin, 1996), suggesting that it is more natural to compare the variant to the
standard than vice versa. For example, when faced with a picture of a person,
we notice that the picture has a likeness to the person rather than the person
having a likeness to the picture. That is, what is relevant is that the picture
has some aspects of, or similitude to, the person. In the same vein, comparing
nonhuman animals to humans may be much more sensible and coherent for
young children than the reverse comparison. As applied to induction, the
claim is that in making an inductive inference, we act as if a comparison in the
natural direction is more likely to involve a shared property than a compar-
ison in the unnatural direction. (See Medin, Coley, Storms, & Hayes, 2003,
for a relevance framework that seems generally compatible with this view.)
For this view to explain human–nonhuman animal asymmetries, we will
need the idea that comparisons go from the target to the base. Hence when
a nonhuman animal is a target and a human is the base, the nonhuman gets
compared to the (prototypic ideal) human, and it is natural to assume that
the property is shared. In the reverse direction, when the standard is being
compared to the variant, differences become more accessible. This position
has some testable consequences. To begin with, if humans aren’t the prototype
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 61

(as they may not be in some populations), then asymmetries should disappear.
All we need is some independent index of prototypicality to tie things down.
Once we have determined a typicality ordering, we should be able to predict
a range of asymmetries. In summary, from this perspective, there is nothing
special about humans with respect to generating asymmetries; the results
follow from the (ideal-based) typicality of humans. If we can establish a
ranking of idealness, then we should be able to predict a corresponding series
of asymmetries.

B. Knowledge and Experience Effects

The explanations included in this section emphasize participants’ knowledge
about the base and target, focusing mainly on the base. In essence, the ar-
gument is that willingness to generalize from a given base to a given target
depends on how much one knows about that base. We describe three main
variants of this view.

1. Knowledge Increases Projection. The first notion is that the more a par-
ticipant knows about some base, the more likely they are to project from it
to some target. This idea is endorsed by Kayoko Inagaki and Giyoo Hatano
(2001) and fits their observation that children who have experience raising
goldfish use goldfish as well as humans as a good base for inductive gener-
alization. The paradigm used by Inagaki and Hatano is somewhat different
from that employed by Carey, so one has to be careful about their interrela-
tionships. Inagaki and Hatano have shown that children who have knowledge
of certain biological properties will extend those properties to near neighbors.
We, however, do not know if we would find asymmetries if a novel property
were attributed to goldfish versus, say, a turtle. That is, we don’t know if an
inference from goldfish to turtle would be stronger than an inference from
turtle to goldfish. If such an asymmetry were found, it would provide strong
support for knowledge of a base driving induction.
There is also some more direct evidence consistent with the idea that knowl-
edge or familiarity with a base increases inductive projection (Atran et al.,
2001), and in the limiting case of complete ignorance about the kind in ques-
tion, it seems plausible. Nonetheless, an element is missing. It is not obvious
that increased familiarity or certainty should lead to a greater tendency to
generalize. After all, there is classic evidence from animal learning and con-
ditioning to suggest that increased experience (in the form of discrimination
training) tends to sharpen generalization gradients.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

62 Douglas L. Medin and Sandra Waxman

2. History of Successes and Failures of Induction Influences Inductive Confi-

dence. There are both philosophical and psychological analyses of why some
categories support induction better than others (Goodman, 1983; Shipley,
1993). The category of “blue things” probably does not permit much by way
of inferences beyond the fact that they are things that are blue, but the category
of “blue jays” is rich in the sense that the members share a large number of
properties besides being blue. There may be innate or rapidly acquired biases
for children to treat some categories as more “entrenched” or more powerful
in inductive inferences than others (Gelman, 2003). Whether or not this is the
case, categories and properties (or predicates) that have a history of successful
inductive potential become further entrenched, or more powerful as induc-
tive bases, over time. Therefore, based on this history, some categories and
properties will come to support more and stronger inferences than others. In
the variant that we consider here, the focus is once again on the categories
and properties of the bases in inductive projections.
Projections are like predictions, and if they are supported by data, they
become reinforced. Here the idea is that learners have made, seen, or heard
previous successful inductive projections from humans to other animals but
are much less likely to have done so from other animals to animals. If children
have a more successful history projecting properties from humans to other
animals than from other animals to humans or to other animals, then they
should be more willing to project a novel property from humans than from
some other animal. One straightforward way to obtain evidence bearing on
this hypothesis would be to examine the input (this could be written text or
speech corpora) to ascertain whether humans are preferred over nonhuman
animals as a base in the comparisons that are available to children. To take
this idea one step further, distinguishing between comparisons involving
physical (biological) versus psychological properties should permit us to make
more detailed predictions. In any event, this view falls into the projectability
category given that the history of inductive inference may give rise to greater
or lesser willingness to generalize from a particular base.

3. Differences in Prior Odds or Surprisingness of Predicates. In virtually all

studies of induction, researchers use “blank” properties; that is, properties
for which participants have no strong a priori beliefs. This design feature
is important because participants’ projection or generalization of the blank
property cannot be guided by a simple application of their prior knowledge.
Instead, the assumption is that their projection of a blank property must
be guided by a genuine inductive inference. There is, however, a wrinkle
in this assumption. Although it is usually assumed that blank properties
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 63

have uniform prior odds, this may not be the case. Specifically, assuming
that people are experts on people, they may have lower prior odds for novel
(blank) properties attributed to people than for the same properties attributed
to another base. That is, participants should be more surprised to hear that
humans have an omentum inside (where omentum is a novel property) than
to hear that a raccoon does.
Two factors (at least) appear to support this intuition. The first involves
a “lack of knowledge inference” (Collins & Michalski, 1989). This would
run something like the following: “If humans had omenta inside, then I
would have heard about it, so it’s unlikely that they do.” The second and
related idea is that there may be a bias for assuming that novel properties are
associated with unfamiliar categories (Testa, 1975). If so, the asymmetries will
follow directly for two reasons. First, because (by hypothesis) the priors are
lower for humans than for other mammals, even if the premise (regarding
the target) were ignored entirely and therefore had no effect whatsoever on
beliefs about the target category, blank properties should be more broadly
attributed when the target is a nonhuman mammal than when it is a human.
(In the limiting case they might ignore the premise and simply base judgments
on prior odds for the conclusion.) Second, it is generally assumed that the
more informative or surprising a premise is, the greater its effect on a target
(Osherson, Blok, & Medin, 2005). We suspect that it is more surprising to
hear that humans have an omentum inside than to hear that some other
mammal does. Taken together, then, these two factors suggest that inferences
from humans to (other) mammals should be higher than inferences from
mammals to humans.

C. Distinctive Categories and Features of the Target

The previous two accounts have focused primarily on the ways in which
properties of the bases lead to asymmetric patterns of induction. Our third
class of explanations brings into sharper focus the contribution of the targets
and the relationship between the target and the base, and represents a strong
departure from prior work. Specifically, it proposes that when bases and
targets are compared, inferences are limited by distinctive properties of targets,
relative to the base.

1. Different Patterns of Category-Label Activation. This explanation draws

on the observation that, for most individuals living in the United States,
humans have a dual status: they are considered members of the inclusive
category “animal” (encompassing both human and nonhuman animals) and
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

64 Douglas L. Medin and Sandra Waxman

are also members of the more specific category “human” (contrasting with
nonhuman animals).1 In contrast, nonhuman animals do not have this dual
status. It is important to point out that the inductive strength of a given target
is diminished by salient categories it has that do not apply to the base. These
observations are relevant to inductive inference in the following way: When
a human is the base and a nonhuman animal is a target, the more inclusive
category “animal” gets activated, which tends to prime the broad sense of
“animal” that includes humans. In contrast, when a nonhuman animal is
the base and a human is the target, the distinctive category of “human” gets
activated in the target, and, as we noted above, this diminishes inductive
confidence to nonhuman animals. (Later on we’ll see that the same sort of
asymmetry may arise for plants at a more superordinate level. We suspect that
because animals may be prototypical living things, plants are more distinctive
members of the superordinate category “living thing” than are animals. As a
result, we suspect that when a plant is the base and an animal is a target, the
more inclusive category “living thing” gets activated, but when an animal is
the base and a plant is the target, the distinctive category of “plant” is more
strongly activated, and this diminishes inductive confidence to animals.)
What might count as evidence for this proposal? First, if this proposal is
correct, it should be evident in participants’ justifications for their inductions.
For example, participants should mention the distinctive feature of humans
(e.g., “people aren’t animals”) as a justification more frequently when a hu-
man is a target than when it is a base for induction. A second piece of evidence
involves cross-cultural comparisons. We know that there are cultures and lan-
guages in which the contrast between humans and animals is more clear-cut
than in urban and suburban English-speaking communities in the United
States. For example, in Indonesia these categories are kept quite distinct, and
the Indonesian names are mutually exclusive (Anggoro, Waxman, & Medin,
2005). That is, the Indonesian term for “animal” cannot be applied to humans.
In such cultural contexts, we would expect to find less inductive generaliza-
tion between humans and nonhuman animals, and, as a consequence, the
asymmetries should be diminished. See Anggoro et al. (2005) for evidence
that this is the case.

2. Different Patterns of Feature or Property Activation. The argument here is

parallel to the previous one, but the focus is on constituent properties rather
than categories. As we have suggested, in urban and suburban communities

1 A further complication is that “animal” sometimes is treated as equivalent to “beast” or

“mammal” or even “quadruped.”
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 65

in the United States, humans certainly can be considered animals, but they
may also be considered to be a special case, namely, as animals with more
distinctive features than other animals (e.g., humans are special because they
talk, have beliefs, construct buildings, cook their food, etc.). The core idea,
then, is that inductive inferences from nonhuman animals to humans should
be diminished whenever the distinctive features of humans are activated, and
such features are activated more when humans are the target than when they
are the base. This follows from the idea that induction begins with a focus on
the target category. By logic, the same should hold for any animal, human or
nonhuman, whenever participants attribute more distinctive features to the
target than to the base. In other words, there is nothing special about humans
(save their distinctive features). If this is the case, then for any distinctive
category (e.g., “goldfish” for Inagaki and Hatano’s goldfish caretakers), gener-
alizing a novel property (e.g., from goldfish to frogs) should be stronger than
generalizations in the opposite direction (e.g., from frogs to goldfish). This
would follow if the young goldfish caretakers knew more distinctive features
about goldfish than about frogs.2
The account in terms of distinctive features of the target concept is not new.
For example, Sloman’s (1993) feature-based induction model assumes that
distinctive features of the target category reduce inductive confidence more
than the distinctive features of premise or base categories. Sloman (1993)
also reports a variety of data from undergraduate populations, including
asymmetries in inductive reasoning, supporting this assumption.
The feature-activation account shares similarities with the typicality ac-
count in that typicality is reduced by distinctive features. But note that the pre-
dictions are opposite in form – the feature activation story has it that distinc-
tive features of the target category reduce induction, and this implies a reverse
typicality effect. That is, when the premise or base is atypical and the target is
typical, inductive confidence should be greater than in the reverse ordering.

3. Relevance and Distinctive Features. This position can be seen as a modest

variation on either or both of the first two explanations in this class. The
idea is that asymmetries of induction are not based on overall similarity but
rather on matches and mismatches of a subset of features that appear to be

2 Note that, as in the assumption for category activation above, we are attributing to this view the
claim that the distinctive features of the target affect induction more than do distinctive features
of the base. This is the opposite of Tversky’s (1977) claim that it is the distinctive features of
the base that receive more weight in asymmetries (in similarity judgment). This point is worth
noting, but it is also worth noting that the induction paradigm is considerably different from a
similarity-judgment paradigm.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

66 Douglas L. Medin and Sandra Waxman

most relevant in the experimental context. The motivating principle is that

similarity comparisons are mainly focused on how things are similar rather
than on how similar they are (Medin, Goldstone, & Gentner, 1993).
This position is compatible with the two others in this section, as long
as one comes down on the side of distinctive features of the target being
more important than those of the base. We cannot think of a compelling
reason for why this should be the case, but, as we shall see, it appears to hold
for our developmental data. In general, one would expect that if a salient
property of some target is also a feature of the base, then inductive confidence
should increase (and if it isn’t, confidence should decrease). For example, an
argument from “skunk” to “zebra” should be stronger than from “zebra” to
“skunk” because the most distinctive property of zebras (being striped) is
also shared by skunks (note that the reverse does not hold). This seems like
a pretty strong (and perhaps incorrect) prediction in that a hypothesis-based
model of induction (McDonald, Samuels, & Rispoli, 1996) would clearly make
the opposite prediction – “skunk” should give rise to two hypotheses (odor,
stripes) and “zebra” only one (stripes). To evaluate this position it would be
important to obtain an independent measure of distinctive features.
Of course, relevant features of both the base and the target might well
matter. There is good evidence that there are at least two different strategies
for comprehending conceptual combinations (see Wisniewski, 1997, for a
review) and that which of the two is more prominent depends on the similarity
of the base and target (more technically on their alignability). In the case of
induction the idea would be that for a base and target within the same lifeform,
the base and target are aligned and induction is therefore limited by distinctive
features of the base. For bases and targets that are from different lifeforms (and
for humans versus other animals), attention would shift to salient properties
and categories of the target. This idea is certainly testable.

4. Summary. We will stop with these three main classes of explanation for
asymmetries in induction, though there certainly are others.3 It should be

3 Here are two additional possibilities:

1. Causal associations produce the asymmetry. Many causal relations have directional implica-
tions for induction. If Grass has Enzyme x, we are more sure that Cows have it than we are
sure that Grass has Enzyme Y given that Cows have it (Medin et al., 2003). For this argument
to work, the causal associations should be more likely in the human to (other) animal case
than in the (other) animal to human case. One major problem with this explanation is that
among populations of children where we see this sort of reasoning (see Ross et al., 2003, for
details), the associations we see involving humans tend to have humans as an effect rather
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 67

clear that human–nonhuman animal asymmetries reported by Carey do not

unambiguously implicate any one theoretical account over others.

.   :    

 
We suggest that the most straightforward means of obtaining some leverage
in choosing among the alternative accounts is to consider additional asym-
metries in inductive reasoning. For example, if asymmetries are based on
relative amounts of knowledge about bases and targets, then they should ap-
pear wherever there are demonstrable differences in knowledge, independent
of whether humans are involved in the comparison. It seems uncontroversial
that children being raised in urban and suburban communities in the United
States know more about mammals than they do about plants, so one should
expect mammal-plant asymmetries in which mammal to plant inductions
are stronger than plant to mammal inductions. In this section we will look
at data from an expanded analysis of asymmetries. We first describe a few
cautionary notes to consider when comparing across different methods and
populations, since a number of studies have examined the generality of Carey’s
results across populations and procedures (e.g., Atran et al., 2001; Ross et al.,
2003) but have not followed Carey’s procedure in every detail. Using this as a
foundation, we go on to present evidence on a wider range of asymmetries.

A. Cautionary Notes in Comparing across Studies and Populations

First, although in Carey’s original design the same kinds appear as both base
and target,4 in most recent instantiations of the category-based induction task
this is not always the case. This difference in design may have consequences

than a cause (e.g., bees stinging humans). If this position has value, it may be in explaining
why human-animal asymmetries are not found in some cases.
2. Asymmetries are mediated by size differences between base and target. There are two moti-
vations for this idea. One is that both Itza’ Maya and tree experts sometimes explicitly refer
to relative size on induction tasks and argue that going from large to small (tree species)
is more likely than going from small to large (Medin et al., 1997). The other impetus for
worrying about size is that children may worry about it. Hatano and Inagaki (2003) observed
that children sometimes deny that ants have hearts on grounds that they are too small to
contain a heart. Note, however, that this example cuts the other way in that a property true
of ants will fit into human dimensions, but the converse is not true. Still, it may be that, in
induction, there is a big-to-small bias.
4 Humans are the exception to this generalization. Although humans have consistently been
presented as both bases and targets, the specific other mammal, insect, or plant that has been a
base is rarely also a target.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

68 Douglas L. Medin and Sandra Waxman

on patterns of performance: A human–nonhuman animal asymmetry may be

a consequence of differences in the similarity of the base versus target animals
to humans (e.g., human to bears inferences may be stronger than squirrel to
human inferences, not because of any inherent asymmetry but because bears
and humans are more similar than bears and squirrels). Ideally, we would
have studies where we equate for base and target similarity and have the same
items as both base and target, at least across participants but ideally within
participants. In our ongoing work, we have designed experiments in this way,
but we do not yet have sufficient data to report on. Still, because we find that
most of our findings consistently hold across studies, across populations, and
across a variety of stimulus materials, the patterns of asymmetries reported
below are sufficiently strong to support our main empirical claims.
There is a second difference between Carey’s design and the more recent
versions. In Carey’s original version of the task, four- and six-year-olds were
trained and tested on separate days, though this apparently was not done for
older children and adults. On the first day, children were taught about “spleen”
or “omentum” within the context of a review/probe of other, presumably
more entrenched properties like “has lungs” or “has a heart.” Moreover,
the teaching was fairly elaborate. Children saw a diagram that indicated the
location of the omentum, but there was no explicit appeal to other kinds
that might or might not have that property. They were then tested on the
novel property (e.g., “Does a dog have an omentum?”) at a later date and
within the context of the more familiar properties (e.g., “Does a dog have a
heart?”). One consequence of this design difference is that children in Carey’s
study may have treated the probes regarding the novel property more as a
test of knowledge than as an invitation to make an inductive inference. This
raises the interesting possibility that there is a fundamental difference between
spontaneous inference (as in Carey’s task) versus prompted inference (as in
Atran et al.’s and Ross et al.’s tasks). In this regard, we note that Inagaki and
Hatano’s work may be conceptually aligned to Carey’s. In their work, they
provided a group of young children with extensive experience with goldfish,
and later examined children’s spontaneous inferences (Inagaki, 2002). In any
case, we suspect that if children view the probes as a test of knowledge, this
would raise the threshold on their willingness to project from a base to a
target – this account would still need to explain why this might happen for
“dog” and “bee” as bases but not for “humans.” A third difference between
Carey’s design and the more recent versions pertains to the inclusion criteria.
Carey’s analyses included only those children who correctly attributed the
novel property (e.g., “omentum” or “spleen”) to the base on which it was
introduced.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 69

Finally, and perhaps most interestingly, there are differences in the types of
participants included in the original Carey task as compared to more recent
work. Several recent studies have moved beyond the urban and suburban
populations to examine how children raised in rural environments, who
enjoy a more intimate contact with nature, reason about the biological world.
We have found that children raised in rural environments sometimes appeal
to “ecological reasoning,” focusing on potential mechanisms of transmission
of a property from one biological kind to another. For example, when they
are told that “bees have sacra inside them,” these children often infer that
bears must also have sacra, and they justify their responses by mentioning
that sacra could be transmitted when a bee stings a bear or when a bear eats
honey (Ross et al., 2003).5 The problem, however, is that when ecological
reasoning is combined with procedures where base and target items are not
counterbalanced, the problems loom larger. For example, “bee” may be a
better base for supporting ecological reasoning than “fly.” The bottom line is
that skeptics might want to defer judgments about the claims we make in the
next section.

B. Additional Asymmetries Help to Constrain the Alternative Theories

In this section we focus on the patterns of asymmetries produced by children
and adults from a range of communities, in urban and rural U.S. settings
and in rural Mexico. The primary data of interest, taken from Ross et al.
(2003),6 appear in Table 3.1. In this study children were told that humans,
wolves, bees, or goldenrod had some novel biological property (e.g., “has
sacra inside”) and then were asked if they thought that other animals, plants,
and inanimate things would also have sacra inside them. We refer the reader
to the original article for methodological details, but we include the table
here to permit readers to check our assertions about patterns of asymmetries.
A number of the individual comparisons no doubt fall short of statistical
reliability, but the trends are generally quite robust across populations.

5 We should also note that in the causal-reasoning literature, there are some well-documented
asymmetries in which it is easier to reason from cause to effect than vice versa (see Kahne-
man & Tversky, 1974, for asymmetries in causal judgment). In principle, at least some of the
explanations that we have been considering might apply here as well.
6 We also have considerable data from a study which follows up and expands on these studies
with additional populations including Yukatek Maya children, majority culture and Menominee
children from rural Wisconsin, and Chicago area urban and suburban children (see also Atran
et al., 2001). In most cases we also have adult data but not always in sufficient quantity to be
reported (yet). Although we do not present these data, they consistently support and expand
the claims we make here from published data.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

70 Douglas L. Medin and Sandra Waxman

table 3.1. Ross et al. (2003) induction probabilities, indicating asymmetries

Ross et al. – Asymmetries

City Rural Menominee
Population
Age 7–9 9–11 5–7 7–9 9–11 5–7 7–9 9–11
Human-Mammal 0.70 0.73 0.52 0.58 0.80 0.58 0.70 0.72
Mammal-Human 0.33 0.35 0.24 0.16 0.47 0.42 0.47 0.52
Mammal-Mammal 0.96 0.90 0.76 0.78 0.92 0.75 0.82 0.88
Human-Insect 0.47 0.35 0.07 0.34 0.63 0.46 0.55 0.48
Insect-Human 0.31 0.08 0.28 0.20 0.53 0.63 0.35 0.38
Human-Plant 0.50 0.27 0.16 0.17 0.30 0.30 0.22 0.34
Plant-Human 0.13 0.12 0.10 0.12 0.30 0.42 0.41 0.38
Mammal-Insect 0.50 0.38 0.38 0.28 0.47 0.57 0.56 0.52
Insect-Mammal 0.50 0.56 0.38 0.25 0.56 0.56 0.56 0.56
Mammal-Plant 0.32 0.12 0.14 0.14 0.22 0.18 0.33 0.24
Plant-Mammal 0.32 0.38 0.24 0.25 0.42 0.50 0.42 0.46
Insect-Plant 0.40 0.44 0.26 0.30 0.58 0.46 0.40 0.42
Plant-Insect 0.31 0.62 0.24 0.34 0.33 0.29 0.41 0.43

Finding 1. Human to Nonhuman Mammal Asymmetries. As we noted before,

studies on this topic have tended to use different mammals as base and target;
humans are essentially the only kind that have appeared as both base and
target. Despite this nontrivial concern, it nonetheless appears as if the human
to nonhuman mammal asymmetry is fairly robust. The key finding is that this
asymmetry never reverses. Although there may be cultural factors moderating
the relationship (compare, for example, the evidence for this asymmetry in
Ross et al., 2003, vs. Atran et al., 20017 ), we suggest that the human–nonhuman
mammal asymmetry is strong.

Finding 2. Human to Mammal vs. Mammal to Mammal Asymmetries. This

result is not an asymmetry, but it constrains explanations of asymmetries. In
Carey’s view, young children lack a theory of folkbiology and rely instead on
a theory of folkpsychology in which humans serve as the prototype. If this
were the case, then inferences from humans to other nonhuman mammals
should be stronger than from one nonhuman mammal (in our case: wolf,
dog, peccary) to another. To the best of our knowledge, Carey is the only
researcher who has found this pattern of results. The results of Table 3.1
reveal the opposite pattern. Although the absolute level of human to mammal

7 The asymmetry was clear for “peccary” as a base but essentially absent with “dogs” as a base. This
provides some support for knowledge or familiarity affecting induction. Nonetheless, familiarity
will not account for the other asymmetries even within the Atran et al. data.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 71

inferences ranges across groups, the (other) mammal to mammal inferences

are always higher. This pattern of reverse typicality, apparent at each age and
in all of our populations, strongly undermines the argument that humans
serve as the prototypic animal for either young children or adults. On the
contrary, the data suggest that throughout development, humans retain a
somewhat atypical status within the animal category, and that it is something
of a conceptual achievement when young children notice that humans are, in
fact, also animals.

Finding 3. Human to Plant Asymmetries. In contrast to the previous result,

here the asymmetries appear to vary across populations, and as a result,
no clear overall pattern emerges. Urban children show a strong asymmetry
favoring humans over plants as a base, rural children show little generalization
and no asymmetry, and Menominee children show a reverse asymmetry. One
possibility is that this pattern is a consequence of a shift to increasing use of
ecological/causal reasoning across groups, but additional research is required
to tie this down.

Finding 4. Mammal to Plant Asymmetries. A review of Table 3.1 reveals that

the rate of generalization between mammals and plants is limited. How-
ever, we do see a consistent trend in favor of reverse asymmetries, with
more inductive inferences from plants to mammals than from mammals
to plants. This reverse asymmetry pattern holds for all but one (seven- to
nine-year-old urban children show no difference) of the age and population
samples.

Finding 5. Human to Insect Asymmetries. This set of comparisons is currently

limited because the data are quite variable and in most cases include only one
target insect (fly). In the Ross et al. (2003) data, asymmetries seem to be
present for the urban children, but they are weak and in a few cases reversed
for the rural and Menominee children. In the Atran et al. (2001) data, the
asymmetries diminish with age and are sharply reversed in adults. This is
consistent with the possibility that ecological reasoning (e.g., generalizing
from a bee to a bear by suggesting that the property might be in honey,
which the bear eats, or transmitted via a bee sting) increases with presumed
intimacy of contact with nature. For more recent data with native and exotic
bases intermixed, where we appear to have eliminated ecological reasoning,
there is a pretty consistent asymmetry. This asymmetry appears to be larger
for the exotic set of bases than the native set (“bee” versus “tarantula” as the
base) for all but the Yukatek children.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

72 Douglas L. Medin and Sandra Waxman

Finding 6. Mammal to Insect Asymmetries. There’s no reliable trend across

populations and there are no really notable differences. Comparing human-
insect arguments with mammal-insect arguments, the main difference is that
insect to mammal inferences are higher than insect to human inferences. Note
also that humans to insects inferences are consistently lower than mammals
to insects inferences.

Finding 7. Insect to Plant Asymmetries. In the Ross et al. (2003) data, where
the trends are a bit variable, there doesn’t seem to be any clear overall asymme-
try. For the newer data with mixed native and exotic bases there is a consistent
pattern of reverse asymmetries (plant to insect being greater than insect to
plant), though in several cases the effects are small because there is little
overall generalization between insects and plants (especially compared to the
generalization seen in Ross et al., 2003).

Other Asymmetries. A careful examination of other asymmetries in Table 3.1

reveals no consistent pattern across populations or age groups. As we noted
earlier, we are suspicious of comparisons involving insects because base (bee)
and target (fly) were not counterbalanced and may differ substantially in the
availability of children’s ecological/relational knowledge. In some follow-up
studies we have unwittingly more or less eliminated ecological reasoning, and
these induction data provide still stronger asymmetry constraints.

C. Bringing the Evidence from CBI to Bear

on the Theoretical Alternatives
The results from this wider range of categories provide us with a stronger foun-
dation from which to choose among the alternative theoretical accounts of the
asymmetries in induction. With respect to potential for asymmetries (greater
inductive confidence when the category is in the base position rather than
the target position), we have an overall ordering of human and plant greater
than insect and mammal (and insect greater than mammal when ecological
reasoning is eliminated). In addition, we consistently find that mammal to
mammal inferences are stronger than human to mammal inferences. As we’ll
see, these orderings place strong constraints on theories of induction. We’ll
now go through them, one by one.

1. Typicality Effects. Neither variant of typicality theory can explain these

patterns of asymmetries. If humans are the prototype, then inferences from
human to (nonhuman) mammal should be stronger than those from another
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 73

mammal to mammals, and this was not the case. Instead, the ordering with
respect to asymmetries may conform better to a reverse typicality gradient,
suggesting that mammals are the most prototypical biological entity and
humans the least. If typicality alone cannot account for the underlying basis
for asymmetries, perhaps an account in which typicality is considered in
conjunction with similarity would fare better. The proposal would be that
humans are the ideal animal but, at the same time, are not especially similar
to other mammals. But this account seems to add a new parameter for every
data point it attempts to explain. Moreover, it provides no account of the plant
to mammal asymmetry (and the simultaneous absence of a human to plant
asymmetry). In short, the evidence from the broadened range of observations
undermines the view that typicality effects can account for the human-animal
asymmetries.

2. Knowledge/Familiarity Effects. The idea that inductive strength increases

with knowledge about the base is also challenged by our observations. It
is plausible to suggest that children know more about humans than other
biological kinds, but it is extremely unlikely that they know more about plants
than insects and more about insects than mammals. Therefore, none of the
knowledge-based accounts would seem to predict the observed ordering.

3. Distinctive Features and Categories. The final class of explanations pro-

vides a nice account of the Ross et al. (2003) data. One can explain most
of the results by claiming that “animal” (as opposed to “plant”) is the most
prototypical member of the category “living thing,” and that most nonhuman
mammals (as opposed to humans) are the most prototypical members of the
category “animal.” As a result, when reasoning about biological entities, hu-
mans are the most distinctive (either in terms of their features or in terms of
categories), followed closely by plants. Insects and nonhuman mammals are
less distinctive. Specifically, according to this account, when a base premise
is given and a target suggested, participants compare the target to the base.
Distinctive features or category memberships of the target reduce inductive
confidence more than distinctive features of the base. So if (as we have sug-
gested), with respect to distinctive features and categories, humans are equally
or slightly more distinctive than plants, and if both are much more distinctive
than insects or mammals, the following implications are straightforward:
1. Human to mammal should be stronger than mammal to human
2. Human to mammal should be weaker than other mammal to mammal
3. Plant to mammal should be stronger than mammal to plant
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

74 Douglas L. Medin and Sandra Waxman

4. Plant to insect should be stronger than insect to plant

5. Human to insect should be stronger than insect to human
6. No clear human/plant asymmetry (trends should favor human to plant)
or insect/mammal asymmetries (trends might favor insect to mammal).

This pattern conforms closely with the data. So far this looks very good for
the distinctive features/categories position. The major limitation in what has
been presented so far is that we have relied on intuitive notions of distinctive
features and categories without providing any direct evidence bearing on
them. Justification data, to be presented next, do provide some converging
evidence for the distinctive features/categories account.

D. Bringing More Evidence to Bear: Justification Data

Children’s justifications for their responses in the inductive inference task
bear directly on the distinctive features/categories view. If the distinctive
features/categories position is correct, then failures to generalize from a base
to a target should be accompanied by justifications focusing on the distinctive
features and categories of targets rather than bases. For example, a child
should justify her failure to generalize from humans to raccoons by saying
that raccoons have a tail or raccoons are animals. In contrast, a failure to
generalize from squirrels to humans should bring with it comments that
humans are not animals or that humans have two legs (and squirrels four).
At this point in our investigation, we have extensive justifications only from
rural, majority culture children in Ross et al. (2003) (with a few additional
justifications from Menominee children; they don’t spontaneously offer justi-
fications with sufficient frequency to justify reporting them). However, these
justifications provide clear support for the distinctive features/categories ac-
count. The pattern about to be described holds for human-mammal and
mammal-insect inferences. Both groups of rural children show the same
qualitative trends. Furthermore, the younger and older rural, majority cul-
ture children show the same pattern, though the younger children focus
relatively less on distinctive categories and relatively more on distinctive
features.
First, we can also examine justifications to see if they focus on distinctive
features/categories of the base versus those of the target. The answer is quite
clear: The justifications overwhelmingly (95% or better, depending on the
sample) focus on distinctive features and categories of the target object.8

8 Importantly, this is true both for what one might call “alignable features” (Markman & Gentner,
1997) that reflect different values on a common dimension (e.g., number of legs) and for
nonalignable features (e.g., presence versus absence of a tail). This not only points to the
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 75

We next consider the extent to which justifications focus on common versus

distinctive features and categories. For this purpose we examined the relative
numbers of common versus distinctive features and converted the numbers
to proportions. For human to mammal inferences, the ratio of common to
distinctive features is 0.64 for young children and 0.68 for older children. In
other words, the majority of the time children mentioned common features.
The corresponding ratio for mammal to human is 0.25 and 0.19 for the
younger and older ages, respectively. That is, relatively speaking, children
mention distinctive features much more often when humans are the target
for induction than when they are the base.
The same pattern holds for shared versus distinctive categories. The ratio
of common to distinctive categories for human to mammal is 0.23 and 0.77
for the younger and older ages, respectively. When the comparison shifts to
mammal to human, these proportions drop to 0.12 and 0.43, respectively.
Again, distinctiveness is more salient when humans are in the target position
for induction. For mammal to insect inferences, exactly the same qualitative
trends are observed in both age groups. For example, the young children show
a common to distinctive features ratio of 0.22 in the insect to mammal case
and a ratio of 0.00 in the reverse direction. The corresponding proportions for
common to distinctive categories are 0.38 and 0.10, respectively. This pattern
is noticeably absent for the human-insect justification data, where the young
children give only distinctive features and distinctive categories regardless of
the direction of comparison. Older children show a very slight tendency to
have more common features and categories in the human to insect direction.

E. Summary
Overall, the induction data coupled with the justifications provide detailed
support for the distinctive features/categories account (see also Sloman, 1993).
They also highlight the importance of focusing on the target, the base, and the
relation between them. On this view, we suggest that humans have the most
distinctive features/categories, followed closely by plants, then insects and
mammals. Note that other explanations for the pattern of asymmetries can,
in principle, describe some of these trends, but they do so only by imposing
some implausible assumptions. For example, consider the alternative view
that differences in knowledge underlie the asymmetries. To capture the full
range of data, this view would have to make the dubious claim that children

robustness of the focus on the target object but also undermines the idea that children might
focus on the base for comparisons of alignable targets and the target for nonalignable targets. A
further problem for this two-process account is that this same focus on the target category also
holds for the mammal-mammal probes.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

76 Douglas L. Medin and Sandra Waxman

have more knowledge about plants than mammals. The same difficulty holds
for accounts that appeal to ideals, histories of induction, or prior odds. Only
the distinctive features/categories proposal can account for the ordering of
asymmetries that emerge in the data.

. 
What are the important implications of these findings? First, we suggest that
the human-animal asymmetries that have been observed in category-based
induction tasks do not bear on the status of bases as being either more
familiar or better examples than targets. More generally, we suggest that
accounts which focus primarily on the base are theoretically and empirically
insufficient, for they cannot account for the range of asymmetries that emerge
in children’s inductive inferences about the biological world.
We have argued that the most promising explanation for the observed
asymmetries in inductive inference is one that focuses on the target, the base,
and the relation between them. More specifically, we suggest that such an
account must incorporate not only the categories and features of the target,
but also the distinctive categories and features that emerge when the tar-
get is compared with the base. This more inclusive view can account for the
full pattern of asymmetries, the observation that nonhuman mammal to non-
human mammal inferences are stronger than human to nonhuman mammal
inferences, and the patterns of justification provided by children in category-
based induction tasks. When it comes to categories, it appears that mammals
are animals and little else, other than a specific kind of animal. Specific plants
may be both plants (as a contrast with animals) and living things. Humans
both contrast with and constitute animals. These categorical distinctions are
paralleled by corresponding differences in distinctive features as a function of
comparison direction.
A review of the literature reveals that, as is often the case, the seeds of this
account are evident in previous work. For example, the idea that distinctive
features of the target may be weighted more than distinctive features of the
base is directly taken from Sloman’s (1993) feature-based induction model.
As another example, we point out that at least one earlier model of induction,
the similarity-coverage model (Osherson et al., 1990), also drew on cate-
gories and relations among categories to account for patterns of inductive
inference.
However, the account that we have offered here, the distinctive cate-
gories/features account, goes beyond its progenitors in (at least) three ways.
First, it broadens the range of categories that are relevant for induction.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 77

Second, it considers the relation between the coordinate categories in the

target and base rather than relying on inclusion relations between bases and
their superordinates. Third, our account highlights a role for language. We
suggest that category names may play a more instrumental role in biological
induction than previous accounts would suggest (see Waxman, 2005, for an
amplification of this point and Anggoro et al., 2005, for an example involv-
ing Indonesian). Waxman has described the various inclusion and contrast
classes associated with different biological kinds and outlined some of the ways
in which these categories may influence children’s learning about biological
properties. Her results and ours suggest that the reasoning task provides a
window on how different categories are conceptualized.9
The analysis that we have offered in this chapter is intriguing and appears to
account well for the phenomena at hand. At the same time, however, the dis-
tinctive features/categories hypothesis would benefit from a fuller treatment
of several issues. For example, it will be important to examine more closely
the patterns of inference across different cultural and language groups. It will
also be important to examine more closely the role of ecological reasoning in
patterns of inductions. Although we have touched only briefly on this issue
here, we do have data suggesting that intermixing native and exotic bases and
targets largely eliminates ecological reasoning. In future work we would like
to prime ecological reasoning and employ a design where categories appear as
both base and target in order to examine the relationship between ecological
reasoning and asymmetries more systematically. We also plan to explore ways
of collecting independent evidence on the distinctive features and categories
of bases and targets that arise in the comparison context.
In summary, we think that asymmetries in induction are theoretically
significant, despite the fact that there are many contending, plausible accounts
of human-animal asymmetries. As is often the case, broadening the set of
empirical constraints allows us to gain leverage on these alternative theoretical
explanations of asymmetries. The main theoretical implication of our work is
that researchers should shift attention away from the inductive base alone in
favor of examining the relationship between target and base. This relationship
gives rise to distinctive features and categories and suggests new avenues for
research on the interplay between categories and inductive reasoning. Finally,

9 It may be possible to translate distinctive features/categories back into typicality language – in

this view mammals are the typical biological organism, followed by insects and then plants and
humans. Next you have to claim that comparing the variant to the standard yields a stronger
induction than vice versa and that it’s better (in terms of induction likelihood) to have distinctive
properties in the base than in the target. Unless there’s some obvious novel prediction that we’re
missing, this seems like a notational variation on the distinctive categories/features idea.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

78 Douglas L. Medin and Sandra Waxman

returning to Carey’s original work, our data suggest that the human-animal
asymmetries observed in the induction paradigm do not indicate that humans
are typical or ideal animals. Instead the full range of evidence converges on the
opposite conclusion – that humans are not singled out for their prototypicality
as animals, but for their marked distinctiveness.

References
Anggoro, F., Waxman, S., & Medin, D. (2005). The effect of naming practices on children’s
understanding of living things. Paper presented at Cognitive Science 2005, Stresa,
Italy.
Atran, S., Medin, D., Lynch, E., Vapnarsky, V., Ucan Ek’, E., & Sousa, P. (2001). Folkbiology
doesn’t come from folkpsychology: Evidence from Yukatek Maya in cross-cultural
perspective. Journal of Cognition and Culture, 1, 1–40.
Baillargeon, R. (1994). How do infants learn about the physical world? Current Directions
in Psychological Science, 3(5), 133–140.
Barsalou, L. W. (1985). Ideals, central tendency, and frequency of instantiation as de-
terminants of graded structure in categories. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 11(4), 629–654.
Carey, S. (1985). Conceptual change in childhood. Cambridge, MA: Bradford Books.
Collins, A., & Michalski, R. (1989). The logic of plausible reasoning: A core theory.
Cognitive Science, 13, 1–49.
Gelman, S. A. (2003). The essential child: Origins of essentialism in everyday thought. New
York: Oxford University Press.
Gelman, S. A., & Markman, E. M. (1986). Categories and induction in young children.
Cognition, 23(3), 183–209.
Gleitman, L. R., Gleitman, H., Miller, C., & Ostrin, R. (1996). Similar, and similar
concepts. Cognition, 58(3), 321–376.
Goodman, N. (1983). Fact, fiction, and forecast. New York: Bobbs-Merrill. (Original work
published 1955.)
Gutheil, G., Bloom, P., Valderrama, N., & Freedman, R. (2004). The role of histor-
ical intuitions in children’s and adults’ naming of artifacts. Cognition, 91(1), 23–
42.
Gutheil, G., Vera, A., & Keil, F. C. (1998). Do houseflies think? Patterns of induction and
biological beliefs in development. Cognition, 66(1), 33–49.
Hatano, G., & Inagaki, K. (1994). Young children’s naive theory of biology. Cognition,
50(1–3), 171–188.
Hatano, G., & Inagaki, K. (2003). The formation of culture in mind: A sociocultural
approach to cognitive develoment. In J. Meheler, S. Carey, & L. L. Bonatti (Eds.),
Cognitive development and conceptual change. Cambridge, MA: MIT Press.
Hatano, G., & Inagaki, K. (2002). Young children’s thinking about the biological world.
New York: Psychology Press.
Inagaki, K., & Hatano, G. (2001). Chidren’s understanding of mind-body relationships.
In M. Siegal & C. Peterson (Eds.), Children’s understanding of biology and health.
Cambridge, UK: Cambridge University Press.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

Interpreting Asymmetries of Projection 79

Johnson, S., & Carey, S. (1998). Knowledge enrichment and conceptual change in
folkbiology: Evidence from Williams syndrome. Cognitive Psychology, 37(2), 156–
200.
Kahneman, D., & Tverksy, A. (1973). On the psychology of prediction. Psychological
Review, 80, 237–251.
Leslie, A. M. (1984). Infant perception of a manual pick-up event. British Journal of
Developmental Psychology, 2(1) (Mar 1984), 287–305.
Markman, A. B., & Gentner, D. (1997). The effects of alignability on memory. Psycho-
logical Science 8(4), 363–367.
McDonald, J., Samuels, M., & Rispoli, J. (1996). A hypothesis-assessment model of
categorical argument strength. Cognition, 59(2), 199–217.
Medin, D. L., & Atran, S. (2004). The native mind: Biological categorization and
reasoning in development and across cultures. Psychological Review, 111(4), 960–
983.
Medin, D. L., Coley, J. D., Storms, G., & Hayes, B. K. (2003). A relevance theory of
induction. Psychonomic Bulletin & Review, 10(3), 517–532.
Medin, D. L., Goldstone, R. L., & Gentner, D. (1993). Respects for similarity. Psychological
Review, 100(2), 254–278.
Medin, D. L., Lynch, E. B., Coley, J. D, & Atran, S. (1997). Categorization and reasoning
among tree experts: Do all roads lead to Rome? Cognitive Psychology, 32, 49–96.
Medin, D. L., & Smith, E. E. (1981). Strategies and classification learning. Journal of
Experimental Psychology: Human Learning & Memory, 7(4), 241–253.
Osherson, D. N., Smith, E. E., Wilkie, O., Lopez, A., & Shafir E. (1990). Category-based
induction. Psychological Review, 97(2), 185–200.
Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in the internal structure
of categories. Cognitive Psychology, 7(4), 573–605.
Ross, N., Medin, D., Coley, J. D., & Atran, S. (2003). Cultural and experimental differences
in the development of folkbiological induction. Cognitive Development, 18(1), 25–
47.
Shipley, E. F. (1993). Categories, hierarchies, and induction. Psychology of Learning and
Motivation, 30, 265–301.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Smith, E. E., & Medin, D. L. (1981). Categories and concepts. Cambridge, MA: Harvard
University Press.
Smith, E. E., Shoben, E. J., & Rips, L. J. (1974). Structure and process in semantic
memory: A featural model for semantic decisions. Psychological Review, 81(3), 214–
241.
Spelke, E. S. (1990). Principles of object perception. Cognitive Science, 14(1), 29–56.
Testa, T. J. (1975). Causal relationships and associative formation. Dissertation Abstracts
International, 35(12-B, Pt 1), 6147.
Tversky, A. (1977). Features of similarity. Psychological Review, 84(4), 327–352.
Waxman, S. R. (2005). Why is the concept “Living Thing” so elusive? Concepts, lan-
guages, and the development of folkbiology. In W. Ahn, R. L. Goldstone, B. C.
Love, A. B. Markman, & P. Wolff (Eds.), Categorization inside and outside the labora-
tory: Essays in honor of Douglas L. Medin. Washington, DC: American Psychological
Association.
P1: JZP
0521672443c03 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:30

80 Douglas L. Medin and Sandra Waxman

Waxman, S. R., Lynch, E. B., Casey, K. L., & Baer, L. (1997). Setters and samoyeds:
The emergence of subordinate level categories as a basis for inductive inference in
preschool-age children. Developmental Psychology, 33(6), 1074–1090.
Wellman, H. M., & Gelman, S. A. (1992). Cognitive development: Foundational theories
of core domains. Annual Review of Psychology, 43(1), 337–375.
Wisniewski, E. J. (1997). When concepts combine. Psychonomic Bulletin & Review, 4(2),
167–183.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning

Bob Rehder

Induction – reasoning to uncertain conclusions – appears in many forms. In

some cases one makes an uncertain inference about a specific object or event.
Given a particular dog, one asks whether it is safe to pet it; given a particular
berry, one asks if it is safe to eat it (e.g., see Chapter 8 by Murphy & Ross,
this volume). But in other cases one makes inductive generalizations that are
intended to characterize an entire class of situations or objects. On the basis of
a finite number of medical cases, one might induce a general causal law, such
as that unsafe sex can cause AIDS or that mosquitoes can cause malaria. Or,
from a few examples one might make a generalization about a property being
displayed by many or most members of a particular category, such as that
koalas eat bamboo, Apple laptops have fire-prone batteries, or Madagascar
Fire Ants have poisonous venom. This chapter is concerned with category-
based generalizations such as these in which new properties are projected to
an entire class of objects.
Let’s consider a couple of examples intended to illustrate two different mod-
els of the reasoning processes which are supposed to underlie category-based
generalizations. Suppose you are introduced to someone from Uzbekistan,
the first person from that country you have ever met. During your otherwise
pleasant conversation, you realize that this person has bad breath, and you are
led to wonder whether bad breath is common among Uzbekistanis.1 If you
knew nothing about Uzbekistan, you’d probably be unlikely to jump to the
conclusion that bad breath was a characteristic national trait.2 But suppose

1 No offense to the people of Uzbekistan, whose country was chosen at random from the too-long
list of countries about which I know virtually nothing. The suggestion that Uzbekistani’s have
bad breath (or, as alleged below, that they eat salted fish, drink vodka, or don’t wear business
suits) is a figment of the author’s imagination.
2 Actually, you would be justified in thinking that its prevalence is slightly higher in Uzbekistan,
because you do have a sample – albeit a very small one of size 1 – of people from Uzbekistan

81
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

82 B. Rehder

that you actually know a little bit more about Uzbekistan. For example, imag-
ine that you know how Uzbekistanis traditionally dress, some characteristic
facial features, common religious beliefs, and so on. According to one promi-
nent theory, you will be more likely to think that bad breath is frequent among
Uzbekistanis if your new friend has many of these typical Uzbekistani traits,
as compared to if he is, say, wearing a Western-style business suit, endorses a
traditional western religion (e.g., Roman Catholicism), and so on.
Do you find this example compelling? Perhaps my intuitions have been
ruined by reading too many papers about induction, but I can’t say that I do. I
might be slightly more willing to endorse the generalization about bad breath
on the basis of typical or characteristic features, but not by much, and I think
the reason has to do with the second model of category-based generalization
illustrated by the next example. You’re chatting with the Uzbekistani, but
now, instead of knowing a few typical properties of people from that country,
you know a few of the causes of bad breath. Let’s say you know that eating
salted fish causes bad breath, and so too does drinking vodka. You then learn
from the Uzbekistani that salted fish is one of the staples of the Uzbekistani
diet and that vodka is consumed at every meal. Now suddenly the idea that
large numbers of Uzbekistanis frequently have bad breath doesn’t seem so far
fetched after all.
The first example suggesting that generalizations are based on an example’s
typicality is the province of the well-known Similarity-Coverage Model of
category-based induction (Osherson, Smith, Wilkie, Lopez, & Shafir, 1990)
discussed in numerous places throughout this volume (also see Sloman, 1993,
for a related model). For example, according to the Similarity-Coverage Model
(hereafter SCM), people will be more confident that all birds have some new
property (e.g., sesamoid bones) when given an example of a sparrow with
sesamoid bones as compared to a penguin, because sparrows are more typical
birds than penguins (and thus, on this account, provide more “coverage”
of the features of the bird category). Likewise, the model would predict
that you would generalize bad breath to Uzbekistanis more strongly if your
one example was typical and thus covers the “Uzbekistani” category more
completely. In addition to coverage, the second component of the SCM –
similarity – allows it to account for generalizations between categories that
are not hierarchically nested. For example, people will be more confident that

with bad breath. After all, if bad breath is as uncommon in Uzbekistan as everywhere else, it
is pretty surprising that the one Uzbekistani you get to meet just happens to be one of the few
that have bad breath. But all depends on your beliefs regarding the prior distribution over the
prevalence of bad breath. If you believe that either everybody in a country has bad breath or no
one does, then the single example would be enough to conclude in favor of the former.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 83

blackbirds have sesamoid bones given the fact that crows do as compared
to sparrows, because (according to the SCM) blackbirds are more similar to
crows than sparrows.
In contrast, the second mode of reasoning, referred to here as generalization
as causal reasoning (hereafter GCR), focuses on the specific causal explana-
tions which might lead one to believe that most or all members of a particular
category display a particular property. Although it may be vague or incom-
plete, one often has at least some general idea of the type of causal mechanisms
which produce or generate particular properties, and this knowledge can be
used to estimate the prevalence of that property in some population (i.e.,
category of object). This view also explains why, despite the predictions of the
SCM, the influence of typicality may be uncompelling in certain cases. For
example, I usually attribute bad breath to a transient property – something
consumed at a recent meal for instance – and this makes me reluctant to
consider bad breath a stable property of even that individual person, to say
nothing of the person’s nationality, ethnicity, or any other group he or she
might belong to. It also leads me to believe that the fact that the person is or
is not displaying a typical property on some other dimension (their clothing
style) is irrelevant to the question of generalizing bad breath. Of course, to be
fair to the SCM, it was never intended to be a model of the mental processes
that arise when one has knowledge of the causal mechanisms which might
generate the to-be-generalized property. Famously, its predictions were in-
tended to apply only to “blank properties” (like sesamoid bones) about which
one has no prior knowledge. Nevertheless, the example illustrates how easy
it may be for its similarity-based processes to be supplanted by explanation-
based ones.
Moreover, it may often be the case that even when more typical examples
do support stronger generalization, they do so not because of typicality per se,
but rather because of the distinct pattern of reasoning that a typical example
may support. For example, if you, unlike me, did think that a Uzbekistani
dressed in traditional clothing justified a stronger generalization, I suspect you
did so because you recognized that (a) bad breath can be attributed to a stable
property such as the person’s diet, (b) because diets and ways of dress covary,
this diet (whatever it is) may be common to many Uzbekistanis, and therefore
(c) so too is bad breath. But although this makes you a deeper reasoner than
me, you are reasoning nonetheless, not just basing your judgment on some
unanalyzed notion of typicality.
This chapter is divided into three sections. The first will present the cur-
rent empirical evidence that people in fact engage in causal reasoning when
generalizing properties. This evidence will come both from studies testing
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

84 B. Rehder

the effects of “background” causal knowledge – knowledge which subjects

possess before the experiment starts – and from those which instruct subjects
on causal knowledge as part of the experiment. The assumption is that both
sorts of studies are useful. Establishing results with background knowledge
helps provide some assurance that the reasoning processes involved are likely
to be evoked in real-world contexts. Experimental materials, in turn, allow
the controls necessary to distinguish specific hypotheses from one another.
Two specific models of how causal knowledge influences category-based gen-
eralizations will be discussed.
The second section will address the relationship between causal reasoning
and the similarity-based effects on property generalization. The claim here is
not that similarity never influences generalizations, or that similarity-based
processes are mischaracterized by the SCM, but rather that they and causal
reasoning processes coexist side-by-side in the mental toolbox of processes
which support induction. But when similarity and causality can each form
the basis of a generalization, do people weigh (equally or unequally) both
sources of information, or does one mode of reasoning override or suppress
the other? I will review evidence suggesting that when causal information is
available, it largely eliminates similarity-based influences.
The first two sections address the influence of the causal relations which di-
rectly relate existing category features and a to-be-generalized novel property.
But what of the influence of causal relations not involving the novel property
which relate features of the category to one another? The final section will
discuss how such relations change the coherence of category, and by so doing
promote the strength with which novel properties are generalized.

    

There are now numerous sources of evidence that people reason causally in
order to generalize properties. I first discuss the studies involving background
causal knowledge, but, because recent reviews of this research already exist
(Heit, 2000; Rips, 2001), only a few example will be covered here. I will then
turn to more recent research testing causal relations which are provided as
part of the experimental session.

Research with Background Causal Knowledge

One of the first studies suggesting that property generalization can depend
on causal knowledge was conducted by Heit and Rubinstein (1994). Recall
that, according to the SCM, a novel property will be generalized from one
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 85

category to another to the extent that the categories are similar.3 In particular,
the SCM does not allow for how such generalizations might depend on the
property itself, by virtue of, say, the specific causal mechanisms it is involved
in. Heit and Rubinstein showed otherwise. They found, for example, that
a behavioral property (e.g., travels in a zig-zag path) was generalized more
strongly from tunas to whales as compared to from bears to whales. This result
may have been expected on the basis of the fact that whales are more similar
to tunas than bears. But when the novel property was biological rather than
behavioral (e.g., a liver with two chambers that acts as one), it was generalized
more strongly from bears to whales instead. Why should this reversal arise
despite the fact that the categories involved (bears, tunas, and whales) were
unchanged, and thus so too were the similarity relations? One explanation
is that participants thought that bears and whales share biological properties
(such as two-chambered livers), because such properties are likely to arise
from causal mechanisms associated with their common biological category,
that is, mammals. Tunas and whales, on the other hand, are more likely to
share a survival behavior (traveling in a zig-zag path) because they are both
prey animals living in a common ecology (also see Gelman & Markman, 1986,
and Springer & Keil, 1989).4
Sloman (1994, 1997) has provided other examples of how a property will be
more strongly generalized when the causal history which leads to a property
in a base category is also present in the target. For example, he found that un-
dergraduates were more willing to project a feature like “hired as bodyguard”
from war veterans to ex-convicts, presumably because the underlying expla-
nation (fighting experience makes for a good bodyguard) for war veterans
also applies to ex-convicts. In contrast, they were less to willing to project the
property “unemployed,” apparently because the reasons for unemployment
of war veterans (physical injury, PTSD) does not apply to ex-convicts. Note
again that these different results obtained despite the fact that similarity was
held constant (in both cases war veterans were the base and ex-convicts were
the target).
Finally, Smith, Shafir, and Osherson (1993) also found that subjects appar-
ently engaged in a form of causal reasoning when generalizing hypothetical

3 Technically, according to the SCM generalizations between non-hierarchically nested categories

are a function of not only their similarity but also the coverage the premise provides for the
lowest-level superordinate category that includes both premise and target categories. This more
precise definition is unimportant for what follows.
4 Note that Heit and Rubinstein themselves attributed these results to subjects computing sim-
ilarity between the base and target differently depending on whether the novel property was
anatomical or behavioral. I’ll return to this issue in the Discussion.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

86 B. Rehder

properties of familiar categories. For example, undergraduates were more

likely to generalize the property “can bite through barbed wire” to German
Shepherds from poodles than from Dobermans. This result obtained de-
spite the fact that German Shepherds are more similar to Dobermans than
to poodles. It seems that the participants were reasoning about the causal
preconditions for the capacity to bite through barbed wire, and judged that
a German Shepherd certainly could do so if a small dog like a poodle could,
but not necessarily if the base category was another powerful dog (also see
Osherson, Smith, Myers, & Stob, 1994).
Thus, there appears to be good reason to suspect that causal reasoning is
intimately involved in the generalization of properties. But although sugges-
tive, studies testing background knowledge are often correlational in nature,
and thus there is always the possibility that the results arise due to factors
other than causal knowledge itself. For this reason, some researchers have
adopted a purely experimental approach in which subjects are presented with
category members that display novel properties which are accompanied by
a description of the (previously unknown) causal factors that led to their
presence. The precision with which this approach allows causal knowledge
to be manipulated has led to the first specific theories regarding how such
knowledge affects property generalization, as I now describe.

Research with Experimentally-Provided Causal Knowledge

The theories which have been advanced thus far have been concerned with
how a to-be-generalized property is related to a category’s existing features.
The first of these appeals to the notion of feature centrality (Hadjichristidis,
Sloman, Stevenson, & Over, 2004). Features are more central to the extent
they have many features which “depend on” them, or, equivalently, to the
extent they are responsible for generating or producing many other features
(Sloman, Love, & Ahn, 1998). One purported role for centrality is in deter-
mining a feature’s importance to an object’s category membership and its
degree of typicality within a category. For example, a bird with all the charac-
teristic features of robins except that it has unusual coloring is a more typical
robin (and more likely to be a robin) than one with all the characteristic
features of robin except for robin DNA. According to the centrality account,
robin coloring is less central than robin DNA (and thus less important for
establishing category membership) because coloring depends on DNA and
not the other way around (Ahn, 1998; Ahn et al., 2000; see Rehder, 2003a, b;
Rehder & Kim, 2006, for an alternative view).
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 87

Hadjichristidis et al. (2004) have proposed that the concept of centrality is

also important to induction, suggesting that novel properties which are central
are more likely to be generalized from one category to another. For example,
suppose one person believes that many of a seal’s physiological functions de-
pend on a particular hormone, whereas another believes that only a few such
functions depend on it. According to the centrality account, the hormone
will be more central for the first person than the second. Hadjichristidis et al.
suggest that the hormone will as a result not only be more important for iden-
tifying an animal as a seal for the first person, but he or she will also be more
likely to generalize it to another category (say, dolphins). But they point out
that the strength of generalization will also depend on whether the hormone
is viewed as being central in the target category, which in turn depends on
whether the hormone’s “dependency structure” (e.g., the physiological func-
tions) also appears in the target category. People will sometimes know enough
about the target category to make this determination directly. But when they
don’t, Hadjichristidis et al. suggest that reasoners use the similarity between
the base and target to estimate the amount of shared dependency structure.
Thus, a hormone which is responsible for a seal’s physiological functions is
more likely to be present in dolphins than sparrows, because the similarity
between seals and dolphins suggests they have many physiological functions
in common (and thus also the hormone). In contrast, the dissimilarity be-
tween seals and sparrows suggests that they share many fewer physiological
functions (and thus not the hormone).
Hadjichristidis et al. (2004, Experiment 2) tested this hypothesis by varying
a novel property’s centrality (e.g., a fictitious hormone was depended on by
either many physiological functions, few such functions, or no information
about dependency was provided) and the similarity of the base and target cat-
egory (high, medium, or low). As predicted, central properties (with many de-
pendents) were projected more strongly than less central properties (with few
dependents), which in turn were stronger than properties with no dependents
(Fig. 4.1). Also as predicted, Figure 4.1 also shows how the strength of this
effect was moderated by the similarity of the two categories: Centrality had its
greatest effect when the target category was very similar to the base category.
These findings provide one important illustration of how interfeature
causal relations influence property generalization. However, one can ask if
characterizing these results in terms of feature centrality alone artificially
limits the scope of the effect of causal knowledge on such generalizations.
This leads to the second theory of property generalization which treats such
generalizations as an instance of causal reasoning (Rehder, 2006). According
to this view, one computes whether a novel feature is likely to appear in a
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

88 B. Rehder

90
High Similarity

Generalization Rating 80 Medium Similarity

Low Similarity
70

60

50

40
Many Few None
figure 4.1. Number of dependents of novel property. Results from Hadjichristidis et al.
(2004), Experiment 2.

target category not on the basis of any one characteristic of the feature (e.g.,
its centrality), but rather on the basis of one’s beliefs about the causal laws
which relate the feature to those of the target category.
The generalization as causal reasoning view (GCR) makes (at least) three
predictions regarding how causal knowledge supports category-based gener-
alizations. The first of these is that property generalization can be an instance
of diagnostic reasoning in which one reasons from the presence of a novel
property’s effects to the presence of the novel property itself. The second
prediction is that generalizations can reflect prospective reasoning, in which
one reasons from the presence of the causes of a novel property to infer the
presence of the property. The third prediction is that generalizations should
exhibit the basic property of extensional reasoning in which a novel prop-
erty will be more prevalent among category members to the extent its causes
and/or effects are prevalent. As it turns out, each of these predictions has
received empirical support.

evidence for diagnostic reasoning. According to GCR, a property will

be generalized to the extent that its presence can be inferred on the basis
of the effects it is likely to produce. That is, generalization can be seen as
that species of causal reasoning in which one “diagnoses” the presence of a
property within category members on the basis of whether its “symptoms”
(effects) are present in those objects.
The results from Hadjichristidis et al. (2004) just presented can be inter-
preted as just such a case of diagnostic reasoning. The fact that physiological
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 89

functions “depend on” a hormone can be understood as the hormone causing

or enabling those functions. Thus, the presence of one or more of the same
physiological functions in another species “diagnoses” the presence of the
hormone in that species. However, one complicating factor is the possibility
that in the target species those physiological functions have different causes
(e.g., different hormones). For this reason, the inference to the hormone is
more certain when multiple of its “symptoms” (physiological functions) are
present.5 This situation is more likely to be true to the extent that the base
and target categories are similar rather than dissimilar (Figure 4.1).

evidence for prospective reasoning. GCR not only reinterprets evi-

dence in favor of the feature centrality view as a case of causal reasoning, it
also makes new predictions which distinguish it from the feature centrality
view. The second prediction of GCR is that a property will be generalized to
the extent that the cause or causes which produce it are present in the target
category. On this view, generalization can be seen an instance of prospective
reasoning in which one reasons from causes to effects rather than from effects
to causes.
To start, note that each of the demonstrations of the effect of background
knowledge presented earlier implicates the role of the causal mechanisms
which produce a novel property, such as when bears and whales are likely
to posses the same kind of liver on the basis of the shared biological mech-
anisms of mammals (which presumably generate their internal organs). The
prospective reasoning view also receives support from studies which manipu-
late causal knowledge experimentally. For example, Rehder (2006) instructed
undergraduates on an artificial, but plausible, category, such as Romanian Ro-
gos (a type of automobile). Rogos were described as having four characteristic
features, each of which was described as occurring in 75% of category mem-
bers: butane-laden fuel, loose fuel-filter gasket, hot engine temperature, and
large amount of carbon monoxide in the exhaust. A single example of a Rogo
was then presented with a novel property, and subjects were asked whether
a second Rogo was likely to have that property. On some trials no causal
explanation of the novel property was provided. For example, subjects would

5 Whether one finds multiple versus few symptoms more diagnostic of an underlying common
cause will depend on the details of the causal beliefs involved. If one believes there is no
other possible causes of the symptoms, then even just one symptom is sufficient to infer the
cause with certainty. Also relevant is whether the cause is deterministic (produces its effects
with probability 1) or probabilistic. A deterministic cause invariably produces all of its effects,
and thus the presence of the cause is ruled out if even one of its effects is absent. For a
probabilistic representation of causality see Rehder (2003a, b) and Cheng (1997). For discussion
of a deterministic versus probabilistic view of causaility, see Thagard (1999).
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

90 B. Rehder

100
90

Generalization Rating
80
70
60
50
40
30
20
10
Present No Cause Absent
figure 4.2. Cause of novel property. Results from Rehder (2006), Experiment 3.

be presented with a Rogo that had ceramic shocks, and then judged what
proportion of all Rogos had ceramic shocks. Properties like ceramic shocks
were chosen because they have no obvious causal connection with the other
known properties of Rogos (e.g., butane-laden fuel, carbon monoxide in the
exhaust, etc). But on other trials the novel property was described as being
caused by one of the Rogo’s characteristic features. For example, subjects
would be presented with a Rogo which had a typical property (e.g., hot en-
gine temperature) and a novel property (e.g., melted wiring) and told that the
typical property caused the novel one (“The melted wiring is caused by the
high-engine temperature”). A second Rogo was presented which either did
or did not have high-engine temperature, and subjects were asked whether it
had melted wiring.
The results, presented in Figure 4.2, shows that the novel property was
generalized very strongly when its cause was present in the target Rogo, very
weakly when it was absent, and with an intermediate rating when no infor-
mation about causal mechanism was provided. The results were the same
regardless of whether the category was an artifact (e.g., Rogos), a biological
kind, or a nonliving natural kind. Clearly, subjects reason forward (prospec-
tively) to infer a novel property’s presence from the presence of its causes as
readily as they reason backward (diagnostically) to infer a property’s presence
from its effects (also see Lassaline, 1996, and Wu & Gentner, 1998).

evidence for extensional reasoning. A third prediction of GCR is that

generalizations should exhibit extentional reasoning in which the prevalence
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 91

of a novel property depends on the prevalence of its causes and/or effects. For
example, the prevalence of a hormone in a target category shouldn’t depend on
just the number of physiological functions it causes, but also how widespread
those functions are amongst the population of target category members.
Similarly, the prevalence of melted wiring caused by hot engines in a brand of
automobile should depend on what proportion of those automobiles in fact
have hot engines.
Maya Nair and I decided to test this prediction by explicitly manipulating
the prevalence of a category feature which was the purported cause (or effect)
of a novel property (Nair, 2005). The same categories as in Rehder (2006) were
used (including Romanian Rogos), and subjects were asked what proportion
of all Rogos had a novel property on the basis of a single example of a Rogo
with that property. Two factors were manipulated as within-subjects variables.
The first was the base rate of the characteristic feature. Two randomly chosen
features of Rogos were described as occurring in 90% of Rogos, whereas the
other two were described as occurring in 60%. The second orthogonal factor
was whether the novel property was described as caused by or as the cause
of one of the characteristic features. The characteristic features, the novel
properties, and the causal relationships between the two are presented in
Table 4.1 for Rogos. For example, some subjects would be presented with
a Rogo which had all four characteristic features and a novel property (e.g.,
zinc-lined gas tank) and told that one of the typical properties caused the novel
one. (“Butane-laden fuel causes a zinc-lined gas tank. The butane interacts
with the chromium in the metal of the gas tank, which results in a thin
layer of zinc on the inside of the tank.”) Other subjects would be told that
the novel property was the cause of the typical property. (“A zinc-lined gas
tank causes the fuel to be butane-laden. The zinc prevents corrosion of the
tank, but interacts chemically with gasoline to produce butane.”) All subjects
would then rate what proportion of all Rogos possessed the novel property.
Each subject performed four generalization trials in which the four novel
properties in Table 4.1 were each presented once, either as a cause and effect
of a characteristic feature, and with the base rate of the characteristic feature
(e.g., butane-laden fuel) described as either 90% or 60%.
Generalization ratings from this experiment are presented in Figure 4.3A
as a function of whether the novel property was the cause or the effect and
the base rate of the typical property. There was a strong effect of base rate, as
subjects’ generalization ratings were much higher when the novel property was
causally related to a characteristic feature with a 90% versus a 60% base rate.
Moreover, this result obtained regardless of whether the characteristic feature
 P1: IBE
0521672443c04

table 4.1. Features and causal relationship for Romanian Rogos, an artifical category

Characteristic Feature Novel Feature Characteristic Feature → Novel Feature Novel Feature → Characteristic Feature
Butane-laden fuel Zinc-lined gas tank Butane-laden fuel causes a zinc-lined gas tank. A zinc-lined gas tank causes the fuel to be
The butane interacts with the chromium in butane-laden. The zinc prevents corrosion
the metal of the gas tank, which results in a of the tank, but interacts chemically with
thin layer of zinc on the inside of the tank. gasoline to produce butane.
CUFX144-Feeney 0 521 85648 5

Loose fuel filter gasket Vibrations during A loose fuel filter gasket causes vibrations Vibration during breaking causes a loose fuel
breaking during braking. The fuel which leaks filter. The rattling caused by the vibrations
through the fuel filter gasket falls on one of eventually leads to the fuel filter gasket
the brake pads, causing abrasion which becoming loose.
results in the car vibrating while braking.

92
Hot engine temperature Thin engine oil Hot engine temperature causes thin engine Thin engine oil causes hot engine
oil. The oil loses viscosity after it exceeds a temperature. Thin oil does not provide
certain temperature. sufficient lubrication for the engine’s
moving parts, and the engine temperature
goes up as a result.
High amounts of Inefficient High amounts of carbon monoxide in the An inefficient turbocharger causes high
carbon monoxide turbocharger exhaust causes an inefficient turbocharger. amounts of carbon monoxide in the
in the exhaust As the exhaust leaves the engine it passes exhaust. An inefficient turbocharger fails to
through the turbocharger. The lower inject enough oxygen into the engine, and
density of carbon monoxide in the exhaust so excess carbon does not undergo
means that the turbocharger is not combustion.
sufficiently pressurized.
July 19, 2007
14:32
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 93

A
90

Generalization Rating 80

70

60

50 Novel As Cause

Novel As Effect
40

60 90

B
90
Generalization Rating

80

70

60

50 Novel As Cause

Novel As Effect
40

62.5 87.5
figure 4.3. Observed base rate. Results from Nair (2005).

was the cause or effect of the typical property. In other words, subjects readily
engage in the extensional reasoning which characterizes causal reasoning
more generally.
In a follow-up experiment, we asked whether these results would obtain
when the feature base rates were observed rather than given through explicit
instruction. Informal post-experiment interviews in the first experiment re-
vealed that the base rate manipulation was very powerful. Although general-
ization ratings were entered by positioning a slider on a computer-displayed
scale which was not labeled numerically (one end was simply labeled “None,”
meaning that no Rogos have the novel property; the other end was labeled
“All,” meaning they all do), a number of subjects reported trying to position
the marker at a point that they felt corresponded to 60% or 90%. That is, if
a novel property was causally related to a typical property with a base rate
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

94 B. Rehder

table 4.2. Category structure from Nair (2005),

Experiment 2, consisting of eight Rogos (R1–8) and
eight non-Rogos (NR1–8).

D1 D2 D3 D4
Rogos
R1 1 1 0 1
R2 1 1 0 1
R3 1 1 0 1
R4 1 1 1 0
R5 1 1 1 0
R6 1 1 1 0
R7 0 1 1 1
R8 1 0 1 1
Non-Rogos
NR1 0 0 1 0
NR2 0 0 1 0
NR3 0 0 1 0
NR4 0 0 0 1
NR5 0 0 0 1
NR6 0 0 0 1
NR7 1 0 0 0
NR8 0 1 0 0

of 90% (or 60%), many subjects felt that that property would be displayed
by exactly 90% (or 60%) of Rogos. To determine if the extensional reasoning
effect obtained only because the base rates were explicit, we performed a
replication of the first experiment in which the feature base rates were learned
implicitly through a standard classification-with-feedback task. In the first
phase of the experiment subjects were asked to learn to distinguish Romanian
Rogos from “some other kind of automobile.” They were presented with the
exemplars shown in Table 4.2, consisting of eight Rogos and eight non-Rogos.
In Table 4.2, a “1” stands for a typical Rogo feature (e.g., loose fuel filter gasket,
hot engine temperature, etc.), whereas each “0” stands for the opposite value
on the same dimension (e.g., tight fuel filter gasket, cool engine temperature,
etc.). The thing to note is that on two of the dimensions (Dimensions 1 and
2 in Table 4.1), the typical Rogo feature is very typical, because it occurs in
seven out of eight Rogos (or 87.5%). On the other two dimensions (3 and 4),
the characteristic feature is less typical, occurring in five out of eight Rogos
(or 62.5%). The structure of the contrast category (the non-Rogos) is the
mirror image of the Rogos.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 95

The items in Table 4.2 were presented in a random order in blocks, subjects
classified each and received immediate feedback, and training continued until
subjects performed two blocks without error or reached a maximum of twenty
blocks. They then performed the same generalization task as in the previous
experiment in which whether the novel property was the cause or effect was
crossed with the base rate of the typical Rogo feature (87.5% or 62.5%). The
results of this experiment are presented in Figure 4.3B. The figure indicates
that subjects in this experiment, just like those in the first one, took the
base rate of the characteristic feature into account when generalizing the new
property. They did this even though the difference in base rates between
features was manifested in observed category members rather than provided
explicitly. That is, people take into account the prevalence of that characteristic
feature in the population of category members when generalizing the novel
property, and they do so regardless of whether a novel property is the effect
or cause of a characteristic feature. Note that in both experiments from
Nair (2005), the results were unaffected by whether the category was a novel
artifact, biological kind, or nonliving natural kind.
In summary, studies provide strong evidence that people are exhibiting
some of the basic properties of causal reasoning when generalizing properties.
When many of a novel property’s effects are present in a category, people
reason (diagnostically) to the presence of the property itself. People also
reason (prospectively) to the presence of the property as a function of whether
its causes are present or absent in the target. Finally, people are sensitive to
how prevalent a novel property’s cause (or effect) is among the population of
category members.

    

  
The preceding research firmly establishes the influence of causal explanations
on how properties are generalized to categories. Another important question
concerns how those explanations interact with the influences of similarity
which have been formalized by the Similarity-Coverage model (SCM). Does
each of these sources of evidence contribute (perhaps unevenly) to the gen-
eralization of a property, or do people choose to generalize on the basis of
one source of evidence alone? As it turns out, research suggests that causal
explanations not only influence generalizations, but they may also largely
undermine, or supplant, similarity-based processes. Once again, I will first
briefly review studies testing background knowledge, and then turn to the
results from studies using experimentally-provided causal knowledge.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

96 B. Rehder

Research with Background Causal Knowledge

To investigate the role of causal knowledge in property generalization, a
number of studies have compared how inductive reasoning varies across
populations which differ in the nature and amount of background knowledge.
For example, Lopez, Atran, Coley, Medin, and Smith (1997) studied category-
based generalizations among the Itza’ Maya (an indigenous population in
central Guatemala), and found that the Itza’ failed to exhibit standard diversity
effects. Diversity refers to the fact that a more diverse set of base examples leads
to stronger generalizations. All else being equal, people are more likely to
conclude that all birds have sesamoid bones given that a dissimilar set of birds
do (e.g., sparrows, hawks, and chickens) as compared to if the birds are more
similar (e.g., sparrows, robins, and bluejays). Diversity effects are predicted
by the SCM because less similar base categories provide better coverage of
the target category. But Lopez et al. found that the Itza’ failed to exhibit
diversity effects, as they frequently appealed to specific causal mechanisms
instead (e.g., judging that the prevalence of a disease in a species depended
on the mechanisms which might spread that disease). In contrast, American
undergraduates exhibited standard diversity effects on the same items (also
see Bailenson, Shum, Atran, Medin, & Coley, 2002).
Of course, these findings could be attributed to some cultural difference
between Americans and the Itza’ other than domain knowledge (such as
a difference in their default reasoning strategy). Contra this interpretation
however, the dominance of causal explanations also holds for Americans who
possess sufficient domain knowledge. Proffitt, Coley, and Medin (2000) tested
three groups of North American tree experts (taxomomists, landscapers, and
maintenance workers) to determine how they generalized a novel property of
trees (an unspecified disease) as a function of the typicality and diversity
of base categories (species of trees). Not only did two of the three groups
of experts (the landscapers and maintenance workers) fail to exhibit diver-
sity effects (just as the Itza’ failed to), none of the groups exhibited standard
typicality effects. Instead, the experts frequently reasoned in a manner which
reflected their causal understanding of disease processes, such as the suscep-
tibility or resistance of certain species to diseases, the geographic distribution
of a tree species (more disperse species infect more trees), and the specific
mechanisms by which diseases are transmitted from one tree to another (also
see Bailenson et al., 2002; Shafto & Coley, 2003).
Finally, there is ample evidence that causal reasoning in support of induc-
tion is not limited to the rarefied world of domain experts (such as the Itza’ or
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 97

landscapers). Testing American undergraduates, Medin, Coley, Storms, and

Hayes (2003) found, for example, that a more diverse set of premise categories
consisting of a mammal and a bird (cats and sparrows) supported weaker gen-
eralizations (e.g., of a blank property “enzyme X”) to lizards as compared to
a less diverse set of two mammals (cats and rhinos). Apparently, the fact that
cats eat sparrows suggested to subjects a possible causal mechanism by which
enzyme X might be shared by them but not by lizards, which are in a different
food chain. Similarly, the more diverse set of categories again consisting of a
mammal and a bird (pigs and chickens) resulted in weaker generalization to
cobras as compared to a less diverse set of two mammals (pigs and whales). It
may have been that pigs and chickens were both recognized as farm animals,
which suggested one or more candidate causes of the novel property (e.g.,
being injected with antibiotics). Inductions were then weaker because those
candidate causes are absent in cobras.
These studies make a strong case that, in the presence of causal knowledge,
similarity-based factors may become largely irrelevant to how people choose
to generalize examples. Can we also find evidence for this conclusion when
causal relations are manipulated experimentally?

Research with Experimentally Provided Causal Knowledge

To answer this question Rehder (2006) conducted a series of experiments
which assessed how causal explanations moderate those effects of similarity
on property generalization predicted by the SCM. Each experiment used
a two-factor design in which one factor was the presence or absence of a
causal explanation, and the second factor was either typicality, diversity, or
similarity itself. This design of these experiments allowed a demonstration of
SCM’s standard similarity-based effects in the absence of causal explanations,
and how those effects are moderated when such explanations are present.

causal explanations versus typicality. American undergraduates were

instructed on artificial categories like Romanian Rogos which possessed four
characteristic features each described as occurring in 75% of category mem-
bers. They were then presented with a particular Rogo that exhibited a novel
property and asked to judge what proportion of all Rogos had that property.
On some trials, the novel property was accompanied by a causal explanation
(e.g., the Rogo was said to have melted wiring, which was caused by its hot
engine temperature), whereas on other trials the novel property was “blank,”
that is, no explanation was provided (e.g., ceramic shocks). The second
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

98 B. Rehder

experimental factor was the typicality of the Rogo, which had either 1, 2,
3, or 4 characteristic features. A Rogo always possessed at least the character-
istic feature which was described as the cause of a novel property (e.g., hot
engine temperature in the case of melted wiring).
Based on the SCM, the prediction of course was that the generalization
of blank properties would strengthen with the number of characteristic fea-
tures. The critical question concerned whether typicality would also affect
the generalization of novel properties when a causal explanation was present.
The results are presented in Figure 4.4A as a function of whether a causal
explanation was provided and the typicality of the exemplar. As expected,
generalization ratings for blank properties increased as the exemplar’s typi-
cality (i.e., its number of characteristic features) increased, indicating that this
experiment replicated SCM’s standard typicality effect for blank properties.
However, Figure 4.4A indicates that this effect of typicality was reduced when
a causal explanation was provided.
One important question is whether the response pattern shown in Figure
4.4A is manifested consistently by all participants, or whether it arose as
a result of averaging over individuals with substantially different response
profiles. In fact, two subgroups of participants with qualitatively different
responses were identified, shown in Figures 4.4B and 4.4C. The subgroup in
Figure 4.4B produced higher induction ratings for the nonblank properties as
compared to the blanks, that is, they were more willing to generalize a novel
property when accompanied by a causal explanation. Importantly, however,
whereas ratings for blanks properties were sensitive to typicality, typicality
had no effect on the generalization of the nonblanks. In contrast, the second
subgroup in Figure 4.4C did not generalize nonblanks more strongly than
blanks, and both types of properties were equally sensitive to typicality. In
other words, when reasoners chose to base their responses on the causal
explanations (as evidenced by the nonblank’s higher ratings in Figure 4.4B),
there was no effect of typicality; when they chose to base their responses on
typicality, there was no effect of causal explanations (Figure 4.4C). Apparently,
the use of a causal explanation versus typicality is an all-or-none matter, with
reasoners using one strategy or the other, but not both.

causal explanations versus diversity. In a second experiment, Rehder

(2006) also assessed how the presence of a causal explanation for a novel prop-
erty moderates the diversity effect. Diversity was manipulated by presenting
two category members with five stimulus dimensions with the same novel
property. The two category members exhibited either low diversity (i.e., they
shared all five features) or high diversity (they shared only one feature). (The
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 99

A
90

Generalization Rating
80
Nonblank
70

60
Blank
50

40

30
1 2 3 4
B
90
Generalization Rating

80 Nonblank

70

60

50 Blank

40

30
1 2 3 4
C
90
Generalization Rating

80

70
Nonblank
60
Blank
50

40

30
1 2 3 4
figure 4.4. Number of features in source example. Results from Rehder (2006),
Experiment 1.

two exemplars were chosen so that both always had three features, so that their
typicality was held constant across low- and high-diversity trials.) Whether
the novel property had a causal explanation was manipulated orthogonally.
On the basis of the SCM, the prediction was that the generalization of blank
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

100 B. Rehder

novel properties would be stronger when displayed by more diverse pairs of

exemplars. The critical issue concerned the effect of causal explanations on
the diversity effect.
In contrast to the previous experiment in which all subjects exhibited
a typicality effect with blank properties, only half of the subjects in this
experiment exhibited a diversity effect with blank properties. This result is
in keeping with the literature demonstrating the diversity heuristic’s general
lack of robustness relative to typicality (Lopez, Gelman, Gutheil, & Smith,
1992; see Heit, 2000, for a review). The important result, however, is that
when a causal explanation for a novel property was provided, any effect of
diversity disappeared entirely. That is, just as in the previous experiment, the
presence of a causal explanation can eliminate the effect of one of the SCM’s
similarity-based heuristic (in this case, diversity). Apparently, explanations
direct attention away from features which would normally contribute to how
well a category is “covered” by the exemplar(s) displaying a novel property.
When this occurs, the typicality or diversity of those exemplars becomes
irrelevant to the generalization.

causal explanations versus similarity. As mentioned, the SCM pre-

dicts that generalizations between items that are not hierarchically nested will
be influenced by their similarity. To determine how causal explanations affect
the influence of similarity, Rehder (2006, Experiment 3) asked participants
to generalize a property from one category member to another. One experi-
mental factor was whether the base and target exemplars shared three of four
features (high similarity condition) or only one (low similarity condition).
(The base and target were chosen so that they always possessed three and two
characteristic features, respectively, so that their typicality was held constant
over similarity conditions.) The other factor was whether a causal explanation
was provided for the novel property. For blank properties, SCM’s prediction
was that the generalization will be stronger for high versus low similarity tar-
gets. The critical question concerned the influence of similarity on nonblank
properties.
The results are presented in Figure 4.5 as a function of the similarity between
the base and target exemplar, and whether the novel property was blank or
nonblank (and, if nonblank, whether the cause feature was present or absent
in the target). As predicted by the SCM, blank properties were more strongly
projected when the target exemplar was more similar to the base exemplar.
Also as expected, nonblank properties were more strongly projected when
the cause of that property appeared in the target exemplar versus when it
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 101

100
Nonblank (Cause in target)

Generalization Rating
80

60
Blank
40

20
Nonblank (Cause not in target)
0
Low High
figure 4.5. Similarity of source and target. Results from Rehder (2006), Experiment 3.

did not appear. The important finding is that the projection of nonblank
properties was much less sensitive to the similarity of the base and target
exemplar. (Unlike the previous two experiments, the group-level pattern of
responding, shown in Figure 4.5, was exhibited by virtually all participants.)
Once again, it appears that the effect of causal explanations is to draw attention
away from features not involved in the explanation, with the result that the
two exemplars’ similarity becomes largely irrelevant to how strongly a novel
property is generalized from one to the other.6
At first glance it may seem that these results conflict with those from Had-
jichristidis et al. (2004) which showed that novel properties were projected
more strongly as a function of the similarity between the base and target
(Figure 4.1). Recall however that Hadjichristidis et al.’s claim is that similarity
will be used to estimate dependency structure when its presence in the target
category is otherwise uncertain. This situation was manifested in their ex-
periments by referring to a hormone’s dependents as nameless “physiological
functions,” making it impossible to determine whether those specific func-
tions were present in the target category. In contrast, in the current experiment
subjects were told exactly which category features were causally related to the
novel feature, a situation which apparently invited them to largely ignore how
the base and target might be similar on other dimensions.

6 Note that in this experiment the effect of similarity was not completely eliminated (as the effects
of typicality and diversity were eliminated in the first two experiments), as the generalization
rating of nonblank properties was on average 6 points higher (on a 100 point scale) when the
base and target were similar as compared to dissimilar, a difference which reached statistically
significance. Nevertheless, note that the magnitude of this similarity effect is vastly lower than
it is for blank properties (34 points). (See Lassaline, 1996, for related results.)
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

102 B. Rehder

In summary then, these three experiments support the claim that when a
causal explanation for a novel property is available, it often supplants simi-
larity as the basis for the generalization of that property. First, in the first ex-
periment all participants exhibited an effect of typicality for blank properties,
but half of those participants showed no sensitivity to typicality when the
novel property was accompanied with a causal explanation (nonblanks). Sec-
ond, in the second experiment half the participants exhibited sensitivity to
diversity for blank properties, but that sensitivity was completely eliminated
for nonblanks for all participants. Finally, in the third experiment virtually
all participants exhibited sensitivity to similarity for blanks, but this effect
was (almost) completely eliminated for nonblanks. Apparently, when people
note the presence of a causal explanation for a novel property, it often draws
attention away from the exemplars’ other features, making their similarity (or
typicality or diversity) largely irrelevant to the inductive judgment.

    

The preceding studies have addressed how the causal relations in which a novel
property is directly involved influence the generalization of that property.
But what of causal relations not involving the property itself, such as those
between the category’s existing features? For example, we know not only that
birds have wings, fly, and build nests in trees, but also that they build nests
in trees because they can fly and fly because they have wings. We know not
only that automobiles have gas, spark plugs, and produce carbon monoxide,
but also that gas and spark plugs interact to produce the carbon monoxide.
How does the coherence a category gains by these causal relations affect the
degree to which it supports the generalization of new blank properties, such as
sesamoid bones (for birds) or ceramic brakes (for automobiles)? Once again,
light on this question has been shed by studies testing both background and
experimentally provided causal knowledge.

Research with Background Causal Knowledge

Numerous studies have demonstrated that natural categories differ in system-
atic ways in the degree to which they support generalization. In one study,
Gelman (1988) found that second graders, but not preschoolers, were more
likely to generalize new properties to natural kinds rather than to artifacts.
She attributed this result to folk theories about the coherence of natural kinds
acquired by the older children. On this account, the coherence provided by
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 103

emerging theories of biological kinds led the older children to expect such
kinds to be more structured and constrained, and hence more homogenous.
Novel properties of biological kinds are generalized more strongly because
this expectation of homogeneity extends to new properties in addition to
existing one (also see Gelman et al., 1988; Shipley, 1993).
Additional evidence comes from studies investigating which level in a taxo-
nomic hierarchy supports the strongest inductions. Coley, Medin, and Atran
(1997) presented both American undergraduates and the Itza’ with a sub-
species (e.g., black vultures) that exhibited an unspecified disease and then
tested the degree to which that disease was generalized to vultures (a category
at the species or folk-generic level), birds (the life form level), or all animals (the
kingdom level). They found that for both groups the strength of generaliza-
tions dropped substantially when the target category was at the life form level
or higher. Again, biological species appear to be especially potent targets of
inductive generalizations (also see Atran, Estin, Coley, & Medin, 1997).
One notable aspect of these latter findings is the fact that American un-
dergraduates, like the Itza’, treated the species level as inductively privileged
despite the well-known result that for these individuals the basic level is nor-
mally one level higher in the hierarchy, namely, at the level of life form (e.g.,
tree or bird) rather than species (e.g., oak or robin) (Rosch, Mervis, Gray,
Johnson, & Boyes-Braem, 1976). One explanation that Coley et al. offer for
this discrepancy is that whereas the explicit knowledge that the American
students possess (which is tapped, for example, by the various tasks used by
Rosch et al., such as feature listing) is sufficient only for life forms to emerge
as identifiable feature clusters, their expectations are nevertheless that individ-
ual species are those biological categories with the most inductive potential.
(In contrast to the Americans, for the Itza’ the species level is both basic
and inductively privileged, a result Coley et al. suggest reflects their greater
explicit ecological knowledge as compared to the Americans.) They further
suggest that these expectations reflect an essentialist bias in which species are
presumed to be organized around a causally potent essence which determines
not only category membership, but also the degree to which category mem-
bers are likely to share (known and to-be-discovered) features (Gelman, 2003;
Medin & Ortony, 1989).
Can studies which experimentally manipulate the knowledge associated
with a category shed any light on the details of how interfeature theoret-
ical knowledge promotes generalizations? In particular, can we find direct
evidence that a category based on an underlying cause (i.e., an essence) pro-
motes generalizations?
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

104 B. Rehder

A.

F2
F4 F3

F4

B.

F1
F3 F4

F2

C.

F1 2 F3 F4

figure 4.6. Network topologies tested in Rehder and Hastie (2004), Experiment 2.
A. Common cause network. B. Common effect network. C. Chain network.

Research with Experimentally Provided Causal Knowledge

Rehder and Hastie (2004) addressed this question by manipulating the causal
relations between known category features in order to create a laboratory
analog of the essentialist knowledge associated with real-world categories.
Once again, subjects were instructed on artificial but plausible categories like
Romanian Rogos. In these experiments, however, the Rogos’ characteristic
features were described as causally related to one another rather than to a to-
be-generalized property. In one condition the causal links formed a common
cause topology in which one characteristic feature (F1 ) was the cause of the
three remaining ones (F2 , F3 , and F4 ), as shown in Figure 4.6A. For example,
for Romanian Rogos subjects would be told that butane-laden fuel (F1 ) caused
the other three characteristic features (e.g., loose fuel-filter gasket, hot engine
temperature, etc.). They were then presented with a series of Rogos which
each possessed a blank novel property (e.g., ceramic brakes) and judged what
proportion of all Rogos possessed that property. The particular combination
of characteristic features that the Rogos possessed was varied.
The question of interest was whether the presence of a common cause
causal structure, which was intended to mimic that of a category organized
around a causally potent essence (a common cause), would result in the same
enhanced inductive potential exhibited by biological species. However, we felt
it important also to ask whether any such enhancement required the specific
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 105

common-cause causal structure. For example, one alternative hypothesis is

that categories support stronger generalizations when they are organized
around any central feature or theme, regardless of whether that feature is a
cause or an effect. Another is that generalizations may be stronger as a result
of the global coherence afforded by interfeature relations, regardless of the
specific arrangement of the causal links. To assess these possibilities, we also
tested a common-effect network in which one feature (F4 ) is caused by the
other three features (F1 , F2 , and F3 ) (Figure 4.6B), and a chain network in
which feature F1 causes F2 which in turn causes F3 which in turn causes F4
(Figure 4.6C). We expected that only a common-cause network would result in
stronger generalizations if they are promoted by a single underlying generative
cause, both the common-cause and common-effect networks would do so if
they are promoted by any central theme, and all three networks would do
so if inductions are promoted by any category exhibiting overall coherence
provided by interfeature causal relations.
The results are presented in Figure 4.7 for each causal network as a function
of the particular combination of characteristic feature that the Rogo with the
to-be-generalized property possessed. Each panel in Figure 4.7 also includes
the results from a control condition which presented no causal links between
characteristic features. There were two notable results. First, Figure 4.7A shows
that, as predicted, common-cause categories supported stronger inductive
generalizations relative to the control condition, but only when the Rogo was
typical (i.e., possessed all four characteristic features, as denoted by “1111”).
When the Rogo was atypical (e.g., when it possessed the common cause
F1 but not the other three features, as denoted by “1000”), generalizations
were weaker than in the control condition. Note that in the common cause
condition this Rogo is not just atypical because it is missing three characteristic
features (F2 , F3 , and F4 ); it is incoherent in the light of the category’s causal
laws (the common cause F1 , butane-laden fuel, is present but its three effects
are absent).
The second important finding is that these effects were not unique to a
common-cause network, because the results were essentially the same for
categories with a common-effect network (Figure 4.7B) or a chain network
(Figure 4.7C). Each network supported stronger generalizations, so long as the
category member with the novel property was typical (1111). But when
the category member was incoherent, generalization were weaker than in
the control condition. Exemplar 0001 is incoherent in the common effect
condition because the common effect F4 is present even though its causes F1 ,
F2 , and F3 are all absent. Exemplar 0101 is incoherent in the chain condition
because F2 and F4 are present even though their causes F1 and F3 are absent
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

106 B. Rehder

A
100 Common Cause
90

Categorization Rating
80
70
60
50
40
30
20 Common Cause

10 Control
0
1000 1111
B
100 Common Effect
90
Categorization Rating

80
70
60
50
40
30
20 Common Effect
10 Control
0
0001 1111
C
100 Chain
90
Categorization Rating

80
70
60
50
40
30
20 Chain

10 Control
0
0101 1111
Exemplar
figure 4.7. Generalization ratings from Rehder and Hastie (2004), Experiment 1.

(and because F3 is absent even though its causes F2 is present). In other words,
it is not coherent categories that support stronger generalizations but rather
coherent category members, and this effect holds regardless of the topology
of the causal relations which make them coherent.7 As in my other studies

7 In a different phase of the experiment, Rehder and Hastie (2004) also asked the subjects to
estimate the degree of category membership of the same exemplars which displayed the novel
properties. Like the generalization ratings, category membership ratings for the most typical
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 107

reported in this chapter, these results held not just for artifacts like Rogos, but
also for biological kinds and for nonliving natural kinds.
These results have three implications for the received view that biological
kinds promote inductions because of the presumed presence of an essence.
First, it may not be a presumption of a single common cause (an essence)
which promotes inductions but rather a presumption of coherence, that is,
causal relations which link category features regardless of the specific topology
of those links. Second, this effect of coherence is not limited to biological kinds
but will apply equally well to other kinds (e.g., artifacts) when generalizers
have reason to believe that one of those kinds is coherent. Third, this effect
can be reversed when the specific category member which displays a novel
feature is incoherent in the light of the category’s causal laws.
Of course, it should be noted that our failure to discover any special in-
ductive potential of a common cause network as compared to the other two
networks may have obtained because our experimental materials failed to
fully elicit our subjects’ essentialist intuitions. For example, perhaps such
intuitions depend on the underlying nature of a category remaining vague
(e.g., even young children have essentialist intuition, even though they know
nothing of DNA; see Gelman, 2003), but in our categories the common
cause was an explicit feature (e.g., butane-laden fuel for Rogos). Or, per-
haps those explicit features clashed with our adult subjects’ own opinions
about the nature of the “essence” of a brand of automobile (the intention of
the automobile’s designer; see Bloom, 1998) or a biological species (DNA).
Nevertheless, these results raise the possibility that the inductive potential of
categories thought to arise from folk essentialism about biological kinds may
in fact arise from a more general phenomenon, namely, theoretical coher-
ence, which potentially applies to kinds of categories in addition to biological
kinds.


This chapter has reviewed three sorts of evidence related to how causal
knowledge is involved in property generalization. The first section presented

exemplars (i.e., 1111) were higher in the causal conditions than in the control condition, and
ratings for incoherent exemplars (e.g., 1000, 0001, and 0101 in the common cause, common
effect, and chain conditions, respectively) were lower than in the control condition. That is,
the causal knowledge which makes exemplars appear more (or less) coherent is reflected in
their perceived degree of category membership in addition to their propensity to support
generalizations. Overall, the correlations between generalization and category membership
ratings were .95 or higher.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

108 B. Rehder

evidence that people can go beyond the similarity-based heuristics for-

malized by the Similarity-Coverage model and generalize on the basis of
causal reasoning. The second section addressed how causal reasoning and
the similarity-based heuristics interact, and it showed that subjects which
showed an effect of causal relations exhibited virtually no effect of similarity,
whereas those which showed no effect of causal relations exhibited a sensitiv-
ity to similarity. Apparently, the use of a causal explanation versus similarity
is an all-or-none matter, with reasoners using one strategy or the other, but
not both. The third section reviewed evidence suggesting that categories pre-
sumed to be based on an underlying cause seem to be the target of strong
generalizations, and it found that stronger generalizations may not be the
responsibility of a single generative essence but rather of the presumed co-
herence that a category gains from the causal knowledge which interrelates
existing features.
The evidence in support of these claims came from studies testing both nat-
ural and artificial categories, approaches which have complimentary strengths
and weaknesses. For example, a critic could dismiss the studies testing arti-
ficial categories on the grounds that the causal knowledge was provided to
participants as part of the experimental session rather than coming from a
participants’ own background knowledge, a practice which was likely to have
made that knowledge especially salient and available (bypassing the need for
the reasoner to retrieve it from their semantic memory). However, this crit-
icism is blocked by the numerous studies testing natural categories, which
show that it is common for reasoners to use causal knowledge in the service of
generalizations. The benefit of testing artificial categories, in turn, is that it has
made it possible to test specific claims regarding how causal knowledge influ-
ences generalization. One view is that properties are more generalizable to the
extent they are more central (i.e., have more dependents in the base and tar-
get categories) (Hadjichristidis et al., 2004). My own view is, in contrast, that
people engage in a kind of causal reasoning in which they evaluate whether
a property is likely to appear in a target category. Data were presented show-
ing that people can reason forwards from the presence of to-be-generalized
property causes to the property itself (prospective reasoning) as well as they
can reason from the property’s effects or dependents (diagnostic reasoning).
Additional experiments also showed that people will judge a new property to
be more prevalent in a population of category members to the extent that its
causes and/or effects are present in that population.
Although this chapter’s central claim is that properties are often generalized
on the basis of causal reasoning, it is important to consider whether some of
the results could be explained by a more sophisticated version of similarity.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 109

For example, recent progress has been made in extending similarity-based

models to account for some results with nonblank properties, such as the Gap2
model proposed by Blok, Osherson, and Medin (this volume). In addition,
one component of Medin et al.’s (2003) relevance theory of induction is that
similarity emerges from a process of comparing the premise and conclusion
categories. For example, comparing skunks and zebras would yield stripedness
as a highly relevant feature, whereas comparing skunks and onions would yield
strong odor (or, equivalently, stripedness becomes available in the context of
skunks and zebras whereas strong odor is available in the context of skunks and
onions; Shafto, Coley, & Vitkin, this volume). However, by themselves neither
the Gap2 model nor comparing premise and conclusion categories are able
to account, for example, for the Heit and Rubinstein (1994) study presented
earlier because similarity doesn’t change as a function of the property involved.
The import of the Heit and Rubinstein results, however, is just that the relative
strength of generalization (between, say, bears and whales as compared to
tunas and whales) varied as a function of whether the property was anatomical
or behavioral.
Heit and Rubinstein themselves proposed that their results could be ex-
plained in terms of similarity, but that similarity was computed flexibly, that
is, in a manner which is influenced by the to-be-generalized property. Indeed,
they showed that judgments regarding the generalization of anatomical and
behavioral properties were well predicted by the rated similarity of the base
and target categories when those ratings were made with respect to anatomy
or behavior, respectively. Heit and Rubinstein suggested that anatomy and be-
havior may be two of a small number of fixed similarity measures which people
can compute, or that people might use the theoretical knowledge invoked by
the to-be-generalized property to dynamically select which dimensions along
which similarity between base and target should be computed. The latter
account can also explain some of the results from Rehder (2006) presented in
the second section. For example, there I showed that the presence of a causal
explanation linking the to-be-generalized property to an existing feature in
the base resulted in reasoners being very sensitive to whether that feature was
present in the target and insensitive to whether the base and target were sim-
ilar on other dimensions (Figure 4.5). I described these results as indicating
that the explanation drew attention away from the other dimensions, which
of course is equivalent to saying that it determined the (single) dimension
along which base-target similarity was computed.
Nevertheless, these more sophisticated views of similarity fail to explain, for
instance, the diagnostic reasoning which apparently occurs during some cases
of generalization. In the Hadjichristidis et al. (2004) study, subjects were asked
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

110 B. Rehder

to generalize a fictitious hormone from one category to another (e.g., from

seals to dolphins). According to Heit and Rubinstein, the fact that the novel
property is a hormone might induce reasoners to adopt an anatomical notion
of similarity. But this fails to explain why the hormone was projected more
strongly from seals to dolphins when it was responsible for many versus few
physiological functions, because in both cases the base category (seals), the
target category (dolphins), and the property (the hormone) are the same, and
thus so too is (anatomical) similarity. Clearly, to explain these results one must
also make reference to the causal network in which the hormone is involved.
The view I promote is that generalizers reasoned backward (diagnostically)
from the presence of the many physiological functions to the presence of the
hormone.
As mentioned, testing artificial categories has made it possible to begin to
develop specific computational theories of how causal knowledge influences
generalization. But while a start has been made, my own work has considered
only the causal knowledge which interrelates a category’s features. But cate-
gories are also related by causal relations to one another, and research reviewed
above documented the influence of one type of inter-category causal link –
ingestion – on induction (e.g., a property displayed by both cats and sparrows
will be attributed to the fact that cats eat sparrows). Indeed, in addition to
comparison-based similarity, a second component of Medin et al.’s (2003)
relevance theory is the assumption that causal relations between premise and
conclusion categories will often form the basis of induction. Progress in pro-
viding a computational account of the influence of ingestion relations as one
type of inter-category relation is reported by Tenenbaum, Kemp, and Shafto
(this volume). In this research, subjects were trained to memorize a particu-
lar “food web” consisting of a number of animal species which prey on one
another. For example, one food web included the facts that sand sharks eat
herring which eat kelp; another that lions eat wolves which eat squirrels. Sub-
jects were then asked to generalize a disease property “D” from one species
to another in the same food web. Their judgments were well-predicted by a
Bayesian model which computed the probability with which the disease was
transmitted through the web from the base species to the target. Interestingly,
when the property was “gene XR-23,” the food-web model performed poorly,
and it was a Bayesian model which operated on a taxonomic representation
of the species that performed well instead. This result is reminiscent of Heit
and Rubinstein’s suggestion that the nature of the to-be-generalized property
influences the mental processes responsible for induction, but it suggests that
this influence amounts to recruiting different complex knowledge structures
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 111

(a food-web versus a taxonomic hierarchy) rather than just different ways of

computing similarity.
More work is needed, because of course diseases can spread in ways other
than ingestion (e.g., physical proximity; see Proffitt et al., 2000). And the
generalization of other types of properties might depend on yet other types
of inter-category relations (properties might be shared between substances
based on their common origin, e.g., between sand and glass; Medin et al.,
2003). But one common thread between the causal reasoning view I have
advocated here and that of Tenenbaum et al. (this volume) is the emphasis
on domain-general reasoning processes. For example, one striking aspect of
the results from Rehder (2006) and Rehder and Hastie (2004) testing artificial
categories is the consistency of the results across biological kinds, nonliving
natural kinds, or artifacts. Such results suggest that domain differences found
with natural categories arise from differences in the type and amount of
knowledge associated with those domains. Of course, this is not to suggest
that such differences are theoretically unimportant, because one can always
ask what causes them to exist. For example, it may be that young children make
unique assumptions about the organization of biological kinds (e.g., that they
are based on some kind of essence, or, as I prefer, that they are presumed to
be coherent). Even so, it is causal reasoning which directly mediates whether
or not a property is generalized to an entire class of objects.

References
Ahn, W. (1998). Why are different features central for natural kinds and artifacts? The
role of causal status in determining feature centrality. Cognition, 69, 135–178.
Ahn, W., Kim, N. S., Lassaline, M. E., & Dennis, M. J. (2000). Causal status as a
determinant of feature centrality. Cognitive Psychology, 41, 361–416.
Atran, S., Estin, P., Coley, J. D., & Medin, D. L. (1997). Generic species and basic levels:
Essence and appearance in folk biology. Journal of Ethnobiology, 17(1), 22–45.
Bailenson, J. N., Shum, M. S., Atran, S., Medin, D. L., & Coley, J. D. (2002). A bird’s eye
view: Biological categorization and reasoning within and across cultures. Cognition,
84, 1–53.
Bloom, P. (1998). Theories of artifact categorization. Cognition, 66, 87–93.
Cheng, P. (1997). From covariation to causation: A causal power theory. Psychological
Review, 104(2), 367–405.
Coley, J. D., Medin, D. L., & Atran, S. (1997). Does rank have its privilege? Inductive
inference within folkbiological taxonomies. Cognition, 64, 73–112.
Gelman, S. A. (1988). The development of induction within natural kinds and artifact
categories. Cognitive Psychology, 20, 65–95.
Gelman, S. A. (2003). The essential child: The origins of essentialism in everyday thought.
New York: Oxford University Press.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

112 B. Rehder

Gelman, S. A., & Markman, E. M. (1986). Categories and induction in young children.
Cognition, 23, 183–208.
Gelman, S. A., & O’Reilly, A. W. (1988). Children’s inductive inferences with superordi-
nate categories: The role of language and category structure. Child Development, 59,
876–887.
Hadjichristidis, C., Sloman, S. A., Stevenson, R., & Over, D. (2004). Feature centrality
and property induction. Cognitive Science, 28(1), 45–74.
Heit, E. (2000). Properties of inductive reasoning. Psychonomic Bulletin & Review, 7(4),
569–592.
Heit, E., & Rubinstein, J. (1994). Similarity and property effects in inductive reasoning.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(2), 411–
422.
Lassaline, M. E. (1996). Structural alignment in induction and similarity. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 22(3), 754–770.
Lopez, A., Atran, S., Coley, J. D., Medin, D. L., & Smith, E. E. (1997). The tree of life:
Universal and cultural features of folkbiological taxonomies and inductions. Cognitive
Psychology, 32(3), 251–295.
Lopez, A., Gelman, S. A., Gutheil, G., & Smith, E. E. (1992). The development of
category-based induction. Child Development, 63, 1070–1090.
Medin, D. L., Coley, J. D., Storms, G., & Hayes, B. K. (2003). A relevance theory of
induction. Psychonomic Bulletin & Review, 10(3), 517–532.
Medin, D. L., & Ortony, A. (1989). Psychological essentialism. In S. Vosniadou & A.
Ortony (Eds.), Similarity and analogical reasoning (pp. 179–196). Cambridge: Cam-
bridge University Press.
Nair, M. (2005). The role of causal reasoning in category-based generalizations. Unpub-
lished honors thesis. New York University.
Osherson, D., Smith, E. E., Myers, T. S., & Stob, M. (1994). Extrapolating human
probability judgment. Theory and Decision, 36, 103–126.
Osherson, D. M., Smith, E. E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Proffitt, J. B., Coley, J. D., & Medin, D. L. (2000). Expertise and category-based induction.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(4), 811–828.
Rehder, B. (2003a). Categorization as causal reasoning. Cognitive Science, 27(5), 709–748.
Rehder, B. (2003b). A causal-model theory of conceptual representation and catego-
rization. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29(6),
1141–1159.
Rehder, B. (2006). When similarity and causality compete in category-based property
induction. Memory & Cognition, 34, 3–16.
Rehder, B., & Hastie, R. (2004). Category coherence and category-based property in-
duction. Cognition, 91(2), 113–153.
Rehder, B., & Kim, S. W. (2006). How causal knowledge affects classification: A generative
theory of categorization. Journal of Experimental Psychology: Learning, Memory, &
Cognition, 32, 659–683.
Rips, L. J. (2001). Necessity and natural categories. Psychological Bulletin, 127(6), 827–
852.
Rosch, E. H., Mervis, C. B., Gray, W., Johnson, D., & Boyes-Braem, P. (1976). Basic
objects in natural categories. Cognitive Psychology, 8, 382–439.
P1: IBE
0521672443c04 CUFX144-Feeney 0 521 85648 5 July 19, 2007 14:32

Property Generalization as Causal Reasoning 113

Shafto, P., & Coley, J. D. (2003). Development of categorization and reasoning in the
natural world: Novices to experts, naive similarity to ecological knowledge. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 29(4), 641–649.
Shipley, E. F. (1993). Categories, hierarchies, and induction. In D. Medin (Ed.), The
psychology of learning and motivation (Vol. 30, pp. 265–301). New York: Academic
Press.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Sloman, S. A. (1994). When explanations compete: The role of explanatory coherence
on judgments of likelihood. Cognition, 52, 1–21.
Sloman, S. A. (1997). Explanatory coherence and the induction of properties. Thinking
and Reasoning, 3(2), 81–110.
Sloman, S. A., Love, B. C., & Ahn, W. (1998). Feature centrality and conceptual coherence.
Cognitive Science, 22(2), 189–228.
Smith, E. E., Shafir, E., & Osherson, D. (1993). Similarity, plausibility, and judgments of
probability. Cognition, 49, 67–96.
Springer, K., & Keil, F. C. (1989). On the development of biologically specific beliefs:
The case of inheritance. Child Development, 60, 637–648.
Thagard, P. (1999). How scientists explain disease. Princeton, NJ: Princeton University
Press.
Wu, M. L., & Gentner, D. (1998). Structure in category-based induction. Paper presented
at the 20th Annual Conference of the Cognitive Science Society, Madison, WI.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction

Patrick Shafto, John D. Coley, and Anna Vitkin

Induction can be supported by many kinds of knowledge. To be effective,

reasoning should be context sensitive; different kinds of knowledge should
be selectively deployed in different situations. For example, in the domain
of biology, when reasoning about the distribution of novel internal proper-
ties over species, taxonomic knowledge should be recruited since we know
that taxonomic membership is not only related to perceptual similarity but
is also a good predictor of shared unobservable anatomical features such as
four-chambered hearts. However, when reasoning about the distribution of
environmental toxins, ecological knowledge should be recruited since such a
toxin would plausibly spread through an ecosystem. In this chapter, we address
the factors that influence the recruitment of different kinds of knowledge in
different contexts. We propose that different kinds of knowledge are differ-
entially available across contexts. Using this concept of availability, we will
address an array of experimental results, arguing for availability as a way to
unite and explain a broad range of phenomena in category-based induction.
In a classic paper, Tversky and Kahneman (1973) discuss availability as a
heuristic “by which people evaluate the frequency of classes or the likelihood
of events” (p. 207). This involves estimating frequency or probability “by
the ease with which instances or associations are brought to mind” (p. 208).
As such, availability on this view is essentially a metacognitive heuristic by
which information is judged more likely or plausible based on an estimate
of the effort involved in retrieving the information; indeed, Tversky and
Kahneman argue that “to assess availability it is not necessary to perform the
actual operations of retrieval or construction. It suffices to assess the ease
with which these operations could be performed” (p. 208). Our goal in this
chapter is to use the spirit of this concept of availability – if not the letter –
to unify a set of seemingly disparate findings on category-based inductive
reasoning.
114
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 115

We see inductive inference as a process by which knowledge of relations

among concepts is used to assess the likelihood that members of two categories
will share a novel property. We will argue that the ease with which specific
knowledge of conceptual relations comes to mind predicts the likelihood
that such knowledge will be used to guide inductive inference. Below, we
will present evidence that the availability of different kinds of knowledge to
inform induction depends on two factors: acute (short-term) influences of
context on availability and chronic (long-term) effects of experience, reflected
in baseline differences in the availability of different knowledge structures. In
this sense, availability can be thought of as a dynamic property of knowledge
in memory that provides a ranking of the likelihood that different kinds of
knowledge will be accessed in a given context (e.g., Horowitz et al., 1966;
Pearlstone & Tulving, 1966).
Thus, like Tversky and Kahneman (1973), we argue that availability is a
variable that mediates between prior knowledge and behavior via the relative
accessibility of different kinds of knowledge in a particular situation. Here we
part ways with Tversky and Kahneman because for our purposes an estimate of
the ease with which information might be accessed is not sufficient. Because
specific knowledge must be brought to mind to guide or assess inductive
inference, it is necessary to “perform the actual operations of retrieval or
construction.”
The structure of this chapter will be as follows. First, we will discuss the
effects of two kinds of context on availability. We will then turn to how expe-
rience influences availability and interacts with context. Finally, we will relate
this approach to other frameworks, and discuss novel predictions motivated
by availability.

  - 

A central problem in the psychology of induction is what kinds of knowledge
we use to guide a particular inductive inference. Given all that we may know
about a concept, which knowledge is relevant to a particular problem? This
problem is beyond the scope of many previous models of inductive reason-
ing, which rely on a single kind of knowledge in a domain (though some
recent progress has been made toward models supporting multiple kinds of
knowledge, see Shafto, Kemp et al., 2005; Shafto et al., in prep). We propose
a framework for understanding reasoning in rich domains, explicitly linking
the acute effects of immediate context and the chronic effects of long-term
experience via the availability of different kinds of knowledge. We suggest that
inductive generalizations are guided by the knowledge that is most available
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

116 P. Shafto, J. D. Coley, and A. Vitkin

in a particular context, given past experience. Moreover, availability is dy-

namic in that changes in availability arise from one of two sources: changes
in context or changes in underlying knowledge.
Context plays a crucial role in acute changes in availability. For any given
inductive task, the presence of a biasing context can change the relative avail-
ability of different kinds of knowledge. For example, if you learn of a new
environmental toxin affecting frogs, ecological knowledge about who may be
exposed or causal knowledge about how the toxin might be transmitted may
become more available, leading you to expect that herons might be at risk
(because they share a habitat with and eat frogs). In contrast, if you learn
of a new enzyme discovered in frogs, you might expect the same enzyme to
be present in toads, due to their close biological affinity, and never consider
herons as a plausible candidate for the enzyme. Even if we know nothing
about the nature of the property in question, any known examples that pos-
sess it may influence the availability of knowledge. For example, if told about
a property that cows, chickens, and pigs have, knowledge about domesticated
animals may become more available.
In the case where context does not provide a bias, availability reduces
to a default way of thinking about a domain, which reflects the knowledge
that has been rendered chronically most available by experience. Experience
may lead to multiple kinds of knowledge all becoming equally available or a
single kind of knowledge becoming more available over time. For example,
the relative salience of the ecological versus taxonomic knowledge that inform
the inferences about frog might differ for an ecologist working for Greenpeace,
a molecular biologist, or a relatively uninformed undergraduate participant
in a psychology experiment. The chronic effects of experience are manifest in
baseline availability of default knowledge in a domain.
Importantly, acute and chronic changes in availability can interact, which
can have important implications for reasoning in a domain. If, at baseline,
multiple kinds of knowledge are equally available, reasoning will be highly
sensitive to acute changes in availability as a result of context. However, if at
baseline one kind of knowledge is a highly available default, much stronger
context will be required to elicit reasoning based on the less-available knowl-
edge. Thus, the interaction between acute and chronic changes in availability
has an important impact on the kinds of knowledge that are available to guide
reasoning.
In the following sections, we will consider several studies that demonstrate
that context and experience result in important changes in inductive reason-
ing. Most important, we will consider studies that show interactions between
experience and context in inductive reasoning. We will argue that these effects
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 117

are best understood in terms of acute and chronic changes in the availability
of different kinds of knowledge.

Context-Based Changes in Availability

We suggest that the context provided by an inductive argument results in
acute changes in the availability of different kinds of knowledge and different
patterns of reasoning. In this section, we will consider a range of previous
results investigating reasoning in knowledge-rich domains. We will reconsider
this evidence in the light of availability, particularly looking for qualitative
changes in reasoning based on immediate inductive context. We consider two
sources of context in these category-based induction tasks: the property in
question, and the set of categories given in the problem.

Property Effects on Induction. One source of context in an inductive problem

is the property to be generalized from premise to conclusion. Knowing some-
thing about the nature of this property can be informative as to what relations
are recruited for making guesses about the distribution of that property. For
example, if we are told that chickens have sesamoid bones we might con-
clude that other flightless birds might share a physiological, internal property
such as bone structure. On the other hand, if we knew that chickens had the
sesamoid flu we might reconsider having chicken for dinner for fear of catch-
ing it ourselves. Although in each example the premise category is chicken, the
conclusion categories we believe to share a property can change dramatically
depending on what that property is. The systematic use of different kinds
of knowledge to inform induction has been termed inductive selectivity. On
our view, the immediate context provided by the property of an inductive
argument can produce acute changes in the availability of different kinds of
knowledge for inductive generalizations.
Heit and Rubinstein (1994, exp. 1 & 2) proposed that different relations
between premise and conclusion categories determined the strength of an
inductive inference depending on the property a premise and conclusion
were said to share. In their study they asked participants to make estimates of
the probability that a pair of animal categories might share a property. Pairs of
animals were chosen to be related by shared anatomy (e.g., whale and bear) or
behavior (e.g., whale and tuna). Participants were asked about the likelihood
of such pairs sharing an anatomical property like having two chambers in their
liver that act as one, or a behavioral property such as traveling in a back-and-
forth or zigzag trajectory. Heit and Rubinstein found that participants made
the highest probability estimates for items where the relation between a pair
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

118 P. Shafto, J. D. Coley, and A. Vitkin

of animals matched the kind of property they were asked to reason about (e.g.,
whales and bears have two-chamber livers, whales and tuna travel in zigzag
trajectories). These results suggest that the property of projection influenced
the kind of knowledge that was recruited to support inductive inferences:
anatomical properties made anatomical knowledge more available, whereas
behavioral properties made behavioral knowledge more available.
Ross and Murphy (1999) have shown similar effects of property on the se-
lective use of conceptual relations to guide inferences in the domain of food.
They established that most participants cross-classified food into two major
knowledge structures – taxonomic, based on shared features or composition,
and script, based on what situations a food is consumed in. Participants in
their study were asked to make biochemical or situational inferences about
triplets of food. Participants were taught that the target food, such as bagels,
had a biochemical property (enzyme) or situational property (eaten at an
initiation ceremony). They were then asked to project the property to one
of two alternatives, a taxonomic alternative such as crackers, or a script al-
ternative such as eggs. Ross and Murphy found that participants made more
taxonomic choices for biochemical properties and more script choices when
considering situational properties. It seems that the nature of the property
increased the availability of relevant knowledge about shared composition or
situational appropriateness of a food, ultimately producing different patterns
of induction for different kinds of properties.
Recent evidence also suggests that, like adults, children’s inductive general-
izations are sensitive to the property in question. Using a similar triad method
to that of Ross and Murphy (1999), Nguyen and Murphy (2003) found that
in the domain of food, seven-year-old children (but not four-year-olds) made
more taxonomic choices when reasoning about a biochemical property and
more script choices when reasoning about a situational property. That is, they
thought that a bagel would be more likely to share an enzyme with crackers
but be more likely to be eaten at a ceremony with eggs. This suggests that the
context provided by a property begins to mediate the differential availability
of knowledge from an early age.
Further evidence of children’s selective use of knowledge based on prop-
erty comes from the domain of biology. Coley and colleagues (Coley, 2005;
Coley, Vitkin, Seaton, & Yopchick, 2005; Vitkin, Coley, & Kane, 2005) asked
school-aged children to consider triads of organisms with a target species,
a taxonomic alternative (from the same superordinate class but ecologically
unrelated), and an ecological alternative (from a different taxonomic class
but related via habitat or predation). Children were taught, for example, that
a banana tree had a property (either a disease or “stuff inside”) and asked
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 119

to choose if a calla lily (taxonomic match) or a monkey (ecological match)

shared the property. Results clearly indicate that children were sensitive to
the kind of property they were asked about, choosing taxonomic matches
at above-chance levels when reasoning about insides, and choosing ecologi-
cal matches at above-chance levels when reasoning about diseases that might
plausibly spread through an ecosystem. In another task, children were asked to
make open-ended predictions about what kinds of things might come to share
properties (Vitkin, Coley, & Hu, 2005). They were taught that a pair of animals
either had a disease or “stuff inside,” and they were then asked to project what
other things might share this property and explain their reasoning. Children’s
responses were characterized as being based on similarity (taxonomy, shared
features, etc.) or interaction (contact through ecosystems). Children made
more responses rooted in similarity when thinking about internal substances
and more interaction-based responses when asked to consider diseases. Con-
sistent with availability, this pattern of response demonstrated that children’s
generalizations were guided by different kinds of knowledge depending on
what properties they were reasoning about.
Overall, we have seen that the property projected in an inductive task
provides context for inferences, changing the kinds of information that are
used make generalizations. Specifically, the property being projected pro-
vides useful clues about what kinds of conceptual relations (disease trans-
mitted through ecology, biochemical composition shared by taxonomic food
classes, etc.) might support the generalization of such a property. This has
been demonstrated in at least two different domains of knowledge (food
and biology) and in both children and adults. These results are consistent
with the idea that the property being projected is one factor that influences
the relative availability of different kinds of knowledge to support inductive
reasoning.

Relations Among Premise and Conclusion Categories. Another line of evi-

dence compatible with the proposal that availability mediates the basis of in-
ductive reasoning examines the effects of relations among premise categories,
or among premise and conclusion categories, as influencing the availability of
different kinds of knowledge for guiding inductive inference. Work by Medin,
Coley, Storms, and Hayes (2003) identified several effects associated with the
relationship between categories presented in the premises and conclusions,
under the rubric of relevance effects. One key component of the relevance
framework, as well as our availability framework, is the idea that salient
relations among premise categories, or between premise and conclusion
categories, may guide or constrain the inference supported by those premises.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

120 P. Shafto, J. D. Coley, and A. Vitkin

Medin et al. (2003) examine this idea with respect to two broad classes of
phenomena, causal relations and property reinforcement. We concentrate on
the latter. Medin et al. (2003) present several examples where increasing the
salience of specific relations among categories in an inductive argument leads
to violations of normative logic or of the predictions of similarity-based
models of inductive reasoning.
The first of these is non-diversity via property reinforcement, which predicts
that an argument with less diverse premises might be perceived to be stronger
than an argument with more diverse premises if the premise categories of the
more diverse argument reinforce a salient relation not shared by the con-
clusion category. For example, consider the following arguments:

A. Polar Bears have Property X.

Antelopes have Property X.
All Animals have Property X.

B. Polar Bears have Property X.

Penguins have Property X.
All Animals have Property X.

From a strictly taxonomic point of view, polar bears (a mammal) and penguins
(a bird) provide better coverage of the conclusion category animal than polar
bears and antelopes, which are both mammals. Thus, a model based only on
taxonomic knowledge must predict Argument B to be stronger. However, the
salient property shared by polar bears and penguins – namely, adaptation to
a cold climate – renders plausible the possibility that Property X is related
to life below zero, and therefore might weaken the inference that all animals
would share the property. Indeed, Medin at al. (2003) find that subjects in
the United States, Belgium, and Australia on average rated arguments like A
stronger than arguments like B.1 This suggests that the salient property shared
by the premise categories in B cancels out the greater coverage they provide
of the conclusion category.
A second related phenomenon discussed by Medin et al. (2003) is conjunc-
tion fallacy via property reinforcement, which predicts that arguments with a
single conclusion category might be perceived to be weaker than arguments
with an additional conclusion category (a violation of normative logic because
a conjunctive statement cannot be more probable than one of the component
statements) if the second conclusion category reinforces a salient relation
shared by all categories. For example, consider the following arguments:

1 Though recent evidence suggests that this effect may not be particularly robust (Heit & Feeney,
2005).
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 121

C. Chickens have Property X.

Cows have Property X.

D. Chickens have Property X.

Cows have Property X.
Pigs have Property X.

Normative logic requires that “cows have Property X” must be more likely
than “cows and pigs have Property X,” but Medin et al. (2003) found that par-
ticipants reliably rated arguments like D as more likely than arguments like
C. The addition of pigs might serve to increase the availability of the knowl-
edge about farm animals, and therefore strengthens Argument D relative to
Argument C.
Finally, Medin et al. (2003) discuss non-monotonicity via property rein-
forcement. Monotonicity is the idea that all else being equal, adding premise
categories that are proper members of the same superordinate as a conclusion
category should strengthen the argument (see Osherson et al., 1990). Medin
et al. (2003) predict that adding premises might weaken an argument if the
added categories reinforce a relation shared by premise categories but not the
conclusion category. For example, consider the following arguments:

E. Brown Bears have Property X.

Buffalo have Property X.

F. Brown Bears have Property X.

Polar Bears have Property X.
Black Bears have Property X.
Grizzly Bears have Property X.
Buffalo have Property X.

Monotonicity predicts that Argument F should be at least as strong as Ar-

gument E because the premise categories in F necessarily cover the lowest
inclusive category mammals at least as well as the premise of E, and are at least
as similar to buffalo. However, Medin et al. (2003) find that – contrary to the
principle of monotonicity – arguments like E are consistently given stronger
ratings than arguments like F, presumably because the premises of F reinforce
the relation of being a bear, and therefore make plausible the inference that
Property X is particularly ursine in nature, and therefore unlikely to be true
of buffalo.
Together, these results suggest that manipulation of relations among pre-
mise and/or conclusion categories can result in violations of normative
logic (in the case of the conjunction fallacy) or violations of predic-
tions derived from similarity-based models (in the case of non-diversity
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

122 P. Shafto, J. D. Coley, and A. Vitkin

table 5.1. Sample items from adult open-ended

projection study

Ecological Taxonomic Distance

Relation Near Far
Related Heron, Duck Hawk, Field Mouse
Unrelated Otter, Deer Chipmunk, Bullfrog

or non-monotonicity). We propose that these manipulations can be seen as

manipulations of the availability of specific conceptual relations. By rendering
specific relations such as polar animals, farm animals, or bears highly available,
these manipulations serve to overcome more general default approaches to
the evaluation of inductive arguments.
Using a more natural paradigm that allows participants to generate their
own conclusions given a set of premises, Baker and Coley (2005) (see also
Coley, Baker, & Kemp, 2004) have investigated whether spontaneous and rel-
atively unconstrained inductive projections are sensitive to manipulations of
relations among premise categories. In this study, thirty college undergrad-
uates were given pairs of premise species, taught a novel property said to be
shared by the pair, and asked what other species might have the property, and
why. Premise pairs were either drawn from the same superordinate category
(taxonomically near pairs) or from different superordinates (taxonomically
far pairs). Pairs were also ecologically related (via habitat, predator-prey rela-
tion, or ecological niche) or unrelated. See Table 5.1 for sample items.
Reponses were coded according to the relationship between the given
premise pair and the conclusion categories generated by the participants.
The basis of an inference was judged to be taxonomic if participants’ expla-
nations emphasized that premise and conclusion categories belonged to the
same class or kind, were similar in appearance, or similar in general. Responses
were coded as ecological if participants’ explanations relied on an interaction
between premise and conclusion categories that highlighted a non-taxonomic
relation such as a predator/prey relation, shared diet, or shared habitat.
Results suggest that projections were sensitive to salient relations among
premise categories. Specifically, taxonomic inferences were more frequent for
taxonomically near pairs than for taxonomically far pairs, and were also more
frequent for ecologically unrelated than for ecologically related pairs. Likewise,
ecological inferences were more frequent for ecologically related pairs than for
unrelated pairs, although frequency of ecological inferences was not affected
by taxonomic distance among premise categories. In sum, relations among
premise categories appear to influence the knowledge recruited to support
generalizations.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 123

These findings suggest that relations among premise categories may im-
pact the availability of different kinds of knowledge for guiding sponta-
neous inferences. Premise categories that share salient taxonomic relations
render such knowledge available and thereby increase the likelihood of
taxonomic generalizations. Likewise, premise categories that share salient
ecological relations increase the availability and likelihood of ecological
inferences.

Availability, Experience, and Default Domain Knowledge

Availability, as we have described it, depends on prior knowledge and context.
In the previous section, we considered acute changes in availability due to the
nature of the property being projected and relations among categories in an
argument. In this section, we present evidence that increased knowledge and
experience in a domain can lead to chronic changes in the relative availability
of different kinds of knowledge for inductive reasoning. We will argue that
experience-related changes in underlying knowledge – such as accrual of
more facts, changes in the frequency with which different knowledge is used,
and even fundamental changes in conceptual structure – are accompanied by
changes in the baseline availability of different kinds of knowledge, and also
by an increased sensitivity to context. In brief, we argue that the concept of
availability is useful in explaining observed effects of experience on inductive
reasoning.

Experience-Related Changes in Availability of Different Kinds of Knowledge.

By definition, experts in a given domain know more than novices. However,
in addition to the accrual of facts, expertise may also result in changes to
the relative availability of different kinds of knowledge. We argue above that
taxonomic knowledge may be a default guide for novices when reasoning
about a given domain. However, several lines of evidence suggest that tax-
onomic relations consistently fail to predict expert reasoning to the same
extent. For instance, López et al. (1997) show that when forced to choose the
stronger of two inductive arguments about local mammals species, Univer-
sity of Michigan undergraduates’ responses are almost unanimously in ac-
cord with the Osherson et al. (1990) taxonomically based similarity-coverage
model. In contrast, responses of Itza’ Maya participants – indigenous people
of Guatamala who live in close contact with nature and depend largely on
hunting and swidden agriculture for subsistence, and who therefore possess
extensive knowledge of local flora and fauna – were at chance when evaluated
against predictions of the Similarity-Coverage Model. Instead of relying on
taxonomic relations, the Itza’ recruited specific causal-ecological knowledge
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

124 P. Shafto, J. D. Coley, and A. Vitkin

to evaluate the arguments. Follow up work by Proffitt, Coley, and Medin

(2000) revealed a pattern of induction among Chicago-area tree experts that
was remarkably similar to that of the Itza’, suggesting that domain-specific
experience, rather than language or culture, is the driving factor. Indeed, this
same general finding – increase in the relative salience of non-taxonomic re-
lations for guiding induction – has been reported for commercial fishermen,
professional musicians, and even undergraduates reasoning about alcohol
(see Coley, Shafto, Stepanova, & Baraff, 2005, for a review).
Not only does experience in general change the relative availability of dif-
ferent kinds of knowledge, but specific kinds of expertise also appear to lead to
differential salience of knowledge among experts. For instance, Proffitt et al.
(2000) report that of three different groups of experts (taxonomists, land-
scapers, and maintenance workers), only taxonomists’ patterns of induction
relied on taxonomic knowledge. Taxonomists were also more likely to explain
their inferences by referring to taxonomic factors such as typicality, family size,
or diversity of premise species. Taxonomists, with their focus on knowledge
about genealogical relations among species, tend to apply taxonomic knowl-
edge even when reasoning about diseases among trees. However, landscapers
and maintenance workers, who have other concerns, were more likely to ap-
ply other kinds of knowledge to the same task. Converging evidence comes
from studies of categorization; Medin, Lynch, Coley, and Atran (1997) found
that these same groups of experts tended to sort tree species on the basis of
relations specific to their area of expertise. For example, taxonomists tended
to utilize botanical families, landscapers utilized categories like specimen tree
that reflected a species’ typical use in landscape design, and maintenance
workers tended to use categories like weed tree, which reflected the ease of
caring for different species. Thus, within a single domain, differential experi-
ence can render different knowledge available. These effects demonstrate how
extensive experience can elicit chronic changes in availability.
Culture is another source of differential experience that could impact the
availability of knowledge for induction. A striking example of culturally in-
duced changes in availability can be found in the work of Choi, Nisbett,
and Smith (1997). They investigated reasoning about biological and social
categories among Americans and Koreans. Some have argued that western
individualist cultures tend to process information analytically and generally
attend to categories and rules, whereas eastern collectivist cultures tend to
consider problems holistically, attending to categories in terms of interac-
tions in a setting (see Nisbett, Peng, Choi, & Norenzayan, 2001, for a review).
One interesting exception to this is in the domain of social categories and
roles, which are argued to be particularly salient and important to eastern
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 125

collectivist cultures. Based on these cultural differences, Choi et al. (1997)

predicted that manipulations of taxonomic category salience (specific versus
general conclusions) would have a differential effect on American and Ko-
rean peoples’ reasoning about biological and social categories. Specifically,
they predicted that because taxonomic categories in general are less cultur-
ally salient to members of a collectivist culture, manipulations of salience
should have a more pronounced effect on Koreans’ biological inferences than
on Americans’, for whom taxonomic knowledge is already highly salient. In
contrast, they predicted the opposite effect for social categories, where taxo-
nomic knowledge is argued to be more salient for Koreans than for Americans.
Consistent with these predictions, they found that when reasoning about bio-
logical categories, Koreans preferred taxonomic responses to arguments with
general conclusions than to those with specific conclusions, whereas this
manipulation had no effect on Americans’ responses. Conversely, when rea-
soning about social categories, Americans preferred taxonomic responses to
arguments with general conclusions than to those with specific conclusions,
whereas this manipulation had no effect on Koreans’ responses. These results
are consistent with the claim that culture can be viewed as a kind of experi-
ence that may result in chronic changes in the availability of different kinds
of knowledge in a domain.
In sum, experimental results consistently reveal differential use of taxo-
nomic versus other knowledge in experienced versus novice populations. We
propose that these differences can be understood as reflecting chronic changes
in the relative availability of different kinds of knowledge that accompany the
acquisition of expertise.

Experience-Related Changes in Sensitivity to Context. In this section, we

focus on experience-based differences in the availability of knowledge to
guide induction as a function of context. We argue that general changes in
baseline availability of different kinds of knowledge also lead experts to be
more sensitive to context. In other words, because more kinds of knowledge
become available with experience, experts can draw on the most appropriate
relation to guide a given inference. Below we present evidence that domain-
specific experience is associated with increased inductive selectivity and with
increased sensitivity to relations among premise categories.

experience-related changes in inductive selectivity. Recent evi-

dence suggests that experience leads to an increase in inductive selectivity. For
example, Shafto and Coley (2003) demonstrate experience-related changes
in inductive selectivity by contrasting how experts (commercial fishermen)
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

126 P. Shafto, J. D. Coley, and A. Vitkin

and novices (university undergraduates) reasoned about marine creatures. In

this experiment, participants were given either a novel blank property (“has a
property called sarca”) or a novel disease (“has a disease called sarca”) to rea-
son about. Participants were given examples of creatures that had the property
and then were asked to infer which other creatures (from a broad array of
fish, sharks, whales, and crustaceans) would have the property. Results indi-
cated marked differences in inductive selectivity between novices and experts.
When told about blank properties, experts tended to generalize to taxonom-
ically related creatures. However, when told about diseases, experts tended
to generalize to creatures related in the food web, specifically by making di-
rectional inferences from prey to predators. In contrast, novices tended to
generalize to taxonomically related creatures regardless of the property. This
result is notable because although novices were unlikely to have the detailed
knowledge about marine food web relations on which commercial fishermen
depend for their livelihood, they undoubtedly possessed some rudimentary
knowledge of marine food web relations (e.g., that sharks eat fish and not
the other way around). This constitutes enough knowledge to make rough
inferences based on food web information. Nevertheless, we observed no such
reasoning among novices.
We interpret these results as suggesting that for novices, knowledge about
food web relations is generally less available and that the context provided
in this experiment (essentially, disease) did not create enough of a change in
the availability of food web knowledge to overcome their taxonomic default.
On the other hand, for experts who rely heavily on knowledge of food web
relations among these creatures on a daily basis, taxonomic and food web
knowledge have relatively similar baseline availabilities, and the experimental
context (disease versus property) was enough to manipulate the availability of
the different kinds of knowledge.
Stepanova and Coley (2003) (see also Coley, Shafto, Stepanova, & Baraff,
2005; Stepanova, 2004) also contrasted reasoning by individuals with ex-
tensive or limited experience in a domain. However, rather than comparing
experts and novices reasoning about a single domain, they compared a sin-
gle population (U.S. college undergraduates) reasoning about a domain they
have extensive experience with (alcoholic drinks) versus a domain with much
less relevance to their daily lives (animals), on the assumption that for the
typical college student, alcohol possesses greater relevance and cultural im-
portance than animals, and undergraduates are likely to have more first-hand
experience of, more frequent exposure to, and richer and more abundant
folk theories about alcohol than animals. The task required participants to
choose which of two pairs of premise categories provided better evidence for
a generalization to any alcohol (or any animal). Sets of premise categories
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 127

were chosen so that one would clearly be stronger via diversity. In addition
to being randomly assigned to the animal or alcohol conditions, participants
were also randomly assigned to evaluate arguments about a chemical com-
ponent or about getting sick. Results showed clear evidence for inductive
selectivity when undergraduates were reasoning about alcohol. Specifically,
participants reasoning about alcohol showed differential use of taxonomic
knowledge as a function of property; these participants were more likely to
make diversity-based inferences about getting sick than about a chemical
component. Moreover, participants reasoning about alcohol also provided
different explanations for their choices as a function of property; they were
more likely to offer causal explanations for inferences about getting sick, but
more likely to offer taxonomic explanations for inferences about a chemical
component. In contrast, there was no evidence of inductive selectivity when
participants were reasoning about animals; neither the relative frequency of
diversity-based choices nor the type of explanations provided for those choices
varied for inferences about a chemical versus getting sick. These results are in
close accord with those of Shafto and Coley (2003) just described, and they
suggest that greater domain-specific experience may increase the potential
availability of multiple conceptual relations and therefore increase inductive
selectivity as a function of property being projected.

experience-related changes in sensitivity to premise relations.

Recent developmental work also suggests that domain-specific experience
may increase children’s sensitivity to relations among premise categories as
potential constraints on induction. In the Vitkin, Coley, and Hu (2005) study
described earlier under “Property effects on induction” relations between
premises as well as property effects were investigated using open-ended in-
ductive projections among elementary school children from urban communi-
ties and rural communities. Relations among animal pairs were manipulated
in a two-by-two design such that each pair was either taxonomically close
(from the same superordinate) or taxonomically far (from different superor-
dinates) and either ecologically related (via habitat or predator-prey relations)
or ecologically unrelated. Responses were coded as being based on similarity
(taxonomy, shared features, etc.) or interaction (contact through ecosystems).
If experience increases sensitivity to context, then the greater opportunity for
direct interaction with plants and animals in relatively intact ecosystems af-
forded by a rural environment may lead to increased sensitivity to relations
among premise categories among rural children.
Indeed, urban and rural children differed strikingly with respect to their
sensitivity to relations among premise categories. Rural children showed
consistent sensitivity to differential relations among premise categories.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

128 P. Shafto, J. D. Coley, and A. Vitkin

Specifically, rural children made more similarity-based projections for taxo-

nomically close pairs than for taxonomically far pairs. They also made more
interaction-based projections for ecologically related pairs than for unrelated
pairs, and for taxonomically far pairs than for taxonomically close pairs. None
of these effects were evident for urban children. These results suggest that ex-
perience may mediate the availability of different conceptual relations for
guiding children’s spontaneous inferences. For biologically experienced rural
children, relations among premise pairs was sufficient context to render tax-
onomic and ecological knowledge differentially available to inform inductive
projections. In contrast, for biologically inexperienced urban children, this
was not the case.
Taken together, these results suggest that the influence of experience on
inductive reasoning can be thought of in terms of changes in the relative avail-
ability of different kinds of knowledge. Experience can be seen as increasing
the relative availability of non-default knowledge for guiding induction, as in
the cases of fishermen reasoning about marine creatures and undergraduates
reasoning about alcohol. Because multiple kinds of knowledge are equally
available, expert reasoning is highly sensitive to context. In contrast, because
one kind of knowledge is a highly available default for novices, much stronger
context is required to elicit reasoning based on the less-available knowledge.
Thus, changes in chronic availability result in increased sensitivity to context
among experts relative to novices.

Summary: Availability in Category-Based Induction

Previous research has demonstrated two main factors that influence recruit-
ment of different kinds of knowledge in category-based induction: experi-
mental context and prior experience. We argue that these factors are best un-
derstood as manipulations of the availability of different kinds of knowledge.
We have reviewed previous work demonstrating that people’s inductive
generalizations are sensitive to the property to be projected. We find that
different properties can lead to qualitatively different patterns of generaliza-
tion, such as in the domain of foods where biochemical properties lead to
taxonomic reasoning and situational properties lead to script-based reason-
ing. We propose that properties that are consistent with a particular kind
of knowledge will increase the availability of that kind of knowledge. For
example, reasoning about what foods would be eaten together at an initiation
ceremony activates script knowledge about what foods typically co-occur,
therefore increasing the likelihood of script-based inferences by rendering
that knowledge temporarily available. Thus, the property of projection can
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 129

lead to an acute change in availability of knowledge and characteristically

different patterns of generalizations.
Similarly, we propose that relations between premise and conclusion cat-
egories are also able to elicit changes in availability. In cases such as non-
monotonicity via property reinforcement as well as open-ended induction, the
presence of known relations among the premises results in characteristically
different patterns of generalization. For example, if you learn a new fact that
is true of turtles and lizards, you might generalize it to other reptiles, whereas
if you learn a new fact that is true of turtles and ducks, you might gener-
alize it to pond creatures instead. We suggest that presenting information
about categories that are united by a particular kind of knowledge leads to an
acute increase in the availability of that kind of knowledge and an increased
likelihood of an inference based on that knowledge.
More compelling evidence derives from the role of experience. In our
framework, availability depends on context and prior knowledge accrued
through experience. We have reviewed research with experts suggesting that
experience can change the baseline availability of different kinds of knowl-
edge. People tend to use knowledge that they have a deep understanding of
and that has proven useful in the past. For example, evidence from tree experts
suggests that a person’s experiential background leads to chronic changes in
the kinds of knowledge that are available for reasoning; by default, tree tax-
onomists tend to think about trees in terms of scientific taxonomies, whereas
maintenance workers tend to think about trees in terms of maintenance con-
cerns. This evidence suggests that experience can lead to chronic changes in
the availability of different kinds of knowledge.
However, we think the most compelling evidence is manifest in interac-
tions between experience and context. People are more likely to demonstrate
robust inductive selectivity in domains where they have extensive experience.
For example, when reasoning about marine creatures, commercial fisher-
men demonstrated more inductive selectivity than university undergraduates;
university undergraduates demonstrated more inductive selectivity when rea-
soning about alcohol than when reasoning about animals. We propose that
experience facilitated inductive selectivity in these studies by causing an in-
crease in the availability of non-default knowledge, essentially “leveling the
playing field” and allowing for an increased sensitivity to context.
In sum, we see availability as a promising explanation for both context
and experience effects in category-based induction. However, at this point,
much of our case remains circumstantial. Though availability seems an apt
explanation of previous results, one might ask, what do we gain by thinking
about these effects in terms of availability?
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

130 P. Shafto, J. D. Coley, and A. Vitkin

  

Traditional models of inductive reasoning have focused on a single kind
of knowledge in a domain, eschewing the effects of context. For example,
the similarity-coverage model (Osherson et al., 1990) focused on taxonomic
knowledge in the domain of biology, using similarity between premises and
conclusions as well as a taxonomic coverage to predict inductive generaliza-
tions. These models account for phenomena such as taxonomic diversity, that
is, the fact that people rate an anatomical property true of diverse premises
such as robins and ostriches more likely to be shared by all birds than a prop-
erty shared by robins and doves and are generally in remarkably close accord
with novices generalizations of anatomical properties.2 Though able to cap-
ture reasoning based on taxonomic knowledge, the similarity-coverage model
does not account for reasoning based on other kinds of knowledge and thus
does not naturally extend to the kinds of reasoning we have been discussing
here (see Smith et al., 1993, for an account of some property effects). Some
context effects, such as inferences about behavioral and anatomical infer-
ences in Heit and Rubinstein (1993), can be handled by allowing similarity
to be defined context sensitively; however, no similarity-based models can
account for the kinds of asymmetric causal inferences in Shafto and Coley
(2003).
Sloman’s feature-based induction model (1993) also predicts inferences
about taxonomic properties, but it differs from the similarity-coverage model
in not assuming a taxonomic structure over objects. Under (the geometric
interpretation of) this model, prior knowledge is represented by an object-
feature matrix, and inferences are generated by considering the proportion
of shared properties between the premise and conclusion, relative to the total
number of properties for the premise. There are many potential ways to
extend the feature-based model to handle context effects, including adding
context-specific feature weights. However, all extensions require the addition
of abstract knowledge not present in the original model.
Rather than focusing on the fact that these models do not account for prop-
erty effects, we think it worthwhile to emphasize commonalities between the
similarity-coverage and feature-based models and our availability-based ap-
proach. The success of both of these models in predicting undergraduates’
judgments suggests that knowledge about taxonomic relations is central to
undergraduates’ conceptualization of biological kinds. We have suggested

2 The authors also propose that category members may be differentially available; for example,
suggesting that robins are more available members of the category bird than turtledoves.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 131

that taxonomic knowledge is chronically more available to undergraduates,

and therefore relied on as a default strategy, while taxonomic and ecological
knowledge are both chronically available to experts (such as commercial fish-
ermen). However, neither the similarity-based nor the feature-based model
includes a means to explain the effects of properties on reasoning, or a natural
way to explain how experience influences reasoning.
Some interesting parallels can be drawn between the availability framework
and some previous work that has addressed aspects of induction discussed
under our approach. Our notion of availability in inductive reasoning is
analogous to Barsalou’s (1982) distinction between context-independent and
context-dependent properties of concepts.3 Context-independent properties
(e.g., basketballs are round) “are activated by the word for a concept on all
occasions” (p. 82), whereas context-dependent properties (e.g., basketballs
float) “are only activated by relevant contexts in which the word appears”
(p. 82). Barsalou demonstrates that, for example, priming with relevant con-
text facilitates the verification of context-dependent properties but has no
effect on the verification of context-independent properties. This distinction
can be applied to our analysis of availability in category-based induction by
granting that relations among concepts, as well as properties of individual
concepts, may vary in their context dependence. Thus, what we have called
acute changes in availability correspond to priming a relevant context, whereas
chronic changes in availability corresponds to representational changes in the
context dependence of classes of relations among concepts (e.g., predator-
prey relations). It remains to be seen whether any class of relations is truly
context independent, but taxonomic relations may be one candidate (e.g.,
Coley, Shafto et al., 2005).
Our notion of availability as applied to category-based induction also fits
nicely into Medin et al.’s (2003) “relevance framework.” A central claim of the
relevance framework is that relations among premise categories, or between
premise and conclusion categories, guide or constrain an inductive inference
to the degree that such relations are deemed relevant. Relevance, in turn, is
conceptualized in terms of effort and effect. The less effort needed to access or
construct a relation, the more likely it is to be deemed relevant to a particular
inference. Conversely, premises that have greater cognitive effect (in the sense
of potentially leading to conceptual changes or allowing the derivation of novel
conclusions) are more likely to be deemed relevant. On this view, our notion of
availability can be seen as a more detailed way to think about the effort compo-
nent of the relevance framework. Specifically, the effort associated with a given

3 We are grateful to Brett Hayes for pointing out this connection.

P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

132 P. Shafto, J. D. Coley, and A. Vitkin

conceptual relation reflects the availability of that knowledge; all else being
equal, more available knowledge requires less effort to access and use. Thus,
both the acute and chronic changes in availability reviewed above reflect –
from the perspective of relevance – acute and chronic changes in the effort
required to access a given set of conceptual relations.
In the spirit of both Barsalou (1982) and Medin et al. (2003), availability
makes two distinctions: chronic changes in availability of different kinds of
knowledge grounded in experience, and acute changes in availability as a
result of context. Chronic changes in availability account for why taxonomic
knowledge is less effortful than ecological knowledge for biological novices
and taxonomists but not for fishermen. Acute changes in availability reflect
the fact that context manipulates effort to access different knowledge (cf.
Heit & Bott, 2000). The interaction between chronic and acute changes in
availability determines the degree to which people show inductive selectivity.
Though we believe availability provides a coherent framework uniting
expertise and property differences in induction, we think that the true merit
of thinking about category-based induction in terms of availability will be
in the guidance it provides in moving research forward. In the next section,
we describe some recent work inspired by availability, and derive additional
novel (but yet untested) predictions from this proposal.

Availability in Action
We see two major challenges in the development of an availability-based
framework. The first is to identify what kinds of predictions can be generated
by understanding the relationship between knowledge and reasoning in terms
of availability, and to begin to test those predictions. The second is to explain
how chronic changes in availability arise with experience. In this section,
we will outline some initial studies addressing the first challenge and some
preliminary ideas which address the second.
To be useful, any framework must generate new hypotheses as well as de-
scribe existing results. In a recent set of studies, we investigated availability as a
possible explanation for the lack of inductive selectivity in novice populations
(Shafto, Coley, & Baldwin, 2005). To implicate availability, it is important to
show that having knowledge is not sufficient for inductive selectivity. Previ-
ous research (Shafto & Coley, 2003) suggests that context (novel diseases and
novel properties) did not elicit the selective use of taxonomic and ecological
knowledge in biological novices. One reason novices may not have demon-
strated inductive selectivity was a baseline difference in the availability of tax-
onomic and ecological knowledge. In a series of experiments (Shafto, Coley, &
Baldwin, 2005), we provided support for this claim by investigating the effects
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 133

of context on novices’ use of taxonomic and ecological knowledge, focusing on

ecological relations that were familiar to novices. Pre-testing ensured that the
novices knew the taxonomic and ecological relations in question. However,
despite demonstrated knowledge of ecological relations, participants consis-
tently rated inductive generalizations between taxonomically related species
stronger than generalizations between ecologically related species. This was
true regardless of whether they were reasoning about a blank property, dis-
ease, or toxin. In other words, possessing requisite knowledge of ecological
relations was not sufficient for the selective use of that knowledge to guide
induction. As Shafto and Coley (2003) found, the property manipulation did
not render ecological knowledge sufficiently available to novices.
In two subsequent experiments, we provided evidence that taxonomic
knowledge was more available than ecological knowledge in this population.
First, we contrasted primed and unprimed similarity judgments for pairs of
taxonomically or ecologically related species (following Ross & Murphy, 1999,
exp. 4). If ecological knowledge was chronically less available, then priming
ecological categories should increase the availability of ecological knowledge,
resulting in increased similarity ratings for ecological pairs in the primed ver-
sus the unprimed condition. In contrast, if taxonomic knowledge was already
highly available, then taxonomic priming should elicit no change in simi-
larity ratings. As predicted, priming was found to increase similarity ratings
for ecologically related pairs but not taxonomically related pairs, consistent
with the suggestion that ecological knowledge is less available than taxonomic
knowledge.
A second experiment provided further evidence by contrasting inductive
judgments with and without time pressure. We predicted that time pressure
would decrease access to less available knowledge by curtailing memory search
but would not affect use of knowledge that was already highly available. In
line with this prediction, likelihood ratings for ecological inferences decreased
under time pressure relative to unspeeded judgments, whereas ratings for tax-
onomic inferences remained unchanged. These results show that for novices
in the domain of biology, taxonomic knowledge is chronically more available
than ecological knowledge and suggest that differences in availability may im-
pede inductive selectivity. More generally, this series of studies is one example
of how the notion of availability can be used to generate novel predictions
about the use of knowledge in category-based induction.
Apart from generating testable hypotheses about inductive reasoning, an-
other important challenge for the availability framework is to elucidate the
mechanisms by which experience elicits changes in availability of different
kinds of knowledge. One route through which chronic changes may be elicited
is the frequency with which knowledge is accessed. Naturally, we expect
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

134 P. Shafto, J. D. Coley, and A. Vitkin

information that gets used frequently in a particular context to be more

available (Horowitz et al., 1966). This assumption reflects the fact that our
past experience should provide a useful reference in the future, a basic princi-
ple of memory (Anderson & Milson, 1989). Extending this idea to availability
merely implements this assumption as a means to sort out what knowledge is
deemed appropriate for a particular inference given our experience and the
current context.
A second potential mechanism for eliciting chronic change in availabil-
ity is representational efficiency. Availability should increase with increased
representational efficiency. Here representational efficiency reflects a com-
promise between the accumulation of facts and a means to summarize the
facts efficiently. For example, one reason that taxonomic knowledge may be
highly available for biological reasoning is the fact that it provides a succinct
summary of a large amount of factual knowledge. Taxonomic knowledge en-
compasses genetic information, anatomical information, information about
shape, behavior, environment, and so forth. On the other hand, knowledge
valued by experts such as ecological knowledge does not have the same imme-
diate payoff that taxonomic knowledge does. Taxonomic structures provide
a simple way of encompassing all living kinds in a representational struc-
ture that is scale invariant; all subsections of the taxonomy are branching
trees. Perhaps an increase in availability of ecological knowledge for experts
represents an increase in the efficiency with which ecological relations cap-
ture relevant knowledge. However, these proposals are speculations that will
require extensive empirical research.
We suggest that the problem of reasoning in knowledge-rich domains is
crucial to understanding human intelligence. We have focused on one aspect
of this problem, how experience is brought into contact with context in
informing inductive reasoning. We have argued that in any given context,
different kinds of knowledge are more or less available, and that availability
predicts how likely the knowledge is to be used in reasoning. Though much
of our evidence at this stage is preliminary, the notion of availability unites
existing work on knowledge and context-specific reasoning and may provide
a useful framework from which to investigate how knowledge is deployed in
specific situations to guide category-based induction.

References
Anderson, J. R., & Milson, R. (1989). Human memory: An adaptive perspective. Psycho-
logical Review, 96, 703–719.
Baker, A., & Coley, J. D. (2005). Taxonomic and ecological relations in open-ended induc-
tion. Paper presented at the 27th Annual Conference of the Cognitive Science Society,
Stresa, Italy.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

Availability in Category-Based Induction 135

Barsalou, L. W. (1982). Context-independent and context-dependent information in

concepts. Memory & Cognition, 10, 82–93.
Choi, I., Nisbett, R. E., & Smith, E. E. (1997). Culture, category salience, and inductive
reasoning. Cognition, 65, 15–32.
Coley, J. D. (2005). Relational properties in folk biological induction. Paper presented at
the Biennial Meetings of the Society for Research in Child Development, Atlanta, GA.
Coley, J. D., Baker, A., & Kemp, C. (2004). Taxonomic and ecological relations in open-
ended induction. Paper presented at the 45th Annual Meeting of the Psychonomic
Society, Minneapolis, MN.
Coley, J. D., Shafto, P., Stepanova, O., & Baraff, E. (2005). Knowledge and category-based
induction. In W. Ahn, R. L. Goldstone, B. C. Love, A. B. Markman, & P. Wolff (Eds.),
Categorization inside and outside the laboratory: Essays in honor of Douglas L. Medin
(69–86). Washington, DC: American Psychological Association.
Coley, J. D., Vitkin, A. Z., Seaton, C. E., & Yopchick, J. E. (2005). Effects of experience on
relational inferences in children: The case of folk biology. Paper presented at the 27th
Annual Conference of the Cognitive Science Society, Stresa, Italy.
Heit, E., & Bott, L. (2000). Knowledge selection in category learning. Psychology of
Learning and Motivation, 39, 163–199.
Heit, E., & Feeney, A. (2005). Relations between premise similarity and inductive
strength. Psychonomic Bulletin & Review, 12, 340–344.
Heit, E., & Rubinstein, J. (1994). Similarity and property effects in inductive reasoning.
Journal of Experimental Psychology: Learning, Memory & Cognition, 20, 411–422.
Horowitz, L. M., Norman, S. A., & Day, R. S. (1966). Availability and associative sym-
metry. Psychological Review, 73, 1–15.
López, A., Atran, S., Coley, J. D., Medin, D., & Smith, E. E. (1997). The tree of life:
Universal and cultural features of folkbiological taxonomies and inductions. Cognitive
Psychology, 32, 251–295.
Medin, D., Coley, J. D., Storms, G., & Hayes, B. (2003). A relevance theory of induction.
Psychonomic Bulletin and Review, 10, 517–532.
Medin, D. L., Lynch, E. B., Coley, J. D., & Atran, S. (1997). Categorization and reasoning
among tree experts: Do all roads lead to Rome? Cognitive Psychology, 32, 49–96.
Nguyen, S. P., & Murphy, G. L. (2003). An apple is more than just a fruit: Cross-
classification in children’s concepts. Child Development, 6, 1783–1806.
Nisbett, R. E., Peng, K., Choi, I., & Norenzayan, A. (2001). Culture and systems of
thought: Holistic versus analytic cognition. Psychological Review, 108, 291–310.
Osherson, D. N., Smith, E. E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97(2), 185–200.
Pearlstone, Z., & Tulving, E. (1966). Availability versus accessibility of information in
memory for words. Journal of Verbal Learning and Verbal Behavior, 5, 381–391.
Proffitt, J. B., Coley, J. D., & Medin, D. L. (2000). Expertise and category-based induction.
Journal of Experimental Psychology: Learning, Memory & Cognition, 26, 811–828.
Ross, B. H., & Murphy, G. L. (1999). Food for thought: Cross-classification and category
organization in a complex real-world domain. Cognitive Psychology, 38, 495–553.
Shafto, P., & Coley, J. D. (2003). Development of categorization and reasoning in the
natural world: Novices to experts, naı̈ve similarity to ecological knowledge. Journal of
Experimental Psychology: Learning, Memory & Cognition, 29, 641–649.
Shafto, P., Coley, J. D., & Baldwin, D. (2005). Availability in inductive reasoning. Paper
presented at the 46th Annual Meeting of the Psychonomic Society, Toronto.
P1: KNQ
0521672443c05 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:51

136 P. Shafto, J. D. Coley, and A. Vitkin

Shafto, P., Kemp, C., Baraff, E., Coley, J. D., & Tenenbaum, J. B. (2005). Context-sensitive
induction. Paper presented at the 27th Annual Conference of the Cognitive Science
Society, Stresa, Italy.
Shafto, P., Kemp, C., Baraff, E., Coley, J. D., & Tenenbaum, J.B. (in prep). Inductive
reasoning with causal knowledge.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Smith, E. E., Shafir, E., & Osherson, D. N. (1993). Similarity, plausibility, and judgments
of probability. Cognition, 49, 67–96.
Stepanova, O., & Coley, J. D. (2003). Animals and alcohol: The role of experience in
inductive reasoning among college students. Paper presented at the 44th Annual Meeting
of the Psychonomic Society, Vancouver, British Columbia.
Stepanova, O. (2004). Vodka and vermin: Naı̈ve reasoning about animals and alcohol.
Unpublished doctoral dissertation, Northeastern University.
Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and
probability. Cognitive Psychology, 5, 207–232.
Vitkin, A. Z., Coley, J. D., & Hu, R. (2005). Children’s use of relevance in open-ended
induction in the domain of biology. Paper presented at the 27th Annual Conference of
the Cognitive Science Society, Stresa, Italy.
Vitkin, A., Coley, J. D., & Kane, R. (2005). Salience of taxonomic and ecological relations
in children’s biological categorization. Paper presented at the Biennial Meetings of the
Society for Research in Child Development, Atlanta, GA.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance

Sergey Blok, Daniel Osherson, and Douglas L. Medin

In reality, all arguments from experience are founded on the similarity which we
discover among natural objects, and by which we are induced to expect effects similar
to those which we have found to follow from such objects. . . . From causes which
appear similar we expect similar effects.
David Hume, An Enquiry Concerning Human Understanding (1772)

.   

1.1 The Property Problem

Ever since Goodman’s (1972) warnings about the “false friend” similarity,
psychologists have been cautious about explaining inductive inference in
terms of resemblance. Innumerable shared properties unite any pair of objects,
inviting their dubious certification as similar along with the inference that
some further property is also shared. Genuine resemblance appears to be a
three-place predicate, relating two objects only in the context of a set S of
properties (often implicit) that determine the respects in which objects may
be compared.1
When it is plausible that the same set S is brought to mind in judging
similarity and inductive strength, it can be revealing to explain the latter in
terms of the former. Thus, famous Linda resembles the average feminist bank
teller more than the average bank teller; and whatever properties support this
judgment are likely also in place when people judge her chance of being one

1 See also Goodman (1955), Quine (1960). In the psychology literature, the matter has been
rehearsed by Tversky & Gati (1978), Osherson, Smith, & Shafir (1986), Medin, Goldstone, &
Gentner (1993), and Keil, Smith, Simons, & Levin (1998), among others.

137
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

138 S. Blok, D. Osherson, and D. L. Medin

or the other.2 Indeed, the probabilities people assign in Linda-like problems

are accurately predicted from the relevant similarities rated independently
(Shafir, Smith, & Osherson, 1990).
A radical means of inclining S to remain constant between similarity and
inference is to limit attention to arguments involving “blank” predicates like
requires biotin for hemoglobin synthesis. The blankness consists in the difficulty
of evaluating the relative credibility of the predicate’s application to different
objects.3 The blank-predicate argument
Bears require biotin for hemoglobin synthesis.
-----------------------------------------------------
Wolves require biotin for hemoglobin synthesis.

evokes a vaguely physiological context but not much more. Thus, one may
have the intuition that cameras and computers do not require biotin but no
a priori sense of whether bears and wolves do. Nonetheless, the assertion that
bears require biotin, coupled with the similarity of wolves to bears, accords
strength to the foregoing argument. Theories that incorporate similarity as a
basis for induction – such as the Similarity-Coverage model – have been
able to account for a range of reasoning phenomena.4
Similarity-Coverage is unsuited to non-blank predicates like that figur-
ing in the following arguments;
Fieldmice often carry the parasite Floxum.
(1) (a)
Housecats often carry the parasite Floxum.

Tigers often carry the parasite Floxum.

(b)
Housecats often carry the parasite Floxum.

2 Linda was born in Tversky & Kahneman (1983). For recent debate about the fallacy she illustrates,
see Kahneman & Tversky (1996), Gigerenzer (1996), and Tentori, Bonini, & Osherson (2004),
along with references cited there.
3 A more cautious formulation would relativize blankness to a particular judge and also to the
objects in play. Degrees of blankness would also be acknowledged, in place of the absolute
concept here invoked. In what follows, we’ll rely on rough formulations (like the present one)
to convey principal ideas.
4 See Osherson, Smith, Wilkie, López, & E. Shafir (1990). An alternative account is offered in
Sloman (1993). Blank predicates first appear in Rips (1975). For examination of Similarity-
Coverage type phenomena in the context of natural category hierarchies, see Coley, Atran, &
Medin (1997) and Medin, Coley, Storms, & Hayes (2003). Developmental perspectives on
Similarity-Coverage are available in López, Gelman, Gutheil, & Smith (1992), Heit & Hahn
(1999), and Lo, Sides, Rozelle, & Osherson (2002). Similarity among biological categories may
be more resistant to contextual shifts than is the case for other categories (Barsalou & Ross,
1986). This would explain their popularity in many discussions of inductive inference.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 139

Many people find (1)a to be stronger than (1)b, no doubt because the relational
property characteristically ingests readily comes to mind (probably evoked by
mention of parasites). In contrast, ingestion is unlikely to be salient when
judging the similarity of fieldmice to housecats, preventing similarity from
predicting strength. The root of the problem is the difference in causal theories
that may occur to the reasoner when diverse sets of properties are evoked in
the two settings.5
For another limitation of Similarity-Coverage, consider the conclusion
that Labradors can bite through wire (of a given thickness and composition)
given the alternative premises (a) Collies can, versus (b) Chihuahuas can. More
people think that (b) provides better evidence for the conclusion than does (a),
despite the fact that Labradors are more similar to Collies than to Chihuahuas
(Smith, Shafir, & Osherson, 1993). This intuition likely derives from the belief
that Chihuahuas are less powerful than Collies, hence (b) is less likely to be
true. The example therefore suggests that a model of induction based solely
on similarity will fail when the probability of predicate-application is not
uniform across the objects in play.
The aim of the chapter is to show how the prior probabilities of state-
ments might be incorporated into a similarity-based model of induction.6
We attempt to achieve this by elaborating a new version of the “Gap Model”
advanced in Smith et al. (1993). We also suggest how probabilities and similar-
ities can be exploited to construct joint distributions over sets of propositional
variables. The first step in our endeavor is to characterize a class of predicates
that are neither blank nor epistemically too rich.

1.2 Towards Non-Blank but Manageable Predicates

In the space between blank and Floxum-like predicates are those that are
meaningful yet mentally evoke the same set S of properties when people
judge similarity compared to argument strength. Let us use Qo to abbreviate
the attribution of a predicate Q to an object or category o, and ¬Qo for
the denial of this attribution. An argument composed of statements like
these will be termed stable just in case the same set of properties come to

5 The Floxum example is based on López, Atran, Coley, Medin, & Smith (1997). It is exploited in
Lo et al. (2002) to argue that the “diversity principle” does not have the normative status often
assumed by psychologists. For the role of causal theories in commonsense reasoning, see Ahn,
Kalish, Medin, & Gelman (1995) and Lassaline (1996), along with the theoretical discussions in
Ortiz (1993) and Turner (1993).
6 Reservations about the role of similarity in theories of reasoning are advanced in Sloman & Rips
(1998). Overviews of theories of inductive reasoning are available in Heit (2000), Sloman &
Lagnado (2005).
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

140 S. Blok, D. Osherson, and D. L. Medin

mind when judging the strength of the argument compared to judging the
similarity of the objects figuring in it. To test the stability of an argument,
the mental potentiation of various predicates would need to be measured
during similarity judgment and inference. We do not offer a recipe for such
measurement but observe that stability is not circularly defined in terms that
guarantee the success of any particular model of inference (like ours, below).
Call a predicate “stable” if it gives rise to stable arguments.
For a further distinction, suppose that a is Dell Computer Corporation,
and c is HP/Compaq. If Q is increases sales next year, then Qa will strike many
reasoners as confirmatory of Qc , whereas if Q is increases market share next
year, thenQa will seem disconfirmatory ofQc . The similarity of a and c can be
expected to have different effects in the two cases. To mark the difference, we
qualify a predicate Q as monotonically increasing [respectively decreasing] for
objects O = {o 1 · · · o n } just in case Prob (Qx : Qy) ≥ Prob (Qx) [respectively,
Prob (Qx : Qy) ≤ Prob (Qx)] for all x, y ∈ O; and call Q monotonic for O if
Q is either monotonically increasing or decreasing for O.
Let us illustrate these ideas. The authors expect that for many people the
following predicates are monotonically increasing, and yield stable arguments
over the class of mammal species7 :

has trichromatic vision

can suffer muscle damage through contact with poliomyelitis
(2) brain/body mass ratio is 2 percent or more
requires at least 5 hours of sleep per day for normal functioning
sex drive varies seasonally
We are now in a position to define a circumscribed yet vast class of argu-
ments.

(3) Definition: An argument A is elementary just in case:

(a) A has the form ±Qa / ± Qc or ±Qa, ±Qb / ± Qc , where a, b, c
occupy the same hierarchical level;
(b) A is stable; and
(c) Q is monotonic for the objects appearing in A.

The notation ±Qo represents either Qo or ¬Qo.

How often do elementary arguments embody a person’s reasoning? The
authors refrain from exaggerated appraisal. We are uncertain, for example, of
the stability of the predicate “can bite through wire,” discussed earlier. Let it

7 For another example, susceptibility to disease is usually perceived to be shared for species of the
same genus (Coley, Atran, & Medin, 1997).
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 141

merely be noted that few explicit, testable theories of inductive strength for
non-trivial classes of arguments are presently on offer. The stable predicates
in (2) suggest the richness of the class of elementary arguments. Success in
predicting their strength is thus a challenging endeavor and may point the
way to more general models.

1.3 Theoretical Goals

The next section advances a theory of the strength of elementary arguments.
Given such an argument, let Prob (±Qc : ± Qa) or Prob (±Qc : ± Qa, ±Qb)
be the conditional probability assigned by a given judge to the argument’s
conclusion given its premises. We attempt to predict these numbers from
(a) the absolute probabilities the judge assigns to premises and conclusion
individually, and (b) the similarities among the objects appearing in the
argument. Letting sim(x, y) represent the judge’s perceived similarity of x
to y, our theory takes the form of a function with the following inputs and
outputs:
type of
elementary
argument inputs output

Prob (±Qc )
(4) ±Qa/ ± Qc Prob (±Qa) Prob (±Qc : ± Qa)
sim(a, c )

Prob (±Qc ) Prob (±Qa)

Prob (±Qc : ± Qa,
±Qa, ±Qb/ ± Qc Prob (±Qb) sim(a, c )
±Qb)
sim(b, c ) sim(a, b)
Similarity judgments are assumed to lie in the unit interval (identity cor-
responding to 1), and to be symmetric.8 Beyond this, our model makes no
assumptions about how similarity assessments are generated; in particular,
complex processes of “feature alignment” may be involved (as in Gentner &
Markman, 1997; Goldstone, 1994). Nor do we require that reasoners know
whether an argument is elementary, specifically, whether its predicate is stable.
Our theory applies when stability holds; other cases lie outside the boundary

8 Claims for asymmetry appear in Tversky (1977) but seem to arise only in unusual circumstances;
see Aguilar & Medin (1999). The assumption sim(o i , o j ) ∈ [0, 1] is substantive; in particular,
dissimilarity might be unbounded. This possibility is ignored in what follows. For a review of
ideas concerning asymmetries in inductive reasoning, see Chapter 3 by Medin & Waxman in
the present volume.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

142 S. Blok, D. Osherson, and D. L. Medin

conditions. The reliance on absolute probabilities as inputs to the theory

[Prob (±Qc ), etc.] underlines the fundamental character of such judgments.
For analysis of their provenance, see Juslin and Persson (2002), which adapts
the classification model of Medin and Schaffer (1978).
The “Gap” model described by Smith, Shafir, and Osherson (1993) was
designed for stable predicates but required strong hypotheses about the fea-
tural decomposition of objects and predicates. The present theory is free of
such assumptions and captures the insights of the Gap model in a different
way. Given the family history of these ideas, we refer to the present model as
“Gap2.”
After presentation of the theory, we describe an initial experimental test and
its results. A second approach to exploiting similarity for predicting inductive
judgments is then advanced.

.     

Our theory takes the form of equations that determine the outputs of (4) from
the inputs. For simplicity in what follows, we consider only predicates that are
monotonically increasing. The decreasing case is presumed to be parallel. As
a preliminary, we list some qualitative conditions that any candidate theory
seems required to satisfy. The conditions express aspects of common sense
for a typical reasoner.

2.1 Qualitative Requirements: The One-Premise Case

Let an elementary argument Qa / Qc be given, and consider Prob (Qc : Qa).
We expect the latter quantity to approach 1 as sim(a, c ) does. For if
sim(a, c ) ≈ 1 then Prob (Qc : Qa) ≈ Prob (Qc : Qc ) = 1. To illustrate, the
conditional probability that pigs have trichromatic vision given that the hogs
do is close to 1 given the similarity of these creatures.
On the other hand, Prob (Qc : Qa) should go to Prob (Qc ) as sim(a, c )
goes to 0, This is because sim(a, c ) ≈ 0 signals the unrelatedness of a and c ,
rendering Qa irrelevant to the estimation of Qc .
Further conditions arise from purely probabilistic considerations. Thus,
Prob (Qa) ≈ 1 should imply Prob (Qc : Qa) ≈ Prob (Qc ). (Consider the
probability that newborn rats typically weigh at least one ounce assuming
that the same is true for newborn elephants.) Conversely, and other things
equal, as Prob (Qa) decreases, Prob (Qc : Qa) should increase. Our equations
should also respect the familiar fact that as Prob (Qc ) goes to 1, so does
Prob (Qc : Qa), and similarly for zero.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 143

Related requirements apply to arguments of the forms (i) ¬Qa / Qc ,

(ii) Qa / ¬Qc , and (iii) ¬Qa / ¬Qc . For example, Prob (Qc : ¬Qa) and
Prob (¬Qc : Qa) should both approach 0 as sim(a, c ) approaches 1, whereas
Prob (¬Qc : ¬Qa) should approach 1.

2.2 Qualitative Requirements: The Two-Premise Case

Now let a two-premise elementary argument Qa, Qb / Qc be given. Common
sense suggests a large number of constraints on Prob (Qc : Qa, Qb). To begin,
as either sim(a, c ) or sim(b, c ) approach 1, Prob (Qc : Qa, Qb) should also
approach 1. To illustrate, the conditional probability that pigs have trichro-
matic vision given that the hogs and squirrels do is close to 1 in view of the
similarity of pigs and hogs.
If both sim(a, c ) and sim(b, c ) go to 0, then Prob (Qc : Qa, Qb) should go to
Prob (Qc ) (since zero similarity signals irrelevance of the conditioning events).
On the other hand, if just sim(a, c ) approaches 0, then Prob (Qc : Qa, Qb)
should approach Prob (Qc : Qb); likewise, if just sim(b, c ) approaches 0, then
Prob (Qc : Qa, Qb) should approach Prob (Qc : Qa). For example, the prob-
ability that wolves are fond of garlic given that the same is true of bears
and bees is close to the probability that wolves are fond of garlic given than
bears are.
Next, as sim(a, b) goes to 1, Prob (Qc : Qa, Qb) should go to Prob (Qc : Qa)
[equivalently, Prob (Qc : Qa, Qb) should go to Prob (Qc : Qb)]. For
sim(a, b) ≈ 1 indicates that Qa, Qb record nearly identical facts. Thus, the
probability that otters can hear ultrasounds given that porpoises and dol-
phins can should be close to the probability that otters can hear ultrasounds
given that porpoises can. On the other hand, as sim(a, b) approaches 0, the
strength of Qa, Qb / Qc should increase (the “diversity effect”). We there-
fore expect Prob (Qc : Qa, Qb) > Prob (Qc : Qa), Prob (Qc : Qb) (except at
the extremes). To illustrate, the probability that geese have a magnetic sense
given that sparrows and eagles do exceeds the probability that geese have a
magnetic sense given that sparrows do, without reference to eagles.
Additional conditions on Prob (Qc : Qa, Qb) involve only probability. They
include the following:

(a) As Prob (Qa) approaches 1, Prob (Qc : Qa, Qb) approaches Prob (Qc :
Qb). Likewise, as Prob (Qb) approaches 1, Prob (Qc : Qa, Qb) ap-
proaches Prob (Qc : Qa).
(b) As Prob (Qa) and Prob (Qb) both approach 1, Prob (Qc : Qa, Qb) ap-
proaches Prob (Qc ).
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

144 S. Blok, D. Osherson, and D. L. Medin

(c) Other things equal, as Prob (Qa) and Prob (Qb) both decrease,
Prob (Qc : Qa, Qb) increases.
(d) As Prob (Qc ) approaches 1, so does Prob (Qc : Qa, Qb); as Prob (Qc )
approaches zero, so does Prob (Qc : Qa, Qb).

Similar constraints apply to the seven types of two-premise arguments with

negated premises or conclusion; their formulation is left to the reader.

2.3 Formulas for One-Premise Arguments

Consider again the elementary argument Qa / Qc . We propose that
Prob (Qc : Qa) is governed by
(5) Prob (Qc : Qa) = Prob (Qc )α , where

1−Prob(Qa)
1 − sim(a, c )
α=
1 + sim(a, c )

The reader can verify that (5) satisfies the qualitative conditions reviewed
sim(a,c )
above for Qa / Qc . To illustrate, as sim(a, c ) goes to 1, 1−
1+sim(a,c )
goes to 0,
α
hence α goes to 0, so Prob (Qc ) goes to 1 [hence, Prob (Qc : Qa) goes to 1].
For another example, as Prob (Qa) goes to 1, α goes to 1, so Prob (Qc )α goes
to Prob (Qc ), hence Prob (Qc : Qa) goes to Prob (Qc ) as desired. Of course,
(5) is not unique with these properties, but it is the simplest formula that
occurs to us [and restricts Prob (Qc : Qa) to the unit interval]. In the next sec-
tion, (5) will be seen to provide a reasonable approximation to Prob (Qc : Qa)
in an experimental context. It can be seen that our proposal is meant for
use with monotonically increasing predicates inasmuch as it guarantees that
Prob (Qc : Qa) ≥ Prob (Qc ).
The following formulas are assumed to govern one-premise elementary
arguments with negations. They satisfy commonsense requirements corre-
sponding to those discussed above.
(6) Prob (Qc : ¬Qa) = 1.0 − (1.0 − Prob (Qc ))α , where


1 − sim(a, c ) Prob(Qa)
α=
1 + sim(a, c )
(7) Prob (¬Qc : Qa) = 1.0 − Prob (Qc )α , where

1−Prob(Qa)
1 − sim(a, c )
α=
1 + sim(a, c )
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 145

(8) Prob (¬Qc : ¬Qa) = (1.0 − Prob (Qc ))α , where


1 − sim(a, c ) Prob(Qa)
α=
1 + sim(a, c )
The formulas also guarantee that Prob (Qc : Qa) + Prob (¬Qc : Qa) = 1 and
Prob (Qc : ¬Qa) + Prob (¬Qc : ¬Qa) = 1.
We do not assume that people perform the calculations corresponding to
Equations (5)–(8). Our claim is merely that the formulas approximate what-
ever underlying process gives rise to judgments about conditional probability.

2.4 Formulas for Two-Premise Arguments

Two-premise arguments likewise require distinguishing multiple cases, all
variations on the same theme. We consider only arguments with positive
conclusions; negated conclusions are handled similarly (and none of our
experiments involve them). We rely on the following concept:
(9) Definition: The confirmation exhibited by an argument of form Qa / Qc
is
Prob (Qc : Qa) − Prob (Qc )
.
1 − Prob (Qc )
The confirmation exhibited by an argument of form ¬Qa / Qc is
Prob (Qc ) − Prob (Qc : ¬Qa)
.
Prob (Qc )
Given an argument ±Qa, ±Qb / Qc , the dominant premise is ±Qa if the
confirmation exhibited by ±Qa / Qc exceeds the confirmation exhibited
by ±Qb / Qc ; otherwise, ±Qb is the dominant premise.

Confirmation is the impact of an argument’s premise on the credibility of its

conclusion. The measure suggested above normalizes by the potential impact
allowed by the prior credibility of the conclusion (namely, the distance to
unity for positive premises and the distance to zero for negative). A variety of
confirmation measures are analyzed in Eells and Fitelson (2002). In Tentori,
Crupi, Bonini, and Osherson (2007), they are compared for their ability
to predict shifts of opinion in an experimental setting involving urns. The
dominant premise in a two-premise argument is the one that yields the one-
premise argument of greatest confirmation. The one-premise probabilities
are derived from the theory of one-premise arguments offered above. We
now present our theory of two-premise arguments.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

146 S. Blok, D. Osherson, and D. L. Medin

(10) Case 1. Arguments of form Qa, Qb / Qc with Qa dominant.

Prob (Qc : Qa, Qb) = Prob (Qc : Qa) +

[(1 − sim(a, b)) × (1 − Prob (Qc : Qa)) × (Prob (Qc : Qb) − Prob (Qc ))].

Arguments of form Qa, Qb / Qc with Qb dominant are treated similarly.

In words, (10) claims that Prob (Qc : Qa, Qb) is given by the probability of
the dominant argument increased by a fraction of the probability 1 − Prob
(Qc : Qa) that the dominant argument “leaves behind.” The size of this frac-
tion depends on two factors, namely, the similarity between a and b (to avoid
redundancy), and the impact of the nondominant premise on the credibility
of Qc .
The constraints outlined earlier are satisfied by (10). For example, the
formula implies that Prob (Qc : Qa, Qb) ≈ 1 if sim(a, c ) ≈ 1. Note that
our proposal ensures that strength increases with extra premises, that is,
Prob (Qc : Qa, Qb) ≥ Prob (Qc : Qa), Prob (Qc : Qb). This feature is plausi-
ble given the restriction to monotonically increasing predicates.
We now list the other two-premise cases. They are predictable from (10)
by switching the direction of similarity and “the probability left behind” as
appropriate.
(11) Case 2. Arguments of form ¬Qa, Qb / Qc with ¬Qa dominant.

Prob (Qc : ¬Qa, Qb) = Prob (Qc : ¬Qa) +

[sim(a, b) × (Prob (Qc : ¬Qa)) × (Prob (Qc : Qb) − Prob (Qc ))].

Arguments of formQa, ¬Qb /Qc with ¬Qb dominant are treated similarly.

(12) Case 3. Arguments of form ¬Qa, Qb / Qc with Qb dominant.

Prob (Qc : ¬Qa, Qb) = Prob (Qc : Qb) +

[sim(a, b) × (1 − Prob (Qc : Qb)) × (Prob (Qc : ¬Qa) − Prob (Qc ))].

Arguments of form Qa, ¬Qb / Qc with Qa dominant are treated similarly.

(13) Case 4. Arguments of form ¬Qa, ¬Qb / Qc with ¬Qa dominant.

Prob (Qc : ¬Qa, ¬Qb) = Prob (Qc : ¬Qa) +

[(1 − sim(a, b)) × (Prob (Qc : ¬Qa)) × (Prob (Qc : ¬Qb) − Prob (Qc ))].

Arguments of form ¬Qa, ¬Qb / Qc with ¬Qb dominant are treated

similarly.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 147

In summary, Gap2 – our theory of elementary arguments – consists of

formulas (5)–(8), and (10)–(13). For a blank predicate Q, we set Prob (Qo) to
a fixed constant for any object o (.5 will often be a reasonable choice). Then
Gap2 still makes intuitively sound predictions, allowing it to be applied to
elementary arguments with blank predicates as a limiting case.

.     

3.1 Stimuli and Procedure

To test Gap2 quantitatively we chose a domain about which undergraduates
were likely to have opinions and interest, namely, post-graduation salaries
from different colleges and universities. The following institutions served as
objects:
(a) Connecticut State University (b) Oklahoma State University
(c) Arkansas State University (d) Yale University
(14)
(e) Harvard University (f) Harvard Divinity School
(g) Texas Technical Institute (h) Texas Bible College

The following predicate was employed:

graduates [of a given institution] earned an average salary of more than $50,000
a year in their first job after graduation.

Its negation was taken to be the following:

graduates [of a given institution] earned an average salary of less than $50,000 a
year in their first job after graduation.

Considering just objects (14)a – e, there are twenty arguments of form

Qa / Qc , and sixty of form Qa, Qb / Qc . Nine undergraduates were recruited
to assess the inputs and outputs in (4) for all twenty arguments of form
Qa / Qc , and half of the sixty arguments of form Qa, Qb / Qc . Another nine
did the same for all twenty arguments of form Qa / Qc , and the other half of
the sixty arguments of form Qa, Qb / Qc .9

9 The predicate was slightly (but inessentially) different for these eighteen participants. Note that
we distinguish the order of two conditioning events; otherwise, there would be only thirty prob-
abilities of form Prob (Qc : Qa, Qb) based on five objects. No student received two arguments
differing just on premise order. The order in which information is presented is an important
variable in many reasoning contexts (Johnson-Laird, 1983) but there was little impact in the
present study.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

148 S. Blok, D. Osherson, and D. L. Medin

Considering just objects (14)e–h, there are twelve arguments of form

Qa / Qc and twelve of form ¬Qa / Qc . Forty-one additional students eval-
uated the inputs and outputs for all twenty-four arguments of these two
forms. The same objects give rise to ninety-six two-premise arguments of
form Qa, Qb / Qc or ¬Qa, Qb / Qc or Qa, ¬Qb / Qc or ¬Qa, ¬Qb / Qc . A
third group of forty-seven students evaluated inputs and outputs for different
halves of these ninety-six arguments (each argument was evaluated by either
twenty-three or twenty-four people).
All participants were undergraduates at Northwestern University. Data
were collected using a computerized questionnaire. Similarity judgments were
elicited first, followed by absolute probabilities, then conditional probabilities.
Within these categories, stimuli were individually randomized.

3.2 Results
Overall, the inputs and outputs for 200 different arguments were evaluated,
each by nine, eighteen, twenty-three, twenty-four, or forty-one undergradu-
ates. The analysis that follows is based on the mean estimates for each input
and output.
We computed the Pearson correlation between the values obtained for
Prob (Qc : ± Qa) or Prob (Qc : ± Qa, ±Qb) with the values predicted by
Gap2. The result is r = 0.94. The regression line has slope = 0.874, and in-
tercept = 0.094. See Figure 6.1. To gauge the role of the prior probability
Prob (Qc ) in Gap2’s performance, we computed the correlation between the
observed probabilities Prob (Qc : ± Qa) and Prob (Qc : ± Qa, ±Qb) versus
the predictions of Gap2 with Prob (Qc ) subtracted out. The correlation re-
mains substantial at r = 0.88 (slope = 0.811, intercept = 0.079).
Twenty arguments of form Qa / Qc were evaluated using objects (14)a–e,
and another twelve using (14)e–h. In thirty-one of these thirty-two cases, the
average probabilities conform to the monotonicity principle Prob (Qc : Qa) ≥
Prob (Qc ) thereby agreeing with Gap2. The situation is different in the two-
premise case. An argument of form Qa, Qb / Qc exhibits monotonicity if the
average responses yield the following:
(15) Prob (Qc : Qa, Qb) ≥ max{Prob (Qc : Qa), Prob (Qc : Qb)}.

In our experiment, objects (14)a–e figured in sixty arguments of form Qa,

Qb / Qc , with the same participants also evaluating Qa / Qc and Qb / Qc . In
these sixty cases, (15) was violated fifty-one times ( p < .001 via a binomial
test).
These results suggest modifying our theory by averaging the impact of in-
dividual premises in a two-premise argument. Let us be cautious, however,
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 149

1.00

0.80

0.60
Observed

0.40

0.20

0.00
0.00 0.20 0.40 0.60 0.80 1.00

Predicted
figure 6.1. Predicted versus observed probabilities for the 200 arguments.

about backing off Gap2’s monotonicity in this way. Averaging depicts the
reasoner as strangely insensitive to the accumulation of evidence, which be-
comes an implausible hypothesis for larger premise sets. Also, the prevalence
of nonmonotonic responding may be sensitive to procedural details. A pre-
liminary study that we have conducted yields scant violation of monotonicity
when premises are presented sequentially rather than two at a time.

3.3 Second Test of the Model

To assess the generality of the foregoing results, we constructed sixteen one-
premise and sixteen two-premise arguments using the following objects and
predicate.

bears wolves cows

Objects:
sheep cougars lions
Predicate: “have at least 18% of their cortex in the frontal lobe”

The sixteen one-premise arguments were equally divided among the forms
Qa / Qc , Qa / ¬Qc , ¬Qa / Qc , and ¬Qa / ¬Qc . The sixteen two-premise ar-
guments were equally divided amongQa, Qb /Qc andQa, ¬Qb /Qc . Objects
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

150 S. Blok, D. Osherson, and D. L. Medin

0.80

0.70

0.60

0.50
Observed

0.40

0.30

0.20

0.10
0.00 0.20 0.40 0.60 0.80 1.00
Predicted
figure 6.2. Predicted versus observed probabilities for thirty-two arguments.

(i.e., the six mammals) were assigned to the thirty-two arguments so as bal-
ance their frequency of occurrence. Twenty college students at Northwestern
University rated the similarity of each of the fifteen pairs of objects drawn
from the list above. A separate group of twenty students estimated the condi-
tional probabilities of the thirty-two arguments plus the probabilities of the
six statements Qa. Data were collected via computer interface.
Since the similarity and probability estimates were collected from separate
groups of subjects there could be no contamination of one kind of judgment
by the other. The data from each group were averaged. The fifteen similarity
averages plus six averages for unconditional probabilities were then used
to predict the thirty-two average conditional probabilities via Gap2. The
correlation between predicted and observed values was .89. The results are
plotted in Figure 6.2.
More thorough test of Gap2 (and comparison to rivals) requires experi-
mentation with a range of predicates and objects. The arguments described in
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 151

here are homogeneous in the sense that the same predicate figures in each. We
expect Gap2 to apply equally well to heterogeneous collections of elementary
arguments (involving a multiplicity of predicates).

.   
It is easy to envision extensions of Gap2 beyond elementary arguments [as
defined in (3)]. For example, the conditional probability associated with a
general-conclusion argument like

Rats have retractable claws.

Squirrels have retractable claws.

All rodents have retractable claws.

might be computed as the minimum of the conditional probabilities for

Rats have retractable claws.

Squirrels have retractable claws.

X have retractable claws.

according to Gap2, where X ranges over the rodent species that come to the
reasoner’s mind.10 Another extension is to arguments in which neither the
object nor predicate match between premise and conclusion. The goal would
be to explain the strength of arguments like the following:

Howler Monkeys will eat cheddar cheese.

Spider Monkeys will eat Swiss cheese.

Both these extensions still require that the argument’s predicate(s) be stable
and monotonic for the single-premise case. Escaping this limitation requires
understanding how the same predicate can evoke different associations when
assessing strength compared to similarity. Specific causal information may
often be responsible for such a state of affairs. We acknowledge not (presently)

10 A “coverage” variable could also be added to the theory, in the style of the Similarity-Coverage
Model. Coverage for specific arguments is motivated by asymmetry phenomena such as the
greater strength of ‘Qcats / Qbats’ compared to ‘Qbats / Qcats’ (see Carey, 1985; Osherson et al.,
1990). Gap2 does not predict this phenomenon for any predicate Q such that Prob(Qcats) =
Prob (Qbats) (e.g., when Q is blank). Asymmetry is known to be a weak effect, however (see
Hampton & Cannon, 2004). If it is robust enough to merit modeling, this could also be achieved
by positing asymmetry in sim.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

152 S. Blok, D. Osherson, and D. L. Medin

knowing how to incorporate causal judgments into an explicit model of

induction.

.    

 
Theories like Gap2 might help to automate the collection of subjective proba-
bilities for artificial expert systems, especially since it is conditional probabil-
ities that are typically needed (Pearl, 1988; Russell & Norvig, 2003). A more
general approach to generating large sets of probability estimates for use in
A.I. would construct an entire joint distribution from simple inputs involv-
ing similarities and probabilities. The remainder of the chapter advances one
method for this purpose.
We limit attention to the following kind of context. Let n propositions
Q(o 1 ) · · ·Q(o n ) be given, where the o i are n objects at the same hierarchical
level, and Q is a stable predicate (in the sense of Section 1.2, above). In the
present setting the requirement that predicates be monotonic is no longer
needed. Thus, for n = 6, the Q(o i ) might correspond to the following:

Hawks have muscle-to-fat ratio at least 10-to-1.

Eagles have muscle-to-fat ratio at least 10-to-1.
Parakeets have muscle-to-fat ratio at least 10-to-1.
(16)
Cardinals have muscle-to-fat ratio at least 10-to-1.
Geese have muscle-to-fat ratio at least 10-to-1.
Ducks have muscle-to-fat ratio at least 10-to-1.

Closing the six statements under boolean operators yields logically complex
statements like the following:

(17) (a) Eagles and Cardinals have muscle-to-fat ratio at least 10-to-1.11
(b) Ducks but not Geese have muscle-to-fat ratio at least 10-to-1.
(c) Either Hawks or Parakeets (or both) have muscle-to-fat ratio at least
10-to-1.

Pairs of statements represent conditionals such as the following:

(18) (a) Cardinals have muscle-to-fat ratio at least 10-to-1 assuming that
Parakeets do.

11 A more literal rendition of closure under boolean conjunction would be “Eagles have muscle-
to-fat ratio at least 10-to-1 and Cardinals have muscle-to-fat ratio at least 10-to-1.” We assume
throughout that the official logical structure of statements is clear from the abbreviations that
make them more colloquial. Note also that negation of predicates may often be expressed
without “not.” Thus, “have muscle-to-fat ratio below 10-to-1” serves as the negation of “have
muscle-to-fat ratio at least 10-to-1.”
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 153

(b) Eagles have muscle-to-fat ratio at least 10-to-1 assuming that Geese
don’t.
(c) Ducks and Geese have muscle-to-fat ratio at least 10-to-1 assuming
that Hawks do.

Judgments about chance begin to fade in the face of increasing logical com-
plexity, but many people have a rough sense of probability for hundreds of
complex and conditional structures like those in (17) and (18). We claim

(19) Estimates of the chances of complex and conditional events defined over
Q(o 1 ) · · ·Q(o n ) are often mentally generated from no more information
than:
(a) probability estimates for each of Q(o 1 ) · · ·Q(o n ); and
(b) the pairwise similarity of the n objects o i to each other.

Just as for Gap2, we here make no claim about the probabilities of

Q(o 1 ) · · ·Q(o n ). The focus is rather on sentences involving logical connec-
tives and conditionalization, predicting their perceived chance on the basis of
(19)a,b. The next section describes an algorithm whose inputs are the prob-
abilities of Q(o 1 ) · · ·Q(o n ) along with the pairwise similarities of the o i , and
whose output is the probability of every complex and conditional statement
over Q(o 1 ) · · ·Q(o n ). In support of (19), experimental data will be presented
to show that the algorithm’s output is a fair approximation of human intuition
about statements simple enough to be evaluated.
The algorithm is not intended as a performance model, that is, as an account
of the processes whereby people convert the algorithm’s inputs into outputs.
The algorithm lends credence to (19) only by demonstrating the sufficiency of
its inputs as a basis for generating its output. The human method for effecting
the same transformation is doubtless different.
Finally, consider a person whose estimates of probability are to be predicted.
We use Ch(·) and Ch(· : ·) to denote the chances that the person assigns to
statements and conditional statements. It is not assumed that Ch is a total
function since the person may have no opinion about the probability of highly
complex statements. As before, we let sim(·, ·) denote the person’s similarity
function over pairs of objects. It is assumed that sim(o i , o j ) ∈ [0, 1], with 1
representing identity and 0 maximal dissimilarity. For simplicity, we again
assume that sim is symmetric in its two arguments.
The goal of our algorithm may now be stated concisely. It is to predict the
values of Ch from sim along with just the values of Ch on Q(o 1 ) · · ·Q(o n ).
We begin by reviewing elements of subjective probability (for more ample
discussion, see Jeffrey, 1983; Nilsson, 1986; and Skyrms, 2000, among other
sources). Then an overview of the algorithm is presented, followed by details.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

154 S. Blok, D. Osherson, and D. L. Medin

5.1 Subjective Probability

Let Q(o 1 ) · · ·Q(o n ) be given, where each statement Q(o i ) is called a
variable. By a complete conjunction we mean a sentence of the form
±Q(o 1 ) ∧ · · · ∧ ±Q(o n ), where each ± is either blank or the negation sym-
bol ¬. For n = 3, one complete conjunction is ¬Q(o 1 ) ∧ Q(o 2 ) ∧ ¬Q(o 3 ). A
complete conjunction relative to (16) might be expressed in English as

Hawks and Eagles and Parakeets but neither Cardinals nor Geese nor Ducks have
muscle-to-fat ratio at least 10-to-1.

When there are n variables, there are 2n complete conjunctions. By a prob-

ability distribution over Q(o 1 ) · · ·Q(o n ) is meant a function that maps each
complete conjunction to a nonnegative number so that all the numbers sum
to one. Such a distribution is extended as follows to the class of boolean
formulas constructed from Q(o 1 ) · · ·Q(o n ). Given a distribution Prob and a
formula ϕ, Prob (ϕ) is defined to be

{Prob (α) : α is a complete conjunction that logically implies ϕ}.

Prob is extended again to pairs (ϕ, ψ) of formulas such that Prob (ψ) > 0;
specifically:

Prob (ϕ ∧ ψ)
Prob (ϕ, ψ) = .
Prob (ψ)

Conforming to the usual convention, we write Prob (ϕ, ψ) as Prob (ϕ : ψ)

(the conditional probability of ϕ assuming ψ). To summarize, any procedure
mapping the complete conjunctions into nonnegative numbers that sum to
one also assigns probabilities and conditional probabilities to formulas of
arbitrary complexity.
The function Ch is called (probabilistically) coherent if there is some dis-
tribution Prob such that (a) for all formulas ϕ, if Ch(ϕ) is defined then
Ch(ϕ) = Prob (ϕ), and (b) for all pairs ϕ, ψ of formulas, if Ch(ϕ : ψ) is
defined then Ch(ϕ : ψ) = Prob (ϕ : ψ). It is widely recognized that human
estimates of chance are easily led into incoherence.12 Since our algorithm
makes its predictions through construction of a (coherent) distribution, its
accuracy is limited by whatever incoherence is manifested in Ch.

12 See the references in note 2. Incoherence can be “fixed” via a method described in Batsell,
Brenner, Osherson, Vardi, & Tsvachidis (2002).
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 155

5.2 Overview of the Algorithm

The algorithm is based on a function f , defined below, that maps binary
conjunctions of form ±Q(o i ) ∧ ±Q(o j ) (i < j ) into the unit interval. Given
a complete conjunction α over n variables, f (α) is defined to be the average
of f (±Q(o i ) ∧ ±Q(o j )) for all binary conjunctions ±Q(o i ) ∧ ±Q(o j ) whose
conjuncts occur in α (i < j ≤ n). To illustrate, if n = 3, then the complete
conjunction ¬Q(o 1 ) ∧ Q(o 2 ) ∧ ¬Q(o 3 ) includes three binary conjunctions,
namely, ¬Q(o 1 ) ∧ Q(o 2 ), ¬Q(o 1 ) ∧ ¬Q(o 3 ), and Q(o 2 ) ∧ ¬Q(o 3 ). f maps
each of them into [0, 1]; and f maps ¬Q(o 1 ) ∧ Q(o 2 ) ∧ ¬Q(o 3 ) into the av-
erage of the latter three numbers. Use of average reflects the difficulty of eval-
uating long conjunctions directly; the binary conjunctions serve as estimates
of the compatibility of all the conjuncts in a given complete conjunction.
We conceive of f as assigning “provisional chances” to the set of complete
conjunctions generated from Q(o 1 ) · · ·Q(o n ). Such provisional chances may
not define a probability distribution since there is no guarantee that

{ f (α) : α is a complete conjunction} = 1.

After the provisional chances are generated via f , the goal of the algorithm
is to build a genuine distribution that comes as close as possible to respecting
them, subject to the constraints imposed by the rated probability Ch(Q(o i ))
of each variable. We therefore solve the following quadratic programming
problem:
(20) Find a probability distribution Prob such that

{( f (α) − Prob (α))2 : α is a complete conjunction}

is minimized subject to these constraints:

(a) Prob (Q(o i )) = Ch(Q(o i )) for i ≤ n,
 (α) ≥ 0 for all complete conjunctions α, and
(b) Prob
(c) {Prob (α) : α is a complete conjunction} = 1.
The last two constraints embody the requirement that Prob be a probability
distribution. It can be shown (Luenberger, 1984) that the solution Prob to
(20) is unique. It is taken to be the output of our algorithm.
We note that (20) can be supplemented with any other equalities Prob (ϕ) =
Ch(ϕ) or Prob(ϕ : ψ) = Ch(ϕ : ψ) where ϕ, ψ are arbitrary formulas; the
only requirement is that the set of equalities appearing in (20) be coherent. The
resulting quadratic program yields a distribution that honors these constraints
while minimizing discrepancy with similarity-based estimates of the chances
of complete conjunctions. Thus, our method may be applied to situations in
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

156 S. Blok, D. Osherson, and D. L. Medin

which the agent has particular information about the probabilities of various
events and conditional events.13

5.3 The Function f

Let c be the conjunction ±Q(o i ) ∧ ±Q(o j ) (i < j ≤ n). We define f (c) via
four cases, depending on the polarity of the two conjuncts of c.

Case 1: c = Q(o i ) ∧ Q(o j ). As sim(o i , o j ) approaches unity, Q(o i ) and

Q(o j ) become the same proposition, hence Ch(c) should approach
min{Ch(Q(o i )), Ch(Q(o j ))}.14 As sim(o i , o j ) approaches zero, Q(o i ) and
Q(o j ) bear no relation to each other, hence should be probabilistically in-
dependent, so that Ch(c) approaches Ch(Q(o i )) × Ch(Q(o j )). Linear inter-
polation between these extremes yields

f (Q(o i ) ∧ Q(o j )) = [sim(o i , o j ) × min{Ch(Q(o i )), Ch(Q(o j ))}]

+ [(1 − sim(o i , o j )) × Ch(Q(o i )) × Ch(Q(o j ))].

Case 2: c = ¬Q(o i ) ∧ ¬Q(o j ). Substituting 1 − Ch(Q(o i )) for Ch(Q(o i )),

and similarly for Q(o j ), transforms this case into the prior one. We therefore
define

f (¬Q(o i ) ∧ ¬Q(o j )) = [sim(o i , o j )×min{1 − Ch(Q(o i )), 1 − Ch(Q(o j ))}]

+ [(1 − sim(o i , o j )) × (1 − Ch(Q(o i ))) × (1 − Ch(Q(o j )))].

Case 3: c = Q(o i ) ∧ ¬Q(o j ). As sim(o i , o j ) approaches unity, Q(o i ) and

Q(o j ) become the same proposition, hence c becomes a contradiction
and Ch(c) should approach zero. As sim(o i , o j ) approaches zero, Q(o i ) and
¬Q(o j ) bear no relation to each other, hence should be probabilistically in-
dependent. Linear interpolation yields

f (Q(o i ) ∧ ¬Q(o j ) = (1 − sim(o i , o j )) × Ch(Q(o i )) × (1 − Ch(Q(o j ))).

13 More generally, the constraints may be weak inequalities rather than equalities.
14 As sim(o i , o j ) approaches unity, Ch(c) should also approach max{Ch(Q(o i )), Ch(Q(o j ))}. Prior
to reaching the limit, however, the maximum might exceed Ch(c), which is probabilistically
incoherent. No such incoherence is introduced by the minimum.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 157

Case 4: c = ¬Q(o i ) ∧ Q(o j ). This case is parallel to the preceding one. Hence

f (¬Q(o i ) ∧ Q(o j ) = (1 − sim(o i , o j )) × (1 − Ch(Q(o i ))) × Ch(Q(o j )).

The foregoing definition of f seems to be among the simplest ways to assign

reasonable probabilities to conjunctions, based on just similarity and the
chances attributed to variables.
Let us denote by QPf the algorithm just described, involving quadratic
programming and the use of f to assign provisional chances to the set of
complete conjunctions generated from Q(o 1 ) · · ·Q(o n ).

5.4 Test of the Algorithm QPf

Materials. As objects o i we chose the following sets of avian categories:

(21) (a) hawk, eagle, parakeet, cardinal, goose, duck

(b) robin, sparrow, chicken, bluejay, pigeon, parrot.

For predicates Q we chose the following:

(22) (a) have muscle-to-fat ratio at least 10-to-1

(b) have detectable testosterone blood levels throughout the year.

Based on no more than intuition, we expected that the predicates would be

stable and monotonic for our subjects. Applying each predicate to each group
yields four sets of six variables Q(o 1 ) · · ·Q(o 6 ).
Similarity was assessed by measuring the distance between each kind of
bird in Figure 1 of Rips, Shoben, and Smith (1973, p. 10). The latter figure
displays the multidimensional scaling solution among the twelve categories
in (21)a,b.15 These distances are shown in Table 6.1. To compute sim(o i , o j ),
we divided each distance by the largest value in the associated matrix, and
subtracted this number from 1. The most separated pair in a given subset is
thus assigned 0, whereas 1 is reserved for self-identity.
Note that the same similarity function is used for all participants in the
study. Since similarity questions were not posed, there is no possibility of re-
spondents interpreting similarity queries as disguised requests for conditional
probability.

15 The raw data for the multidimensional scaling algorithm were rated semantic similarities for
each pair of categories. Distance in the output reflects dissimilarity. (Included in the ratings was
semantic similarity to the superordinate concept bird.)
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

158 S. Blok, D. Osherson, and D. L. Medin

table 6.1. Distances (in millimeters) in Figure 1 of Rips,

Shoben, & Smith (1973)

robin sparrow chicken bluejay pigeon

sparrow 4.0
chicken 39.0 38.5
bluejay 10.3 6.4 38.1
pigeon 17.5 17.4 21.5 18.6
parrot 15.5 16.0 23.5 17.9 2.3
hawk eagle parakeet cardinal goose
eagle 4.8
parakeet 41.0 43.0
cardinal 36.2 36.9 13.4
goose 40.0 44.8 34.1 42.6
duck 40.5 44.9 31.5 40.5 3.2
Note: To illustrate, sim(robin, parrot) = 1.0 − (15.5/39.0)

Statements. Each of the four combinations available from (21) and (22) gives
rise to a set of variables Q(o 1 ) · · ·Q(o 6 ). Thirty-six complex statements were
generated from the variables. The logical forms of the complex statements
were the same across the four sets of stimuli; they are shown in Table 6.2. To
illustrate, relative to (21)a and (22)a, the complex statement of form Q(o 6 ) ∧
¬Q(o 5 ) is (17)b, above. For the same set of stimuli, the form Q(o 2 )|¬Q(o 5 ) is
(18)b. In this way, each of the four stimulus-sets was associated with forty-two
statements (six variables plus thirty-six complex statements).

Procedure. Forty Princeton undergraduates assessed the probabilities of the

forty-two statements associated with one of the four sets of stimuli. It was
explained that “probability” was to be interpreted in the sense of personal
conviction. Brief explanation was also provided for the logical connectives.
Probabilities were collected via a computer interface, which presented the
forty-two statements in individualized random order.
Three data sets were discarded without further analysis; two manifested
stereotyped answers (one or two values chosen for almost all questions), the
other included a majority of extreme responses (either zero or one). Of the
remaining thirty-seven participants, ten worked with the stimuli (21)a, (22)a,
ten with (21)a, (22)b, nine with (21)b, (22)a, and eight with (21)b, (22)b.

5.5 Results
The average estimates for the variables are shown in Table 6.3. The algorithm
(20) was used to derive a separate distribution Prob for each participant.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 159

table 6.2. The 36 Complex Statements Figuring in the Experiment

Structure # Formulas
p∧q 7 Q(o 1 ) ∧ Q(o 2 ) Q(o 5 ) ∧ Q(o 6 ) Q(o 1 ) ∧ Q(o 5 )
Q(o 1 ) ∧ Q(o 6 ) Q(o 2 ) ∧ Q(o 5 ) Q(o 2 ) ∧ Q(o 6 )
Q(o 3 ) ∧ Q(o 4 )
p ∧ ¬q 7 Q(o 1 ) ∧ ¬Q(o 3 ) Q(o 1 ) ∧ ¬Q(o 4 ) Q(o 2 ) ∧ ¬Q(o 5 )
Q(o 2 ) ∧ ¬Q(o 6 ) Q(o 6 ) ∧ ¬Q(o 1 ) Q(o 1 ) ∧ ¬Q(o 2 )
Q(o 6 ) ∧ ¬Q(o 5 )
¬ p ∧ ¬q 7 ¬Q(o 4 ) ∧ ¬Q(o 3 ) ¬Q(o 6 ) ∧ ¬Q(o 3 ) ¬Q(o 6 ) ∧ ¬Q(o 5 )
¬Q(o 2 ) ∧ ¬Q(o 1 ) ¬Q(o 6 ) ∧ ¬Q(o 1 ) ¬Q(o 2 ) ∧ ¬Q(o 5 )
¬Q(o 3 ) ∧ ¬Q(o 5 )
p∨q 3 Q(o 2 ) ∨ Q(o 3 ) Q(o 4 ) ∨ Q(o 5 ) Q(o 6 ) ∨ Q(o 4 )
p:q 5 Q(o 1 ) : Q(o 2 ) Q(o 2 ) : Q(o 3 ) Q(o 3 ) : Q(o 2 )
Q(o 4 ) : Q(o 1 ) Q(o 5 ) : Q(o 2 )
p : ¬q 4 Q(o 1 ) : ¬Q(o 5 ) Q(o 2 ) : ¬Q(o 6 )
Q(o 3 ) : ¬Q(o 4 ) Q(o 6 ) : ¬Q(o 5 )
¬p : q 3 ¬Q(o 4 ) : Q(o 3 ) ¬Q(o 5 ) : Q(o 1 ) ¬Q(o 6 ) : Q(o 3 )
Note: Indices are relative to the orderings in (21)a, b. The symbols ∧ and ∨ denote
conjunction and disjunction, respectively; ¬ denotes negation. Conditional probabilities
are indicated with a colon (:).

We then compared Prob to the subject’s thirty-six assignments of chance to

complex statements.16 The median correlations obtained for the four sets of
stimuli appear in the first column of Table 6.4. Next, for a given stimulus set
we averaged the probabilities assigned to each of the forty-two statements,
and repeated the analysis for this “average subject.” (There was thus one
average subject for each stimulus-set; its forty-two “judgments” were the mean
responses from the eight, nine, or ten subjects who worked with those stimuli.)
The correlations between predicted and observed values for the four average
subjects are shown in the second column of Table 6.4; scatter plots are provided
in Figure 6.3. When random numbers are substituted for the provisional
chances provided by f , the quadratic program (20) yields correlations reliably
lower than those shown in Figure 6.3 ( p < .05); see the third column of
Table 6.4.
Recall that our algorithm returns a genuine distribution of probability.
Its predictive success thus requires a modicum of coherence in the input
estimates of chance. Allowing a 10% margin of error, we counted the number

16 Prob perfectly “predicts” the probabilities assigned to Q(o 1 ) · · ·Q(o 6 ) since the quadratic pro-
gramming problem (20) takes the latter values as constraints on the solution.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

160 S. Blok, D. Osherson, and D. L. Medin

table 6.3. Average Probabilities Assigned to Variables

(21)a, (22)a (21)a, (22)b (21)b, (22)a (21)b, (22)b

# (N = 10) (N = 10) (N = 9) (N = 8)
1 .79 .75 .54 .35
2 .79 .74 .61 .43
3 .39 .36 .30 .53
4 .45 .42 .61 .40
5 .28 .59 .40 .44
6 .29 .47 .48 .36
Note: The numbers under # are relative to the orderings in (21)a,b.

of times the inputs violated the following consequences of the probability

calculus17 :

(23) (a) Prob (ϕ) + Prob (ψ) − 1.05 ≤ Prob (ϕ ∧ ψ) ≤ Prob (ϕ) +
.05, Prob (ψ) + .05
(b) Prob (ϕ) − .05, Prob (ψ) − .05 ≤ Prob (ϕ ∨ ψ) ≤ Prob (ϕ) +
Prob (ψ) + .05
(c) Prob (ϕ ∧ ψ)/Prob (ψ) − .05 ≤ Prob (ϕ : ψ) ≤ Prob (ϕ ∧ ψ)/
Prob (ψ) + .05.

All twenty-four complex, absolute statements (see Table 6.2) are either
conjunctions or disjunctions, hence accountable to one of (23)a,b. Only four
of the twelve conditional statements could be tested against (23)c since in
eight cases the required conjunction ϕ ∧ ψ does not figure among the absolute
statements. For the averaged data, there are few violations of 23a,b – 6, 0, 1, and
1, respectively, for the four stimulus sets. In contrast, (23)c was violated all four
times in each stimulus set (averaged data), confirming the well documented
incomprehension of conditional probability by college undergraduates (see,
e.g., Dawes, Mirels, Gold, & Donahue, 1993). Incoherence was greater when
tabulated for each subject individually. Participants averaged 10.3 violations
of 23a,b out of 24 possible, and 3.8 violations of (23)c out of 4 possible.
Our results support Thesis (19) inasmuch as the only inputs to the algo-
rithm are similarities and the chances of variables. The algorithm’s predictions
are not perfect, however, and suggest that revised methods might prove more
successful. In particular, accuracy may improve when similarities are elicited
from respondents to the probability questions, instead of derived globally
from an ancient text (Rips et al., 1973).

17 For their derivation, see Neapolitan (1990).

P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 161

table 6.4. Correlations Between Predicted and Observed Probabilities

(1) (2) (3) (4)

Correlations Correlations
Median Correlation with random based on 32
individual for the average provisional provisional
correlation subject chances chances
(21)a, (22)a .811 .915 .757 .795
(21)a, (22)b .763 .899 .539 .872
(21)b, (22)a .609 .753 .531 .601
(21)b, (22)b .595 .773 .238 .754
Note: Each correlation involves the set of 36 complex observations. All coefficients in
columns (1), (2), and (4) are reliable ( p < .01). Column (3) shows the correlation between
predicted and observed probabilities when random numbers in the unit interval are
substituted for the provisional chances provided by f in the quadratic program (20),
using averaged data. In each case, the correlations in columns (2) and (3) differ reliably
( p < .05). The correlations in the column (4) are the average of six computations with 32
randomly chosen complete conjunctions.

The construction of a (coherent) probability distribution over the set of

complete conjunctions renders our algorithm unrealistic psychologically, but
enhances its potential role in the design of autonomous, intelligent agents.
Indeed, the probability calculus provides an attractive canonical form for
reasoning, and is often considered to be the “faithful guardian of common
sense” (Pearl, 1988).

5.6 Application to the Design of Autonomous Intelligent Agents

Our algorithm allows numerous probabilities to be computed from a small
reservoir of stored values. For example, a distribution for thirty variables
of form Q(o i ) requires more than a billion probabilities but is specified by
only 465 numbers within the scheme described here (namely, 30 probabilities
for variables plus 435 pairwise similarities). The empirical results described
above suggest that the 465 numbers can be chosen in a way that generates
reasonable – that is human-like – estimates of chance. Moreover, the bulk of
these numbers (namely, the similarity coefficients) will be of potential use for
other purposes, for example, when the agent must decide whether one object
can be substituted for another in pursuit of a larger goal.
Although 465 numbers suffice to define a distribution involving thirty
variables, recovering the distribution via our algorithm requires quadratic
programming with a billion-dimensional matrix [see (20), above]. To avoid
such computations, a small random sample of complete conjunctions can
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

162 S. Blok, D. Osherson, and D. L. Medin

1.00 1.00

0.80 0.80

Observed estimates
Observed estimates

0.60 0.60

0.40 0.40

0.20 0.20

0.00 0.00
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
Predicted estimates Predicted estimates

Stimuli (21)a, (22)a Stimuli (21)a, (22)b

0.80 0.80

0.70 0.70
Observed estimates

0.60
Observed estimates

0.60

0.50 0.50

0.40 0.40

0.30 0.30

0.20 0.20

0.10 0.10
0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00
Predicted estimates Predicted estimates
Stimuli (21)b, (22)a Stimuli (21)b, (22)b
figure 6.3. Predicted versus observed probabilities for the four sets of stimuli.

be chosen to carry all positive probabilities, and quadratic programming

undertaken just with them. At least one sampled complete conjunction must
imply each variable in order for the variables’ estimated chances to be pre-
served in the solution; but this condition on samples is easy to check.
To assess the potential of the foregoing scheme, for each set of stimuli we
drew six samples of thirty-two complete conjunctions, out of the sixty-four
that are generated by six variables.18 The resulting quadratic program – which

18 A few times it was necessary to draw a new sample because one of the variables was not implied
by any of the chosen complete conjunctions.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 163

is one quarter the size of the original – was then used to predict the subjects’
average estimates of chance. The last column in Table 6.4 shows the mean of
the six correlations produced for each set of stimuli. Comparing these values
to the second column reveals some loss of predictive accuracy; substantial
correlations nonetheless remain. In a sophisticated implementation of our
algorithm, loss of accuracy could be partially remedied by applying quadratic
programming to several samples of complete conjunctions and averaging the
results.

.  
Both Gap2 and QP f rely on similarity to predict the probabilities assigned
to complex statements. We hope that the performance of the models demon-
strates that similarity can play an explanatory role in theories of induction
provided that attention is limited to stable predicates. Extension beyond the
confines of stability is not trivial but might be facilitated by comparison to
the simpler case.
The importance of similarity to estimates of probability was already ar-
ticulated in the 17th and 18th century (see Cohen, 1980; Coleman, 2001).
Bringing quantitative precision to this idea would be a significant achieve-
ment of modern psychology.


Thanks to Robert Rehder for careful reading and helpful suggestions. Re-
search supported by NSF grants 9978135 and 9983260. Contact information:
[email protected], [email protected], [email protected].

References
Aguilar, C. M., and D. L. Medin (1999). “Asymmetries of comparison,” Psychonomic
Bulletin & Review, 6, 328–337.
Ahn, W., C. W. Kalish, D. L. Medin, & S. A. Gelman (1995). “The role of covariation
versus mechanism information in causal attribution,” Cognition, 54, 299–352.
Barsalou, L. W., and B. H. Ross (1986). “The roles of automatic and strategic process-
ing in sensitivity to superordinate and property frequency,” Journal of Experimental
Psychology: Learning, Memory, & Cognition, 12(1), 116–134.
Batsell, R., L. Brenner, D. Osherson, M. Vardi, and S. Tsvachidis (2002). “Eliminating in-
coherence from subjective estimates of chance,” Proceedings of the Ninth International
Workshop on Knowledge Representation. Morgan Kaufmann.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

164 S. Blok, D. Osherson, and D. L. Medin

Blok, S. V. (2004). “Modeling induction as conditional probability judgment,” Disserta-

tion Abstracts International, 65(1), 457, Northwestern University.
Carey, S. (1985). Conceptual change in childhood, MIT Press, Cambridge MA.
Cohen, L. J. (1980). “Some historical remarks on the Baconian conception of probability,”
Journal of the History of Ideas, 61, 219–231.
Coleman, D. (2001). “Baconian probability and Hume’s theory of testimony,” Hume
Studies, 27(2), 195–226.
Coley, J. D., S. Atran, and D. L. Medin (1997). “Does rank have its privilege? Inductive
inference within folk biological taxonomies,” Cognition, 64, 73–112.
Dawes, R., H. L. Mirels, E. Gold, and E. Donahue (1993). “Equating inverse probabilities
in implicit personality judgments,” Psychological Science, 4(6), 396–400.
Eells, E., and B. Fitelson (2002). “Symmetries and asymmetries in evidential support,”
Philosophical Studies, 107(2), 129–142.
Gentner, D., and A. Markman (1997). “Structure mapping in analogy and similarity,”
American Psychologist, 52(1), 45–56.
Gigerenzer, G. (1996). “Reply to Tversky and Kahneman,” Psychological Review, 103(3),
592–593.
Goldstone, R. (1994a). “Similarity, interactive activation, and mapping,” Journal of Ex-
perimental Psychology: Learning, Memory, & Cognition, 20(1), 3–28.
Goldstone, R. L. (1994b). “The role of similarity in categorization: Providing a ground-
work,” Cognition, 52, 125–157.
Goodman, N. (1955). Fact, fiction, and forecast, Bobbs-Merrill, Indianapolis IN.
Goodman, N. (1972). “Seven strictures on similarity,” in Problems and projects. Bobbs-
Merrill, New York NY.
Hampton, J., and I. Cannon (2004). “Category-based induction: An effect of conclusion
typicality,” Memory & Cognition, 32(2), 235–243.
Heit, E. (2000). “Properties of inductive reasoning,” Psychonomic Bulletin & Review, 7,
569–592.
Heit, E., and U. Hahn (1999). “Diversity-based reasoning in children age 5 to 8,” in
Proceedings of the Twenty-First Annual Conference of the Cognitive Science Society,
Erlbaum, Mahawah NJ.
Jeffrey, R. C. (1983). The logic of decision (2nd edition). The University of Chicago Press,
Chicago IL.
Johnson-Laird, P. N. (1983). Mental models. Harvard University Press, Cambridge
MA.
Juslin, P., and M. Persson (2002). “PROBabilities from EXemplars (PROBEX): A ‘lazy’
algorithm for probabilistic inference from generic knowledge,” Cognitive Science, 26,
563–607.
Kahneman, D., and A. Tversky (1996). “On the reality of cognitive illusions,” Psychological
Review, 103(3), 582–591.
Keil, F., W. C. Smith, D. J. Simons, and D. T. Levin (1998). “Two dogmas of conceptual
empiricism: Implications for hybrid models of the structure of knowledge,” Cognition,
65, 103–135.
Lassaline, M. E. (1996). “Structural alignment in induction and similarity,” Journal of
Experimental Psychology : Learning, Memory, and Cognition, 22, 754–770.
Lo, Y., A. Sides, J. Rozelle, and D. Osherson (2002). “Evidential diversity and premise
probability in young children’s inductive judgment,” Cognitive Science, 26, 181–206.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

From Similarity to Chance 165

López, A., S. Atran, J. Coley, D. Medin, and E. Smith (1997). “The tree of life: Uni-
versal and cultural features of folkbiological taxonomies and inductions,” Cognitive
Psychology, 32, 251–295.
López, A., S. A. Gelman, G. Gutheil, and E. E. Smith (1992). “The development of
category-based induction,” Child Development, 63, 1070–1090.
Luenberger, D. G. (1984). Linear and nonlinear programming (2nd edition), Addison-
Wesley, Reading MA.
Medin, D. L., J. D. Coley, G. Storms, and B. K. Hayes (2003). “A relevance theory of
induction,” Psychonomic Bulletin & Review 10(3), 517–532.
Medin, D. L., R. L. Goldstone, and D. Gentner (1993). “Respects for similarity,” Psycho-
logical Review, 100, 254–278.
Medin, D. L., and M. M. Schaffer (1978). “Context model of classification learning,”
Psychological Review, 85, 207–238.
Neapolitan, R. (1990). Probabilistic reasoning in expert systems: Theory and algorithms.
John Wiley & Sons, New York NY.
Nilsson, N. (1986). “Probabilistic logic,” Artificial Intelligence, 28(1), 71–87.
Ortiz, C. L. (1999). “A commonsense language for reasoning about causation and rational
action,” Artificial Intelligence, 111, 73–130.
Osherson, D., E. E. Smith, and E. Shafir (1986). “Some origins of belief,” Cognition, 24,
197–224.
Osherson, D., E. E. Smith, O. Wilkie, A. López, and E. Shafir (1990). “Category-based
induction,” Psychological Review, 97(2), 185–200.
Pearl, J. (1988). Probabilistic reasoning in intelligent systems. Morgan Kaufmann, San
Mateo, CA.
Quine, W. V. O. (1960). Word and Object. MIT Press, Cambridge MA.
Rips, L. (1975). “Inductive judgments about natural categories,” Journal of Verbal Learn-
ing and Verbal Behavior, 14, 665–681.
Rips, L., E. Shoben, and E. Smith (1973). “Semantic distance and the verification of
semantic relations,” Journal of Verbal Learning and Verbal Behavior, 12, 1–20.
Russell, S. J., and P. Norvig (2003). Artificial intelligence: A modern approach (2nd edition).
Prentice Hall, Upper Saddle River NJ.
Shafir, E., E. Smith, and D. Osherson (1990). “Typicality and reasoning fallacies,” Memory
and Cognition, 18(3), 229–239.
Skyrms, B. (2000). Choice & chance: An introduction to inductive logic, Wadsworth,
Belmont CA.
Sloman, S. A. (1993). “Feature based induction,” Cognitive Psychology, 25, 231–280.
Sloman, S. A., and D. Lagnado (2005). “The problem of induction,” in Cambridge
handbook of thinking & reasoning, ed. by K. Holyoak and R. Morrison. Cambridge
University Press, Cambridge.
Sloman, S. A., and L. J. Rips (1998). “Similarity as an explanatory construct,” Cognition,
65, 87–101.
Smith, E. E., E. Shafir, and D. Osherson (1993). “Similarity, plausibility, and judgments
of probability,” Cognition, 49, 67–96.
Tentori, K., N. Bonini, and D. Osherson (2004). “The conjunction fallacy: A misunder-
standing about conjunction?” Cognitive Science, 28(3), 467–477.
Tentori, K., V. Crupi, N. Bonini, and D. Osherson (2007). “Comparison of confirmation
measures,” Cognition, 103, 107–119.
P1: d
0521672443c06 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:53

166 S. Blok, D. Osherson, and D. L. Medin

Turner, H. (1999). “A logic of universal causation,” Artificial Intelligence, 113, 87–

123.
Tversky, A. (1977). “Features of similarity,” Psychological Review, 84, 327–352.
Tversky, A., and I. Gati (1978). “Studies of similarity,” in Cognition and categorization,
ed. by E. Rosch and B. Lloyd. Erlbaum, Hillsdale NJ.
Tversky, A., and D. Kahneman (1983). “Extensional versus intuitive reasoning: The
conjunction fallacy in probability judgment,” Psychological Review, 90, 293–315.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning

Joshua B. Tenenbaum, Charles Kemp, and Patrick Shafto

. 
Philosophers since Hume have struggled with the logical problem of induc-
tion, but children solve an even more difficult task – the practical problem
of induction. Children somehow manage to learn concepts, categories, and
word meanings, and all on the basis of a set of examples that seems hopelessly
inadequate. The practical problem of induction does not disappear with ado-
lescence: adults face it every day whenever they make any attempt to predict an
uncertain outcome. Inductive inference is a fundamental part of everyday life,
and for cognitive scientists, a fundamental phenomenon of human learning
and reasoning in need of computational explanation.
There are at least two important kinds of questions that we can ask about
human inductive capacities. First, what is the knowledge on which a given
instance of induction is based? Second, how does that knowledge support gen-
eralization beyond the specific data observed: how do we judge the strength
of an inductive argument from a given set of premises to new cases, or infer
which new entities fall under a concept given a set of examples? We provide
a computational approach to answering these questions. Experimental psy-
chologists have studied both the process of induction and the nature of prior
knowledge representations in depth, but previous computational models of
induction have tended to emphasize process to the exclusion of knowledge
representation. The approach we describe here attempts to redress this im-
balance by showing how domain-specific prior knowledge can be formalized
as a crucial ingredient in a domain-general framework for rational statistical
inference.
The value of prior knowledge has been attested by both psychologists and
machine learning theorists, but with somewhat different emphases. Formal
analyses in machine learning show that meaningful generalization is not

167
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

168 J. B. Tenenbaum, C. Kemp, and P. Shafto

possible unless a learner begins with some sort of inductive bias: some set of
constraints on the space of hypotheses that will be considered (Mitchell, 1997).
However, the best known statistical machine-learning algorithms adopt rela-
tively weak inductive biases and thus require much more data for successful
generalization than humans do: tens or hundreds of positive and negative
examples, in contrast to the human ability to generalize from just one or
few positive examples. These machine algorithms lack ways to represent and
exploit the rich forms of prior knowledge that guide people’s inductive in-
ferences and that have been the focus of much attention in cognitive and
developmental psychology under the name of “intuitive theories” (Murphy &
Medin, 1985). Murphy (1993) characterizes an intuitive theory as “a set of
causal relations that collectively generate or explain the phenomena in a do-
main.” We think of a theory more generally as any system of abstract principles
that generates hypotheses for inductive inference in a domain, such as hy-
potheses about the meanings of new concepts, the conditions for new rules, or
the extensions of new properties in that domain. Carey (1985), Wellman and
Gelman (1992), and Gopnik and Meltzoff (1997) emphasize the central role
of intuitive theories in cognitive development, both as sources of constraint
on children’s inductive reasoning and as the locus of deep conceptual change.
Only recently have psychologists begun to consider seriously the roles that
these intuitive theories might play in formal models of inductive inference
(Gopnik & Schulz, 2004; Tenenbaum, Griffiths, & Kemp, 2006; Tenenbaum,
Griffiths, & Niyogi, in press). Our goal here is to show how intuitive theo-
ries for natural domains such as folk biology can, when suitably formalized,
provide the foundation for building powerful statistical models of human
inductive reasoning.
Any familiar thing can be thought about in a multitude of ways, and
different kinds of prior knowledge will be relevant to making inferences
about different aspects of an entity. This flexibility poses a challenge for
any computational account of inductive reasoning. For instance, a cat is a
creature that climbs trees, eats mice, has whiskers, belongs to the category
of felines, and was revered by the ancient Egyptians – and all of these facts
could potentially be relevant to an inductive judgment. If we learn that cats
suffer from a recently discovered disease, we might think that mice also have
the disease; perhaps the cats picked up the disease from something they ate.
Yet if we learn that cats carry a recently discovered gene, lions and leopards
seem more likely to carry the gene than mice. Psychologists have confirmed
experimentally that inductive generalizations vary in such ways, depending
on the property involved. Our computational models will account for these
phenomena by positing that people can draw on different prior knowledge
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 169

structures – or different intuitive theories – within a single domain, and by

showing how very different patterns of inference can arise depending on
which of these theories is triggered.
Our models aim for both predictive and explanatory power. As in any
mathematical modeling, we seek accounts that can provide close quantitative
fits to human judgments across a range of different tasks or contexts, with
a minimum number of free parameters or ad hoc assumptions. At the same
time, we would like our models to explain why people make the inductive
generalizations that they do make, and why these judgments are mostly suc-
cessful in the real world – how people can reliably come to true beliefs about
the world from very limited data. In the spirit of rational analysis (Anderson,
1990; Oaksford & Chater, 1998), or Marr’s (1982) computational-theory level
of analysis, we will assume that people’s inductive capacities can be charac-
terized as approximations to optimal inferences, given the structure of the
environments and the task contexts that they have adapted to over the course
of evolution and development. Our mission as modelers is then to characterize
formally the nature of the optimal inference mechanism, the relevant aspects
of environmental structure and task context, and the interaction between
these components.
Our core proposal has two components. First, domain-specific knowledge
that supports induction in different contexts can be captured using appropri-
ate families of probabilistic models defined over structured representations:
in particular, relational systems of categories such as taxonomic hierarchies
or food webs. These structured probabilistic models are far from being com-
plete formalizations of people’s intuitive domain theories; they are minimalist
accounts, intended to capture only those aspects of theories relevant for the
basic inductive inferences we study. Second, knowledge in this form can
support inductive generalization by providing the prior probabilities for a
domain-general Bayesian inference engine. Both of these claims are necessary
for explaining how people’s inductive inferences can be so successful, and
perhaps approximately optimal, with respect to the world that we live in. The
structured representations of intuitive domain theories are important because
the world contains genuine structure: a tree-structured representation of bi-
ological species is useful, for example, because it approximates the structure
of the evolutionary tree. Bayesian inference is important because it provides a
normative and general-purpose procedure for reasoning under uncertainty.
Taking these two components together – rational domain-general statistical
inference guided by appropriately structured intuitive domain theories – may
help explain the uniquely human capacity for learning so much about the
world from so little experience.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

170 J. B. Tenenbaum, C. Kemp, and P. Shafto

Our work goes beyond previous formal models of induction, which either
do not address the rational statistical basis of people’s inferences or find
it difficult to capture the effects of different kinds of knowledge in different
inductive contexts, or both. In one representative and often-cited example, the
similarity-coverage model of Osherson, Smith, and colleagues, the domain-
specific knowledge that drives generalization is represented by a similarity
metric (Osherson et al., 1990). As we will see below, this similarity metric
has to be defined in a particular way in order to match people’s inductive
judgments. That definition appears rather arbitrary from a statistical point of
view, and arbitrarily different from classic similarity-based models of other
cognitive tasks such as categorization or memory retrieval (Nosofsky, 1986;
Hintzman et al., 1978). Also, the notion of similarity is typically context
independent, which appears at odds with the context-dependent nature of
human inductive reasoning. Even if we allow some kind of context-specific
notion of similarity, a similarity metric seems too limited a representation
to carry the richly structured knowledge that is needed in some contexts, or
even simple features of some reasoning tasks such as the strong asymmetry of
causal relations. In contrast, the knowledge that drives generalization in our
theory-based Bayesian framework can be as complex and as structured as a
given context demands.
The plan of this chapter is as follows. Section 2 provides a brief review
of the specific inductive tasks and phenomena we attempt to account for,
and Section 3 briefly describes some previous models that have attempted
to cover the same ground. Section 4 introduces our general theory-based
Bayesian framework for modeling inductive reasoning and describes two spe-
cific instantiations of it that can be used to model inductive reasoning in two
important natural settings. Section 5 compares our models and several alter-
natives in terms of their ability to account for people’s inductive judgments on
a range of tasks. Section 6 discusses the relation between our work and some
recent findings that have been taken to be problematic for Bayesian models
of induction. Section 7 concludes and offers a preview of ongoing and future
research.

.  
This section reviews the basic property induction task and introduces the core
phenomena that our models will attempt to explain. Following a long tradition
(Rips, 1975; Carey, 1985; Osherson et al., 1990; Sloman, 1993; Heit, 1998), we
will focus on inductive arguments about the properties of natural categories,
in particular biological species categories. The premises of each argument state
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 171

that one or more specific species have some property, and the conclusion (to
be evaluated) asserts that the same property applies to either another specific
species or a more general category (such as all mammals). These two kinds of
arguments are called specific and general arguments, respectively, depending
only on the status of the conclusion category.
prop
We use the formula P1 , . . . Pn −−→ C to represent an n-premise argument
where Pi is the i th premise, C is the conclusion, and prop indicates the
property used. We will often abbreviate references to these components of an
argument. For example, the argument
Gorillas have T4 hormones
Squirrels have T4 hormones
All mammals have T4 hormones
T4
might be represented as gorillas, squirrels −→ mammals.
The most systematic studies of property induction have used so-called
“blank properties.” For arguments involving animal species, these are proper-
ties that are recognized as biological but about which little else is known – for
example, anatomical or physiological properties such as “has T4 hormones”
or “has sesamoid bones.” As these properties are hardly “blank” – it is impor-
tant that people recognize them as deriving from an underlying and essential
biological cause – we will instead refer to them as “generic biological proper-
ties.”
For this class of generic properties, many qualitative reasoning phenomena
have been described: Osherson et al. (1990) identify thirteen and Sloman
(1993) adds several others. Here we mention just three. Premise-conclusion
similarity is the effect that argument strength increases as the premises be-
T4
come more similar to the conclusion: for example, horses −
→ dolphins is
T4
weaker than seals −
→ dolphins. For general arguments, typicality is the effect
that argument strength increases as the premises become more typical of
T4
the conclusion category. For example, seals −
→ mammals is weaker than
T4
horses −→ mammals, since seals are less typical mammals than horses. Finally,
diversity is the effect that argument strength increases as the diversity of the
T4
premises increases. For example, horses, cows, rhinos −
→ mammals is weaker
T4
than horses, seals, squirrels −
→ mammals.
Explaining inductive behavior with generic biological properties is a chal-
lenging problem. Even if we find some way of accounting for all the phenom-
ena individually, it is necessary to find some way to compare their relative
weights. Which is better: an argument that is strong according to the typicality
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

172 J. B. Tenenbaum, C. Kemp, and P. Shafto

criterion, or an argument that is strong according to the diversity criterion?

The problem is especially difficult because arguments that are strong ac-
cording to one criterion may be weak according to another: for example,
T4
seals, squirrels −
→ mammals has premises that are quite diverse but not very
typical of the conclusion. For this reason, rather than trying to account for
isolated qualitative contrasts between pairs of arguments, we will assess the
performance of computational models in terms of how well they can predict
relative argument strengths across multiple datasets each containing a large
number of arguments of the same form.
The strength of an argument depends critically on the property in-
volved, because changing the property will often alter the inductive con-
text. Many researchers have described related effects (Gelman & Markman,
1986; Heit & Rubinstein, 1994; Shafto & Coley, 2003; Smith et al., 1993),
and we mention just three of them. Gelman and Markman (1986) showed
that children reason differently about biological properties (e.g., “has cold
blood”) and physical properties (e.g., “weighs one ton”) – for example,
coldblood oneton
brontosaurus −−−−→ triceratops is relatively strong, but brontosaurus −−−→
triceratops is relatively weak. Heit and Rubinstein (1994) showed that anatom-
ical or physiological properties and behavioral properties are treated dif-
ferently by adults. While anatomical or physiological properties typically
support default, similarity-like patterns of inductive reasoning, behavioral
properties may depend less on generic similarity and more on shared eco-
logical roles. Finally, Shafto and Coley (2003) argue that disease properties
may draw on causal knowledge about predator-prey interactions, and thus
may be treated differently from arguments about generic biological proper-
leptospirosis
ties. For example, salmon −−−−−→ grizzlybears may be judged stronger than
leptospirosis
grizzlybears −−−−−→ salmon, where leptospirosis stands for “carry leptospiro-
sis bacteria.” This asymmetry has no justification in terms of the similarity
between salmon and grizzly bears, which is presumably symmetric or nearly
so, but it seems sensible from the perspective of causal reasoning: knowing
that grizzly bears eat salmon, it seems more likely that grizzly bears would
catch some disease from salmon than that any specific disease found in grizzly
bears necessarily came from the salmon that they eat.
Our aim has been to develop a unifying computational framework that
can account for many of the phenomena mentioned above. We will focus in
this chapter on modeling reasoning about two kinds of properties: the classic
setting of generic biological properties, and causally transmitted properties
such as diseases (Shafto & Coley, 2003) that give rise to very different patterns
of judgment. Before introducing the details of our framework, we summarize
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 173

several existing models of property induction and describe how we hope to

improve on them.

.  
The tradition of modeling property induction extends at least as far back as
the work of Rips (1975). Here we summarize a few of the more prominent
mathematical models that have been developed in the intervening thirty years.

3.1 Similarity-Coverage Model

The similarity-coverage model (SCM) of Osherson et al. (1990) is perhaps
the best known mathematical model of property induction. It predicts the
strength of inductive arguments as a linear combination of two factors,
the similarity of the conclusion to the premises and the extent to which
the premises “cover” the smallest superordinate taxonomic category includ-
ing both premises and conclusion. The SCM has some appealing properties.
It makes accurate predictions for generic biological properties, and it uses
a simple equation that predicts many different kinds of judgments with a
minimum of free parameters. Yet the SCM has two major limitations. First,
it can only use domain knowledge that takes the form of pairwise similari-
ties or superordinate taxonomic categories. The model is therefore unable to
handle inductive contexts that rely on knowledge which cannot be expressed
in this form. Second, the SCM lacks a principled mathematical foundation:
the accuracy of its predictions depends critically on certain arbitrary choices
which specify the mathematical form of the model.
This arbitrariness shows up most clearly in the formal definition of cover-
age: the average over all instances i in the superordinate class of the maximal
similarity between i and the examples in the premise set. We refer to this
(standard) version of SCM as “MaxSim.” Osherson et al. (1990) also con-
sider a variant we call “SumSim,” in which coverage is defined by averaging
the summed similarity to the examples over all instances of the superordi-
nate class. Generalization based on the summed similarity to exemplars or
weight traces is the foundation for many other successful models of catego-
rization, learning, and memory (Nosofsky, 1986; Kruschke, 1992; Hintzman
et al., 1978) and can be interpreted in rational statistical terms as a version of
nonparametric density estimation (Ashby & Alfonso-Reese, 1995; Silverman,
1986). Yet despite these precedents for using a summed-similarity measure,
Osherson et al. (1990) advocate MaxSim, or some weighted combination
of MaxSim and SumSim – perhaps because SumSim performs dramatically
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

174 J. B. Tenenbaum, C. Kemp, and P. Shafto

worse than MaxSim in judging the strength of general arguments (see Section
5). Since Osherson et al. (1990) do not explain why different measures of
setwise similarity are needed in these different tasks, or why SumSim per-
forms so much worse than MaxSim for inductive reasoning, the SCM is less
principled than we might like.

3.2 Feature-Based Models

As Goodman (1972) and Murphy and Medin (1985) have argued, similarity
is a vague and elusive notion, and it may be meaningless to say that two
objects are similar unless a respect for similarity has been specified. Instead
of founding a model directly on similarity judgments, an alternative is to
start with a collection of object features, which might plausibly be observable
perceptually or have been previously learned.1 In some settings, it will be
necessary to assume that the features are extracted from another kind of input
(linguistic input, say), but in general the move from similarity to features is a
move towards models that can learn directly from experience.
The feature-based model of (Sloman, 1993) computes inductive strength
as a normalized measure of feature overlap between conclusion and example
categories. Sloman (1993) presents a quantitative comparison with the SCM:
the results are not conclusive but suggest that the model does not predict hu-
man judgments as accurately as the SCM. The model, however, predicts some
qualitative phenomena that the SCM cannot explain. More recently, Rogers
and McClelland (2004) have presented a feature-based approach to semantic
cognition that uses a feedforward connectionist network with two hidden
layers. This connectionist approach is more ambitious than any of the others
we describe, and Rogers and McClelland (2004) apply their model to a diverse
set of semantic phenomena. One of the applications is a property induction
task where the model makes sensible qualitative predictions, but there has
been no demonstration so far that the model provides good quantitative fits
to human judgments.
From our perspective, both feature-based models share the limitations of
the SCM. Despite the range of applications in Rogers and McClelland (2004),
it is not clear how either model can be extended to handle causal settings
or other inductive contexts that draw on sophisticated domain knowledge.
Both models also include components that have been given no convincing

1 Note that a feature-based version of the SCM is achieved if we define the similarity of two objects
as some function of their feature vectors. Section 5 assesses the performance of this model.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 175

justification. The model of Sloman (1993) uses a particular mathematical

measure of feature overlap, but it is not clear why this should be the right
measure to use. Rogers and McClelland (2004) provide no principled expla-
nation for the architecture of their network or their strategy for computing the
strength of inductive arguments, and their model appears to rely on several
free parameters.

3.3 A Bayesian Analysis

Heit (1998) presented a computational theory where property induction
is modeled as Bayesian inference. This inference engine is essentially the
same as we describe later in Section 4.1. Applying a Bayesian analysis to
any specific case of property induction requires a prior distribution over
hypotheses about the extension of the property in question. Heit does not
specify a formal method for generating priors, nor does he test his model
quantitatively against any specific judgments. He shows that it captures several
qualitative phenomena if it is supplied with the right kinds of priors, and
that appropriate priors could allow it to handle both blank and non-blank
properties. He also suggests how priors could be extracted from long-term
memory: the probability of a hypothesis could be proportional to the number
of familiar features that can be retrieved from memory and that have the same
extension as that hypothesis. But it is far from clear that this suggestion would,
if implemented, yield appropriate priors; as we show below, a simple version
of this idea does not perform nearly as well as the SCM’s gold standard in
predicting human judgments.
Our framework adopts a Bayesian approach to inference like Heit’s, but
we emphasize the importance of modeling the form and the origins of ap-
propriate priors. A formal account of how the learner’s prior is structured
and where it comes from provides two distinct advantages. First, it leads to
strong quantitative models, predicting people’s inductive judgments as well
or better than any previous approach. More important, it adds genuine ex-
planatory power. Most of the knowledge that supports induction is captured
by the prior, and a computational theory should be as explicit as possible
about the knowledge it assumes and how that knowledge is used. It has long
been argued that different inductive contexts lead to quite different patterns
of generalization behavior, but whether this is due to the operation of differ-
ent kinds of knowledge, different mechanisms of reasoning, or both has not
been so clear. We will argue that a single general-purpose Bayesian reasoning
mechanism may be sufficient, by showing explicitly how to generate priors
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

176 J. B. Tenenbaum, C. Kemp, and P. Shafto

that can capture two important and very different kinds of domain knowledge
and that can strongly predict people’s judgments in appropriately different
inductive contexts.

.  -  

Our framework includes two components: a Bayesian engine for inductive
inference, and a language for specifying relevant aspects of domain theo-
ries and using those theories to generate prior probability distributions for
the Bayesian inference engine. The Bayesian engine reflects domain-general
norms of rational statistical inference and remains the same regardless of the
inductive context. Different domain theories may be appropriate in different
inductive contexts, but they can often be formalized as instances of a single
unifying scheme: a probabilistic process, such as diffusion, drift or trans-
mission, defined over a structured representation of the relevant relations
between categories, such as taxonomic or ecological relations.

4.1 The Bayesian Inference Engine

Assume that we are working within a finite domain containing n categories.
We are interested in a novel property, Q, that applies to some unknown subset
of these categories. Let H be the hypothesis space of all logically possible
extensions for Q – the set of all possible subsets h of categories in the domain,
each of which could a priori be the extension of the novel property. Since
there are n categories, the number of hypotheses is 2n . To each hypothesis we
assign a prior probability p(h), where p(h) is the probability that h includes
all and only the categories with property Q.
Suppose now that we observe X, a set of m labeled objects where the labels
indicate whether each category in X has property Q. We want to compute
p(y has Q|X), the probability that object y has propertyQ given the examples
X. Summing over all hypotheses in H, we have


p(y has Q|X) = p(y has Q, h|X) (1)
h∈H

= p(y has Q|h, X) p(h|X). (2)
h∈H
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 177

Now p(y has Q|h, X) equals one if y ∈ h and zero otherwise (independent
of X). Thus


p(y has Q|X) = p(h|X) (3)
h∈H:y∈h
 p(X|h) p(h)
= (4)
h∈H:y∈h
p(X)

where the last step follows from Bayes’s rule.

The numerator in Equation 4 depends on the prior p(h), as well as on the
likelihood p(X|h), the probability of observing the labeled examples X given
that h is the true extension of Q. The likelihood p(X|h) should in general
depend on the process assumed to generate the observations in X. Here, for
simplicity we will assume that p(X|h) ∝ 1 for all hypotheses consistent with
X, and p(X|h) = 0 otherwise.2 A hypothesis h for the extension of property
Q is consistent with a set of labeled examples X if h includes all positively
labeled categories in X and excludes all negatively labeled categories in X.
Then Equation 4 is equivalent to


p(h)
h∈H:y∈h,h consistent with X
p(y has Q|X) =  (5)
p(h)
h∈H:h consistent with X

which is the proportion of hypotheses consistent with X that also include y,

where each hypothesis is weighted by its prior probability p(h). The proba-
bility of generalizing to y will thus be high to the extent that it is included in
most of the high-prior-probability hypotheses that also include the observed
examples X.
Other inferences can be formulated similarly. For example, the probability
that all categories in a larger class Y (e.g., all mammals) have property Q

2 More complex sampling models could be appropriate in other circumstances, and are discussed
in Tenenbaum & Griffiths (2001) and Kemp & Tenenbaum (2003). For instance, an assumption
that examples are randomly drawn from the true extension of Q might be particularly important
when learning word meanings or concepts from ostensive examples (Tenenbaum & Xu, 2000;
Xu & Tenenbaum, in press).
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

178 J. B. Tenenbaum, C. Kemp, and P. Shafto

could be formalized as

p(h)
h∈H:Y ⊂h,h consistent with X
p(Y has Q|X) =  (6)
p(h)
h∈H:h consistent with X

Note that a Bayesian approach needs no special purpose rules for deal-
ing with negative evidence or arguments with multiple premises. Once the
prior distribution p(h) and the likelihood p(X|h) have been specified, com-
puting the strength of a given argument involves a mechanical application
of the norms of rational inference. Since we have assumed a simple and
domain-general form for the likelihood above, our remaining task is to spec-
ify appropriate domain-specific prior probability distributions.

4.2 Theory-Based Priors

Generating the prior distributions used in Equations 5 and 6 appears to be a
difficult problem – for either the cognitive modeler or the human reasoner.
Somehow we need to specify 2n numbers: p(h) for each of the logically pos-
sible hypotheses h in H. We cannot simply assign all hypotheses equal prior
probability; without any inductive biases, meaningful generalization would
be impossible (Mitchell, 1997). Explicitly enumerating the priors for all 2n
hypotheses is also not an option. This would introduce far more degrees of
freedom into the model than we could ever hope to test empirically. More
importantly, it would fail to capture the most interesting aspect of these
priors – that they are not just lists of numbers but rather the products of
abstract systems of knowledge, or intuitive theories. Induction with different
kinds of properties – such as anatomical features, behavioral tendencies, or
disease states of animal species – will require different kinds of priors because
we have qualitatively different kinds of knowledge that we bring to bear in
those contexts. Our priors for induction can change when we learn new facts,
but the biggest changes come not from statistical observations that might sim-
ply favor one hypothesis over another. Priors can change most dramatically,
and can change globally across a large slice of the hypothesis space, when we
acquire qualitative knowledge that alters our intuitive domain theories: when
we learn about a new species with unexpected characteristics, such as whales
or ostriches, or we learn something surprising about how various species
might be related, such as that whales and dolphins are mammals, or we learn
some new principle about how properties are distributed over species, such
as that diseases tend to spread through physical contact or food.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 179

The heart of our proposal is a way to understand formally how intuitive

domain theories can generate the prior probabilities needed for induction.
Two aspects of intuitive theories are most relevant for constructing priors for
property induction: representations of how entities in the domain are related
to each other, and processes or mechanisms operating over those relational
structures that give rise to the distribution of properties over entities. To
be concrete, we will assume that each of the n categories in a domain can
be represented as a node in a relational structure, such as a directed or
undirected graph. Edges in the graph represent relations that are relevant
for determining inductive potential, such as taxonomic or causal relations
among categories. Priors are generated by a stochastic process defined on this
graph, such as a diffusion process, a drift process, or a noisy transmission
process. These processes can be used to capture general beliefs about how
properties of various types tend to be distributed over related categories in
the domain. Once we have sufficiently characterized the relational structure
and the stochastic generating process, that will fully specify the 2n numbers in
the prior. By choosing different kinds of structures and stochastic processes,
we can capture different kinds of knowledge and account for qualitatively
different patterns of inductive reasoning. In this chapter we describe and test
two such models, one for reasoning about generic biological properties such
as anatomical and physiological features, and another for reasoning about
causally transmitted properties such as diseases.

4.2.1 A Theory For Generic Biological Properties, The prior distribution for
default biological reasoning is based on two core assumptions: the taxonomic
principle and the mutation principle. The taxonomic principle asserts that
species belong to groups in a nested hierarchy, and more precisely, that the
taxonomic relations among species can be represented by locating each species
at some leaf node of a rooted tree structure. Tree-structured taxonomies
of species appear to be universal across cultures (Atran, 1998), and they
also capture an important sense in which species are actually related in the
world: genetic relations due to the branching process of evolution. Outside of
intuitive biology, tree-structured taxonomies play a central role in organizing
knowledge about many systems of natural-kind and artifact categories (Rosch,
1978), as well as the meanings of words that label these categories (Markman,
1989; Tenenbaum & Xu, 2000).
The structures of people’s intuitive taxonomies are liable to deviate from
scientific phylogenies in non-negligible ways, since people’s theories are based
on very different kinds of observations and targeted towards predicting dif-
ferent kinds of properties. Hence we need some source of constraint besides
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

180 J. B. Tenenbaum, C. Kemp, and P. Shafto

(a) (b) (c)

(d) (e)
elephant
rhino
horse
cow
chimp
gorilla
mouse

dolphin
seal
squirrel

figure 7.1. (a) A folk taxonomy of mammal species. (b–e) Examples of mutation his-
tories.

scientific biology in order to generate the particular tree-structured taxon-

omy that our model will use. We have explored several different approaches to
reconstructing taxonomic trees that best characterize people’s intuitive theo-
ries. One possibility is to perform hierarchical clustering on people’s explicit
judgments of similarity for all pairs of species in the domain. Hierarchical
clustering could also be applied to more implicit measures of psychological
distance between species: for example, we could represent each animal using a
set of behavioral and morphological features (e.g., “lives in water,” “has a tail”)
and set the distance between two animals to be the distance between their
feature vectors. We could also use more structured domain knowledge that
might obviate the need for any bottom-up clustering. Both direct judgments
of pairwise similarity and ratings of feature-category associations have been
collected for many of the standard domains of animal species used in study-
ing reasoning about generic biological properties (Osherson et al., 1990), and
these data sets provide a convenient basis for comparing different modeling
frameworks on equal grounds. Figure 7.1a shows a taxonomic tree that was
reconstructed for ten mammal species, based on hierarchical clustering over a
set of features collected by Osherson and colleagues (and slightly augmented
as described later in Section 5.1).
Some simple and intuitive prior distributions p(h) can be generated using
the taxonomic principle alone. For instance, we could assign a uniform prob-
ability to each hypothesis corresponding to one of the nineteen taxonomic
clusters (including singleton species) shown in the tree, and zero probability
to all other sets of species. We call this the “strict-tree model.” It corresponds
roughly to some informal accounts of taxonomically driven induction (Atran,
1995), and it can qualitatively reproduce some important phenomena, such
as diversity-based reasoning. Essentially this prior has also been used success-
fully to explain how people learn words that label taxonomic object categories.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 181

But to see that it is not sufficient to explain biological property induction,

T4 T4
compare the arguments seals, squirrels − → horses and seals, cows −→ horses.
The second appears stronger than the first, yet under the intuitive taxon-
omy shown in Figure 7.1, the strict-tree model assigns them both the same
strength, since each set of premises is compatible with only one hypothesis,
the set of all mammals.
The solution to this problem comes from realizing that in scientific biology,
the assumption that every property is strictly confined to a single taxon is
false. Properties arise randomly via mutations, and through a combination
of chance and natural selection, two species may share a property even if it
occurs in no common ancestor. Convergent evolution is particularly likely
for survival-significant traits of interest to people (e.g., being warm-blooded,
having an eye, being able to fly or swim, forming long-term monogamous
pair bonds). A more veridical folk theory of generic biological properties
should thus generate a prior that allows some probability (perhaps small) of
a property occuring in two or more disjoint taxa.
To capture these patterns in how generic biological properties are dis-
tributed, our theory will assume that novel properties are generated by a
mutation-like stochastic process defined over the tree structure specified by
the taxonomic principle. We refer to this second assumption as the mutation
principle and the Bayesian model of induction that uses the resulting prior as
the “evolutionary model.” Intuitively, we can imagine a property that arises at
the root of the tree and spreads out towards the leaves. The property starts out
with some value (on or off) at the root, and at each point in the tree there is a
small probability that the property will mutate, or switch its value. Whenever
a branch splits, both lower branches inherit the value of the property at the
point immediately above the split, and mutations thereafter occur indepen-
dently along the lower branches. For example, if a property is absent at the
root of the tree in Figure 7.1a, but switches on at the two points marked in
Figure 7.1d, then it would apply to just horses, cows, dolphins, and seals.
More formally, the mutation process is characterized by a single parameter
λ, specifying the average rate at which mutations occur. Mutations are mod-
eled as transitions in a two-state Markov chain defined continuously over the
tree, with infinitesimal matrix [−λ, λ; λ, −λ]. The probability that two points
in the tree separated by a branch of length t will have different values (present
−2λt
or absent) for a given property is then 1−e2 . For simplicity we assume here
that any property is equally likely to be present or absent at the tree root,
and that mutations are equally likely in both directions (present → absent,
absent → present), but more generally, prior knowledge about the nature
or distribution of a feature could bias these probabilities. A mutation history
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

182 J. B. Tenenbaum, C. Kemp, and P. Shafto

for a property Q is an assignment of zero or more mutations to branches of

the tree, together with a specification of the state of the property (present or
absent) at the root of the tree.
The Markov mutation model allows us to assign a probability to any hypo-
thetical mutation history for a property, based only on whether the property
changes state between each pair of branch points on the tree. This is almost
but not exactly what we need. Bayesian property induction requires a prior
probability p(h) that some novel property Q applies to any possible subset
h of species. A mutation history for property Q induces a labeling of all leaf
nodes of the tree – all species in the domain – according to whether Q is
present or absent at each node. That is, each mutation history is consistent
with exactly one hypothesis for Bayesian induction, but the mapping from
mutation histories to hypotheses is many-to-one: many different mutation
histories, with different probabilities, will specify the same distribution of a
property Q over the set of species. We define the prior probability p(h) that
a new property Q applies to some subset h of species to be the sum of the
probability of all mutation histories consistent with h. This prior (and all
the resulting Bayesian computations) can be computed efficiently using belief
propagation over the tree, as described more formally in Kemp, Griffiths, et al.
(2004). For small trees, it can also be approximated by taking many samples
from a simulation of the mutation process. We randomly generate a large
number of hypothetical mutation histories with probability proportional to
their likelihood under the Markov mutation model, by first choosing the
property’s state at the tree root and then following the causal direction of the
mutation process down along all branches of the tree to the leaf nodes. We
can estimate the prior p(h) for each hypothetical labeling h of the leaf nodes
as the frequency with which mutation histories consistent with h occur in this
sample.
The prior generated by this mutation principle has several qualitative fea-
tures that seem appropriate for our problem. First, unlike the strict-tree model
described above, the mutation process induces a non-zero prior probability
for any logically possible extension of a novel property (i.e., any of the 2n hy-
potheses in the full hypothesis space H). The prior is “smooth” with respect
to the tree: the closer two species lie in the tree, the more likely they are to
share the same value for a novel property. Properties are more likely to switch
on or off along longer branches (e.g., the mutation history in Figure 7.1b
is more likely than in 7.1c). Hence p(h) will be higher for properties that
hold only for a highly distinctive taxonomic group of species, such as the
aquatic mammals or the primates, than for properties that hold only for a
less distinctive taxonomic group, such as the “farm animals” ({horses, cows}).
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 183

Multiple independent occurences of the same property will be rare (e.g., the
mutation history in Figure 7.1b is more likely than in 7.1d, which is more
likely than in 7.1e). Hence the prior favors simpler hypotheses corresponding
to a single taxonomic cluster, such as {dolphins, seals}, over more complex
hypotheses corresponding to a union of multiple disjoint taxa, such as the
set {dolphins, seals, horses, cows, mice}. The lower the mutation rate λ, the
greater the preference for strictly tree-consistent hypotheses over discon-
nected hypotheses. Thus this model captures the basic insights of simpler
heuristic approaches to taxonomic induction (Atran, 1995) but embeds them
in a more powerful probabilistic model that supports fine-grained statistical
inferences.
Several caveats about the evolutionary model are in order. The mutation
process is just a compact mathematical means for generating a reasonable
prior for biological properties. We make no claim that people have con-
scious knowledge about mutations as specifically biological phenomena, any
more than a computational vision model which appeals to an energy func-
tion claims that the visual system has explicit knowledge about energy. It is
an open question whether the biological principles guiding our model are
explicitly represented in people’s minds or only implicitly present in the in-
ference procedures they use. We also do not claim that a mutation process is
the only way to build a prior that can capture generalizations about generic
biological properties. The key idea captured by the mutation process is that
properties should vary randomly but smoothly over the tree, so that categories
nearby in the tree are more likely to have the same value (present or absent)
for a given property than categories far apart in the tree. Other stochas-
tic processes including diffusion processes, Brownian motion, and Gaussian
processes will also capture this intuition and should predict similar patterns
of generalization (Kemp & Tenenbaum, submitted). The scope of such “prob-
abilistic taxonomic” theories is likely to extend far beyond intuitive biology:
there may be many domains and contexts where the properties of categories
are well described by some kind of smooth probability distribution defined
over a taxonomic-tree structure, and where our evolutionary model or some
close relative may thus provide a compelling account of people’s inductive
reasoning.

4.2.2 A Theory for Causally Transmitted Properties. The evolutionary model

in the previous section is appropriate for reasoning about many kinds of bi-
ological properties, and perhaps some kinds of nonbiological but still tax-
onomically distributed properties as well, but other classes of properties
give rise to very different patterns of inductive generalization and will thus
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

184 J. B. Tenenbaum, C. Kemp, and P. Shafto

require differently structured priors. For instance, in some contexts inductive

generalization will be asymmetric: the probability of generalizing a property
from category A to category B will not be the same as the probability of gen-
eralizing the same property from B to A. Earlier we described one biological
context where induction is frequently asymmetric: reasoning about disease
properties, such as the probability that grizzly bears will have a disease given
that salmon do, or vice versa. In this section we show how to capture this sort
of inductive reasoning within our theory-based Bayesian framework.
The formal model we present could be appropriate for many properties
whose distributions are governed by asymmetric causal relationships among
categories, but for concreteness, we will assume that the domain comprises a
set of species categories, and the novel property is a disease spread by predator-
prey relations between species. Two abstract principles of an intuitive theory
are relevant in this context; these principles are analogous to the taxonomic
and mutation principles underlying the evolutionary model in the last section.
First, a structured representation captures the relevant relations between
entities in the domain: in this context, we posit a set of directed predator-prey
relations. An example of such a food web is shown in Figure 7.4a. Second,
a stochastic process defined over that directed-network structure generates
prior probabilities for how novel properties are distributed over species: here,
the process is designed to capture the noisy arrival and transmission of disease
states.
As with the mutation process for generic biological properties presented
above, we can describe this noisy-transmission process by explaining how to
generate a single hypothetical property. If we draw a large sample of hypo-
thetical properties by repeating this procedure many times, the prior proba-
bility for each hypothesis about how to generalize a particular novel property
will be proportional to the number of times it appears in this sample. The
transmission process has two parameters: b, the background rate, and t, the
transmission probability. The first parameter captures the knowledge that
species can contract diseases from causes external to the food web. For each
species in the web, we toss a coin with bias b to decide whether that species
develops the disease as a result of an external cause. The second parameter is
used to capture the knowledge that diseases can spread from prey to predator
up the food web. For each edge in the graph, we toss a coin with bias t to
determine whether it is active. We stipulate that all species reachable by active
links from a diseased animal also contract the disease. We refer to the Bayesian
model using this prior as the “causal transmission model.”
Figure 7.2 shows one possible outcome if we sample a property from the
causal transmission process. We see that two of the species develop the disease
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 185

(a) (b)

figure 7.2. One simulated sample from the causal-transmission model, for the foodweb
shown in Figure 7.4a. (a) Initial step showing species hit by the background rate (black
ovals) and active routes of transmission (black arrows). (b) Total set of species with
disease via background and transmission.

for reasons unrelated to the foodweb, and that four of the causal links are
active (Figure 7.2a). An additional three species contract the disease by eating
a disease-ridden species (Figure 7.2b). Reflecting on these simulations should
establish that the prior captures two basic intuitions. First, species that are
linked in the web by a directed path are more likely to share a novel property
than species which are not directly linked. The strength of the correlation
between two species’ properties decreases as the number of links separating
them increases. Second, property overlap is asymmetric: a prey species is more
likely to share a property with one of its predators than vice versa.
Although the studies we will model here consider only the case of disease
transmission in a food web, many other inductive contexts fit the pattern of
asymmetric causal transmission that this model is designed to capture. Within
the domain of biological species and their properties, the causal model could
also apply to reasoning about the transmission of toxins or nutrients. Outside
of this domain, the model could be used, for example, to reason about the
transmission of lice between children at a day care, the spread of secrets
through a group of colleagues, or the progression of fads through a society.

4.2.3 Common Principles of Theory-Based Priors. It is worth noting several

deep similarities between these two theory-based Bayesian models, the evo-
lutionary model and the causal transmission model, which point to more
general aspects of our approach. In each case, the underlying intuitive theory
represents knowledge on at least two levels of abstraction. The lower, more
concrete level of the theory specifies a graph-structured representation of the
relevant relations among species (taxonomic neighbors or predators), and a
stochastic process that distributes properties over that relational structure,
which is characterized by one or two numerical parameters controlling its
degree of stochasticity. At a higher level, each theory specifies the form of
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

186 J. B. Tenenbaum, C. Kemp, and P. Shafto

the structure and stochastic process that are appropriate for reasoning about
a certain domain of entities and properties: generic biological properties,
such as anatomical and physiological attributes, are distributed according to
a noisy mutation process operating over a taxonomic tree; diseases are dis-
tributed according to a noisy transmission process operating over a directed
food web.
For a fixed set of categories and properties, the lower level of the theory is
all we need to generate a prior for inductive reasoning. But the higher level
is not just a convenient abstraction for cognitive modelers to talk about – it
is a critical component of human knowledge. Only these abstract principles
tell us how to extend our reasoning when we learn about new categories,
or a whole new system of categories in the same domain. When European
explorers first arrived in Australia, they were confronted with many entirely
new species of animals and plants, but they had a tremendous head start
in learning about the properties of these species because they could apply
the same abstract theories of taxonomic organization and disease transmis-
sion that they had acquired based on their European experience. Abstract
theories appear to guide childrens’ conceptual growth and exploration in
much the same way. The developmental psychologists Wellman and Gelman
(1992) distinguish between “framework theories” and “specific theories,” two
levels of knowledge that parallel the distinction we are making here. They
highlight framework theories of core domains – intuitive physics, intuitive
psychology, and intuitive biology – as the main objects of study in cognitive
development. In related work (Tenenbaum, Griffiths & Kemp, 2006; Tenen-
baum, Griffiths, & Niyogi, in press), we have shown how the relations between
different levels of abstraction in intuitive theories can be captured formally
within a hierarchical probabilistic model. Such a framework allows us to
use the same Bayesian principles to explain both how theories guide induc-
tive generalization and how the theories themselves might be learned from
experience.
Each of our theory-based models was built by thinking about how some
class of properties is actually distributed in the world, with the aim of giving
a rational analysis of people’s inductive inferences for those properties. It
is therefore not surprising that both the evolutionary model and the causal
transmission model correspond roughly to models used by scientists in rel-
evant disciplines – formalisms like the causal transmission model are used
by epidemiologists, and formalisms like the evolutionary model are used in
biological classification and population genetics. The correspondence is of
course far from perfect, and it is clearest at the higher “framework” level of
abstraction: constructs such as taxonomic trees, predator-prey networks, the
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 187

mutation process, or the transmission process may in some sense be shared

across intuitive theories and scientific theories, even while the specific tree or
foodweb structures, or the specific mutation or transmission rate parameters,
may differ in important ways.
Even though it may be imperfect and abstract, this correspondence between
the world’s structure and our models’ representations provides an important
source of constraint on our approach. If we were free to write down just any
sort of probabilistic model as the source of a prior probability distribution,
it would be possible to give a “rational analysis” of any coherent inductive
behavior, but it is not clear how much explanatory value that exercise would
have. By deriving priors from intuitive theories that at some deep level reflect
the actual structure of the world, it becomes clearer why these priors should
support useful generalizations in real-world tasks and how they might them-
selves be acquired by a rational learner from experience in this world. Our
approach can be extended to other inductive contexts by formally specifying
how the properties covered by those contexts are distributed in the world;
Kemp and Tenenbaum (submitted) consider two other important contexts in
addition to those discussed here.
Our primary goal here has been to characterize with the knowledge that
guides generalization (theory-based priors) and the input-output mapping
that allows this knowledge to be converted into judgments of inductive
strength (Bayesian inference). Our work is located at the most abstract of
Marr’s levels – the level of computational theory (Marr, 1982) – and we make
no commitments about the psychological or neural processes by which people
make inductive judgments. Inference in both of our models can be imple-
mented using efficient approximate methods that have appealing correlates in
the traditional toolkit of cognitive processes: for instance, belief propagation
over Bayesian networks, which can be seen as a kind of rational probabilistic
version of spreading activation. We find it encouraging that efficient and psy-
chologically plausible implementations exist for our models, but we are not
committed to the claim that inference in these Bayesian networks resembles
cognitive processing in any detailed way.
Finally, it is worth emphasizing that our models have not attempted to
capture all or most of the content and structure of people’s intuitive theories.
We are modeling just those aspects of theories that appear necessary to sup-
port inductive reasoning about properties in fairly specific contexts. We are
agnostic about whether people’s intuitive theories contain much richer causal
structures than those we attempt to model here (Carey, 1985), or whether
they are closer to light or skeletal frameworks with just a few basic principles
(Wellman & Gelman, 1992).
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

188 J. B. Tenenbaum, C. Kemp, and P. Shafto

.      

We now describe two series of experimental tests for the theory-based Bayesian
models introduced in the previous section. In each case, we consider multiple
sets of inductive arguments whose strengths have been rated or ranked by
human judges, and we compare these subjective argument strengths with the
theoretical predictions of the models. We will also compare with several alter-
native models, including the classic similarity-coverage model and multiple
variants within our Bayesian framework. The latter comparisons allow us to
illustrate the distinctive importance of both ingredients in our approach: an
appropriate representation of the relevant aspects of a domain’s structure, and
an appropriate probabilistic model for how properties of the relevant type are
distributed over that structure.

5.1 Reasoning about Generic Biological Properties

We begin by looking at the classic datasets of Osherson, Smith, and their
colleagues on inductive reasoning with generic biological properties. Five
datasets will be considered. The two “Osherson” datasets are taken from
Osherson et al. (1990). The “Osherson specific” set contains thirty-six two-
premise arguments, of the form gorillas, squirrels −→ horses. The conclusion
category of each of these arguments is the same: horses. The two premise
categories vary across arguments, but are always drawn from the set of ten
mammal species shown in Figure 7.1. Various generic biological predicates
are used across different arguments; in this section we drop references to
the specific property assuming it is one of these generics. The “Osherson
general” set contains forty-five three-premise general arguments, of the form
gorillas, squirrels, dolphins −→ mammals. The conclusion category of each of
these arguments is always mammals, while the three premise categories again
vary across arguments and are drawn from the set of ten mammal species
shown in Figure 7.1. The three “Smith” data sets are similar, but they draw
on different and fewer mammal categories (Smith et al., 1993).
We compare people’s judgments of argument strength in these five datasets
with the predictions of five computational models for induction with generic
biological properties. Two are theory-based Bayesian models: the evolutionary
model, in which the prior is generated by a mutation process defined on a tax-
onomic tree of species categories, and the strict-tree model, in which the prior
is simply a uniform distribution over taxonomic clusters of species (without
the possibility of a property arising in multiple disconnected branches of
the taxonomy). Another model is inspired by Heit’s proposal for a Bayesian
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 189

analysis (see Section 3.3), in which the prior probability of a hypothesis is

based on the number of familiar properties that can be retrieved from mem-
ory and that have the same extension as that hypothesis. We call this the
“raw-feature model,” because it embodies a prior that is based directly on raw
experience, without the benefit of a structured domain theory that might help
people to reason sensibly in cases that go substantially beyond their experi-
ence. The details of the raw-feature model are explained below. Finally, we
consider two versions of the similarity-coverage model, MaxSim and SumSim,
which respectively compute similarity to the set of premise categories in terms
of the maximum or summed similarity to those categories (see Section 3.1).
In order to predict the strengths of arguments about a particular set of
species, each of these models requires some way to represent people’s prior
knowledge about those species. We can compare the models on equal footing
by grounding the knowledge representations for each model in a matrix
of judged species-feature associations collected by Osherson and colleagues.
Participants were given forty-eight familiar species and eighty-five familiar
features (mostly anatomical or ecological properties, such as “has a tail”
or “lives in water”) and asked to rate the relative “strength of association”
between each species and each feature. Participants gave ratings on a scale
that started at zero and had no upper bound. In order to model the behavioral
judgments described below, we supplemented these data with feature ratings
for two additional species, cows and dolphins, to give a total of fifty species. We
also substituted the judged features of collies for dogs (because dogs appeared
as a premise category in some of the argument judgment tasks, but not in the
feature rating task). Ratings were linearly transformed to values between 0 and
100, then averaged. Let F be the resulting 50 × 85 matrix of average species-
feature associations. We also consider analogous binary matrices F θ obtained
by thresholding the average ratings at some level θ. Let S(F ) be a 50 × 50
matrix of species-species similarities, computed based on Euclidean distances
between the two feature vectors in F representing each pair of species.
For evaluating the predictions of the SCM models, we used the entries of
S(F ) to determine the necessary similarities between species. The taxonomic
tree for the evolutionary and strict-tree models was constructed by running
hierarchical agglomerative (“average linkage”) clustering on S(F ), restricted
to just the species categories involved in each experiment. The prior for the
raw-feature model was defined from the thresholded species-feature matrices
F θ , inspired by the proposal of Heit (1998). (We treat the threshold θ as a free
parameter of the model, to be optimized when fitting judgments of argument
strength.) We assume that the features participants retrieve from memory
correspond to the columns of F θ , plus an additional feature corresponding to
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

190 J. B. Tenenbaum, C. Kemp, and P. Shafto

r=0.93 r=0.92 r=0.89 r=0.98 r=0.8

Evolutionary

r=0.56 r=0.5 r=0.56 r=0.85 r=0.67

Strict tree

r=0.7 r=0.27 r=0.77 r=0.95 r=0.68

Raw features

r=0.71 r=0.79 r=0.93 r=0.94 r=0.84

MaxSim

r=0.74 r=0.34 r=0.85 r=0.92 r=0.018

SumSim

OshSpec OshGen SmithSpecFox SmithSpecEl SmithGen

figure 7.3. Comparing models of property induction with human judgments, for rea-
soning in a default biological context with generic anatomical or physiological properties.
Each row of plots shows the performance of a single model over all data sets; each column
shows the performance of all models over a single data set. In each plot, individual data
points represent the strengths of individual inductive arguments. The x-value of each
point represents the argument’s predicted strength according to a given model, while the
y-value represents the argument’s subjective strength according to the mean judgments
of human experimental participants.

all the animals. The prior p(h) assigned to any hypothesis h for the extension
of a novel property is proportional to the number of columns of F θ (i.e., the
number of familiar features) that are distributed as h specifies – that apply to
just those categories that h posits. Hypotheses that do not correspond to any
features in memory (any column of F θ ) receive a prior probability of zero.
All of these models (except the strict-tree model) include a single free pa-
rameter: the mutation rate in the evolutionary model, the balance between
similarity and coverage terms in MaxSim and SumSim, or the feature thresh-
old θ in the raw-feature model. Each model’s free parameter was set to the
value that maximized the average correlation with human judgments over all
five datasets.
Figure 7.3 compares the predictions for all five of these models on all
five datasets of argument strength judgments. Across these datasets, the
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 191

predictions of the evolutionary model are better than or comparable to the

best of the other models. This success provides at least some evidence that
the model’s core assumptions – a taxonomic-tree structure over species and a
mutation-like distribution of properties over that tree – do in fact characterize
the way people think and reason about generic biological properties. More
revealing insights come from comparing the performance of the evolutionary
model with that of the other models.
The strict-tree model captures the general trends in the data but does not
predict people’s judgments nearly as accurately as the evolutionary model.
This is because, without the mutation principle, the strictly taxonomic hy-
pothesis space is too rigid to capture the graded degrees of support that
more or less diverse premise sets provide for inductive generalization. For
example, in the Osherson-general experiment, the strict-tree model assigns
the same probability (100%) to cows, dolphins, squirrels −→ mammals and
seals, dolphins, squirrels −→ mammals, because in both cases the set of all
mammals is the only hypothesis consistent with the examples. The evolution-
ary model correctly distinguishes between these cases, recognizing that the
first premise set is better spread out over the tree and therefore provides better
evidence that all mammals have the novel property. An intuitive explanation
is that it feels very difficult to imagine a property that would be true of cows,
dolphins, and squirrels but not all mammals, while it seems more plausi-
ble (if unlikely) that there could be some characteristic property of aquatic
mammals (seals and dolphins) that might, for some unknown reason, also
be true of squirrels or rodents, but no other animals. The mutation principle
matches this intuition: the highly specific hypothesis {seal, dolphin, squirrel}
can be generated by only two mutations, one on a very long branch (and thus
relatively likely), while the hypothesis {cows, dolphins, squirrels} could arise
only from three mutations, all on relatively short branches.
This problem with the strict-tree model is hardly restricted to the specific
pair of arguments cited above, as can be seen clearly by the dense vertical
groupings of datapoints on the right-hand side of the plots in Figure 7.3.
Each of those vertical groupings corresponds to a set of arguments that are
judged to have very different strengths by people but that all receive maximum
probability under the strict-tree prior, because they are consistent with just
the single hypothesis of generalizing to all mammals.
The raw-feature model is more flexible than the strict-tree model, but it is
still not sufficient to capture the diversity effect, as can be seen by its dramati-
cally worse performance on the datasets with general arguments. Consider the
premise sets dolphins, chimps, squirrels and dolphins, seals, horses. It is diffi-
cult to think of anatomical or physiological properties that apply to all of the
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

192 J. B. Tenenbaum, C. Kemp, and P. Shafto

animals in the first set but only some of the animals in the second set. None
of the features in our dataset is strongly associated with dolphins, chimps,
and squirrels, but not also seals and horses. The raw-feature model therefore
finds it hard to discriminate between these two sets of premises, even though
it seems intuitively that the first set provides better evidence that all mammals
have the novel property.
More generally, the suboptimal performance of the raw-feature model
suggests that people’s hypotheses for induction are probably not based strictly
on the specific features that can be retrieved from memory. People’s knowledge
of specific features of specific animals is too sparse and noisy to be the direct
substrate of inductive generalizations about novel properties. In contrast, a
principal function of intuitive domain theories is to generalize beyond people’s
limited specific experiences, constraining the kinds of possible situations
that would be expected to occur in the world regardless of whether they
have been previously experienced (McCloskey et al., 1980; Murphy & Medin,
1985; Carey, 1985). Our framework captures this crucial function of intuitive
theories by formalizing the theory’s core principles in a generative model for
Bayesian priors.
Taken together, the performance of these three Bayesian variants shows
the importance of both aspects of our theory-based priors: a structured rep-
resentation of how categories in a domain are related and a probabilistic
model describing how properties are distributed over that relational struc-
ture. The strict-tree model incorporates an appropriate taxonomic structure
over categories but lacks a sufficiently flexible model for how properties are
distributed. The raw-feature model allows a more flexible prior distribution
for properties, but lacking a structured model of how categories are related,
it is limited to generalizing new, partially observed properties strictly based
on the examples of familiar, fully observed properties. Only the prior in the
evolutionary model embodies both of these aspects in ways that faithfully
reflect real-world biology – a taxonomic structure over species categories and
a mutation process generating the distribution of properties over that tree –
and only the evolutionary model provides consistently strong quantitative fits
to people’s inductive judgments.
Turning now to the similarity-based models, Figure 7.3 shows that their
performance varies dramatically depending on how we define the measure
of similarity to the set of premise categories. MaxSim fits reasonably well,
somewhat worse than the evolutionary model on the two Osherson datasets
but comparably to the evolutionary model on the three Smith datasets. The
fits on the Osherson datasets are worse than those reported by Osherson et al.
(1990), who used direct human judgments of similarity as the basis for the
model rather than the similarity matrix S(F ) computed from people’s feature
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 193

ratings. We used the species-feature associations here in order to compare all

models (including the raw-feature model) on equal terms, but we have also
compared versions of the evolutionary model and the similarity-coverage
models using the similarity judgments of Osherson et al. (1990) to build the
taxonomic tree or compute MaxSim or SumSim scores (Kemp & Tenenbaum,
2003). In that setting, too, the evolutionary model consistently performs at
least as well as MaxSim.
SumSim is arguably the least successful model we tested. Its predictions
for the strengths of general arguments are either uncorrelated with or nega-
tively correlated with people’s judgments. Although the good performance of
MaxSim shows that a similarity-based model can describe people’s patterns
of inductive reasoning, the poor performance of SumSim calls into question
the explanatory value of similarity-based approaches. As mathematical ex-
pressions, the SumSim and MaxSim measures do not appear very different,
beyond the presence of a nonlinearity in the latter case. This nonlinearity
turns out to make all the difference; it is necessary for similarity-based mod-
els to predict diversity-based inductive reasoning. Because the total inductive
strength of an argument under SumSim is a linear function of the inductive
strength associated with each premise category, the model assigns highest
strength to arguments in which each of the premise categories would in-
dividually yield the strongest one-premise argument. This preference goes
against diversity, because the categories that make for the best single-premise
arguments tend to be quite similar to each other. For instance, unlike peo-
ple, SumSim assigns a higher strength to horses, cows, rhinos −→ mammals
than to horses, seals, squirrels −→ mammals, because the strength of a gen-
eralization from the individual premise categories horses, cows, or rhinos to
mammals are some of the strongest in the domain – significantly higher than
from less typical mammals such as seals or squirrels.
We can see why SumSim fails, but there is no principled a priori justification
within the similarity-based paradigm for adopting the nonlinear MaxSim
rather than the linear SumSim as a model of inductive generalization. The
SumSim measure has if anything greater precedent in other similarity-based
models of learning, categorization, and memory (Nosofsky, 1986; Kruschke,
1992; Hintzman et al., 1978). Osherson et al. (1990) do not attempt to justify
the preference for MaxSim – it just seems to fit people’s intuitions better in the
particular task of property induction. That is fine as far as a descriptive model
goes, but not very satisfying if one of our goals is to explain why people’s
intuitions work the way that they do.
In contrast, the theory-based Bayesian approach we have presented of-
fers a principled and rational explanation for why people make the judg-
ments that they do on these inductive reasoning tasks. People’s judgments are
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

194 J. B. Tenenbaum, C. Kemp, and P. Shafto

approximately optimal, with respect to our evolutionary model that combines

the inferential optimality of Bayesian principles with a prior based on how
the properties of natural species are distributed in the real world. This ratio-
nal approach can also explain why some Bayesian model variants fare better
than others. The raw-feature and strict-tree models yield substantially poorer
descriptive fits to human judgments, and each is based on a prior that ne-
glects a key principle of how natural categories and properties are structured
which the evolutionary model properly incorporates. Finally, our approach
could explain why the most successful similarity-based models of induction
work the way they do, and in particular, why they are based on the MaxSim
mechanism rather than the more standard SumSim operation. In Kemp and
Tenenbaum (2003) and Kemp, Griffiths, et al. (2004) we show that under
certain conditions, MaxSim (but not SumSim) provides an efficient heuristic
approximation to the ideal computations of the evolutionary Bayesian model.
Hence at some level of analysis, and under certain circumstances, human in-
ductive reasoning could be well described as a similarity-based computation –
but which similarity-based computation that is, and why it works the way
that it does, would still be best explained by an analysis in our theory-based
Bayesian framework.

5.2 Reasoning about Causally Transmitted Properties

A second set of studies was intended to evaluate the descriptive power of
the Bayesian causal-transmission model, our model for inductive reasoning
about diseases and other causally transmitted properties, and also to show how
different theory-based Bayesian models can be used to account for different
patterns of inductive reasoning that arise with different kinds of properties.
We work with data from experiments by Shafto, Kemp, Bonawitz, Co-
ley, and Tenenbaum (submitted) (see also Shafto et al., 2005). Participants
were first trained to memorize the structure of the food webs shown in Fig-
ure 7.4. They were also familiarized with the correct taxonomic groupings for
these species categories. After this initial training, participants were asked to
judge the strength of inductive arguments about one of two kinds of prop-
erties: a disease property (“has disease D”) or a genetic property (“has gene
XR-23”). All arguments had a single premise and a specific conclusion, and
the stimuli exhaustively explored all arguments of this form. That is, every
pair of categories appeared as a premise-conclusion pair in some argument.
Participants also provided pairwise similarity ratings between the species in
each food web, and we again used hierarchical clustering to recover repre-
sentations of people’s taxonomic trees for these categories. The recovered
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 195

(a) (b)

human

mako

dolphin tuna sandshark

herring
tuna
mako

sandshark
dolphin
human
kelp
herring

kelp

(c) (d)

lion

wolf wolverine
squirrel
woodchuck

wolverine

wolf
fox

lion

bobcat
woodchuck bobcat squirrel fox

figure 7.4. Multiple relational structures over the same domains of species. (a) A di-
rected network structure capturing food web relations for an “island” ecosystem. (b) A
rooted ultrametric tree capturing taxonomic relations among the same species. (c) A
directed network structure capturing food web relations for a “mammals” ecosystem.
(d) A rooted ultrametric tree capturing taxonomic relations among the same species.

taxonomies are shown in Figure 7.4. Free parameters for all models (the mu-
tation rate in the evolutionary model, the background and transmission rates
for the causal transmission model) were set to values that maximized the
models’ correlations with human judgments.
We hypothesized that participants would reason very differently about dis-
ease properties and genetic properties, and that these different patterns of
reasoning could be explained by theory-based Bayesian models using appro-
priately different theories to generate their priors. Specifically, we expected
that inductive inferences about disease properties could be well approximated
by the causal-transmission model, assuming that the network for causal trans-
mission corresponded to the food web learned by participants. We expected
that inferences about genetic properties could be modeled by the evolutionary
model, assuming that the taxonomic structure over species corresponded to
the tree we recovered from participants’ judgments. We also tested MaxSim
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

196 J. B. Tenenbaum, C. Kemp, and P. Shafto

r=0.83 r=0.33 r=0. 8 r=0.075

Causal
transmission

r=0.19 r=0.88 r=0.11 r=0.88

Evolutionary

r=0.34 r=0. 9 r=0.11 r=0.83

MaxSim

Island Island Mammal Mammal

Disease Gene Disease Gene
figure 7.5. Comparing models of induction with human judgments, for two kinds of
properties: disease properties and genetic properties. Both kinds of property-induction
tasks were studied for two different systems of species categories, an “Island” ecosystem
and a “Mammals” ecosystem, as shown in Figure 7.4. Plotting conventions here are the
same as in Figure 7.3.

and expected that it would perform similarly to the evolutionary model, as

we found for the previous set of studies.
Figure 7.5 shows that these predictions were confirmed. High correlations
were found between the causal-transmission model and judgments about dis-
ease properties, and between the evolutionary model and judgments about
genetic properties. Morever, we observed a double dissociation between prop-
erty types and model types. The causal-transmission model correlates weakly
or not at all with judgments about genetic properties, while there is no sig-
nificant correlation between the evolutionary model and judgments about
disease properties. This double dissociation is the clearest sign yet that when
our Bayesian models fit well, it is not simply because they are using sophisti-
cated general-purpose inference principles; their success depends crucially on
using a theory-generated prior that is appropriately matched to the domain
structure and inductive context.
The performance of MaxSim was also as hypothesized: highly correlated
with inductive judgments about genetic properties (like traditional “blank”
properties), but poorly correlated with judgments about disease proper-
ties when participants were familiar with relevant predator-prey relations.
This result should not be surprising. It is well-known that similarity-based
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 197

approaches have difficulty accounting for inductive reasoning beyond the

context of generic biological properties, when some other relevant knowl-
edge is available to people (Smith et al., 1993; Medin et al., 2003; Shafto &
Coley, 2003). The usual interpretation of similarity’s shortcomings expresses a
general pessimism about computational models of commonsense reasoning:
different kinds of properties or inductive contexts just call for fundamentally
different approaches to reasoning, and cognitive scientists should not hope to
be able to give a principled general-purpose account for all – or even a large
class – of our everyday inductive inferences.
Our theory-based Bayesian framework offers a more optimistic view. We
have separated out the general-purpose inferential mechanisms from the
context-specific knowledge that guides those mechanisms, and we have intro-
duced a potentially quite general way of modeling relevant aspects of contex-
tual knowledge in terms of relational structures over categories and stochastic
processes for generating distributions of properties over those structures. It is
at least a start towards explaining more precisely how induction in different
contexts operates, and how apparently quite different patterns of reason-
ing in different contexts or domains could be given a unifying and rigorous
explanatory account.

.       

Bayesian models have recently been shown to provide strong accounts of many
core inductive capacities, not just the property induction tasks we consider
here. Examples include Bayesian models of concept learning and catego-
rization (Anderson, 1991; Tenenbaum, 2000), similarity judgment (Shepard,
1987; Tenenbaum & Griffiths, 2001), rule induction (Oaksford & Chater,
1994; McKenzie & Mikkelsen, 2000), word learning (Tenenbaum & Xu, 2000;
Xu & Tenenbaum, in press), causal judgment and covariation assessment
(Anderson, 1990; McKenzie, 1994; Griffiths & Tenenbaum, 2005; Steyvers
et al., 2003), detection of suspicious coincidences (Griffiths & Tenenbaum,
2007) and predictions about the magnitudes or durations of everyday events
(Griffiths & Tenenbaum, 2006). Yet the generality of Bayesian approaches has
not gone unchallenged.
Over the last decade, Murphy and Ross (e.g., Murphy & Ross, 1994, this
volume) have explored a set of natural inductive prediction tasks for which
people’s judgments often do not seem to follow Bayesian principles. In their
studies, participants are presented with partial descriptions of objects or
individuals and asked to predict other aspects of these entities. Murphy and
Ross focus on cases where the partially described entity could belong to one of
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

198 J. B. Tenenbaum, C. Kemp, and P. Shafto

several mutually exclusive categories, each of which suggests different predic-

tions about the properties to be predicted. For instance, a person walking up
to a house could be either a real estate agent, a burglar, or a cable TV worker,
and knowing that the individual belonged to one of those categories would
make it seem more or less likely that he would examine the doors of the house.
Murphy and Ross find that participants often base their predictions on only
the single most probable category (given a partial entity description), rather
than averaging the predictions of all categories weighted by their probability,
as would be prescribed by a Bayesian analysis of prediction with category
uncertainty (Anderson, 1991).
At first glance, these results might appear to be in conflict with our proposed
Bayesian mechanism for property induction (in particular, Equations 5 and
6). However, important computational differences exist between our tasks and
analyses and those of Murphy and Ross. In our models, we posit that learners
implicitly consider a very large number of alternative hypotheses for how to
generalize a novel property. These hypotheses need not – and in most cases,
almost surely could not – be represented consciously and explicitly in peo-
ple’s minds. More fundamentally, the hypotheses in our models do not reflect
mutually exclusive alternatives for categorizing an entity but rather mutually
exclusive candidate extensions for a novel property. We use a Bayesian frame-
work to model learners’ uncertain beliefs about a property to be predicted –
what subset of entities the property holds for – rather than uncertain beliefs
about the categories that entities belong to. Each entity typically falls under
many such hypotheses for the extension of a novel property, and which hy-
potheses an entity falls under is not in doubt; that is given by the structure of
the learner’s domain theory. The uncertainty in our tasks comes from provid-
ing incomplete information about the property to be predicted – participants
are given only one or a few examples of entities that have the property –
rather than from providing incomplete information about the entities over
which predictions are to be made. Hence there is no direct tension between
the Murphy and Ross findings and our work.
It remains an open question why human reasoning appears to incorporate
uncertainty in a fully Bayesian fashion for some kinds of inductive inferences
but not others. Murphy and Ross (Chapter 8 in this volume) suggest several
possibilities, and they argue against explanations based on simple psycho-
logical processing factors such as the complexity or naturalness of the task.
Another possibility that should be explored is a rational analysis that takes
into account key computational differences in different inductive tasks – a
“rational metanalysis,” so to speak. Perhaps human reasoners are more likely
to consider multiple alternative hypotheses in prediction tasks where doing
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 199

so is more likely to lead to a substantial improvement in accuracy. Previ-

ous studies of Bayesian concept learning are consistent with this conjecture
(Tenenbaum, 2000), but much more systematic investigation is required to
assess the importance of this factor relative to others.

.    

Conventional models of induction focus on the processes of inference rather
than the knowledge that supports those inferences. Inference mechanisms are
typically given fully explicit and mathematical treatments, while knowledge
representations are either ignored or boiled down to place-holders like a simi-
larity metric or a set of features that at best only scratch the surface of people’s
intuitive theories. This focus probably stems from the natural modeler’s drive
towards elegance, generality, and tractability: the knowledge that supports
induction is likely to be messy, complex, and specific to particular domains
and contexts, while there is some hope that one or a few simple inference
principles might yield insight across many domains and contexts. Yet this
approach limits the descriptive power and explanatory depth of the models
we can build, and it assigns a second-class status to crucial questions about
the form and structure of knowledge. Our goal here is to return questions of
knowledge to their appropriate central place in the study of inductive infer-
ence, by showing how they can be addressed rigorously and profitably within
a formal Bayesian framework.
Every real-world inference is embedded in some context, and understand-
ing how these different contexts work is critical to understanding real-world
induction. We have argued that different contexts trigger different intuitive
theories, that relevant aspects of these theories can be modeled as gen-
erative models for prior probability distributions in a Bayesian reasoning
framework, and that the resulting theory-based Bayesian models can explain
patterns of induction across different contexts. We described simple theo-
ries for generating priors in two different inductive contexts, a theory for
generic biological properties such as genetic, anatomical, or physiological
features of species, and a theory for causally transmitted properties such as
diseases, nutrients, or knowledge states. We showed that Bayesian models
based on a prior generated by each theory could predict people’s inductive
judgments for properties in the appropriate context, but not inappropriate
contexts.
Intriguingly, both of these generative theories were based on principles
analogous to those used by scientists to model analogous phenomena in
real biological settings. In showing their success as descriptive models of
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

200 J. B. Tenenbaum, C. Kemp, and P. Shafto

people’s intuitive judgments, we begin to provide an explanation for people’s

remarkable ability to make successful inductive leaps in the real world as the
product of rational inference mechanisms operating under the guidance of
a domain theory that accurately reflects the true underlying structure of the
environment.
Of course, just specifying these two theory-based models is far from giving
a complete account of human inductive reasoning. It will be a challenging
long-term project to characterize the space of theories that people can draw
on and the processes by which they are acquired and selected for use in par-
ticular contexts (see Shafto, Coley, & Vitkin, Chapter 5, this volume). Perhaps
the most immediate gap in our model is that we have not specified how to
decide which theory is appropriate for a given argument. Making this deci-
sion automatically will require a semantic module that knows, for example,
that words like “hormone” and “gene” are related to the generic biological
theory, and words like “disease” and “toxin” are related to the theory of causal
transmission. How best to integrate some form of this semantic knowledge
with our existing models of inductive reasoning is an open question.
We have discussed theories that account for inductive reasoning in two
contexts, but it is natural and necessary to add more. Some of our ongoing
work is directed towards developing and testing models beyond the tree- and
network-based models described here. For example, in Kemp and Tenen-
baum (submitted), we show how reasoning about properties such as “can
bite through wire” can be modeled as Bayesian inference over a linear struc-
ture representing a dimension of strength, rather than a tree or a network
representing taxonomic or foodweb relations (Smith et al., 1993). In other
work, we are extending our approach to account for how multiple knowl-
edge structures can interact to guide property induction; in particular, we
are looking at interactions between networks of causal relations among prop-
erties and tree structures of taxonomic relations among species. Finally, in
addition to broadening the scope of theory-based Bayesian models, we have
begun more fundamental investigations into how these probabilistic theories
can themselves be learned from experience in the world (Kemp, Perfors, &
Tenenbaum, 2004) and how learners can infer the appropriate number and
complexity of knowledge structures that best characterize a domain of cat-
egories and properties (Shafto et al., 2006). These problems are as hard to
solve as they are important. Yet we believe we are in a position to make real
progress on them – to develop genuine insights into the origins and nature
of common sense – using models that combine the principles of Bayesian
inference with structured knowledge representations of increasing richness
and sophistication.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 201


This work has benefited from the insights and contributions of many col-
leagues, in particular Tom Griffiths, Neville Sanjana, and Sean Stromsten. Liz
Bonawitz and John Coley were instrumental collaborators on the food web
studies. JBT was supported by the Paul E. Newton Career Development Chair.
CK was supported by the Albert Memorial Fellowship. We thank Bob Rehder
for helpful comments on an earlier draft of the chapter.

References
Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ:
Erlbaum.
Anderson, J. R.(1991). The adaptive nature of human categorization. Psycho-
logical Review, 98(3), 409–429.
Ashby, F. G., & Alfonso-Reese, L. A. (1995). Categorization as probability
density estimation. Journal of Mathematical Psychology, 39, 216–233.
Atran, S. (1995). Classifying nature across cultures. In E. E. Smith & D. N.
Osherson (Eds.), An invitation to cognitive science (Vol. 3) (pp. 131–174).
MIT Press.
Atran, S. (1998). Folkbiology and the anthropology of science: Cognitive
universals and cultural particulars. Behavioral and Brain Sciences, 21, 547–
609.
Carey, S.(1985). Conceptual change in childhood. Cambridge, MA: MIT Press.
Gelman, S., & Markman, E. (1986). Categories and induction in young chil-
dren. Cognition, 23(3), 183–209.
Goodman, N.(1972). Seven strictures on similarity. In Problems and projects.
Indiana: Bobbs-Merrill.
Gopnik, A., & Meltzoff, A. (1997). Words, thoughts, and theories. Cambridge,
MA: MIT Press.
Gopnik, A., & Schulz, L. (2004). Mechanisms of theory formation in young
children. Trends in Cognitive Sciences, 8, 371–377.
Griffiths, T. L., & Tenenbaum, J. B. (2005). Structure and strength in causal
induction. Cognitive Psychology, 51, 354–384.
Griffiths, T. L., & Tenenbaum, J. B. (2007). From mere coincidences to
meaningful discoveries. Cognition.
Griffiths, T. L., & Tenenbaum, J. B. (2006). Optimal predictions in everyday
cognition. Psychological Science, 17(9), 767–773.
Heit, E. (1998). A Bayesian analysis of some forms of inductive reasoning. In
M. Oaksford & N. Chater (Eds.), Rational models of cognition (pp. 248–274).
Oxford University Press.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

202 J. B. Tenenbaum, C. Kemp, and P. Shafto

Heit, E., & Rubinstein, J. (1994). Similarity and property effects in induc-
tive reasoning. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 20, 411–422.
Hintzman, D. L., Asher, S. J., & Stern, L. D. (1978). Incidental retrieval and
memory for coincidences. In M. M. Gruneberg, P. E. Morris, & R. N. Sykes
(Eds.), Practical aspects of memory (pp. 61–68). New York: Academic Press.
Kemp, C., Griffiths, T. L., Stromsten, S., & Tenenbaum, J. B. (2004). Semi-
supervised learning with trees. In Advances in neural processing systems 16,
pp. 257–264. Cambridge, MA: MIT Press.
Kemp, C., Perfors, A., & Tenenbaum, J. B.(2004). Learning domain structures.
In Proceedings of the 26th annual conference of the Cognitive Science Society.
Mahwah, NJ: Erlbaum.
Kemp, C., & Tenenbaum, J. B.(2003). Theory-based induction. In Proceedings
of the 25th annual conference of the Cognitive Science Society.
Kemp, C., & Tenenbaum, J. B. (submitted). Structured statistical models of
inductive reasoning.
Kruschke, J. K. (1992). Alcove: An exemplar-based connectionist model of
category learning. Psychological Review, 99, 22–44.
Markman, E.(1989). Naming and categorization in children. Cambridge, MA:
MIT Press.
Marr, D. (1982). Vision. New York: W. H. Freeman.
McCloskey, M., Caramazza, A., & Green, B. (1980). Curvilinear motion in
the absence of external forces: Naive beliefs about the motion of objects.
Science, 210(4474), 1139–1141.
McKenzie, C. R. M. (1994). The accuracy of intuitive judgment strategies:
Covariation assessment and Bayesian inference. Cognitive Psychology, 26,
209–239.
McKenzie, C. R. M., & Mikkelsen, L. A. (2000). The psychological side of
Hempel’s paradox of confirmation. Psychonomic Bulletin and Review, 7,
360–366.
Medin, D. L., Coley, J. D., Storms, G., & Hayes, B. K. (2003). A relevance
theory of induction. Psychological Bulletin and Review, 10, 517–532.
Mitchell, T. M. (1997). Machine learning. McGraw-Hill.
Murphy, G. L. (1993). Theories and concept formation. In I. V. Meche-
len, J. Hampton, R. Michalski, & P. Theuns (Eds.), Categories and con-
cepts: Theoretical views and inductive data analysis. London: Academic
Press.
Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual
coherence. Psychological Review, 92, 289–316.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

Theory-Based Bayesian Models of Inductive Reasoning 203

Murphy, G. L., & Ross, B. H. (1994). Predictions from uncertain categoriza-

tions. Cognitive Psychology, 27, 148–193.
Nosofsky, R. M. (1986). Attention, similarity, and the identification-cate-
gorization relationship. Journal of Experimental Psychology: General, 115,
39–57.
Oaksford, M., & Chater, N. (1994). A rational analysis of the selection task as
optimal data selection. Psychological Review, 101, 608–631.
Oaksford, M., & Chater, N. (Eds.) (1998). Rational models of cognition. New
York: Oxford University Press.
Osherson, D., Smith, E. E., Wilkie, O., López, A., & Shafir, E.(1990). Category-
based induction. Psychological Review, 97(2), 185–200.
Rips, L. J. (1975). Inductive judgements about natural categories. Journal of
Verbal Learning and Verbal Behavior, 14, 665–681.
Rogers, T. T., & McClelland, J. L. (2004). Semantic cognition: A parallel dis-
tributed processing approach. Cambridge, MA: MIT Press.
Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B. Lloyd
(Eds.), Cognition and categorization (pp. 27–48). Hillsdale, NJ: Lawrence
Erlbaum.
Shafto, P., & Coley, J. D.(2003). Development of categorization and reasoning
in the natural world: Novices to experts, naive similarity to ecological
knowledge. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 29, 641–649.
Shafto, P., Kemp, C., Baraff, E., Coley, J. D., & Tenenbaum, J. B. (2005).
Context-sensitive induction. In Proceedings of the 27th annual conference
of the Cognitive Science Society. Mahwah, NJ: Erlbaum.
Shafto, P., Kemp, C., Mansinghka, V., Gordon, M., & Tenenbaum, J. B.
(2006). Learning cross-cutting systems of categories. In Proceedings of
the 28th annual conference of the Cognitive Science Society. Mahwah, NJ:
Erlbaum.
Shafto, P., Kemp, C., Bonawitz, E. B., Coley, J. D. & Tenenbaum, J. B.(submit-
ted). Inductive reasoning about causally transmitted properties.
Shepard, R. N. (1987). Towards a universal law of generalization for psycho-
logical science. Science, 237, 1317–1323.
Silverman, B. (1986). Density estimation for statistics and data analysis.
London: Chapman and Hall.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25,
213–280.
Smith, E. E., Safir, E., & Osherson, D. (1993). Similarity, plausibility, and
judgements of probability. Cognition, 49, 67–96.
P1: JZP
0521672443c07 CUFX144-Feeney 0 521 85648 5 July 20, 2007 8:32

204 J. B. Tenenbaum, C. Kemp, and P. Shafto

Steyvers, M., Tenenbaum, J. B., Wagenmakers, E. J., & Blum, B. (2003). In-
ferring causal networks from observations and interventions. Cognitive
Science, 27, 453–489.
Tenenbaum, J. B. (2000). Rules and similarity in concept learning. In S. A.
Soller, T. K. Leen, & K. R. Müller (Eds.), Advances in neural processing
systems 12 (pp. 59–65). Cambridge, MA: MIT Press.
Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalization, similarity, and
Bayesian inference. Behavioral and Brain Sciences, 24, 629–641.
Tenenbaum, J. B., Griffiths, T. L., & Kemp, C. (2006). Theory-based Bayesian
models of inductive learning and reasoning. Trends in Cognitive Sciences,
10(7), 309–318.
Tenenbaum, J. B., Griffiths, T. L., & Niyogi, S. (in press). Intuitive theories as
grammars for causal inference. In A. Gopnik & L. Schulz (Eds.), Causal
learning: Psychology, philosophy, and computation. Oxford: Oxford Univer-
sity Press.
Tenenbaum, J. B., & Xu, F. (2000). Word learning as Bayesian inference. In
Proceedings of the 22nd annual conference of the Cognitive Science Society
(pp. 517–522). Hillsdale, NJ: Erlbaum.
Wellman, H. M., & Gelman, S. A. (1992). Cognitive development: Founda-
tional theories of core domains. Annual Review of Psychology, 43, 337–375.
Xu, F., & Tenenbaum, J. B.(in press). Sensitivity to sampling in Bayesian word
learning. Developmental Science.
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories

in Category-Based Induction
Gregory L. Murphy and Brian H. Ross

Uncertainty is a basic fact of life. Despite uncertainty, people must make pre-
dictions about the world. Will the car you are considering buying be reliable?
Will you like the food you order? When you see an animal in the woods, what
should you do? One source of information that reduces uncertainty is cate-
gory membership. Although all Toyota Camrys are not exactly the same, they
are similar enough that you can predict with some confidence that the new
Camry you are considering will be reliable. Kansas City style barbecue ribs
are not identical, but they taste more similar to one another than they do to
roast chicken or “tofu surprise.” Knowing the category of an entity therefore
serves to reduce the uncertainty associated with it, and the category reduces
uncertainty to the degree that the category members are uniform with respect
to the prediction you want to make. This category-based induction is one of
the main ways that categories are useful to us in everyday life.
Unfortunately, this reduction of uncertainty is limited by the uncertainty
of category membership itself. If you go to the Toyota dealership and order
a Camry, you can be close to 100% sure that your new car is going to be a
Camry. But in many other situations, you cannot be 100% sure. Is the animal
you see in the woods a squirrel, or was your glimpse too fleeting for you to
rule out the possibility it was really a raccoon? Is the person you are dating
honest? You know how you would react to a squirrel or an honest person, but
when you are not sure if these things are in their respective categories, then
you cannot be sure of their other properties.
The central issue we address is whether (and, if so, how and when) multiple
categories are used in induction. If you are uncertain whether the animal is
a squirrel or a raccoon, do you make a different prediction than if you are
sure that the animal is a squirrel? Do you use what you know about both
categories, or only one? For items that are not uncertain, but are members of
multiple categories, the question is whether only one or a number of those
205
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

206 G. L. Murphy and B. H. Ross

categories are involved in the induction. If Alice is a physician and a mother,

do you use your knowledge of both categories in making a prediction about
her, or only one? If only one, how do you choose which one?
In a series of studies, we have looked at this problem from a number of
angles, using naturalistic and artificial categories, manipulating uncertainty
in a number of ways. The main questions we have asked are the following.
(1) Do people change their inductions based on the certainty of an object’s
classification? (2) When people are not sure which of two categories an object
is in, do they use both categories to make inductions? (3) What contextual
or task variables moderate the effect of uncertainty? In particular, do people
treat uncertainty differently depending on just how they make the judgment?
(4) What happens when an entity is a member of multiple categories? Are all
categories used, or only one of them?
This chapter presents an overview of our work in this area, attempting to
integrate our findings from various paradigms with work of other researchers
to provide a more complete account of how multiple categories are used in
induction. These issues are important, because they get to the heart of what
categories are used for. We need to know the categories of objects so that
we can make predictions about properties that are not otherwise available to
us. If categories could tell us only about visible properties of an object, they
would not add any information to help direct our behavior towards the object.
Thus, induction is central to category use. Finally, we believe that our results
speak to important issues of the limits of reasoning. When people fail to use
multiple categories, is it because they are unable to do so, because they don’t
understand the need to do so, or because of some performance problem? The
answer will speak to issues of inductive processing more generally.

  :   should

   
In an important analysis, Anderson (1991) proposed that uncertainty about
a category should be taken into account when making inductions about it.
If you are not certain whether the strange fruit in front of you is an apple
or a pear, then you should not be certain in predicting what it tastes like.
Indeed, your prediction about the taste should be proportional to your belief
in these two categories. If you are 90% sure it is an apple, then you can make a
strong prediction that it will have an apple taste, but if you are only 55% sure,
then you can make only a weak prediction. Anderson therefore suggested that
when you make a prediction about an object, you access the categories that it
might be in, and you assign it the properties of each category proportionally
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 207

to its likelihood of being in that category (for more detail, see Anderson, 1991;
Murphy & Ross, 1994). For example, if you are trying to predict whether this
fruit will be mushy, you would estimate the likelihood that apples are mushy
and that pears are mushy (and other fruits if they are also possible) and
combine these likelihoods. Thus, this procedure entails that a category-based
induction is made only to the degree that one is certain about the category;
and if multiple categories are possible, then each is used proportionally to its
likelihood.
This model appears normatively correct to us, and it is also consistent with
later Bayesian approaches to classification and prediction (e.g., Heit, 2000;
Tenenbaum 2000). However, following the general rule that anything that is
normatively correct people do not actually do, we might well be skeptical of
this proposal. And in fact, we have found consistent evidence that people do
not in fact follow this rule, which we will summarize next.

Evidence from Natural Categories

Some of our research showing single-category induction has been conducted
with real-world categories, taking advantage of people’s prior knowledge
of categories and properties. In one kind of experiment, we constructed
scenarios in which entities had an uncertain classification. For example, we
described a case in which an unfamiliar person walking up the driveway was
likely to be a real estate agent (the primary category). (All the experiments
reported here had multiple categories and induction properties, but we will
mention only one example each, for the sake of exposition.) Depending on
the condition, it was also possible, though less likely, that the person could
be one of the following secondary categories – a burglar or cable TV worker.
We then asked subjects who they thought the person most likely was and
how likely they thought this person was to examine the doors of the house.
Pretesting had shown that our subject population thinks that real estate agents
and burglars both will pay attention to the doors of a house, but that a cable
worker will not. Given the uncertainty about the person’s identity, how should
people make this prediction? If they are using both categories, they should
be more certain that the unknown person will examine the doors when
the secondary category is the burglar than when it is the cable TV worker.
Because burglars confirm the prediction but cable TV workers do not, people’s
predictions should differ in the two cases. However, they do not in fact differ
(Malt, Ross, & Murphy, 1995), even when considerable emphasis is laid on
the uncertainty of the person’s (or object’s) classification (Ross & Murphy,
1996).
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

208 G. L. Murphy and B. H. Ross

A different kind of experiment varied whether the initial categorization was

described as being certain or uncertain. When it is certain, only one category
should be used in the induction; when it is uncertain, both possible categories
should be used. However, this manipulation also had no effect (Malt et al.,
1995).
More recently, we used an analogous manipulation with pictures that were
ambiguous between two different categories. By morphing two images, we
were able to construct items like a peach-apple or a robin-blackbird (Murphy
& Ross, 2005). Subjects were told to imagine that they were exploring an
abandoned island to survey its flora and fauna. They were able to send samples
of the objects they found to a lab for a quick analysis that would identify the
most likely categories. The category list they were given mentioned either one
or both of the categories. For example, for the peach-apple, they would be told
either that it might be a peach or cherry or raspberry – or else it might be either
a peach or apple or cherry or raspberry. Subjects were more certain about
the object’s identity in the first case than in the second (when the alternative
category was mentioned as a likely possibility), yet this manipulation had no
reliable effect on people’s inductions.
Readers might well be thinking at this point, “Apparently, these guys cannot
get any experiment to work. Everything they report is a null result.” However,
that is fortunately not the case. In the same experiments a number of reliable
results showed that subjects were carefully reading the scenarios and were
paying attention to the category information. However, rather than repeat
those results here (which are not germane to the main theoretical issues), we
will simply mention now that we later discuss results in which people do use
multiple categories in induction, even with similar materials and questions.
Thus, the null results of these studies are not due to dull subjects or insensi-
tive measures, because positive results are possible, under some very specific
conditions.

Evidence from Artificial Categories

Natural categories have the advantage that they are, well, natural, but they do
not allow a precise manipulation of the probabilities of features within the
categories. For better control, we have carried out many experiments using
categories of geometric shapes (or other artificial stimuli). People first studied
the categories. We then told subjects about a feature of another geometric
stimulus (e.g., its shape) and asked them what category it most likely was
in; then we asked them to make a prediction about one of its other features
(e.g., its color or pattern). They also rated the probability that their answers
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 209

were correct, a measure of inductive strength. Because subjects were allowed

to examine the categories even while answering the questions, we were able
to eliminate problems of learning and memory that might have influenced
the results. Instead, people could directly see that the item’s classification was
uncertain (e.g., many triangles, but by no means all, were in one particular
category), and they could also examine the distribution of features in order
to make predictions. Thus, if they chose to use multiple categories to answer
their questions, they could do so without great difficulty. But they did not.
For example, imagine that one category had a lot of triangles and also
mostly solid black figures (see Figure 8.1). Depending on condition, other
categories with some triangles also had some black items or else did not.
In the first condition, people should be more confident in predicting that a
new triangle would be black than in the second. That is, across secondary
categories, there is evidence for the solid black feature in one condition but
not the other. Yet, these two conditions did not differ (Murphy & Ross, 1994).
Again, this null result is not simply a matter of people’s inattention to the
categories. For example, when we varied the distribution of features within the
critical category, it strongly influenced people’s inductions (Murphy & Ross,
1994, 2005). That is, varying the number of solid black items in the triangle’s
most likely category changed people’s judgments; varying the number of black
items in other categories had no effect. Clearly, people were using one category
rather than all the possible categories.
In a somewhat different design, we kept constant the predicted features
but varied the certainty of the main category. For example, we reduced the
number of triangles in the critical category so that people would be less certain
that the triangle was in that category. Although this did in fact change their
categorization certainty, it did not change their predictions. That is, people
lowered their ratings that the triangle was in the category but did not lower
their ratings that the triangle was black – even though this prediction should
depend on the triangle being in the critical category (Murphy & Ross, 2005).
As a final note, researchers who hear about these studies often wonder
whether asking what category the item is in is critical for obtaining these
results. Perhaps explicitly choosing a category leads people to focus on that
category. To address that question, one of our early experiments replicated
the Figure 8.1 experiment but asked only about the property (Murphy &
Ross, 1994, Expt. 2). This change did lead to an overall decrease in giving
the property from the critical category (i.e., black), but the distribution of
the number of black items in the other categories still had no effect on their
certainty. Thus, the initial question does seem to get people to use the critical
category more often, but does not influence whether they are likely to use more
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

210 G. L. Murphy and B. H. Ross

Liz Mary

Jessica Rachel
figure 8.1. An example similar to the categories used in Murphy and Ross (1994),
illustrating the main comparison of that set of studies. In one condition, subjects would
be asked to predict the shading of a new triangle drawn by one of the children. Triangles
were most often drawn by Jessica, so Jessica was the target category. Note that most of
her drawings were black. Liz and Rachel also each drew a triangle, and they also drew
black items. Thus, information from these categories might increase one’s estimate that
a new triangle would be black (if one used multiple categories). The control question
might be to make a prediction about a new square. Rachel is most likely to have drawn a
square, and most of her figures are empty. Mary and Jessica also drew squares but have no
empty figures, and so the secondary categories do not strengthen this induction. Thus,
if people were using multiple categories, their inductions of the first example (triangle-
solid black) should be stronger than those in the second example (square-empty). But
they are not.

than one category. Similarly, Murphy and Ross (2005, Expt. 3) repeated their
experiment asking the categorization question after the induction question,
so that it would not influence the induction judgment. The pattern of results
was identical to the reverse order of questions.

      

Together, these experiments give surprisingly convincing evidence that people
do not use multiple categories in making inductions. Even though we found
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 211

that they were not certain in their initial categorization (either through rat-
ings or else by observing considerable variance in that categorization across
subjects), this uncertainty did not lower the certainty of their induction, nor
did it make them attend to other categories that the item might have been
in. One particularly striking aspect of these results is that on one question
people might admit that they are not sure what category an item is in but
on the next question they do not show any sign of this uncertainty on the
induction using this category. It is as if these two judgments take place in
compartmentalized phases. First one decides what category the object is in.
Perhaps it is not certain, but the object seems most likely to be a robin. Then
one makes an induction about the object, given that it is actually a robin.
In this second phase, the uncertainty about the initial categorization has no
effect – it is as if the uncertain categorization has been walled off and is now
treated as if it were certain.
Why is this? One possibility is that people simply cannot carry out the
computations necessary to use category uncertainty when making inductions.
It is possible that keeping track of different categories, their properties, and
their uncertainty is simply more than working memory can do. A second
possibility is that people do not recognize the principle involved. That is, they
may not realize that if you are uncertain about an object’s identity, you should
reduce the certainty of predictions about the object. This might be consistent
with findings from the decision-making literature such as the lack of using
base rates in making predictions. Just as people do not attend to the frequency
of green cabs when deciding how likely it is that a green cab was involved in
an accident (Tversky & Kahneman, 1980), they may not attend to how likely
the object is to be a robin or blackbird when making a prediction about it.
However, neither of these possibilities has survived empirical testing. We
have been able to find certain conditions in which people are sensitive to
category uncertainty, which means that they must have the computational
ability to use it in making decisions, and that they must understand the
importance of using it, at least in the abstract. Ross and Murphy (1996)
constructed a new induction question to test the limits of the earlier results.
In our original work, for cases in which the real estate agent was the primary
category, we used questions like “What is the probability that the man walking
up the driveway will pay attention to the sturdiness of the doors on the house?”
This is something that both real estate agents and burglars would do and cable
TV workers would not (according to a norming study), yet the possibility that
the person might be a burglar instead of a cable TV worker did not raise the
rated probability of this property. In the new experiment, we made questions
that were directly associated to the secondary category (a burglar, in this
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

212 G. L. Murphy and B. H. Ross

case), for example, “What is the probability that the man walking up the
driveway will try to find out whether the windows are locked?” This activity is
directly associated with burglars, whereas cable TV workers have little interest
in windows. Our hypothesis was that in reading this question, subjects would
be reminded of the possibility that the man might be a burglar. They might
then increase their prediction (compared to the cable TV worker condition)
as a result of being reminded of this category. And this is what happened: The
predictions were 21% higher when a burglar was the secondary category than
when a cable TV worker was.
Apparently, if the induction question makes people think of the secondary
category, they can use information about that category in making their predic-
tion. Perhaps using multiple categories in that question would raise people’s
awareness of the importance of considering secondary categories in general,
and so they might use both categories to answer subsequent questions. But in
fact, that did not happen. In two experiments, we found that when people an-
swered the induction questions about properties associated to the secondary
category, they seemed to be using two categories, but on subsequent ques-
tions about the same scenario (with unassociated properties), they reverted
to using a single category (Ross & Murphy, 1996).
These results suggest the following, slightly informal explanation: If you hit
people over the head with the fact that there are two possible categories, they
can and do use both categories; if you don’t hit them over the head, then they
do not. This pattern of results shows that people do have the computational
ability to deal with two categories (at least) and that they also can recognize
the principle that both categories are relevant. But this ability and knowledge
are put to work only in fairly circumscribed situations. As will be seen, the
critical element in whether people do or do not use multiple categories seems
to be whether the secondary category is brought to their attention while they
are answering the question. If it is not, the general principle of hedging one’s
bets in uncertain cases (Anderson, 1991) does not seem strongly enough
represented to cause people to use multiple categories.
Once again, we are left with the observation that classification seems to
be a largely separate process from induction, even though they are logically
strongly connected. One possible reason for this is that in many cases, the
two are in fact done separately. We may classify entities without making an
interesting induction about them: the squirrels on the lawn as we walk to
work, the cars on the street, the trees in front of the house, the buildings, the
people, the chairs in our office, and so on. We often do not need to make
any predictions about the squirrels, trees, or chairs. And when we do make
predictions, the prediction is often not category-based (e.g., to decide how to
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 213

avoid something on your path, it is often more useful to know the object’s size
than its category). Furthermore, in many cases when we do make inductions,
the category is certain. If we are worried about a dog prowling around the
yard, we are likely not worried about whether it is a dog – that may be obvious.
We are worried about what it might do. If we try to predict how our doctor
will treat our condition, we do not worry about whether she is or is not really
a doctor. We do not know exactly how often people are confronted with the
situation in which something has an uncertain categorization and yet they
attempt to make a prediction about it. Although we believe that induction with
categorical uncertainty is a common event in an absolute sense, its relative
frequency may be dwarfed by thousands of occasions in which categorization
is clear or we do not draw an induction. Thus, when induction must be
done when categories are uncertain, people may carry over habits from their
experiences with certain categories.
We may be a bit better at dealing with uncertainty when we know almost
nothing about the entity’s categorization. For example, if we stop someone on
the street to ask a question, we can have only a very vague idea of the person’s
profession or religion or hobbies. Under such circumstances, people take
into account the uncertainty. In an experiment using the geometric figure
categories, we placed one triangle (say) in all four categories and then asked
people to make predictions about a new triangle. We obtained evidence that
people used all four categories in this situation (see Murphy & Ross, 1994, for
the details). We assume that the reason for this is that people have no preferred
category in this situation and so cannot engage in the two-stage process we
have described. That is, they could literally see that the triangle was equally
likely to be in each of four categories and thus had no basis for picking one
of the categories, and so they considered all four of them. However, this
situation would be quite rare in real life, and we suspect that when there is
only very weak evidence for a number of categories, people simply do not
engage in category-based induction. That is, even if you believed profession is
a good predictor of helpfulness, if you cannot tell what profession a stranger
is likely to be in, you may not make a guess about his profession to predict his
willingness to help you. Instead, you will simply not make the induction or
will try to find another basis on which to make it.
However, even in situations where there is no clear reason that a category-
based induction would help, people may use a single category that is brought
to mind. Lagnado and Shanks (2004) described an interesting situation like
this, using hierarchically organized categories of persons. They had two super-
ordinate categories of readers of different types of newspapers, which we will
just call A and B, each with two subcategories of particular newspapers. There
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

214 G. L. Murphy and B. H. Ross

were more readers of the newspapers in category A than of the newspapers in

B, but one subcategory of B – let’s call it B1 – was the most frequently read
paper. Lagnado and Shanks taught subjects about the different categories by
exposing them to exemplars and teaching them a property of each person
(what party he or she voted for).
In their test problems, Lagnado and Shanks asked their subjects whether
a randomly chosen person from this population was more likely to read the
type of paper in A or B. Subjects generally guessed A, because of its greater
frequency. Then they made a prediction about what party this random person
voted for. They generally chose the party most voted for by A readers. On
other trials, subjects were asked which of the subcategories a randomly chosen
person was in (i.e., which particular paper they read). They usually guessed
B1, which was the most frequently read paper subcategory, and then they
predicted that the random person would choose the party most voted for by
members of B1. However, this prediction was different from that made by
the first set of subjects, because B1 members were somewhat unusual in their
voting choices. What is very peculiar about this is that both subjects were
given the exact same information about the person’s category, namely, none:
They were told that the person was randomly chosen and were not told any of
the person’s properties. However, by first guessing that the person might be
in A or might be in B1, the subject then answered the second question using
this likely category as the basis for the answer. That is, simply by asking (not
telling) which specific or general category the person was in, Lagnado and
Shanks changed their subjects’ answers to the subsequent induction question.
What is striking about Lagnado and Shanks’s demonstration is that with
truly minimal evidence for what category something is in, subjects still used
that category for making their predictions. Thus, their results are a striking
demonstration that once people choose a category for an object, the uncer-
tainty they experienced in classifying it has no further effect on induction.
This demonstration is important for evaluating the possibility we raised
earlier that people may not understand the general principle of using multiple
categories when categorization is uncertain. In our studies, subjects often
rated category membership, and we could verify that it was uncertain; in
Lagnado and Shanks’s study, membership was clearly uncertain, because no
information was given about the person at all. Thus, if people generally used
this principle, they should have done so in these cases. Nonetheless, some of
our other studies have shown that people do use multiple categories (Ross
& Murphy, 1996). It may be that people “understand” the principle in some
sense, but they do not actually put it to use in most cases. This is reminiscent
of findings in cognitive development that children can put a principle into
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 215

action when the task is easy or the application is obvious, but they do not
follow the principle in more difficult conditions (e.g., Baillargeon, 1998). In
our case, when the induction property itself brings up the alternative category
(and when the subject has already identified the other category as most likely),
both categories are together in working memory, and the person may realize
the importance of using both of them. But when the question does not
bring to mind an alternative category, only the selected category is in working
memory, and the principle itself is not accessed. Presumably, people who make
predictions for a living, such as weather forecasters, poker players, and the
like, have more practice in dealing with their uncertainty and may be better at
answering these questions. (Note to granting agencies: Anyone who would like
to fund a case study of induction in poker playing, please contact the authors.)

-
Multiple categories are important when an object’s identity is uncertain. But
multiple categories may be important in a different situation as well, namely,
when an object has multiple “identities.” We are not referring to a questionable
personality disorder but to the fact that categories are cross-cutting. Rather
than a single hierarchy in which every entity fits into exactly one category at
a given level of specificity, categories are a hodge-podge of groupings with
different bases, overlapping to various degrees. For example, within categories
pertaining to people, there are classifications based on personality, size, eth-
nicity, profession, gender, sexual orientation, areas of interest and expertise,
age, educational level, financial status, and so on. These categories are not
all independent, but their bases may be quite different (e.g., gender vs. pro-
fession), and so they change the way that one thinks about the person being
categorized. A large, outgoing forty-two-year-old Asian-American female ac-
countant who is married, likes sewing and football, has a B.A. and M.B.A, and
is upper-middle class is a member of a few dozen person categories. If you have
to make a prediction about this person, how will you decide which category
to use? Will you use just one? If so, how do you choose which category to use?
Or will you use some or all of these categories? Research in social psychology
has addressed this problem specifically for social categories and stereotypes
(e.g., Macrae, Bodenhausen, & Milne, 1995; Nelson & Miller, 1995; Patalano,
Chin-Parker, & Ross, 2006), and we have investigated it for object categories.
Cross-classification raises issues that are very analogous to those we con-
sidered in the multiple-categorization problems above. However, here the
problem is not of uncertainty but of multiplicity. That is, we are not uncer-
tain whether the person is female or an accountant – we know she is both.
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

216 G. L. Murphy and B. H. Ross

Nonetheless, that does not mean that people will use both categories in making
a prediction. If one category is more relevant than the others, then people
tend to use that as the basis for prediction. For example, Heit and Rubinstein
(1994) showed that when making inductions about a biological property from
one animal to another, people had more confidence when the two animals
were biologically related (e.g., were both mammals). However, when making
inductions about a behavioral property, they were more confident when the
two animals shared an ecological niche (e.g., were both ocean dwellers; also
see Ross & Murphy, 1999). This selective use of relevant categories is not
limited to adults. Kalish and Gelman (1992) found that preschool children
made inductions about the physical properties of a new object based on its
physical make-up (e.g., being made of glass) but made functional inductions
based on the object’s artifact category (e.g., being a knife) (though see Nguyen
& Murphy, 2003).
This research shows that people make inductions at least partly based on
their beliefs about whether the category supports induction of the property
(see also Proffitt, Medin, & Coley, 2000). Thus, these beliefs provide one way
of narrowing down the many categories that an object or person might belong
to. Although Cheri may be a mother, this does not tell you what her highest
educational degree is or whether she is wealthy. Thus, categories like mother
may not be used in making these inductions about Cheri, whereas profession
might be used. What is not so clear from these studies is what people do when
multiple categories are relevant and potentially useful. The above studies were
specifically designed so that different categories were differentially relevant,
and so they could not answer this question. Therefore, we addressed this
issue in a manner analogous to our previous work on multiple-category
induction.
The problem with studying cross-classification is that we cannot as eas-
ily manipulate the categories experimentally, since learning complex cross-
cutting categories would be rather difficult. Instead, we used real-world cat-
egories of food items, because we had already found that foods are very
cross-classified (Ross & Murphy, 1999). For example, a bagel is a bread and
a breakfast food. Because we were interested in induction, we created a novel
predicate for each problem that we attributed to different categories (as in
classic work on category-based induction such as Osherson, Smith, Wilkie,
López, & Shafir, 1990; Rips, 1975). We showed subjects a table of different
categories with these properties in them, which might include information
such as breads contain chloride salts 55% of the time, and breakfast foods
contain chloride salts 25% of the time. These were only two of the facts listed,
so that subjects would not know which predicates were of special interest.
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 217

Then we asked the subjects to judge (among other things) what percentage of
time a bagel had chloride salts. The question was whether they would use one
of the categories (breads or breakfast foods) or both in arriving at an answer.
We found (Murphy & Ross, 1999, Expt. 1) that the majority of subjects used
only one category, but that a minority consistently used both categories in
their induction. Thus, as in our earlier work on uncertain categorizations, it
appeared that people had the ability to combine information from different
categories but did not generally see the necessity of doing so. The choice to use
multiple cross-cutting categories appeared to be a fairly high-level strategic
decision.
Unfortunately, the story is a bit more complex. In later experiments, we
simplified the display, presenting only the critical two categories (e.g., bread
and breakfast food in the bagel example) and asking fewer questions. With
these changes, the majority of people used both categories. Furthermore,
when we constructed a version of this procedure using ambiguous categories
(like the real estate agent/burglar cases considered in our previous work),
we found that people again used multiple categories most of the time (see
Murphy & Ross, 1999, for details).
These results complicate the picture slightly, but they can be explained in
terms of the conclusions drawn earlier. In this series of studies, we presented
the critical categories along with information about predicates (e.g., how of-
ten items in the category have chloride salts). This procedure had the effect
of bringing to people’s attention exactly the relevant categories and predicate.
That is, the presentation format focused people on the information relevant
to making the induction, removing other information that was not relevant,
and so they accessed both categories. When the task did not focus people’s
attention on the critical categories (as in Experiment 1 of the series, when the
critical categories were just two of four presented, and more questions were
asked), they generally used only one category. Thus, these results are theo-
retically consistent with the very different manipulation of using a property
associated to the category (as in the burglar–locked window example) – both
serve the function of bringing an alternative category to the person’s atten-
tion. Surprisingly (to us), we found almost identical results in this paradigm
for the cross-cutting categories (like breakfast food/bagel) and uncertain cat-
egories. Thus, these results suggest that in cases where it is crucial to attend
to different possible categories (e.g., medical decision making with uncertain
diagnosis), explicitly presenting the categories and the relevant information
may encourage people to make more accurate inductions.
Further work needs to be done on induction of cross-cutting categories. It
is clear that people can choose which category is relevant when only one of
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

218 G. L. Murphy and B. H. Ross

them is related to the induction property (as in Heit & Rubinstein, 1994). It
is also clear that people can and will use two cross-cutting categories when
it is obvious that both are relevant, and the property information is easy to
retrieve (as in Murphy & Ross, 1999). What is not so clear is what people do in
more naturalistic situations in which the categories and induction feature are
not explicitly presented. Unfortunately, this situation is exactly the one that is
hardest to study, because it is difficult to identify and control the information
known about each category. That is, to show use of multiple categories, we
need to know the person’s beliefs about each category so that we can tell
whether he or she is relying on one or both of them. In the ambiguous-
category case, we could control the categories by constructing scenarios in
which different categories were mentioned. But in cross-classification, the
item itself determines its categories, and we cannot make a bagel a dinner
entree versus a breakfast food in different conditions. So, this topic is awaiting
a clever idea for how to investigate it.

  

We end this chapter by considering some other cases in which it appears people
are not basing their inductions on a single category, both to better understand
the conditions under which single and multiple categories might be used and
to extend our analysis of induction. A number of Bayesian approaches to
categorization and inference have been proposed in the past decade, and they
seem to come to a different conclusion about multiple category use. One
notable example is the work of Joshua Tenenbaum (see Chapter 7 in this
volume), who has shown very good fits of his Bayesian model to a task in
which people generalize from known examples to novel ones. (See also Heit,
2000, for a different Bayesian model.) In one experiment (Tenenbaum, 1999),
subjects viewed dots in a two-dimensional Cartesian grid that represented
levels of insulin and cholesterol of healthy people. Subjects were asked to find
the rectangle (i.e., the limits on both variables) that defined the healthy levels,
based on these positive examples. Tenenbaum’s model analyzes this situation
as a collection of hypotheses about the range of healthy levels for each variable
(e.g., perhaps 45 to 55 is the range for healthy insulin, or 43 to 55, or 46 to 60,
etc.). Thus, if a subject received data for one variable showing healthy levels of
46, 52, 53, and 55, then hypotheses of 46 to 55, 45 to 55, 40 to 61, and so forth
would all be consistent with the observed data. In making predictions about
new data, the hypotheses consistent with the data were weighted so that more
specific hypotheses were considered more likely. That is, hypotheses 46 to 55
and 40 to 60 would both be used in making predictions about a new item, but
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 219

the first would be more heavily weighted than the second, because it is more
specific. Tenenbaum (2000) reports a related experiment.
Tenenbaum found that both of these experiments were fit well by a Bayesian
rule that considered each hypothesis, weighted by its likelihood. Assuming
that the number ranges are like categories and the given data are like exemplars,
his results suggest that people do after all use multiple categories in making
predictions, just as Anderson (1991) said they should, contrary to our results
in analogous situations.
What are we to make of this apparent contradiction? There are a number of
differences between Tenenbaum’s problems and the ones we have used in our
experiments, and it is likely that many of them contribute to the (apparent)
difference in results. One major difference is that our paradigm has focused
on fairly fixed object categories, whereas the number ranges of Tenenbaum
(1999) are rather ad hoc categories that do not have any independent existence.
That is, the range 46 to 55 is considered only because it happens to include all
the presented data points and not because there is any special coherence or
similarity to this set of numbers. In contrast, our geometric shape categories
were supposedly drawn by different children, were spatially separated, and
had some family resemblance. Our natural categories, such as real estate agent
or breakfast foods, were familiar categories that people use in everyday life.
It is possible that only such entrenched categories show the bias for single-
category use that we have found in most experiments.
Furthermore, it is possible that Tenenbaum’s subjects are not considering
each range of categories as an explicit category but instead are using a similarity
metric comparing the novel data point to the presented data. Certainly, in
doing this task, it does not seem that subjects are consciously considering “Is
it 43 to 55 or 46 to 56 or 44 to 56 or 46 to 55 or. . . .?”(at least, we don’t, when
we try this task). However, in our experiments with a very small number
of possible categories, we suspect that people are consciously choosing the
category the item is in. Then, because we require them to provide an induction
strength judgment, they must give a numerical response. This response may
further encourage the use of a single category, since subjects must come up
with an actual number, and that becomes increasingly difficult to do when
the number of possible categories increases, each with different information
in it. Although we do not have direct evidence that these factors account for
the apparent differences in results, we do have some recent evidence that use
of multiple categories depends on the experimental task, including the type
of decision being made.
In a recent study with Michael Verde (Verde, Murphy, & Ross, 2005), we
compared our past paradigm in which subjects provide probabilities as a
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

220 G. L. Murphy and B. H. Ross

measure of induction strength to a different paradigm in which we looked at

speeded yes-no decisions, using reaction time (RT) as a measure of induction
strength. Our reasoning was that if people are required to quickly make an
induction about a novel item, they may be influenced by a number of cate-
gories that the item might be in. People may not be very good at segregating
the information of different categories when pressured to make a prediction
about an item. In order to test this possibility, however, we had to first teach
the people the categories so that any activation of different categories would
happen fairly automatically. If people had to consciously choose which cate-
gories to examine, as in all our previous work, they would probably follow a
strategy that excluded all but the most likely category.
We will not describe the task in much detail here (see Verde et al., for a com-
plete report). The basic design was somewhat like that shown in Figure 8.1, in
which we asked people to make an induction about an item that was partially
described. But rather than ask the subject to produce an answer, we presented
the answer, and the subject confirmed or disconfirmed it. For example, if
subjects had learned categories like Figure 8.1 (which we are reusing here just
for simplicity), they might then have been given the item triangle followed
a second later by black, which meant that they were to verify whether a new
triangle was most likely to be black. In one condition, this judgment was
consistent with the target category only, and in another, it was also consistent
with a secondary category. If people responded more quickly in the second
condition than the first, that would indicate that they were influenced by two
categories rather than just one. And in fact, they did respond more quickly
when the two categories were consistent, suggesting that both influenced the
decision. (This result was not due to a speed-accuracy trade-off in that accu-
racy was as high or higher in the consistent condition.) Interestingly, when we
used the exact same materials and asked people to provide probability ratings
without time pressure, as in our past experiments, this effect disappeared.
That is, we found evidence for multiple-category use in speeded RTs but not
in explicit probability judgments with the exact same materials.
It may be that in certain kinds of decisions, multiple categories influence
the speed or strength of an induction without necessarily influencing its
outcome. To the degree that a secondary category is consistent with the
decision suggested by the target category, that decision may be easier to make.
However, when making numerical estimates or other deliberate judgments,
people may use information from only one category to arrive at their answer.
The Verde et al. (2005) results also suggest a somewhat different way of
thinking about the multiple-category issue. In Anderson’s (1991) initial pro-
posal, as well as in Tenenbaum’s approach, an explicit calculation is done
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 221

in which each category’s induction is weighted by its likelihood in deriving

the grand induction. Those proposals are intended as computational-level
analyses rather than algorithmic hypotheses – that is, they are not claiming
that people carry out these calculations in their heads to come up with the
answer but instead are claiming that the information in different categories is
somehow combined in making the induction. Our results with the speeded
task suggest that spreading activation might be a possible mechanism under
those circumstances. That is, when you are presented with a triangle, the
categories with triangles are all activated to some degree, and the categories’
other properties (green, blue, black, small) could in turn be activated propor-
tionally. This is one way in which information from different categories could
be combined without explicitly doing the Bayesian calculation.
But a more radical possibility is that the categories themselves are not ac-
tually involved in these predictions. Perhaps thinking about a triangle simply
activates exemplars of the studied triangles, and their properties are thereby
activated as well, without any category information being considered. In
our early work with numerical judgments, it seemed clear that people were
using categories, if only because their responses were often exactly the pro-
portion of items in a category that had the given feature (Murphy & Ross,
1994). That is, if half of Category A items were solid black, many people
responded 50% for their answer. Furthermore, varying properties within the
target category had a reliable effect, and varying them outside the target cat-
egory did not, showing again that subjects were paying close attention to the
categories.
But we do not have this evidence for the speeded task, and this makes
us wonder about similar situations both in experiments and the real world.
For example, in Tenenbaum’s numerical judgment task, we wonder if the
tested number might simply activate the exemplars of the category (e.g.,
the healthy insulin levels) to varying degrees, leading to varying degrees of
confidence in the induction. Such a mechanism might be able to account for
the results without the introduction of the ad-hoc categories (46 to 55, 45 to
56, etc.). In a more naturalistic example, if you are walking in the woods and
something brown and fast-moving starts to rustle in the brush to your left,
your decision about whether to run away may be based on categories such as
“bunny rabbit,” “wolf,” “coyote,” “beaver,” “deer,” and so on, but it may also
rely on a simpler association between the observed properties and an outcome
(danger, in this case). That is, any brown, fast-moving exemplars associated
with danger or its absence might influence your response now, regardless of
their category membership. However, if you are able to identify the object’s
category membership in time (“it’s only a duck”), then that information may
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

222 G. L. Murphy and B. H. Ross

control your induction even if other brown, fast-moving stimuli have had
different properties (e.g., you have had bad experiences with coyotes sneaking
up on you).
In short, in some cases, induction may be based not on category-level in-
formation at all but instead on remembered exemplars. Such inductions have
many of the properties of inductions that Bayesian models such as Tenen-
baum’s and Anderson’s have. That is, they would “weight” information from
the most likely category most of all (because most of the exemplars would
come from that category), and information from other categories would
be weighted proportionally to their likelihood (based on the number of re-
membered exemplars from each category). But the induction itself might
be somewhat different from the one envisioned in Anderson’s model, where
category-level information is the basis of induction. That is, if you identify the
brown animal as a deer, then information about deer in general would be used
in Anderson’s model. In the exemplar-based approach we are considering, it
might be only the properties of the particular remembered exemplars that af-
fect the decision, and if the remembered exemplars are not very typical of their
categories, then the prediction would deviate from a categorical induction.
A third possibility is that people might use both category and exemplar
information in making the prediction. The activated properties provide some
exemplar-based information that might combine with the category-level in-
formation, depending on the circumstances (time pressure, complexity of
prediction, etc.). One can think of this combination as a type of special-
ization, where its usefulness depends on the consistency of category and
exemplar-level information. It allows the greater statistical basis of the cate-
gory to influence the decision but also allows sensitivity to the specifics of the
item about which the prediction is made (see, e.g., Brooks, 1987; Medin &
Ross, 1989; Murphy, 2002, for specificity effects). As Murphy and Ross (2005)
argued, even when one makes category-based inductions, properties of the
specific item involved may matter as well.
So far as we know, there has been no attempt to distinguish the exemplar-
based and category-based induction processes for coherent categories. One
reason for this is that in the classic category-based induction work, subjects are
presented with entire categories and/or an undescribed exemplar. Consider
the example, “If sparrows have sesamoid bones, what is the probability that
eagles have sesamoid bones?” Here, only entire categories are mentioned. The
questions asked by Rips (1975), Osherson et al. (1990), and the more recent
work of Proffitt et al. (2000) all focus on category-level problems.
An important topic for future research, then, is to investigate how item-level
and category-level information is used and integrated in making predictions
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 223

(Murphy & Ross, 2005). If you believe that dogs get sick from eating chocolate,
will you change your prediction for a particularly robust, voracious dog who
is known to eat any semi-edible garbage that it can grab? How about a yappy,
pampered toy dog of weak disposition? Proffitt et al. (2000) make a strong
argument that inductions are based on causal knowledge when it is available.
The usual use of blank predicates is an attempt to eliminate such knowledge
as a factor. But in everyday life, people do have beliefs about the predicates
they make inductions about, and they also have knowledge about the specific
individuals they are interested in (e.g., they can see the dog, they know their
doctor, they have ridden in this car, and so on). The items are not unknown
objects described only by their category name: “a dog,” “a doctor,” or “a car.”
So, to expand our understanding of induction further, we need to begin to
address the interaction between category-based induction and individual-
based induction.


In summary, we would often be better off if we considered multiple categories
when making inductions about entities. When categorization is uncertain,
our inductions ought to take our uncertainty into account. However, in a
number of paradigms manipulating a number of variables, we have found
that people generally do not spontaneously consider multiple categories when
making inductions. Perhaps surprisingly, this does not appear to be due to
unfamiliarity with the principle involved, as shortcomings in decision making
often seem to be. Nor is the problem that people are not able to combine the
predictions drawn from different categories. Instead, the problem appears to
be a kind of production deficit in which the necessity of considering multiple
categories does not occur to people except in situations that strongly bring
the secondary categories to attention. When the property is associated to the
secondary category, or when that category is presented in a way that is difficult
to ignore (Murphy & Ross, 1999), people integrate information from multiple
categories.
On the negative side, people’s reluctance to spontaneously combine predic-
tions from categories suggests that they will often make suboptimal predic-
tions. But on the positive side, when the necessity for using multiple categories
is brought to the fore, people are willing and able to use them. Thus, our results
give some hope for improvement in category-based inductions in situations
where those inductions are important. By presenting the secondary category
prominently or asking just the right question, people can be encouraged to
take multiple categories into account.
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

224 G. L. Murphy and B. H. Ross

Finally, our most recent work suggests the possibility that performance in
tasks that we think of as category-based induction may not necessarily be
based on categories. Performance that seems to involve the use of multiple
categories might be accomplished if people skip the categories entirely, relying
on remembered exemplars. Although this is speculative, we see the possibility
that people make use of both categories and exemplar information when
making inductions to be an interesting possibility that also relates research
on category-based induction to current work on classification.

 
Please address all correspondence to Gregory L. Murphy, Department of
Psychology, New York University, 6 Washington Place, 8th floor, New York,
NY 10003; or email correspondence to [email protected].

References
Anderson, J. R. (1991). The adaptive nature of human categorization. Psychological
Review, 98, 409–429.
Baillargeon, R. (1998). Infants’ understanding of the physical world. In M. Sabourin,
F. Craik, & M. Robert (Eds.), Advances in psychological science, Vol. 2: Biological and
cognitive aspects (pp. 503–529). London: Psychology Press.
Brooks, L. R. (1987). Decentralized control of categorization: The role of prior processing
episodes. In U. Neisser (Ed.), Concepts and conceptual development: Ecological and
intellectual factors in categorization (pp. 141–174). Cambridge: Cambridge University
Press.
Heit, E. (2000). Properties of inductive reasoning. Psychonomic Bulletin & Review, 7,
569–592.
Heit, E., & Rubinstein, J. (1994). Similarity and property effects in inductive reasoning.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 411–422.
Kalish, C. W., & Gelman, S. A. (1992). On wooden pillows: Multiple classification and
children’s category-based inductions. Child Development, 63, 1536–1557.
Lagnado, D. A., & Shanks, D. R. (2003). The influence of hierarchy on probability
judgments. Cognition, 89, 157–178.
Macrae, C. N., Bodenhausen, G. V., & Milne, A. B. (1995). The dissection of selection in
person perception: Inhibitory processes in social stereotyping. Journal of Personality
and Social Psychology, 69, 397–407.
Malt, B. C., Ross, B. H., & Murphy, G. L. (1995). Predicting features for members of nat-
ural categories when categorization is uncertain. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 21, 646–661.
Medin, D. L., & Ross, B. H. (1989). The specific character of abstract thought: Cate-
gorization, problem solving, and induction. In R. J. Sternberg (Ed.), Advances in the
psychology of human intelligence, Vol. 5. Hillsdale, NJ: Erlbaum.
Murphy, G. L. (2002). The big book of concepts. Cambridge, MA: MIT Press.
P1: JZP
0521672443c08 CUFX144-Feeney 0 521 85648 5 July 20, 2007 6:56

Use of Single or Multiple Categories 225

Murphy, G. L., & Ross, B. H. (1994). Predictions from uncertain categorizations. Cogni-
tive Psychology, 27, 148–193.
Murphy, G. L., & Ross, B. H. (1999). Induction with cross-classified categories. Memory
& Cognition, 27, 1024–1041.
Murphy, G. L., & Ross, B. H. (2005). The two faces of typicality in category-based
induction. Cognition, 95, 175–200.
Nelson, L. J., & Miller, D. T. (1995). The distinctiveness effect in social categorization:
You are what makes you unusual. Psychological Science, 6, 246–249.
Nguyen, S. P., & Murphy, G. L. (2003). An apple is more than just a fruit: Cross-
classification in children’s concepts. Child Development, 74, 1783–1806.
Osherson, D. N., Smith, E. E., Wilkie, O., López, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Patalano, A. L., Chin-Parker, S., & Ross, B. H. (2006). The importance of being coher-
ent: Category coherence, cross-classification, and reasoning. Journal of Memory and
Language, 54, 407–424.
Proffitt, J. B., Coley, J. D., & Medin, D. L. (2000). Expertise and category-based induction.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 811–828.
Rips, L. J. (1975). Inductive judgments about natural categories. Journal of Verbal Learn-
ing and Verbal Behavior, 14, 665–681.
Ross, B. H., & Murphy, G. L. (1996). Category-based predictions: Influence of uncertainty
and feature associations. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 22, 736–753.
Ross, B. H., & Murphy, G. L. (1999). Food for thought: Cross-classification and category
organization in a complex real-world domain. Cognitive Psychology, 38, 495–553.
Tenenbaum, J. B. (1999). Bayesian modeling of human concept learning. Advances in
Neural Information Processing Systems, 11, 59–68.
Tenenbaum, J. B. (2000). Rules and similarity in concept learning. Advances in Neural
Information Processing Systems, 12, 59–65.
Tversky, A., & Kahneman, D. (1980). Causal schemas in judgments under uncertainty.
In M. Fishbein (Ed.), Progress in social psychology (pp. 49–72). Hillsdale, NJ: Erlbaum.
Verde, M. F., Murphy, G. L., & Ross, B. H. (2005). Influence of multiple categories on
property prediction. Memory & Cognition, 33, 479–487.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference: From Philosophical Analysis

to Neural Mechanisms
Paul Thagard

1.   
In the 1890s, the great American philosopher C. S. Peirce (1931–1958) used
the term “abduction” to refer to a kind of inference that involves the gen-
eration and evaluation of explanatory hypotheses. This term is much less
familiar today than “deduction,” which applies to inference from premises
to a conclusion that has to be true if the premises are true. And it is much
less familiar than “induction,” which sometimes refers broadly to any kind
of inference that introduces uncertainty, and sometimes refers narrowly to
inference from examples to rules, which I will call “inductive generalization.”
Abduction is clearly a kind of induction in the broad sense, in that the gener-
ation of explanatory hypotheses is fraught with uncertainty. For example, if
the sky suddenly turns dark outside my window, I may hypothesize that there
is a solar eclipse, but many other explanations are possible, such as the arrival
of an intense storm or even a huge spaceship.
Despite its inherent riskiness, abductive inference is an essential part of
human mental life. When scientists produce theories that explain their data,
they are engaging in abductive inference. For example, psychological theories
about mental representations and processing are the result of abductions
spurred by the need to explain the results of psychological experiments. In
everyday life, abductive inference is ubiquitous, for example when people
generate hypotheses to explain the behavior of others, as when I infer that
my son is in a bad mood to explain a curt response to a question. Detectives
perform abductions routinely in order to make sense of the evidence left by
criminal activity, just as automobile mechanics try to figure out what problems
are responsible for a breakdown. Physicians practice abduction when they try
to figure out what diseases might explain a patient’s symptoms. Table 9.1
summarizes the kinds of abductive inference that occur in various domains,

226
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 227

table 9.1. Abductive inference in five domains, specifying what needs to be

explained and the kinds of hypotheses that provide explanations

Targets to be Explanatory
Domains Explained Hypotheses
science experimental results theories about structures and processes
medicine symptoms diseases
crime evidence culprits, motives
machines operation, breakdowns parts, interactions, flaws
social behavior mental states, traits

involving both targets that require explanation and the hypotheses that are
generated to explain them. Abduction occurs in many other domains as well,
for example, religion, where people hypothesize the existence of God in order
to explain the design and existence of the world.
The next section will briefly review the history of the investigation of
abduction by philosophers and artificial intelligence researchers, and discuss
its relative neglect by psychologists. First, however, I want to examine the
nature of abduction and sketch what would be required for a full psychological
theory of it. I then outline a neurocomputational theory of abductive inference
that provides an account of some of the neural processes that enable minds to
make abductive inference. Finally, I discuss the more general implications of
replacing logic-based philosophical analyses of human inference with theories
of neural mechanisms.
Here are the typical stages in the mental process of abduction. First, we no-
tice something puzzling that prompts us to generate an explanation. It would
be pointless to waste mental resources on something ordinary or expected.
For example, when my friends greet me with the normal “Hi,” I do not react
like the proverbial psychoanalyst who wondered “What can they mean by
that?” In contrast, if a normally convivial friend responds to “Good morn-
ing” with “What’s so good about it?”, I will be prompted to wonder what is
currently going on in my friend’s life that might explain this negativity. Peirce
noticed that abduction begins with puzzlement, but subsequent philosophers
have ignored the fact that the initiation of this kind of inference is inherently
emotional. Intense reactions such as surprise and astonishment are partic-
ularly strong spurs to abductive inference. Hence the emotional initiation
of abductive inference needs to be part of any psychological or neurological
theory of how it works. An event or general occurrence becomes a target for
explanation only when it is sufficiently interesting and baffling. I know of no
general experimental evidence for this claim, but Kunda, Miller, and Claire
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

228 P. Thagard

(1990) found that surprise triggered causal reasoning in cases of conceptual

combination.
Second, the mind searches for possible hypotheses that could explain the
target. Sometimes, the search is easily completed when there is a prepackaged
hypothesis waiting to be applied. For example, if you know that your friend
Alice gets stressed out whenever she has a deadline to meet, you might explain
her grumpy behavior by the conjecture that she has a project due. In more
deeply puzzling cases, the search for an explanatory hypothesis may require
a much longer search through memory, or even the use of analogy or other
constructive processes to generate a highly novel hypothesis. This generation is
what happens in science when a genuinely new theory needs to be developed.
Generation of a candidate explanatory hypothesis is the usual third stage
of abductive inference. If one is epistemically lazy, abductive inference may
end with the generation of a single candidate. But scientists and careful
thinkers in general are aware of the perils of abductive inference, in particular
that one should not accept an explanatory hypothesis unless it has been
assessed with respect to competing hypotheses and all the available evidence.
Philosophers call this fourth, evaluative stage of abductive reasoning inference
to the best explanation (Harman, 1973; Thagard, 1988; Lipton, 2004). Ideally,
the reasoner correctly decides that it is legitimate to infer a hypothesis because
it really is the best explanation of all the available evidence. Thus, generation
of an explanatory hypothesis blends into its evaluation.
Just as abduction originates with an emotional reaction, it ends with one,
because formation and acceptance of explanatory hypotheses usually produce
positive emotions. Gopnik’s (1998) comparison of explanations with orgasms
is exaggerated, but it is nevertheless important that finding an explanation
for something puzzling is often very satisfying. Hence we can mark emotional
satisfaction as the final stage of abductive inference, as shown in Figure 9.1.
This diagram will be fleshed out substantially in later sections in terms of
neurological processes. The result will be an account of abduction that goes
far beyond the philosophical account that takes abduction to be a kind of
inference of the form “q, if p then q, so maybe p.” (Somebody once called
this “modus morons.”) To foreshadow, the main differences include not only
the crucial involvement of emotion, but also the allowance that both targets
and hypotheses can be multimodal rather than purely verbal representations.
Moreover, I will contend that the relation between a target and an explanatory
hypothesis is that the target phenomenon is caused by the factors invoked by
the hypothesis, and that people’s understanding of causality is inherently
non-verbal because it is rooted in visual and kinesthetic perception. Hence
abduction, instead of looking like a feeble-minded cousin of the deductive
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 229

puzzlement search for generation of evaluation and

or surprise explanation hypothesis acceptance of
hypothesis

pleasure or
satisfaction

figure 9.1. The process of abductive inference.

principle modus ponens, is actually a far richer and more powerful form of
thinking.

2.   ,  ,

 
Although the term “abduction” emerged only in the nineteenth century,
philosophers and scientists have been aware of inference to explanatory hy-
potheses at least since the Renaissance (Blake, Ducasse, and Madden, 1960).
Some thinkers have been skeptical that a hypothesis should be accepted merely
on the basis of what it explains, for example Isaac Newton, John Herschel,
August Comte, and John Stuart Mill. But others, such as David Hartley, Joseph
Priestley, and William Whewell have argued that such inference is a legitimate
part of scientific reasoning. In the twentieth century, there are still philo-
sophical skeptics about abduction (e.g., van Fraassen, 1980), but many others
contend that abduction, construed as inference to the best explanation, is an
essential part of scientific and everyday reasoning (Magnani, 2001; Psillos,
1999; Thagard, 1992, 1999, 2000).
The generation of explanatory hypotheses has also interested philosophers
concerned with how scientific discoveries are made. Hanson (1958) tried to
develop Peircean ideas into a “logic” of discovery. More recently, Darden
(1991), Magnani (2001), and Thagard (1988) have examined cognitive pro-
cesses that are capable of producing new hypotheses, including scientific
theories. The work of Shelley (1996) and Magnani (2001) shows how abduc-
tion can involve visual as well as verbal representations, which is important
for the multimodal theory developed in Section 3.
In the field of artificial intelligence, the term “abduction” is usually applied
to the evaluation of explanatory hypotheses, although it sometimes also in-
cludes processes of generating them (Charniak and McDermott, 1985; Joseph-
son and Josephson, 1994). AI models of abductive inference have primarily
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

230 P. Thagard

been concerned with medical reasoning. For example, the RED system takes
as input descriptions of cells and generates and evaluates hypotheses about
clinically significant antibodies found in the cells (Josephson and Josephson,
1994). More recently, discussions of causal reasoning in terms of abduction
have been eclipsed by discussions of Bayesian networks based on probability
theory, but later I will describe limitations of purely probabilistic accounts
of causality, explanation, and abduction. Abduction has also been a topic of
interest for researchers in logic programming (Flach and Kakas, 2000), but
there are severe limitations to a characterization of abduction in terms of
formal logic (Thagard and Shelley, 1997).
Some AI researchers have discussed the problem of generating explanatory
hypotheses without using the term “abduction.” The computational models
of scientific discovery described by Langley, Simon, Bradshaw, and Zytkow
(1987) are primarily concerned with the inductive generalization of laws from
data, but they also discuss the generation of explanatory structure models in
chemistry. Langley et al. (2004) describe an algorithm for “inducing explana-
tory process models,” but it is clear that their computational procedures for
constructing models of biological mechanisms operate abductively rather
than via inductive generalization.
Psychologists rarely use the terms “abduction” or “abductive inference,”
and very little experimental research has been done on the generation and
acceptance of explanatory hypotheses. Much of the psychological literature
on induction concerns a rather esoteric pattern of reasoning, categorical
induction, in which people express a degree of confidence that a category has
a predicate after being told that a related category has the predicate (Sloman
and Lagnado, 2005). Here is an example:
Tigers have 38 chromosomes.
Do buffaloes have 38 chromosomes?
Another line of research involves inductive generalizations about the behavior
of physical devices (Klahr, 2000). Dunbar (1997) has discussed the role of
analogy and other kinds of reasoning in scientific thinking in real-world
laboratories. Considerable research has investigated ways in which people’s
inductive inferences deviate from normative standards of probability theory
(Gilovich, Griffin, and Kahneman, 2002).
Experimental research concerning causality has been concerned with top-
ics different from the generation of causal explanations, such as how people
distinguish genuine causes from spurious ones (Lien and Cheng, 2000) and
how knowledge of the causal structure of categories supports the ability to
infer the presence of unobserved features (Rehder and Burnett, 2005). Social
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 231

psychologists have investigated the important abductive task of attribution, in

which people generate explanations for the behavior of others (Kunda, 1999;
Nisbett and Ross, 1980). Read and Marcus-Newhall (1993) tested the appli-
cability of Thagard’s (1992) computational theory of explanatory coherence
to the evaluation of social explanations. Generally, however, psychologists
have had little to say about the mental mechanisms by which new hypotheses
are formed and evaluated. My review of interdisciplinary research on abduc-
tive inference has been very brief, for I want to move on to develop a new
neurocomputational theory of abduction.

3.  
The structure of abduction is roughly this:

There is a puzzling target T that needs explanation.

Hypothesis H potentially explains T.
So, H is plausible.
H is a better explanation of T and other phenomena than competing
hypotheses.
So H is acceptable.

It would be psychologically unrealistic, however, to assume, as philosophers

and AI researchers have tended to do, that T and H must be sentences or
propositions (the meanings of sentences). A broader view of mental repre-
sentation is required.
As already mentioned, abductive inference can be visual as well as verbal
(Shelley, 1996; Magnani, 2001). For example, when I see a scratch along the
side of my car, I can generate the mental image of a grocery cart sliding into
the car and producing the scratch. In this case both the target (the scratch)
and the hypothesis (the collision) are visually represented. Other sensory
modalities can also provide explanation targets. For example, in medical
diagnosis the perception of a patient’s symptoms can involve vision (rash),
touch (swelling), sound (heart murmur), smell (infection), and even taste
(salty, in patients with cystic fibrosis). An observant cook may similarly be
prompted to generate hypotheses by various kinds of sensory experiments,
asking such questions as “Why does the cheese have blue stuff on it?” (vision),
“Why is the broccoli soggy?” (touch), “Why is the timer buzzing?” (hearing),
“Why is the meat putrid?” (smell), and “Why is the soup so salty” (taste). Thus
all of the senses can generate explanation targets that can initiate abductive
inference.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

232 P. Thagard

It is an interesting question whether hypotheses can be represented using all

sensory modalities. For vision the answer is obvious, as images and diagrams
can clearly be used to represent events and structures that have causal effects.
And the answer appears to be yes when one is explaining one’s own behavior:
I may recoil because something I touch feels slimy, or jump because of a loud
noise, or frown because of a rotten smell, or gag because something tastes
too salty. Hence in explaining my own behavior, my mental image of the
full range of examples of sensory experiences may have causal significance.
Applying such explanations of the behavior of others requires projecting onto
them the possession of sensory experiences that I think are like the ones that
I have in similar situations. For example, when I see people wrinkle up their
noses in front of a garbage can, I may project onto them an experience similar
to what I experience when I smell rotting garbage. In this case, my image
of a smell is the olfactory representation of what I see as the cause of their
behavior. Empathy works the same way, when I explain people’s behavior
in a particular situation by inferring that they are having the same kind of
emotional experience that I have had in similar situations. For example, if a
colleague with a recently rejected manuscript is frowning, I may empathize
by remembering how annoyed I felt when a manuscript of mine was rejected,
and my mental image projected onto the colleague constitutes a non-verbal
representation that explains the frown. Of course, I may operate with verbal
explanations as well, but these complement the empathetic ones. Hence there
is reason to believe that abductive inference can be fully multimodal, in that
both targets and hypotheses can have the full range of verbal and sensory
representations. In addition to words, sights, sounds, smells, touches, and
tastes, these can include emotional feelings, kinesthetic experiences such as
feeling the body arranged in a certain way, and other feelings such as pain.
A narrowly verbal account of abduction such as those favored by logicians
would clearly not be able to do justice to the multimodal character of ab-
ductive inference. But from a neurological perspective, there is no problem
in postulating representations that operate in all relevant modalities. Let me
define a neural structure as a complex <neurons, connections, spiking behav-
iors> that consists of a set of neurons, a set of synaptic connections among
them, and a set of behaviors of individual neurons that specifies their pat-
terns of spiking determined by the spiking behaviors of all those neurons to
which they are connected. If the neurons are thought of as a dynamical system
governed by a set of differential equations, then the spiking behaviors can be
thought of as the state space of the system. In contrast to standard connec-
tionist models of neural networks, it is important to specify the behavior of
neurons as more than just patterns of activation, because there is evidence that
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 233

spiking patterns can be both neurologically and computationally important

(Eliasmith and Anderson, 2003; Maass and Bishop, 1999; Rieke et al., 1997).
On the plausible assumption that all mental representations are brain struc-
tures, we can conjecture that verbal and sensory representations are neural
structures of the sort just described. Hence we can reconceptualize abduc-
tion neurologically as a process in which one neural structure representing
the explanatory target generates another neural structure that constitutes a
hypothesis. Two major problems need to be solved in order to construct a
neurological model of abduction consistent with the flow chart presented
earlier in Figure 9.1. The first is how to characterize the emotional in-
puts and outputs to the abductive process, that is, how to mark the target
as puzzling and the hypothesis as satisfying. The second is how to repre-
sent the explanatory relation between the neurally represented target and
hypothesis.
The first problem can be dealt with by supposing that emotions are also
neural structures in the sense just defined. I do not mean to suggest that
for each emotion there is a constrained set of neurons that encodes it. Emo-
tions involve complex interaction among sensory processes involving bodily
states, and cognitive processes involving appraisal of a person’s situation (see
Thagard, 2005, ch. 10). Hence the neural structure corresponding to an emo-
tional experience is not a neuronal group situated in a particular brain area,
but a complex of neurons distributed across multiple brain areas.
Now the question becomes: what is the relation between the neural struc-
ture for the explanation target and the neural structure for an emotion such
as puzzlement? There are two possibly compatible answers involving different
aspects of neural structure: connections and spiking behavior. The target neu-
ral structure may be interrelated with an emotional neural structure because
some of the neurons in the first structure have direct or indirect synaptic con-
nections with the second structure. Hence the two neural structures are part
of a larger neural structure. Moreover, these interconnections may establish
temporal coordination between the two neural structures, so that the spiking
behavior of the target neural structure is synchronized or approximately co-
ordinated with the spiking behavior of emotional neural structure. Through
one or both of these means – physical connectivity and temporal behavior –
the brain manages to mark the target as puzzling and in need of explanation.
Similarly, the hypothesis can be represented by a neural structure which op-
erates in any verbal or sensory modality, and which can be associated with a
neural structure corresponding to the emotional experience of satisfaction or
pleasure. Thus part of the flow chart in Figure 9.1 translates into the follow-
ing neurological diagram, Figure 9.2. What remains to be investigated is the
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

234 P. Thagard

neural neural
structure structure
(target) (hypothesis)

neural neural
structure structure
(puzzled) (satisfied)

figure 9.2. Abduction as a neural process.

relation of explanation and inference marked by the thin arrow. We need an

account of explanation as a neurological process.

4.   

Abductive inference is from a target to a hypothesis that explains it, but
what is explanation? In philosophy and cognitive science, there have been at
least six approaches to the topic of explanation (Thagard, 1992, pp. 118ff.).
Explanations have been viewed as deductive arguments, statistical relations,
schema applications, analogical comparisons, causal relations, and linguistic
acts. All of these approaches have illuminated some aspects of the practice
of explanation in science and everyday life, but the core of explanation,
I would argue, is causality: a hypothesis explains a target if it provides a
causal account of what needs to be explained. In medicine, a disease explains
symptoms because the abnormal biological state that constitutes the disease
produces the symptoms through biological mechanisms (Thagard, 1999). In
ordinary social life, attributing a mental state such as an emotion explains
a person’s behavior because the mental state is the assumed cause of the
behavior. In science, a theory explains phenomena such as the motions of the
planets by providing the causes of such phenomena. Deductive, statistical,
schematic, analogical, and linguistic aspects of explanation can all be shown
to be subordinate to the fundamental causal aspect of explanation.
Hence the problem of describing the explanatory relation between hy-
potheses and targets is largely the problem of describing the causal relation
between what neural structures for hypotheses represent and what neural
structures for targets represent. To put it most simply, abduction becomes:
T, H causes T, so maybe H, where H and T are not propositions but what is
represented by neural structures. Hence it is crucial to give a philosophically
and psychologically plausible account of causality.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 235

The philosophical literature on causality is huge, but here is a quick sum-

mary of extant positions on the nature of causality:
1. Eliminativist: Causality is an outmoded notion no longer needed for
scientific discourse.
2. Universalist: Causation is a relation of constant conjunction: the effect
always occurs when the cause occurs.
3. Probabilistic: Causation is a matter of probability: the effect is more
probable given the cause than otherwise.
4. Causal powers: the cause has a power to produce the effect.
Each of these positions is problematic. The eliminativist position runs afoul of
the fact that talk of causal mechanisms still abounds in science. The universalist
position is untenable because most causal relations are statistical rather than
constant: infection by a mycobacterium causes tuberculosis, but many people
infected by it never develop the disease. Similarly, the probabilistic position
cannot easily distinguish between cases where a cause actually makes an effect
more likely, as in infection by the mycobacterium, and cases where the effect
is made more likely by some other cause. For example, the probability that
people have tuberculosis given that they take the drug Isoniazid is much
greater than the probability that they have tuberculosis, but this is because
Isoniazid is a commonly prescribed treatment for the disease, not because
it causes the disease. (See Pearl, 2000, for a broad and deep discussion of
causality and probability.) People have an intuitive sense of the difference
between causal and purely statistical relations. In part, this arises from an
understanding of how mechanisms connect causes with their effects (see
Section 7). But it also arises from a natural perceptual inclination to see certain
kinds of occurrences as causally related to each other. This inclination does not
depend on the postulation of occult causal powers that causes must have in
relation to their effects, but on fundamental features of our perceptual systems.
Evidence that causal relations can be perceived comes from three kinds of
psychological evidence: cognitive, developmental, and neurological. Michotte
(1963) performed a large array of experiments with visual stimuli that suggest
that adults have a direct impression of cause-effect relations: when people see
an image of one ball moving into another that begins to move, the first ball is
perceived to cause the second one to move. Such reactions are found even in
young infants: Leslie and Keeble (1987) provided experimental evidence that
even 27-week old infants perceive a causal relationship. Baillargeon, Kotovsky,
and Needham (1995) report that infants as young as 2.5 months expect a
stationary object to be displaced when it is hit by a moving object; by around
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

236 P. Thagard

6 months, infants believe that the distance traveled by the stationery object is
proportional to the size of the moving object. Thus at a very primitive stage of
verbal development children seem to have some understanding of causality
based on their visual and tactile experiences. According to Mandler (2004),
infants’ very early ability to perceive causal relations need not be innate but
could arise from a more general ability to extract meaning from perceptual
relationships. Whether or not it is innate, infants clearly have an ability to
extract causal information that develops long before any verbal ability.
Recent work using functional magnetic resonance imaging has investigated
brain mechanisms underlying perceptual causality (Fugelsang et al., 2005).
Participants imaged while viewing causal events had higher levels of relative
activation in the right middle frontal gyrus and the right inferior parietal lob-
ule compared to those viewing non-causal events. The evidence that specific
brain structures are involved in extracting causal structure from the world
fits well with cognitive and developmental evidence that adults and children
are able to perceive causal relations, without making inferences based on uni-
versality, probability, or causal powers. It is therefore plausible that people’s
intuitive grasp of causality, which enables them to understand the distinc-
tion between causal relations and mere co-occurrence, arises very early from
perceptual experience. Of course, as people acquire more knowledge, they
are able to expand this understanding of causality far beyond perception,
enabling them to infer that invisible germs cause disease symptoms. But this
extended understanding of causality is still based on the perceptual experience
of one event making another happen, and does not depend on a mysterious,
metaphysical conception of objects possessing causal powers. For discussion
of the role of causality in induction, see Bob Rehder’s Chapter 4 in this
volume.
Now we can start to flesh out in neurological terms what constitutes the re-
lation between a target and an explanatory hypothesis. Mandler (2004) argues
that CAUSED-MOTION is an image schema, an abstract, non-propositional,
spatial representation that expresses primitive meanings. Lakoff (1987) and
others have proposed that such non-verbal representations are the basis for
language and other forms of cognition. Feldman and Narayan (2004) have de-
scribed how image schemas can be implemented in artificial neural systems. I
will assume that there is a neurally encoded image schema that establishes the
required causal relation that ties together the neural structure of a hypothesis
and the neural structure of the target that it explains. We would then have
a neural representation of the explanatory, causal relation between hypothe-
ses and targets. This relation provides the abductive basis for the inferential
process described in the next section.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 237

The model of abductive inference sketched in Figures 9.1 and 9.2 has
been implemented in a computer simulation that shows in detail how neural
processes can generate emotional initiation and causal reasoning (Thagard
and Litt, forthcoming). The details are too technical to present here, but the
simulation is important because it shows how causal and emotional informa-
tion distributed over thousands of artificial neurons can produce a simple
form of abductive inference.

5. 
On the standard philosophical view, inference is the movement from one
or more propositions taken to be true to another proposition that follows
from them deductively or inductively. Here a proposition is assumed to be an
abstract entity, namely, the meaning content of a sentence. Belief and other
mental states such as doubt, desire, and fear are all propositional attitudes,
that is, relations between persons and propositions. An inference is much like
an argument, which is the verbal description of a set of sentential premises
that provide the basis for accepting a conclusion. Most philosophical and
computational accounts of abductive inference have assumed this kind of
linguistic picture of belief and inference.
There are many problems with this view. It postulates the existence of
an infinite number of propositions, including an infinite number that will
never be expressed by any uttered sentence. These are abstract entities whose
existence is utterly mysterious. Just as mysterious is the relation between
persons and propositions, for what is the connection between a person’s body
or brain and such abstract entities? The notion of a proposition dates back
at least to Renaissance times when almost everyone assumed that persons
were essentially non-corporeal souls, which could have some non-material
relation to abstract propositions. But the current ascendancy of investigation
of mental states and operations in terms of brain structures and processes
makes talk of abstract propositions as antiquated as theories about souls or
disease-causing humors. Moreover, philosophical theories of propositional
belief have generated large numbers of insoluble puzzles, such as how it can
be that a person can believe that Lewis Carroll wrote Alice in Wonderland,
but not that Charles Dodgson did, when the beliefs seem to have the same
content because Carroll and Dodgson are the same person.
Implicit in my account of abductive inference is a radically different ac-
count of belief that abandons the mysterious notion of a proposition in favor
of biologically realistic ideas about neural structures. In short, beliefs are
neural structures consisting of neurons, connections, and spiking behavior;
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

238 P. Thagard

and so are all the other mental states that philosophers have characterized as
propositional attitudes, including doubt, desire, and fear. This view does away
with the metaphysical notion of a proposition, but does not eliminate stan-
dard mental concepts such as belief and desire, which are, however, radically
reconstrued in terms of structures and processes in the brain (cf. Churchland,
1989).
This view of mental operations makes possible an account of the nature of
inference that is dramatically different from the standard account that takes
inference to operate on propositions the same way that argument operates
on sentences. First, it allows for non-verbal representations from all sensory
modalities to be involved in inference. Second, it allows inferences to be
holistic in ways that arguments are not, in that they can simultaneously take
into account a large amount of information before producing a conclusion.
How this works computationally is shown by connectionist computational
models such as my ECHO model of explanatory coherence (Thagard 1992).
ECHO is not nearly as neurologically realistic as the current context requires,
since it uses localist artificial neurons very different from the groups of spiking
neurons that I have been discussing, but it at least shows how parallel activity
can lead to holistic conclusions.
What then is inference? Most generally, inference is a kind of transforma-
tion of neural structures, but obviously not all such transformations count
as inference. We need to isolate a subclass of neural structures that are rep-
resentations, that is, ones that stand for something real or imagined in the
world. Roughly, a neural structure is a representation if its connections and
spiking behavior enable it to relate to perceptual input and/or the behavior
of other neural structures in such a way that it can be construed as standing
for something else such as a thing or concept. This is a bit vague, but it is
broad enough to cover both cases where neural structures stand for concrete
things in the world, for example, George W. Bush, and for general categories
that may or may not have any members, for example, unicorns. Note that
one neural structure may constitute many representations, because different
spiking behaviors may correspond to different things. This capability is a
feature of all distributed representations.
Accordingly, we can characterize inference as the transformation of represen-
tational neural structures. Inference involving sentences is a special case of such
transformation where the relevant neural structures correspond to sentences.
My broader account has the great advantage of allowing thinking that uses
visual and other sensory representations to count as inference as well. From
this perspective, abduction is the transformation of representational neural
structures that produces neural structures that provide causal explanations.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 239

This neural view of inference is open to many philosophical objections.

It remains to be shown that neural structures have sufficient syntactic and
semantic complexity to qualify as sentential representations. The syntactic
problem is potentially solved by theories of neural combinatorics such as
the tensor product theory of Smolensky (1990) which show how vectors
representing neural activity can be combined in ways that capture syntactic
structure. The semantic problem is potentially solved by providing more
detail about how neural structures can relate to the world and to each other
(Eliasmith, 2005). But much more needs to be said about what enables a
neural structure to constitute a meaningful representation.
Another objection to my account of neural structures and inference is that
it requires some way of specifying groups of neurons that are part of identi-
fiable neural structures. In the worse case, one might be forced to conclude
that there is only one neural structure in the brain, consisting of all the neu-
rons with all their connections and spiking behaviors. This problem becomes
especially acute in cases of inference that involve multiple kinds of verbal, per-
ceptual, and emotional representation, which require that multiple brain areas
be involved. In practice, however, the problem of isolating neural structures
does not seem to be insurmountable. Neuroscientists often talk of groups,
populations, or assemblies of neurons that are identifiable subsets of the 100
billion or so neurons that constitute an entire brain. Groups are identifiable
because they have far more connections with each other than they do with
neurons in other parts of the brain. Hence even though there is a great deal
of interconnectivity in the brain, we can still identify groups of neurons with
high degrees of connection to each other and spiking behaviors that enable
them to constitute representations. So the view that inference is transforma-
tion of neural structures does not devolve into the much less precise claim
that inference is just brain transformation.
Another philosophical objection to the neural theory of inference is that it is
unduly narrow in that it does not apply to inferences by robots or by extrater-
restrial beings with brains radically different from ours. My response is that I
am only concerned here to provide a special theory of human inference and
will leave the problem of developing a general theory of inference for the oc-
casion when we actually encounter robots or aliens that are intelligent enough
that we want to count what they do as inference. The general theory would
consist of a broader account of a representational structure, <parts, relations,
behaviors>, analogous to the <neurons, connections, spiking behaviors> of
humans and other terrestrial animals. It is an open question what the degree of
similarity will be between the mental mechanisms of human and non-human
thinkers, if there are any of the latter.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

240 P. Thagard

6.  
I described earlier how abductive inference is initiated by emotional reactions
such as surprise and puzzlement, but other forms of inference also have
affective origins. The influence of emotions on decision making has often
been noted (Damasio, 1994; Mellers et al., 1999; Thagard, 2001, 2006). But
less attention has been paid to the fact that inferences about what to do
are usually initiated by either positive or negative emotions. Decisions are
sometimes prompted by negative emotions such as fear: if I am afraid that
something bad will happen, I may be spurred to decide what to do about it.
For example, a person who is worried about being fired may decide to look for
other jobs. It would be easy to generate examples of cases where other negative
emotions such as anger, sadness, envy, and guilt lead people to begin a process
of deliberation that leads to a practical inference. More positively, emotions
such as happiness can lead to decisions, as when someone thinks about how
much fun it would be to have a winter vacation and begins to collect travel
information that will produce a decision about where to go. Just as people do
not make abductive inferences unless there is some emotional reason for them
to look for explanations, people do not make inference about what to do unless
negative or positive emotional reactions to their current situation indicate
that action is required. In some cases, the emotions may be tied to specific
perceptual states, such as when hunger initiates a decision about what to eat.
Deductive inference might be thought to be impervious to emotional influ-
ences, but there is neurological evidence that even deduction can be influenced
by brain areas such as the ventromedial prefrontal cortex that are known to
be involved in emotion (Houdé et al., 2001; Goel and Dolan, 2003). It is hard
to say whether deduction is initiated by emotion, because I think it is rarely
initiated at all outside the context of mathematical reasoning: readers should
ask themselves when was the last time they made a deductive inference. But
perhaps deduction is sometimes initiated by puzzlement, as when one won-
ders whether an object has a property and then retrieves from memory a rule
that says that all objects of this type have the property in question. This kind
of inference may be so automatic, however, that we never become aware of
the making of the inference or any emotional content of it.
Analogical inference often involves emotional content, especially when it
is used in persuasion to transfer negative affect from a source to a target
(Blanchette and Dunbar, 2001; Thagard and Shelley, 2001). For example,
comparing a political leader to Hitler is a common way of motivating people to
dislike the leader. Such persuasive analogies are often motivated by emotional
reactions such as dislike of a person or policy. Because I dislike the leader, I
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 241

compare him or her to Hitler in order to lead you to dislike the leader also.
Practical analogical inferences are prompted by emotions in the same way
that other decisions are: I want to go on vacation, and remember I had a
good time at a resort before, and decide to go to a similar resort. Analogical
abductions in which an explanatory hypothesis is formed by analogy to a
previous explanation are prompted by the same emotional reactions (surprise,
puzzlement) as other abductive inferences.
Are inductive generalizations initiated by emotional reactions? At the least,
emotion serves to focus on what is worth the mental effort to think about
enough to form a generalization. As a social example, if I have no interest in
Albanians, I will probably not bother to form a stereotype that generalizes
about them, whereas if I strongly like or dislike them I will be much more
inclined to generalize about their positive or negative features. I conjecture
that most inductive generalizations occur when there is some emotion-related
interest in the category about which a rule is formed.
It is unfortunate that no one has collected a corpus that records the kinds of
inferences that ordinary people make every day. I conjecture that such a corpus
would reveal that most people make a large number of practical inferences
when decisions are required, but a relatively small number of inductive and
deductive inferences. I predict that deduction is very rare unless people are
engaged in mathematical work, and that inductive inferences are not very
frequent either. A carefully collected corpus would display, I think, only the
occasional inductive generalization or analogical inference, and almost none
of the categorical inductions studied by many experimental psychologists.
Abductive inferences generating causal explanations of puzzling occurrences
would be more common, I conjecture, but not nearly as common as practical
inferences generating decisions. If the inference corpus also recorded the
situations that prompt the making of inferences, it would also provide the
basis for testing my claim that most inference, including practical, abductive,
inductive, analogical, and deductive, is initiated by emotions. For further
discussion of the relation between deduction and induction, see Chapters 10
and 11 in this volume by Rips and Asmuth and by Oaksford and Hahn.

7. 
Some psychological research on inductive inference has pointed to the ten-
dency of people to assume the presence of underlying mechanisms associated
with categories of things in the world (Rehder and Burnett, 2005; Ahn, Kalish,
Medin, and Gelman, 1995). Psychologists have had little to say about what
mechanisms are, or how people use representations of mechanisms in their
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

242 P. Thagard

inferences. In contrast, philosophers of science have been productively ad-

dressing this issue, and the point of this section is to show the relevance of
this understanding of mechanisms to the problem of abductive inference. The
relevance is double, in that the neural structures I have been describing are
clearly mechanisms, and abductive inferences often involve the generation or
application of new hypotheses about mechanisms.
Machamer, Darden, and Craver (2000, p. 3) characterize mechanisms as
“entities and activities organized such that they are productive of regular
changes from start or set-up to finish or termination conditions.” A mech-
anism can also be described as a system whose behavior produces a phe-
nomenon in virtue of organized component parts performing coordinated
component operations (Bechtel and Abrahamsen, 2005). I think these ways
of describing mechanisms are mostly equivalent, and offer my own termino-
logical variant of a mechanism as consisting of a complex <objects, relations,
changes>, consisting of a set of objects (entities, parts) that have properties
and physical relations to each other that cause changes to the properties of
the objects and changes to the relations the objects have to each other and
to the world. For example, a bicycle is a mechanism consisting of wheels,
pedals, and other parts that are connected to each other, with regular changes
to these parts and their relation to the world arising from external inputs
such as a person pedaling and the internal organization of the machine.
Similarly, a neural structure <neurons, connections, spiking behaviors> is
clearly a mechanism where the objects are neurons, their relations are synap-
tic connections, and their changes are spiking behaviors. I conjecture that
whenever people think of categories in terms of underlying mechanisms, they
have something like the pattern of <objects, relations, changes> in mind. In
human minds, mechanisms can be represented verbally, as in “the pedal is
bolted to the frame,” but visual and kinesthetic representations are also used
in science and everyday thinking (Thagard, 2003). Neural structures as well as
the inferences that transform them are clearly mechanisms, and mechanisms
can be mentally represented by combinations of different sorts of neural
structures.
I doubt that all abductive inference is based on representation of mech-
anisms, because abduction requires only a single causal relation, not full
knowledge of a mechanism. If all I know about electric lights is that when
you push the switch, the light comes on, then I can abduce from the fact that
the light is on that someone pushed the switch. But such knowledge hardly
constitutes awareness of a mechanisms because I know nothing about any in-
teracting system of parts. However, when I do know a lot about an underlying
mechanism, my abductive inferences can be much richer and more plausible.
For example, an electrician who knows much about the objects that constitute
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 243

a house’s electrical systems (wires, switches, fuses, etc.) is in a much better

position to explain the normal occurrences and breakdowns of the system.
Similarly, a physician who is familiar with the biological mechanisms that op-
erate in a patient’s body can generate diagnoses about what might have gone
wrong to produce various symptoms. The general structure of mechanism-
based abductive inference is therefore:

Mechanism <objects, relations, changes> is behaving in unexpected ways.

So maybe there are unusual properties or relations of objects that are responsible
for this behavior.

Mechanism-based abduction differs from the simple sort in that people mak-
ing inferences can rely on a whole collection of causal relations among the
relevant objects, not just a particular causal relation.
So far, I have been discussing abduction from mechanisms, in which rep-
resentations of a mechanism are used to suggest an explanatory hypothesis
about what is happening to the objects in it. But abductive inference is even
more important for generating knowledge about how the mechanism works,
especially in cases where its operation is not fully observable. With a bicycle,
I can look at the pedals and figure out how they move the chains and wheels,
but much of scientific theorizing consists of generating new ideas about unob-
servable mechanisms. For example, medical researchers develop mechanistic
models of how the metabolic system works in order to explain the origins
of diseases such as diabetes. Often, such theorizing requires postulation of
objects, relations, and changes that are not directly observed. In such cases,
knowledge about mechanisms cannot be obtained by inductive generalization
of the sort that works with bicycles, but depends on abductive inference in
which causal patterns are hypothesized rather than observed. This kind of
abduction to mechanisms is obviously much more difficult and creative than
abduction from already understood mechanisms. Often it involves analogical
inference in which a mechanism is constructed by comparing the target to be
explained to another similar target for which a mechanism is already under-
stood. In this case, a mechanism <objects, relations, changes> is constructed
by mapping from a similar one. For more on analogical discovery, see Holyoak
and Thagard (1995, ch. 8).
In order to show in more detail how abductive inference can be both to and
from mechanisms, it would be desirable to apply the neurocomputational
model of abductive inference developed by Thagard and Litt (forthcoming).
That model has the representational resources to encode complex objects and
relations, but it has not yet been applied to temporal phenomena involving
change. Hence neural modeling of inferences about mechanisms is a problem
for future research.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

244 P. Thagard

8. 
In sum, abduction is multimodal in that can operate on a full range of per-
ceptual as well as verbal representations. It also involves emotional reactions,
both as input to mark a target as worthy of explanation and as output to signal
satisfaction with an inferred hypothesis. Representations are neural structures
consisting of neurons, neuronal connections, and spiking behaviors. In ab-
duction, the relation between hypotheses and targets is causal explanation,
where causality is rooted in perceptual experience. Inference is transforma-
tion of representational neural structures. Such structures are mechanisms,
and abductive inference sometimes applies knowledge of mechanisms and
more rarely and valuably generates new hypotheses about mechanisms.
Much remains to be done to flesh out this account. Particularly needed
is a concrete model of how abduction could be performed in a system of
spiking neurons of the sort investigated by Eliasmith and Anderson (2003)
and Wagar and Thagard (2004). The former reference contains valuable ideas
about neural representation and transformation, while the latter is useful
for ideas about how cognition and emotion can interact. Thagard and Litt
(forthcoming) combine these ideas to provide a fuller account of the neural
mechanisms that enable people to perform abductive inference. Moving the
study of abduction from the domain of philosophical analysis to the realm of
neurological mechanisms has made it possible to combine logical aspects of
abductive inference with multimodal aspects of representation and emotional
aspects of cognitive processing. We can look forward to further abductions
about abduction.


I am grateful to Jennifer Asmuth, Aidan Feeney, Abninder Litt, and Dou-
glas Medin for helpful comments on an earlier draft. Funding for research
was provided by the Natural Sciences and Engineering Research Council of
Canada.

References
Ahn, W., Kalish, C. W., Medin, D. L., & Gelman, S. (1995). The role of covariation versus
mechanism information in causal attribution. Cognition, 54, 299–352.
Baillargeon, R., Kotovsky, L., & Needham, A. (1995). The acquisition of physical knowl-
edge in infancy. In D. Sperber, D. Premack, & A. J. Premack (Eds.), Causal cognition:
A multidisciplinary debate (pp. 79–116). Oxford: Clarendon Press.
Bechtel, W., & Abrahamsen, A. A. (2005). Explanation: A mechanistic alternative. Studies
in History and Philosophy of Biology and Biomedical Sciences, 36, 421–441.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 245

Blake, R. M., Ducasse, C. J., & Madden, E. H. (1960). Theories of scientific method: The
renaissance through the nineteenth century. Seattle: University of Washington Press.
Blanchette, I., & Dunbar, K. (2001). Analogy use in naturalistic settings: The influence
of audience, emotion, and goals. Memory & Cognition, 29, 730–735.
Charniak, E., & McDermott, D. (1985). Introduction to artificial intelligence. Reading,
MA: Addison-Wesley.
Churchland, P. M. (1989). A neurocomputational perspective. Cambridge, MA: MIT Press.
Damasio, A. R. (1994). Descartes’ error. New York: G. P. Putnam’s Sons.
Darden, L. (1991). Theory change in science: Strategies from Mendelian genetics. Oxford:
Oxford University Press.
Dunbar, K. (1997). How scientists think: On-line creativity and conceptual change in
science. In T. B. Ward, S. M. Smith, & J. Vaid (Eds.), Creative thought: An investi-
gation of conceptual structures and processes (pp. 461–493). Washington: American
Psychological Association.
Eliasmith, C. (2005). Neurosemantics and categories. In H. Cohen & C. Lefebvre
(Eds.), Handbook of categorization in cognitive science (Vol. 1035–1054). Amsterdam:
Elsevier.
Eliasmith, C., & Anderson, C. H. (2003). Neural engineering: Computation, representation
and dynamics in neurobiological systems. Cambridge, MA: MIT Press.
Feldman, J., & Narayan, S. (2004). Embodied meaning in a neural theory of language.
Brain and Language, 89, 385–392.
Flach, P. A., & Kakas, A. C. (Eds.). (2000). Abduction and induction: Essays on their
relation and integration. Dordrecht: Kluwer.
Fugelsang, J. A., Roser, M. E., Corballis, P. M., Gazzaniga, M. S., & Dunbar, K. N.
(2005). Brain mechanisms underlying perceptual causality. Cognitive Brain Research,
24, 41–47.
Gilovich, T., Griffin, D., & Kahneman, D. (Eds.). (2002). Heuristics and biases: The
psychology of intuitive judgment. Cambridge: Cambridge University Press.
Goel, V., & Dolan, R. J. (2003). Reciprocal neural response within lateral and ven-
tral medial prefrontal cortex during hot and cold reasoning. NeuroImage, 20, 2314–
2321.
Gopnik, A. (1998). Explanation as orgasm. Minds and Machines, 8, 101–118.
Hanson, N. R. (1958). Patterns of discovery. Cambridge: Cambridge University Press.
Harman, G. (1973). Thought. Princeton: Princeton University Press.
Holyoak, K. J., & Thagard, P. (1995). Mental leaps: Analogy in creative thought. Cam-
bridge, MA: MIT Press/ Bradford Books.
Houdé, O., Zago, L., Crivello, F., Moutier, F., Pineau, S., Mazoyer, B., et al. (2001).
Access to deductive logic depends on a right ventromedial prefrontal area devoted
to emotion and feeling: Evidence from a training paradigm. NeuroImage, 14, 1486–
1492.
Josephson, J. R., & Josephson, S. G. (Eds.). (1994). Abductive inference: Computation,
philosophy, technology. Cambridge: Cambridge University Press.
Klahr, D. (2000). Exploring science: The cognition and development of discovery processes.
Cambridge, MA: MIT Press.
Kunda, Z. (1999). Social cognition. Cambridge, MA: MIT Press.
Kunda, Z., Miller, D., & Claire, T. (1990). Combining social concepts: The role of causal
reasoning. Cognitive Science, 14, 551–577.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

246 P. Thagard

Lakoff, G. (1987). Women, fire, and dangerous things. Chicago: University of Chicago
Press.
Langley, P., Shrager, J., Asgharbeygi, N., Bay, S., & Pohorille, A. (2004). Inducing explana-
tory process models from biological time series. In Proceedings of the ninth workshop
on intelligent data analysis and data mining. Stanford, CA.
Langley, P., Simon, H., Bradshaw, G., & Zytkow, J. (1987). Scientific discovery. Cambridge,
MA: MIT Press/Bradford Books.
Leslie, A. M., & Keeble, S. (1987). Do six-month-old infants perceive causality? Cognition,
25, 265–288.
Lien, Y., & Cheng, P. W. (2000). Distinguishing genuine from spurious causes: A coher-
ence hypothesis. Cognitive Psychology, 40, 87–137.
Lipton, P. (2004). Inference to the best explanation (2nd ed.). London: Routledge.
Maass, W., & Bishop, C. M. (Eds.). (1999). Pulsed neural networks. Cambridge, MA: MIT
Press.
Machamer, P., Darden, L., & Craver, C. F. (2000). Thinking about mechanisms. Philoso-
phy of Science, 67, 1–25.
Magnani, L. (2001). Abduction, reason, and science: Processes of discovery and explanation.
New York: Kluwer/Plenum.
Mandler, J. M. (2004). The foundations of mind: Origins of conceptual thought. Oxford:
Oxford University Press.
Mellers, B., Schwartz, A., & Ritov, I. (1999). Emotion-based choice. Journal of Experi-
mental Psychology: General, 128, 332–345.
Michotte, A. (1963). The perception of causality (T. R. Miles & E. Miles, Trans.). London:
Methuen.
Nisbett, R. E., & Ross, L. (1980). Human inference: Strategies and shortcomings of social
judgement. Englewood Cliffs, NJ: Prentice Hall.
Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge: Cambridge
University Press.
Peirce, C. S. (1931–1958). Collected papers. Cambridge, MA: Harvard University Press.
Psillos, S. (1999). Scientific realism: How science tracks the truth. London: Routledge.
Read, S., & Marcus-Newhall, A. (1993). The role of explanatory coherence in the con-
struction of social explanations. Journal of Personality and Social Psychology, 65, 429–
447.
Rehder, B., & Burnett, R. C. (2005). Feature inference and the causal structure of
categories. Cognitive Psychology, 50, 264–314.
Rieke, F., Warland, D., de Ruyter van Steveninick, R. R., & Bialek, W. (1997). Spikes:
Exploring the neural code. Cambridge, MA: MIT Press.
Shelley, C. P. (1996). Visual abductive reasoning in archaeology. Philosophy of Science,
63, 278–301.
Sloman, S. A., & Lagnado, D. (2005). The problem of induction. In R. Morrison &
K. Holyoak, J. (Eds.), Cambridge handbook of thinking and reasoning (pp. 95–116).
Cambridge: Cambridge University Press.
Smolensky, P. (1990). Tensor product variable binding and the representation of symbolic
structures in connectionist systems. Artificial Intelligence, 46, 159–217.
Thagard, P. (1988). Computational philosophy of science. Cambridge, MA: MIT
Press/Bradford Books.
Thagard, P. (1992). Conceptual revolutions. Princeton: Princeton University Press.
P1: JZP
0521672443c09 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:0

Abductive Inference 247

Thagard, P. (1999). How scientists explain disease. Princeton: Princeton University Press.
Thagard, P. (2000). Coherence in thought and action. Cambridge, MA: MIT Press.
Thagard, P. (2003). Pathways to biomedical discovery. Philosophy of Science, 70, 235–254.
Thagard, P. (2005). Mind: Introduction to cognitive science (2nd ed.). Cambridge, MA:
MIT Press.
Thagard, P. (2006). Hot thought: Mechanisms and applications of emotional cognition.
Cambridge, MA: MIT Press.
Thagard, P., & Litt, A. (forthcoming). Models of scientific explanation. In R. Sun (Ed.),
The Cambridge handbook of computational cognitive modeling. Cambridge: Cambridge
University Press.
Thagard, P., & Shelley, C. P. (1997). Abductive reasoning: Logic, visual thinking, and
coherence. In M. L. Dalla Chiara, K. Doets, D. Mundici, & J. van Benthem (Eds.),
Logic and scientific methods (pp. 413–427). Dordrecht: Kluwer.
Thagard, P., & Shelley, C. P. (2001). Emotional analogies and analogical inference. In
D. Gentner, K. H. Holyoak, & B. K. Kokinov (Eds.), The analogical mind: Perspectives
from cognitive science (pp. 335–362). Cambridge, MA: MIT Press.
van Fraassen, B. (1980). The scientific image. Oxford: Clarendon Press.
Wagar, B. M., & Thagard, P. (2004). Spiking Phineas Gage: A neurocomputational theory
of cognitive-affective integration in decision making. Psychological Review, 111, 67–79.
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

10

Mathematical Induction and Induction in Mathematics

Lance J. Rips1 and Jennifer Asmuth

However much we may disparage deduction, it cannot be denied that the laws estab-
lished by induction are not enough.
Frege (1884/1974, p. 23)

At the yearly proseminar for first-year graduate students at Northwestern, we

presented some evidence that reasoning draws on separate cognitive systems
for assessing deductive versus inductive arguments (Rips, 2001a, 2001b). In
the experiment we described, separate groups of participants evaluated the
same set of arguments for deductive validity or inductive strength. For ex-
ample, one of the validity groups decided whether the conclusions of these
arguments necessarily followed from the premises, while one of the strength
groups decided how plausible the premises made the conclusions. The re-
sults of the study showed that the percentages of “yes” responses (“yes” the
argument is deductively valid or “yes” the argument is inductively strong)
were differently ordered for the validity and the strength judgments. In some
cases, for example, the validity groups judged Argument A to be valid more
often than Argument B, but the strength groups judged B inductively strong
more often than A. Reversals of this sort suggest that people do not just see
arguments as ranging along a single continuum of convincingness or proba-
bility but instead employ different methods when engaged in deductive versus
inductive reasoning.
Earlier imaging evidence by Goel, Gold, Kapur, and Houle (1997) and
by Osherson et al. (1998) had implicated separate brain regions when
1 This paper was written while the first author was a Fulbright fellow at the Catholic University of
Leuven, Belgium. We thank the Fulbright Foundation and the University for their support. We
also thank Stella Christie, Mike Oaksford, Claes Strannegård, and the editors of this volume for
comments and discussion of an earlier draft. Please send correspondence about this paper to
Lance Rips, Psychology Department, Northwestern University, 2029 Sheridan Road, Evanston,
IL 60208. Email: [email protected].

248
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 249

participants evaluated arguments for validity versus plausibility, and these

earlier data had inspired our own experiment. All these results cast doubt on
the view that there’s a homogeneous “analytic” reasoning system responsible
for correctly solving deductive and probabilistic problems.
But an incident that followed the proseminar alerted us that not every-
one was buying into our reasoning distinctions. A faculty colleague who had
attended the proseminar stopped us later and questioned whether deduc-
tion and induction were as distinct as we had claimed. Wasn’t mathematical
induction a counterexample to the separation between these two forms of
reasoning? Mathematical induction is essential in many deductive proofs in
mathematics, as any high school or college student knows. So how could in-
duction and deduction be isolated subsystems if mathematicians freely used
one to support the other?
At the time, we thought we had a perfectly good answer to our colleague’s
question. Mathematical induction is really nothing like empirical induction
but is instead a deductive technique, sanctioned by the definition of the
natural numbers (and certain other number systems). It’s true that both
types of “induction” often attempt to establish the truth of a generalization
(All cyclic groups are Abelian; All freshwater turtles hibernate). So there is a
loose analogy between them that might suggest labeling both “induction.”
But this is really no more than a case of ambiguity or polysemy. Mathematical
induction in its usual form requires showing that the generalization holds of
a base case (e.g., that the generalization is true for 0) and then showing that if
the generalization is true for an arbitrary number k, it is also true for k + 1.
Hence, by strict deductive reasoning, the generalization must be true for all
natural numbers (0, 1, 2, . . . ). The conclusion of such a proof shares with all
deductively valid conclusions the property that it is necessarily true or true in
all possible worlds in which the givens are true.
A typical high school example of mathematical induction is the proof
that the sum of the first n natural numbers is 1/2 n (n + 1). Certainly this
relationship holds when n = 0. If we assume that the relationship holds for
the first k natural numbers (i.e., the sum of 0 through k is 1/2 k (k + 1)), then
the sum of the first k + 1 numbers must be:

k+1
k(k + 1)
i= + (k + 1)
i =0
2
k 2 + 3k + 2
=
2
(k + 1)(k + 2)
=
2
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

250 L. J. Rips and J. Asmuth

The last expression is also of the form 1/2 n (n + 1). So this sum formula
necessarily holds for all natural numbers.
By contrast, empirical induction leads to conclusions that are not nec-
essarily true but only probably true, even when these conclusions involve
mathematical relationships. For example, we might hope to establish induc-
tively that people forget memories at a rate equal to y = b − m ln(t), where
t is time since the remembered event and m and b are constants. To support
such a conclusion, we might examine forgetting rates for many people and
many event types and show that forgetting follows the proposed function in a
convincing range of cases (Rubin & Wenzel, 1996). Even if the function fit the
data perfectly, though, no one would suppose that the evidence established
the conclusion as necessarily true. It’s not difficult to imagine changes in our
mental make up or changes in the nature of our everyday experience that
would cause us to forget events according to some other functional form. The
best we can hope for is a generalization that holds for people who are similar
to us in causally important ways (see Heit’s Chapter 1 in this volume for an
overview of empirical induction).
We were able to convince our colleague that this line of thinking was correct.
Or perhaps he was just too polite to press the point. In fact, we still believe
that this standard answer isn’t too far off the mark. But we also think that
we dismissed our colleague’s objection too quickly. One worry might be that
the theoretical distinction between math induction and empirical induction
is not as clear as we claimed. And, second, even if the theoretical difference
were secure, it wouldn’t follow that the psychological counterparts of these
operations are distinct. In this chapter we try to give a better answer to the
objection by examining ways that induction could play a role in mathematics.
In doing so, we won’t be presenting further empirical evidence for a dissocia-
tion between inductive and deductive reasoning. We will be assuming that the
previous evidence we’ve cited is enough to show a difference between them
in clear cases. Our hope is to demonstrate that potentially unclear cases –
ones in which people reach generalizations – don’t compromise the distinc-
tion (see Oaksford & Hahn, Chapter 11 in this volume, for a more skeptical
view of the deduction/induction difference).
We consider first the relationship between math induction and empirical
induction to see how deep the similarity goes. Although there are points of
resemblance, we argue that there are crucial psychological differences between
them. This isn’t the end of the story, however. There are other methods of
reaching general conclusions in math besides math induction, and it is possible
that one of them provides the sort of counterexample that our colleague was
seeking. One such method involves reasoning from an arbitrary case – what is
called universal generalization in logic. Several experiments suggest that people
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 251

sometimes apply this method incorrectly, and we examine these mistakes for
evidence of a continuum between inductive and deductive methods. We
claim, however, that although the mistakes show that people use inductive
procedures when they should be sticking to deductive ones, they provide no
evidence against a qualitative deduction/induction split.
Of course, we won’t be claiming that there’s no role for induction in math-
ematics. Both students and professional mathematicians use heuristics to
discover conjectures and methods that may be helpful in solving problems
(see Polya, 1954, for a classic statement on the uses of induction in mathemat-
ics). Our claim is simply that these inductive methods can be distinguished
from deductive proof, a point on which Polya himself agreed: “There are two
kinds of reasoning, as we said: demonstrative reasoning and plausible reason-
ing. Let me observe that they do not contradict each other; on the contrary
they complete each other” (Polya, 1954, p. vi).

    

In their The foundations of mathematics, Stewart and Tall (1977) provide an
example of a proof by induction similar to the one we just gave of the sum
formula. They then comment:
Many people regard this as an ‘and so on . . . ’ sort of argument in which the truth
of the statement is established for n = 1; then, having established the general
step from n = k to n = k + 1, this is applied for k = 1 to get us from n = 1 to
n = 2, then used again to go from n = 2 to n = 3, and so on, as far as we wish to
go . . . The only trouble is that to reach large values of n requires a large number
of applications of the general step, and we can never actually cover all natural
numbers in a single proof of finite length. (p. 145)
This remark echoes Russell (1919, p. 21) in a similar context: “It is this ‘and
so on’ that we must replace by something less vague and indefinite.”
Perhaps this feeling of “and so on . . .” is what was bothering our colleague.
If there is no more to a proof by induction than a sort of promissory note
that we can keep going, then perhaps such a proof isn’t much different from
ordinary empirical induction. Just as we can’t hope to enumerate all cases
that fall under an inductive generalization in science, we can’t hope to cover
all natural numbers to which a mathematical generalization applies. Mathe-
matical induction may only be able to give us a boost in confidence that the
generalization holds in all cases, not an iron-clad proof.
Stewart and Tall (1977), however, offer a solution to show that mathematical
induction is a rigorously deductive technique after all. “The way out of this
dilemma is to remove the ‘and so on . . .’ part from the proof and place it
squarely in the actual definition of the natural numbers” (p. 145). What
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

252 L. J. Rips and J. Asmuth

follows in Stewart and Tall is an exposition of a variation on the Dedekind-

Peano axioms for the natural numbers, with the following induction axiom:
(IAx) If S ⊆ N0 is such that 0 ∈ S; and n ∈ S ⇒ s (n) ∈ S for all n ∈ N0 , then
S = N0 ,
where S is a set, N0 are the natural numbers, and s (·) is the successor function
(i.e., the function that yields n + 1 for any natural number n). For example,
in the sample proof we gave earlier, S corresponds to the set of all natural
numbers n such that the sum of 0 through n is 1/2 n (n + 1), and the two parts
of the proof correspond to the two clauses in (IAx). We first showed that 0 is in
this set (0 ∈ S) and then that if n is in the set, so is n + 1 (n ∈ S ⇒ s (n) ∈ S).
The conclusion is therefore that the sum formula is true for all natural num-
bers (S = N0 ). Since (IAx) is an axiom, this conclusion now has a kind of
deductive legitimacy.2

Worry Number 1: The Justification of (IAx)

Stewart and Tall are right that (IAx) speaks to one sort of doubt about math
induction. “And so on . . .” complaints about our proof of the sum formula
could center around the fact that what’s been demonstrated is a universal
statement of the form: For all natural numbers n, if n satisfies the sum formula,
so does n + 1. This generalization, together with the fact that 0 satisfies the
formula, allows you to conclude by a modus ponens inference that 1 satisfies
the formula:
(∀n) Sum-formula(n) ⊃ Sum-formula(n + 1).
Sum-formula(0).
---------------------------
Sum-formula(1).
This conclusion in turn allows you to make another modus ponens inference
to 2:
(∀n) Sum-formula(n) ⊃ Sum-formula (n + 1).
Sum-formula(1).
---------------------------
Sum-formula(2).

2 Use of math induction as a proof technique was, of course, well established before the intro-
duction of (IAx) as a formal axiom. Medieval Arabic and Hebrew sources use implicit versions
of math induction to prove theorems for finite series and for combinations and permutations
(Grattan-Guinness, 1997). According to Kline (1972, p. 272), “the method was recognized ex-
plicitly by Maurolycus in his Arithmetica of 1575 and was used by him to prove, for example,
that 1 + 3 + 5 + · · · + (2n − 1) = n2 .”
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 253

And so on. In general, this modus ponens strategy requires k modus ponenses
to conclude that k satisfies the sum formula, and it will be impossible to
reach the conclusion for all natural numbers in a finite number of steps.
(IAx) sidesteps this worry by reaching the conclusion for all n in one step.
However, putting the “and so on . . . ” idea “squarely in the actual definition
of the natural numbers,” as Stewart and Tall suggest, isn’t a very comforting
solution if this means hiding the “and so on . . . ” pea under another shell.
What justifies such an axiom?
There are several ways to proceed. First, we could try to eliminate the “and
so on . . . ” doubt by showing that (IAx) is inherently part of people’s thinking
about the natural numbers and, in this sense, is “squarely in” their definition.
According to this idea, any kind of intuitive understanding of these numbers
already allows you to conclude in one go that if 0 is in S and if k + 1 is in S
whenever k is, then all natural numbers are in S. This is close to Poincaré’s
(1902/1982) view. Although he acknowledged that there is “a striking analogy
with the usual procedures of [empirical] induction,” he believed that math-
ematical induction “is the veritable type of the synthetic a priori judgment.”
By contrast with empirical induction, mathematical induction “imposes itself
necessarily, because it is only the affirmation of a property of the mind itself ”
(pp. 39–40). Thus, any attempt to found math induction on a formal axiom
like (IAx) can only replace an idea that is already intuitively clear with one
that is more obscure. Goldfarb (1988) points out that the psychological na-
ture of math induction was irrelevant to the goals of those like Frege, Hilbert,
and Russell, who were the targets of Poincaré’s critique (see Hallett, 1990, for
debate about these goals). But perhaps Poincaré’s theory is sufficient to settle
specifically psychological doubts. Another possibility, one more consistent
with Frege’s (1884/1974) and Russell’s (1919) viewpoint, is that (IAx) is the
result of a principle that applies in a maximally domain-general way and is
thus a precondition of rationality. We’ll look more closely at the source of this
generality in just a moment in examining Frege’s treatment of (IAx).
We favor a view in which math induction is inherent in the concept of the
natural numbers – whether this is specific to the natural-number concept or
is inherited from more general principles – and this view preserves a crisp
distinction between deductive and inductive reasoning. But can we really
rule out the possibility that the “and so on . . .” doubt is a symptom of a true
intermediate case? Maybe (IAx) merely plasters over a gap like that in ordinary
induction – the “and so on . . .” gap that is left over from the modus ponens
strategy.
One thing that seems clear is that in actually applying math induction –
for example, in our use of it to prove the sum formula – the modus ponens
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

254 L. J. Rips and J. Asmuth

strategy is never performed. In our original proof, for instance, we didn’t

(explicitly or implicitly) grind through the first few steps of the strategy (from
0 to 1 to 2 . . .) before concluding that the sum formula is true for all n. So
the way in which the modus ponens strategy gives us reason to believe in
math induction must be indirect. Perhaps by reflecting on the modus ponens
strategy, we somehow come to see that this strategy is unnecessary and that we
can use the short-cut math induction procedure in its place. But if this is true,
it can’t be the success of the modus ponens strategy that convinces us, since
this strategy is an infinite process that is never completed. We don’t learn
math induction from the modus ponens strategy in the usual generalizing
or transferring sort of way. Instead, it must be our realizing that the modus
ponens strategy could potentially yield the right result under some idealized
conditions (e.g., infinite time and other resources) that persuades us that
math induction will work. If we push further and ask what’s responsible
for such a realization, however, the answer seems to lead right back to our
understanding of the structure of the numbers themselves. At best, the modus
ponens strategy plays a mediating role in reminding us of a property of the
natural number system but provides no independent justification.
Why is (IAx) inherent in the number concept? One way to see this comes
from Frege’s (1984/1974) treatment of math induction. (The modernized ver-
sion that follows is from Quine, 1950, chap. 47. A somewhat similar attempt
appears in Dedekind, 1888/1963, especially Theorems 59 and 89.) We can
think of the natural numbers as a set that contains 0 and that contains all im-
mediate successors of each of its members. Such a set, however, could contain
other objects besides numbers. It could also contain, for example, Frege’s hat,
since Frege’s hat does not have a successor. What we need to define the natural
numbers is a minimality rule that says that the natural numbers are the small-
est set containing 0 and containing all immediate successors of its members.
Put slightly differently, the natural numbers are the elements that belong to
all such sets, including the set containing only 0 and the successors of (the
successors of . . . ) 0. This definition of the natural numbers appears in (1):

(1) n ∈ N0 ↔ (∀u) ((0 ∈ u & (∀k) (k ∈ u ⊃ s (k) ∈ u)) ⊃ n ∈ u),

where N0 is the set of natural numbers and s (·) is the successor function, as
before.
From this definition, mathematical induction follows, just as Stewart and
Tall (1977) claimed. For suppose S is a set that fulfills the two conditions on
(IAx), as in (2):

(2) 0 ∈ S & (∀k) (k ∈ S ⊃ s (k) ∈ S).

P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 255

If n is a natural number, (3) then follows from the definition in (1):

(3) (∀u) ((0 ∈ u & (∀k) (k ∈ u ⊃ s (k) ∈ u)) ⊃ n ∈ u),
and in particular (i.e., by instantiating u to S):
(4) (0 ∈ S & (∀k) (k ∈ S ⊃ s (k) ∈ S)) ⊃ n ∈ S.
It then follows by modus ponens from (2) and (4) that n ∈ S. Hence, every
natural number is an element of S.
Frege’s method helps to bring out what’s important about (IAx) as a part
of the natural-number concept. In using (IAx) to prove a theorem, as we did
with the sum formula, we take an “insider’s view” of the number system in
which its familiar properties are already in place. For these purposes, (IAx) is
the usual tool we pick up in high school algebra. But to see the role that (IAx)
plays in the definition of the natural numbers – the reason that it appears
among the Dedekind-Peano axioms – we need to adopt an “outsider’s view.”
From this perspective, (IAx) is a kind of closure principle. It says, in effect,
that nothing can be a natural number that isn’t either 0 or a successor of (a
successor of . . . ) 0. This is the importance of the minimality constraint we’ve
just looked at. For these purposes, we could substitute for (IAx) other rules
that have equivalent effects, given the rest of the axioms – rules that may
seem less artificial and less similar to empirical induction. An example is the
Least-Number Principle (see Kaye, 1991) that we can spell out in a way that
can be compared to (IAx):
(LNP) If S ⊆ N0 and n ∈ S, then there is an m ∈ S such that for all k < m, k ∈
/ S.
In other words, if S is a nonempty set of the natural numbers, there’s a smallest
number in S. For other number systems, other types of math induction may
be required (see, e.g., Knuth, 1974).

Worry Number 2: The Justification of (UG)

But there could be another hang up about proofs like the example we gave for
the sum formula. As part of the proof of this formula, we showed that for any
particular k, if k satisfies the sum formula then so does k + 1. The second sort
of doubt is whether this is enough to prove that the same is true for all natural
numbers. Clearly, (IAx) is completely powerless to quell this second kind of
doubt. The axiom requires us to show that n ∈ S entails s (n) ∈ S for all
natural numbers n, and this is precisely where the second sort of doubt arises.
This kind of doubt surfaces in some psychological treatments of math
induction. For example, in one of the very few empirical studies on this
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

256 L. J. Rips and J. Asmuth

topic, Smith (2002, p. 6) cites a definition similar to (IAx) and then goes
on to remark, “The definition secures a generalization (universalization).
The reference in the premises is to a property true of any (some particular)
number. But the reference in the conclusion is to any number whatsoever,
any number at all. Note that the term any is ambiguous here since it covers
an inference from some-to-all.” The word any does not, in fact, appear in the
definition that Smith cites. The relevant clause is “ . . . and if it is established
that [the property] is true of n + 1 provided it is true of n, it will be true
of all whole numbers.” But what Smith seems to mean is that the proof is a
demonstration for some particular case, k and k + 1, whereas the conclusion
is that the property is true for all natural numbers n. Similarly, in discussing
earlier work by Inhelder and Piaget, Smith states:

[mathematical induction] runs from what is shown to be true of a specified

number n (n’importe quel nombre) and what in that case is also shown to be true
of its successor (n + 1) to a conclusion about any number whatsoever (nombre
quelconque). It is just such an inference which is at issue in mathematical induction
and it was Inhelder and Piaget’s (1963) contention that their evidence showed
that children were capable of this type of reasoning (p. 23).

The inference that Smith is pointing to, however, is not a feature of math
induction per se but of the universal generalization that feeds it. Universal
generalization, roughly speaking, is the principle that if a predicate can be
proved true for an arbitrary member of a domain, it must be true for all
members. A formal statement of this rule can be found in most textbooks
on predicate logic. We have adapted slightly the following definition from
Lemmon (1965):

(UG) Let P (e) be a well-formed formula containing the arbitrary name e, and
x be a variable not occurring in P (e); let P (x) be the propositional function in
x which results from replacing all and only occurrences of e in P (e) by x. Then,
given P (e), UG permits us to draw the conclusion (∀x)P (x), provided that e
occurs in no assumption on which P (e) rests.

For example, when we proved that the sum formula holds for k + 1 if it
holds for k, we used no special properties of k except for those common to
all natural numbers. In this sense, k functioned as an arbitrary name. Hence
(by universal generalization), for all natural numbers n, the sum formula
holds of n + 1 if it holds for n. This generalization is needed to satisfy the
second condition of (IAx), but it is also needed to supply the major premises
for the modus ponens strategy and, of course, for much other mathematical
reasoning. The confusion between (IAx) and (UG) is an understandable one.
We could rephrase (IAx) in a way that refers to an arbitrary case: if S is a
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 257

subset of the natural numbers, if 0 is an element of S, and if whenever an

arbitrary natural number k is in S so is k + 1, then S includes all natural
numbers. This rephrasing is valid, but it combines (IAx) and (UG) in a way
that obscures their separate contributions.
(UG), like (IAx), may seem to lend deductive legitimacy to what looks
at first like empirical induction. Unlike the case of (IAx), though, there’s
not much temptation to see this rule as sanctioning a kind of induction-by-
enumeration. Informal use of mathematical induction seemed to fall short
because it required an infinite number of modus ponens inferences, and (IAx)
was introduced to avoid this. But the informal use of universal generalization
in mathematics doesn’t have the same kind of “and so on . . .” gap. Once
we’ve shown that P is true for an arbitrary element e,we needn’t apply this
result to get P (0), P (1), P (2), and so on. In general, informal universal
generalization (e.g., in the proof of the sum formula) doesn’t raise the same
kind of sophisticated issues about the nature of proof that informal use of math
induction does (although it does take some sophistication to formulate (UG)
correctly for a logic system, as the definition from Lemmon, 1965, suggests).
Instead, the worry about (UG) stems from its similarity to an inductive
heuristic in which you generalize from a specific number. (Kahneman and
Tversky’s [1972] representativeness heuristic is a related example in the con-
text of probability estimation.) In the case of the sum formula, for instance,
it’s clearly not mathematically kosher to proceed like this:

Let’s pick an arbitrary number, say, 7.

Simple calculations show that 0 + 1 + 2 + 3 + 4 + 5 + 6 + 7 = 28 and 1/2 × 7 ×
(7 + 1) = 28.
Hence, i = 1/2 n (n + 1).

An example of this sort may increase your confidence that the sum formula
is true, but it doesn’t produce a deductively valid proof. However, there’s an
intuitive kinship between this pseudo-proof and the earlier correct proof of the
sum formula. The phrase arbitrary number captures this kinship, since it can
mean either a specific number that we select in some haphazard way or it can
mean a placeholder that could potentially take any number as a value. Sliding
between these two meanings yields an inductive, example-based argument,
on one hand, and a deductively valid (UG)-based proof, on the other. For the
sum formula, it’s the difference between the argument from seven that we just
gave and the correct argument with k.
(UG) is pervasive in mathematics, in part because most algebraic ma-
nipulations depend on it. Variables like k turn up everywhere, not just in
the context of mathematical induction. Thus, confusion about the nature of
variables is apt to have wide-spread consequences and blur the distinction
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

258 L. J. Rips and J. Asmuth

between inductive arguments and deductive proof. We consider the empirical

evidence for this confusion in the next section, including actual pseudo-proofs
from students that are quite similar to our argument from seven. We then
return to the issue of whether this evidence shows that the difference between
deduction and induction in math is merely one of degree.

  
In Smith’s (2002) study of children’s arithmetic, the key question to partici-
pants was: “If you add any number [of toys] at all to one pot and the same
number to the other, would there be the same in each, or would there be
more in one than in the other?” A correct answer to this question seems to
depend on the children realizing that for all numbers x and quantities m
and n, x + m = x + n if and only if m = n. In the experiment, m and n are
the initial quantities of toys in the pots (m = n = 0 in one condition of the
study, and m = 1 and n = 0 in another). It is certainly possible to prove this
principle using (IAx), but it is far from clear that this is the way children were
approaching the problem. Reasoning along the lines of (UG) would mean
going from “for arbitrary k, (m + k = n + k) if and only if (m = n)” to the
above conclusion for all x. But again it is unclear that this is how children
went about answering this question. Among the children’s justifications that
Smith quotes on this problem, the only one that directly suggests (UG) is “I’m
putting like a number in the orange pot and a number in the green pot, and
it’s got to be the same.”
There is plenty of reason, however, to suspect that children in this experi-
ment were using specific examples to justify their decisions. Some simply cited
a specific number, for example, “If you added a million in there, there would
be a million, and if you added a million in there, there would be a million,”
or “it would be just a million and six in there and a million and six in there.”
Others gave a more complete justification but also based on a specific number
of items: “If we started putting 0 in there and 1 in there, and then adding 5
and 5, and so it wouldn’t be the same – this would be 5 and 6.” Instead of the
(UG) strategy of thinking about an arbitrary addition, these children seem
to generalize directly from a particular number, such as 5 or 1,000,006, to
all numbers. The numbers may be selected arbitrarily but they aren’t capable
of denoting any natural number in the way (UG) requires. It is possible, of
course, that these children were using a correct general strategy and simply
citing specific cases to aid in their explanations, but the experiments we are
about to review suggest than even college students evaluate “proof by specific
examples” as a valid method.
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 259

Variations on this kind of instance-based strategy are probably common

in mathematics. One traditional way of thinking about this is that people use
these strategies as heuristics to suggest the right generalization before they
apply more rigorous mathematical methods to produce a genuine proof. This
is similar to the distinction from Polya (1954) that we mentioned earlier (see,
also, Thagard’s Chapter 9 on abduction in this volume). There might be a
“context of discovery” in which such informal methods are appropriate, and
a “context of justification” where only proof-based methods such as (IAx)
and (UG) apply. We were once given the advice that you should try to prove
a theorem only when you were 95% sure it was right. This advice separates
neatly the informal, confidence-based aspect of mathematics from its formal,
proof-based side. So does the usual classroom advice to experiment with
different values to find a formula before you try to prove it with (IAx).
But whatever validity the discovery/justification division has in mathemat-
ics, it is clear that students do not always honor it. Even when asked to prove
a proposition (so that justification is required), students often produce what
look to be instance-based strategies. We give some examples here from exper-
iments on college students and then try to draw some implications about the
relation between induction and deduction in math contexts.

Algebra
Asking college student to prove simple propositions in number theory often
results in “proof by multiple examples,” such as the following one (Eliaser,
2000):
Show that every odd multiple of 5 when squared is not divisible by 2.
Proof:
Odd multiples of 5: 5, 15, 25 . . .
52 = 25 → not divisible by 2.
152 = 225 → not divisible by 2.
Therefore shown.
In the experiment from which we took this excerpt, Northwestern University
undergraduates had to prove a group of simple statements, half of which had
a universal form, like the one above, requiring an algebraic proof. Although
nearly all these students were able to use algebra (as they demonstrated in a
post test), only 41% initially employed it in proving universal propositions.
Instead, 33% used multiple examples (as in the sample “proof ”) and 17%
used single examples.
Of course, students may not always see the right proof strategy immediately
and may fall back on examples instead. For instance, the students may not
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

260 L. J. Rips and J. Asmuth

remember or may not know how to set up a problem that requires represent-
ing the parity of a number (i.e., that all even numbers have the form 2x and
all odd ones 2x + 1). But this doesn’t seem to be the whole story behind the
students’ dismal performance. In a second study, Eliaser (2000) gave potential
proofs to a new group of undergraduates and asked them to rate the quality
of the proof on a four-point scale from “shows no support” to “completely
proves.” The “proofs” included single-example, multiple-example, and alge-
braic versions. But despite the fact that students didn’t have to produce their
own proofs and only had to evaluate the proofs of others, they nevertheless
rated multiple-example pseudo-proofs higher than either single-example or
(correct) algebraic proofs, for universal claims, such as the one about odd
multiples of 5.
Depressing performance of a similar kind comes from a study of students
who were training to be elementary math teachers (Martin & Harel, 1989).
These students rated proofs and pseudo-proofs for two number-theoretic
statements (e.g., If a divides b and b divides c , then a divides c ). The pseudo-
proofs included a typical example (e.g., “12 divides 36, 36 divides 360, 12
divides 360”) and an example with a big, random-looking number:

Let’s pick any three numbers, taking care that the first divides the second, and the
second divides the third: 49 divides 98, and 98 divides 1176. Does 49 divide 1176?
(Computation shown to left.) The answer is yes.

The pseudo-proofs also included an instantiation of the more general, correct

proof:

Take 4, 8, and 24. 4 divides 8, which means that there must exist a number, in
this case 2, such that 2 × 4 = 8. 8 divides 24, which means there must exist a
number, in this case 3, such that 3 × 8 = 24. Now substitute for 8 in the previous
equation, and we get 3 × (2 × 4) = 24. So we found a number (3 × 2), such that
(3 × 2) × 4 = 24. Therefore, 4 divides 24.

Students rated these pseudo-proofs, along with other pseudo- and correct
proofs on a scale from 1 (not considered a proof) to 4 (a mathematical proof).
If we take a rating of 4 to be an endorsement of the argument as a proof, then
28% endorsed the correct proof, 22% the instantiated pseudo-proof, 39% the
specific example, and 38% the big random-looking example.

Geometry
Students’ use of example-based proof strategies isn’t limited to algebra.
It’s a common classroom warning in high school geometry that students
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 261

shouldn’t draw general conclusions from the specific cases that they see in the
diagrams accompanying a problem. Diagrams represent one way in which
the givens of a problem could be true, not all possible ways. Despite these
warnings, students do generate answers to test questions based on single
diagrams.
Evidence on this point comes from a study by Koedinger and Anderson
(1991). This study presented students with yes/no questions about geometric
relationships, accompanied by one of two diagrams. For example, one prob-
lem asked, “If ∠BCD = ∠ACD, must AB ⊥ CD?” The two possible diagrams
for this problem appear as Figures 10.1a and 10.1b. For half the problems
the correct answer was “yes,” as in this example. The remaining problems
had “no” as the correct answer, for instance, “If AC = CB, must AB⊥CD?”
The two possible diagrams for this problem are Figures 10.1c and 10.1d. For
each of these pairs of problems, the given information and the conclusion
were true of one diagram (Figures 10.1a and 10.1c) and were false of the
other (Figures 10.1b and 10.1d). Individual participants saw only one of the
possible diagrams per problem.
The results showed that students were usually able to get the right answer:
they were correct on about 68% of trials. For problems where the correct
answer was “yes” (so that a proof was possible), the students performed more
accurately when the corresponding proof was easy than when it was difficult,
suggesting that they were attempting such a proof. Nevertheless, they tended to
be swayed by the diagrams: they made more “yes” responses in all conditions
when the diagram pictured the given information and the conclusion (as
in Figures 10.1a and 10.1c) than when it didn’t (Figures 10.1b and 10.1d).
Koedinger and Anderson take this pattern of results to mean that the students
were attempting to find a proof but were using the diagram as a backup
inductive strategy if no proof was forthcoming (a “modified misinterpreted
necessity model” in the jargon of reasoning research; see Evans, Newstead, &
Byrne, 1993).
Further evidence that students rely on diagrams for inductive support
comes from one of our own studies of non-Euclidean geometry. In this
experiment, we gave participants a short introduction to hyperbolic geometry,
which differs from Euclidean geometry by the following hyperbolic axiom: for
any given line  and point p not on , there are multiple lines through p
parallel to . The participants then solved a number of problems in the new
geometry while thinking aloud. Figure 10.2 illustrates one such problem. To
avoid confusing participants about the meaning of parallel, we substituted the
nonsense term cordian, which we defined according to the hyperbolic axiom.
For the problem in Figure 10.2, we told participants that lines  and m are
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

262 L. J. Rips and J. Asmuth

a. D

If ∠ BCD = ∠ ACD,
must AB ⊥ CD?

A B
C

b.
D
If ∠ BCD = ∠ ACD,
must AB ⊥ CD?

A B
C

c. D

If AC = CB ,
must AB ⊥ CD?

A B
C

d.
D
If AC = CB ,
must AB ⊥ CD?

A B
C
figure 10.1. Stimulus problems from a study of geometry problem solving (adapted
from Koedinger and Anderson, 1991, Figure 2).

cordian, that ∠BAC and ∠BDC were 90˚ angles, and asked them to decide
whether triangle ABC was congruent to triangle DCB.3

3 If the two triangles ABC and DCB are congruent, corresponding angles of these triangles must
also be congruent. In Figure 10.2, ∠BAC and ∠CDB are given as congruent, but what about the
pair ∠DCB and ∠ABC and the pair ∠BCA and ∠CBD? In Euclidean geometry, the congruence
of each of these pairs follows from the fact that alternate interior angles formed by a transversal
and two parallel lines are congruent. In hyperbolic geometry, however, the alternate interior
angles formed by a transversal and two parallel (“cordian”) lines are not congruent. For example,
in Figure 10.2, ∠ABC and ∠DCB are both the alternate interior angles formed by the cut of
transversal BC across parallel lines  and m. But because these corresponding angles cannot be
proved congruent, neither can the two triangles.
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 263

B
D l

A C
figure 10.2. Stimulus problem for a study of non-Euclidean geometry.

Diagrams for non-Euclidean problems are apt to be misleading. Although

there are several conventions for drawing such diagrams (such as the one
adopted in Figure 10.2; see Greenberg, 1993), they can nevertheless suggest
Euclidean relationships that don’t hold in the non-Euclidean domain. Partici-
pants quickly recognize this difficulty, but they still fall back on the appearance
of a diagram. The following transcript is from a participant working on the
problem in Figure 10.2:
Must be true, ABC is congruent to DCB. If ∠BAC is 90, must they be congruent?
It does not mention a line will diverge, so how can I judge how far apart these
lines will be at any given point, since they’re curving? Angle ABD is split, maybe in
half, if you dissect it, it appears as though it’s dissected . . . I have no information
about the angles ACD or ABD. I must assume that there can be less than 360˚ [in
a quadrilateral]. That the angle B could be equal to angle C. Angle B equals angle
ACD. Therefore, if those angles are equal, and the line joining them is drawn,
angle ABC is equal to angle DBC. Okay. The dotted line will be line q . Line
segment q bisects angle ABD. . . . Angle, angle, angle . . . theorem states that ABD
is congruent with – ABC is congruent with DCB. Must be true? Yes.
Although this participant seems initially unsure of the angle relationships,
he or she is gradually seduced by the appearance that ∠ABD and ∠ACD are
equal and that the dashed line bisects both these angles.

Implications
These examples show that people sometimes use inductive strategies where de-
ductive methods are necessary. More surprisingly, people sometimes evaluate
the inductive methods more favorably than the correct deductive ones. We
can explain the first of these findings on the assumption that people resort to
inductive methods when they run out of deductive ones – either because no
proof is possible or because they can’t find one – as Koedinger and Anderson
(1991) propose. But the evaluation results make it likely that this isn’t the
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

264 L. J. Rips and J. Asmuth

whole story. Why do people find inductive pseudo-proofs better than true
proofs when they’re able to assess them directly?
The answer may lie in the relative complexity of the true proofs versus the
inductive lures. Examples are likely to be easier to understand than proofs,
especially if the proofs are lengthy. Similarly, there’s a difference in concrete-
ness between proofs with individual numbers and those with arbitrary names
(i.e., mathematical variables) in algebra. Geometry diagrams also have a con-
creteness advantage over the proofs that refer to them. Even among correct
deductive proofs, we evaluate the simpler, more elegant ones as superior to
their more complex competitors, and it is possible that students’ incorrect
evaluations are a more extreme version of the same tendency.
Explanations along these lines do no harm to the thesis that deductive and
inductive reasoning make use of qualitatively different mechanisms. These
explanations take the induction/deduction difference for granted and account
for students’ errors on the basis of factors, such as complexity or abstractness,
that swamp deductive methods or make them seem less attractive. We know
of no experimental results that can eliminate these explanations. But it’s
also interesting to consider the possibility of a more continuous gradient
between inductive and deductive strategies. One proposal along these lines
might be based on the idea that people search for increasingly nonobvious
counterexamples to an argument or theorem. If they are unable to find any
counterexamples, either because there are none or because the individuals
can no longer hold enough examples in working memory, they declare the
argument valid. In the case of algebra, we can imagine students first choosing
a specific number that the problem suggests and that meets the theorem’s
givens. For example, in the problem mentioned earlier of showing that the
squares of odd multiples of 5 are not divisible by 2, the participant began
by verifying that 52 is not divisible by 2. If the first instance does not yield
a counterexample, the students try another (e.g., (3 × 5)2 ) until they either
discover a counterexample or run out of working-memory resources. In the
case of geometry, students might first verify that the conclusion is true of
the diagram that appears with the problem. They may then consider new
diagrams of their own in the same search for counterexamples.
In the present context, one difficulty with this line of thinking, however,
is that it doesn’t account for Koedinger and Anderson’s (1991) evidence that
students also attempt to apply a separate deductive strategy and resort to
diagrams only when the first strategy fails. The counterexample hypothesis
predicts instead that students should begin by considering the given diagram,
look for other diagrams that might falsify the conclusion, and end by accepting
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 265

the theorem as valid if it correctly applies to all the diagrams they have
examined. In addition, there is a theoretical problem. In the case of a universal
proposition – such as the sum formula that applies to all natural numbers –
there’s no endpoint at which you can stop and be sure that no counterexamples
will be found. There will always be additional numbers that have features
other than the ones you’ve so far considered. Rather than provide a continuum
between inductive and deductive reasoning, this strategy never provides more
than weak inductive support. Of course, the fact that a pure counterexample
strategy is inadequate as an explanation of the data doesn’t mean that there
aren’t other methods intermediate between induction and deduction that
may be more reasonable. But we know of no convincing cases of this type.
Another way to put this point is to say that there is no number or diagram
that is “completely arbitrary” in a way that could prove a universal proposition.
“Each number has its own peculiarities. To what extent a given particular
number can represent all others, and at what point its own special character
comes into play, cannot be laid down generally in advance” (Frege, 1884/1974,
p. 20). The sort of arbitrariness that does prove such a proposition – and that
(UG) demands – is a matter of abstractness, not randomness or atypicality.
Arbitrariness in the latter sense drives you to more obscure cases as potential
counterexamples. But the sort of arbitrariness that is relevant to proving a
universal proposition pulls in the opposite direction of inclusiveness – an
instance that could potentially be any number.

  

We now have a more complicated answer to give to our colleague. Our original
response was that mathematical induction is a deductive technique with
a confusing name. When people use math induction to prove a theorem,
they aren’t using induction and they aren’t sacrificing deductive validity; the
theorem is still necessarily true, just as with other standard proof techniques.
Although there is debate about the role of probabilistic proofs in mathematics
(see Fallis, 1997) – methods that truly are inductive – math induction isn’t
such a method.
But while we still think this is correct, math induction does loosely resemble
empirical induction. From a formal point of view, there is a tension between
math induction and the usual constraint that proofs be finite. Math induction
seems to rely on an infinite process of iterating through the natural numbers –
the modus ponens strategy that we discussed earlier. From a psychological
point of view, of course, no such infinite process is remotely possible. It’s this
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

266 L. J. Rips and J. Asmuth

gap that creates the similarity with empirical induction, since in both cases
the conclusion of the argument seems beyond what the premises afford. An
induction axiom like (IAx) makes up for this gap in formal contexts, but why
is math induction a deductively correct proof tool if it can’t be backed up by
the infinite iteration it seems to require?
We’ve argued elsewhere (Rips, Bloomfield, & Asmuth, 2007) that math
induction is central to knowledge of mathematics: it seems unlikely that people
could have correct concepts of natural number and other key math objects
without a grip on induction. Basic rules of arithmetic, such as the associative
and commutative laws of addition and multiplication, are naturally proved
via math induction. If this is right, then it’s a crucial issue how people come
to terms with it. We’ve tried to argue here that the analogy we’ve just noted –
enumeration : conclusion by empirical induction :: modus ponens strategy :
conclusion by math induction – is misleading. Math induction doesn’t get its
backing from the modus ponens strategy. Rather the modus ponens strategy
plays at most a propaedeutic role, revealing an abstract property of the natural
number system that is responsible for the correctness of math induction. It is
still possible, of course, that math induction picks up quasi-inductive support
from its fruitfulness in proving further theorems, but in this respect it doesn’t
differ from any other math axiom.
Although we think that math induction doesn’t threaten the distinction
between deductive and inductive reasoning, there is a related issue about
generalization in math that might. Math proofs often proceed by selecting
an “arbitrary instance” from a domain, showing that some property is true
of this instance, and then generalizing to all the domain’s members. For this
universal generalization to work, the instance in question must be an abstrac-
tion or stand-in (an “arbitrary name” or variable) for all relevant individuals,
and there is no real concern that such a strategy is not properly deductive.
However, there’s psychological evidence that students don’t always recognize
the difference between such an abstraction and an arbitrarily selected exem-
plar. Sometimes, in fact, students use exemplars in their proofs (and evaluate
positively proofs that contain exemplars) that don’t even look arbitrary but
are simply convenient, perhaps because the exemplars lend themselves to con-
crete arguments that are easy to understand. In these cases, students are using
an inductive strategy, since the exemplar can at most increase their confidence
in the to-be-proved proposition.
It’s no news, of course, that people make mistakes in math. And it’s also
no news that ordinary induction has a role to play in math, especially in
the context of discovery. The question here is whether these inductive intru-
sions provide evidence that the deduction/induction split is psychologically
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

Mathematical Induction and Induction in Mathematics 267

untenable. Does the use of arbitrary and not-so-arbitrary instances show that
people have a single type of reasoning mechanism that delivers conclusions
that are quantitatively stronger or weaker, but not qualitatively inductive ver-
sus deductive? We’ve considered one way in which this might be the case.
Perhaps people look for counterexamples, continuing their search through
increasingly arbitrary (i.e., haphazard or atypical) cases until they’ve found
such a counterexample or have run out of steam. The longer the search, the
more secure the conclusion. We’ve seen, however, that this procedure doesn’t
extend to deductively valid proofs; no matter how obscure the instance, it
will still have an infinite number of properties that prohibit you from gen-
eralizing from it. It is possible to contend that this procedure is nevertheless
all that people have at their disposal – that they can never ascend from their
search for examples and counterexamples to a deductively adequate method.
But although the evidence on proof evaluation paints a fairly bleak picture
of students’ ability to recognize genuine math proofs, the existence of such
proofs shows they are not completely out of reach.

References
Dedekind, R. (1963). The nature and meaning of numbers (W. W. Beman, Trans.).
In Essays on the theory of numbers (pp. 31–115). New York: Dover. (Original work
published 1888.)
Eliaser, N. M. (2000). What constitutes a mathematical proof? Dissertation Abstracts
International, 60 (12), 6390B. (UMI No. AAT 9953274)
Evans, J. St. B. T., Newstead, S. E., & Byrne, R. M. J. (1993). Human reasoning. Hillsdale,
NJ: Erlbaum.
Fallis, D. (1997). The epistemic status of probabilistic proofs. Journal of Philosophy, 94,
165–186.
Frege, G. (1974). The foundations of arithmetic (J. L. Austin, Trans.). Oxford, England:
Blackwell. (Original work published 1884.)
Goldfarb, W. (1988). Poincaré against the logicists. In W. Aspray & P. Kitcher (Eds.),
History and philosophy of modern mathematics (pp. 61–81). Minneapolis: University
of Minnesota Press.
Goel, V., Gold, B., Kapur, S., & Houle, S. (1997). The seats of reason? An imaging study
of deductive and inductive reasoning. NeuroReport, 8, 1305–1310.
Grattan-Guinness, I. (1997). The Norton history of the mathematical sciences. New York:
Norton.
Greenberg, M. J. (1993). Euclidean and non-Euclidean geometries (3rd ed.). New York:
Freeman.
Hallett, M. (1990). Review of History and philosophy of modern mathematics. Journal of
Symbolic Logic, 55, 1315–1319.
Kahneman, D., & Tversky, A. (1972). Subjective probability: A judgment of representa-
tiveness. Cognitive Psychology, 3, 430–454.
Kaye, R. (1991). Models of Peano arithmetic. Oxford, England: Oxford University Press.
P1: JZP
05216772443c10 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:41

268 L. J. Rips and J. Asmuth

Kline, M. (1972). Mathematical thought from ancient to modern times. Oxford, England:
Oxford University Press.
Knuth, D. E. (1974). Surreal numbers. Reading, MA: Addison-Wesley.
Koedinger, K. R., & Anderson, J. R. (1991). Interaction of deductive and inductive rea-
soning strategies in geometry novices. Proceedings of the Thirteenth Annual Conference
of the Cognitive Science Society, 780–784.
Lemmon, E. J. (1965). Beginning logic. London: Nelson.
Martin, W. G., & Harel, G. (1989). Proof frames of preservice elementary teachers.
Journal for Research in Mathematics Education, 20, 41–51.
Osherson, D., Perani, D., Cappa, S., Schnur, T., Grassi, F., & Fazio, F. (1998). Distinct
brain loci in deductive versus probabilistic reasoning. Neuropsychologia, 36, 369–376.
Poincaré, H. (1982). Science and hypothesis (G. B. Halsted, Trans.). In The foundations
of science. Washington, DC: University Press of America. (Original work published
1902.)
Polya, G. (1954). Induction and analogy in mathematics. Princeton, NJ: Princeton Uni-
versity Press.
Quine, W. V. (1950). Methods of logic (4th ed.). Cambridge, MA: Harvard University
Press.
Rips, L. J. (2001a). Two kinds of reasoning. Psychological Science, 12, 129–134.
Rips, L. J. (2001b). Reasoning imperialism. In R. Elio (Ed.), Common sense, reasoning,
and rationality (pp. 215–235). Oxford, England: Oxford University Press.
Rips, L. J., Bloomfield, A., & Asmuth, J. (2007). Number sense and number nonsense.
Manuscript submitted for publication.
Rubin, D. C., & Wenzel, A. E. (1996). One hundred years of forgetting. Psychological
Review, 103, 734–760.
Russell, B. (1919). Introduction to mathematical philosophy. New York: Dover.
Smith, L. (2002). Reasoning by mathematical induction in children’s arithmetic. Amster-
dam: Pergamon.
Stewart, I., & Tall, D. (1977). The foundations of mathematics. Oxford, England: Oxford
University Press.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

11

Induction, Deduction, and Argument Strength

in Human Reasoning and Argumentation
Mike Oaksford and Ulrike Hahn

Recently researchers in the psychology of reasoning seem to be lining up on

two sides of a divide. On the one hand, some view human inference to be
sensitive to the deductive relationships between propositions (e.g., Johnson-
Laird & Byrne, 1991, 2002; Rips, 1994). For example, the natural language
sentence birds fly can be encoded as the logical proposition that if x is a bird,
then x flies. It is a matter of pure deductive logic that if this proposition is
true, then when someone comes to learn that Tweety is a bird, they should
conclude that Tweety can fly. That is, this inference is simply an instantiation
of the logical rule of modus ponens:
p → q, p
(1)
∴ q
This inference schema reads that if the propositions above the line are true,
then it can be inferred that the propositions below the line are true. If people
endorse this inference maximally, then it would seem that they are sensitive
to the deductive correctness of this argument.
On the other hand, there are those who believe that in evaluating an argu-
ment like this people are more sensitive to inductive strength than deductive
correctness. On some accounts (Oaksford & Chater 2007; Oaksford & Chater,
2003a–d; Oaksford, Chater, & Larkin, 2000), what matters is what probability
the premises, if x is a bird, then x flies and Tweety is a bird, confer on the
conclusion. If one identifies the inductive strength of the conditional with the
conditional probability of things flying given they are birds, then the transfer
of inductive strength to the conclusion is particularly straightforward.1 One

1 Inductive strength, as applied to arguments, is interpreted as the probability of the conclusion

given the premises (see also Rips, 2001). A rational calculation of this probability requires that
the individual propositions that make up the premises can also be assigned inductive strengths,
here interpreted as probabilities.

269
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

270 M. Oaksford and U. Hahn

simply uses the categorical premise Tweety is a bird, P (Tweety is a bird) = 1, to

conditionalize on the conditional premise. So if the inductive strength of the
conditional is .9, that is P (x flys|x is a bird) = .9, then the inductive strength
of the conclusion that Tweety can fly is also .9.
“Inductive strength” (Skyrms, 1975) is an appropriate description because
on this account one’s degree of belief in the conditional, P (if x is a bird, then
x flys), is P (x flys|x is a bird), which can be calculated from the evidence for
(birds that fly) and against (birds that don’t fly) the hypothesis that if x is
a bird, then x flys. Of course inductive strength or argument strength (Rips,
2001) does not have to be established by enumerative induction. For example,
the strength of someone’s belief that the train they are about to board stops
at the destination to which they want to travel but to which they have never
travelled before is based on their confidence in the railway guard who told
them to board the train and not on past inductions from having boarded the
train before.
Why have some researchers proposed that people assess arguments in pu-
tatively deductive tasks by appeal to inductive strength? The reasons are
theoretical and empirical. Empirically, there is much evidence that people do
not respond on a variety of tasks in a way that is consistent with a standard
logical account. For example, according to standard logic, if people under-
stand the task correctly, that is, they interpret the premises as true, then
they should always endorse the conclusion that Tweety can fly in the above
argument. In many tests of this hypothesis using abstract materials, where,
for example, participants are asked to judge inferences like modus ponens,
endorsement rates are indeed very high (see Schroyens & Schaeken, 2003, for
a meta-analysis). However, when real world material is used, endorsement
rates can be much lower, that is, significantly less than 1 (e.g., Oaksford et al.,
2000). Moreover, other inferences which participants should always endorse
according to standard logic, for example, modus tollens:

p → q , ¬q
(2)
∴ ¬p

are always endorsed at much lower rates. One interpretation of these phe-
nomena is that for these inferences people are assessing inductive strength,
on the assumption that the proportion of people endorsing an inference is a
reflection of the underlying subjective degree of belief that people have about
the conditional (Evans & Over, 2004).
Theoretically, one good reason to believe that people should not be partic-
ularly sensitive to deductive relations is that standard logic is monotonic; in
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 271

other words the law of strengthening the antecedent holds that is,
p→q
(3)
∴ (p ∧ r) → q

However, it seems that for most inferences that people are likely to encounter
in their everyday lives, this logical law fails.2 For example, if x is a bird, then x
flys does not entail that if (x is a bird and x is an Ostrich), then x flys. As the
logical standard against which deductive correctness is normally assessed in
psychology is monotonic, that is strengthening of the antecedent is valid, per-
haps people should not be sensitive to deductive relations. Non-monotonicity
is the sine qua non of inductive reasoning. That is, each observation, bird A
can fly, bird B can fly, and so forth can be viewed as a premise in the induc-
tive argument to the conclusion that birds fly. However, this conclusion can
be defeated (logically) by the observation of one non-flying bird. Thus the
observation that most everyday arguments seem to be non-monotonic seems
to suggest that they are mainly assessed for argument strength, as for these
arguments deductive correctness may be unattainable.
Rips (2001a, b; 2002; see also Rips [this volume] for further discussion)
has argued that trying to make inductive strength do service for deductive
correctness (Oaksford & Chater, 1998, 2001) or trying to make deductive
correctness do service for inductive strength (Johnson-Laird, 1994; Johnson-
Laird, Legrenzi, Girotto, Legrenzi, & Caverni, 1999) is misguided, and he refers
to these attempts as reasoning imperialism. In this chapter, we argue that our
position is not an imperialist one. Rather, it is that while people may well be
capable of assessing deductive correctness when explicitly asked to, this is rarely,
if ever, their focus of interest in evaluating an argument. Thus we will conclude
that inductive strength is probably more important in determining people’s
behaviour than deductive correctness even on putative deductive reasoning
tasks. However, despite our emphasis on inductive strength, we still think that
deductive relations, or structure, may be crucial to human inference, as we
point out in Section 4.
In Section 1, we begin by examining Rips’ (2001) own evidence and argu-
ments for a clear distinction between these two ways of evaluating inferences.
We argue (i) that the data are perhaps not as convincing as they may first
seem, (ii) that establishing the claim of sensitivity to deductive correctness
can not be established by examining isolated inferences but only for a whole

2 This can be readily established by a simple thought experiment. Think of any conditional
statement about the everyday world you like and see how easy it is to generate exceptions to the
rule (see also Chater & Oaksford, 1996).
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

272 M. Oaksford and U. Hahn

logic, (iii) that the generality of a concept of a logical system makes it difficult
to refute the claim that people are sensitive to deductive relations in some
logic, but (iv) that given (ii), it is clear that people probably should not be
sensitive to standard deductive validity because it leads to many paradoxes.
In Section 2, we move on to look at conditional inference, as in the examples
of modus ponens and modus tollens above. Here we argue that an account based
on probability logic (Adams, 1975, 1998) which uses an account of inductive
strength resolves many of the paradoxes observed for standard deductive
validity and moreover provides a better account of the data. We also show
that inductive processes, that is, updating beliefs in the light of evidence by
Bayesian revision, may be directly involved in drawing conditional inferences
because of possible violations of what is called the rigidity condition for
Jeffrey conditionalization. We also evaluate the inductive strength of some
conditional argument forms. In particular, we show that people may judge
some conditional reasoning fallacies as being as strong as a deductively correct
inference.
Finally, we show in Section 3 how some very similar proposals address infor-
mal argument fallacies, for example, the argument from ignorance, circularity,
and the slippery slope argument. That is, invoking the concept of argument
or inductive strength can resolve the paradox of why some instances of in-
formal argument fallacies seem perfectly acceptable and capable of rationally
persuading someone of a conclusion.
The theme of this paper could be characterized as an attempt to demonstrate
that an account of argument strength helps resolve many paradoxes and
fallacies in reasoning and argumentation. A logical paradox (see Section 1.2
on the paradoxes of material implication) is usually taken to mean an example
of a logically valid inference, for which there seem to be instances that are
considered obviously unsound.3 Following logicians like Adams (1975, 1998),
we suggest that this may be because these inferences are inductively weak. To
commit a logical fallacy is to endorse an argument that is not logically valid.
What is paradoxical about the logical fallacies is that there seem to be many for
which there are instances that seem to be perfectly good arguments. Indeed
this point extends to a whole range of what are typically called informal

3 There is a potentially important distinction between valid but unsound inferences with true
premises and inferences with false premises that are unsound. It has been argued (Ferguson,
2003) that the defeasibility revealed by violations of strengthening the antecedent is not prob-
lematic because the violations are instances where the premises are strictly false, for example, it
is known that not all birds fly. However, the paradoxes of material implication, as we will see,
seem to be instances where not only is the inference intuitively unsound, but the premises are
also considered true.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 273

reasoning fallacies in logic text books (e.g., Copi & Cohen, 1990). Although
people have been warned against committing these fallacies since Aristotle’s
Sophistical refutations, examples continue to crop up that seem like perfectly
convincing arguments. We suggest that a probabilistic notion of argument
strength allows instances of fallacies that are convincing to be distinguished
from instances that are not convincing.
The upshot of all these arguments is not to deny that people possess some
understanding of deductive validity, but rather to establish that in determining
human inferential behaviour, argument strength is the more central concept.
Thus, in the conclusion, we point out, by considering some recent approaches
in AI, that without some structure it would not be possible to appropriately
transmit inductive strengths from the premises to the conclusions of com-
plex arguments. Consequently, deductive correctness and inductive strength
should be seen as working together rather than as independent systems.

.    

If deductive correctness and inductive strength are psychologically one and
the same thing, then it would be expected that deductive correctness should
track inductive strength. However, when people are asked to rate inferences
for deductive correctness or inductive strength, they do not seem to be treat-
ing deductive correctness simply as at the extreme end of a scale of inductive
strength. Rips (2001) had participants categorize arguments as deductively
correct or not and as inductively strong or not. The arguments fully crossed
deductive correctness with causal consistency. If deductive correctness were
simply the extreme end of the scale of inductive strength, then there should be
a monotone relationship between these judgements. However, this relation-
ship was violated in Rips’ data. This was revealed by a cross-over interaction.
Participants judged deductively correct but causally inconsistent arguments
to be deductively correct more often than they judged them to be inductively
strong. However, participants judged deductively incorrect but causally con-
sistent arguments to be inductively strong more often than they judged them
to be deductively correct.
Rips (2001, 2002) interprets these results as arguing against reasoning impe-
rialism, that is, the tendency to generalise either the inductive-strength view or
the deductive-correctness view outside their proper domains. So, for example,
proponents of the probabilistic view (e.g., Oaksford & Chater, 1998, 2001)
seem to regard most inference as reflecting a probabilistic judgement, that is,
they seem to endorse inductive strength as the sole measure of an inference.
Similarly, proponents of mental-models theory (Johnson-Laird, 1994, 1999)
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

274 M. Oaksford and U. Hahn

adopt a logical approach to inductive inference and probability judgement.

Assessing probabilities or inductive strength is based on the proportion of pos-
sible states of the world, or “mental models,” in which a proposition is true.
We first observe that although we have argued for the importance of in-
ductive strength in assessing arguments in the past, this has usually been as a
corrective to the commonly held view that it was all deduction. The possibility
that people have mechanisms for evaluating deductive correctness has been
conceded in several places. For example, Chater and Oaksford (1999) argued
that in syllogistic reasoning, while possible conclusions may be generated
by probabilistic heuristics, these may then be tested by processes that assess
deductive correctness.
However, in a recent paper (Oaksford & Chater, 2002) we posed the ques-
tion, “How much deduction is there?” and our response, after surveying
several possible ways of answering this question, was “very little”. It is im-
portant to bear in mind that our concerns have always been with the kinds
of inferences that people could be expected to routinely perform in their
everyday lives. Thus while people may well be capable of assessing deductive
correctness when explicitly asked to, this is rarely, if ever, their focus of interest in
evaluating an argument – unless they are mathematicians or logicians.
We argue that Rips (2001) does not provide a strong test of whether de-
ductive correctness and inductive strength dissociate. We concentrate on two
cases. The first is the case of a deductively incorrect but inductively strong ar-
gument, where we argue that the case for a dissociation is weak. This is because
the materials for this case did not provide any plausible logical structure. The
second is the case of a deductively correct but inductively weak argument,
where we argue that there is a much stronger case for a dissociation. However,
it could not be observed because of the restricted range of arguments used.

1.1 Deductively Incorrect and Inductively Strong Arguments

We assume the standard of deductive correctness intended by Rips (2001)
is standard truth-functional propositional logic, as can be found in most
undergraduate texts (e.g., Enderton, 1972). As we will argue in the next
section, this choice is important, but it goes uncommented on in Rips (2001).
The case where the dissociation was at its most extreme in Rips (2001) was
the deductively incorrect but causally consistent cases, for example

Car X10 runs into a brick wall (4)

Car X10 stops

The premise above the line does not logically entail the conclusion below the
line, but the probability of the conclusion given the premise is high because it
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 275

is causally consistent. In constructing the deductively incorrect but causally

consistent cases, Rips (2001) used four inference forms, modus ponens (if
p then q, p, therefore q ), conjunctive syllogism (not[p and q], p, therefore
not q), disjunctive syllogism (p or q, not p, therefore q ), and conjunction
elimination (p and q, therefore q ). The example in (4) was explicitly presented
as the deductively incorrect but causally consistent case for modus ponens. The
deductively correct and causally consistent case included the major premise,
if Car X10 runs into a brick wall, then Car X10 stops.
The premises in examples like (4) have no logical structure, at least not at the
propositional level that we are considering. So invocations such as, “Consid-
ering just the form of the sentences (and not their specific content), does this
form ensure that if the sentences above the line are true so is the sentence be-
low the line?” (Rips, 2001, p. 131)4 seem misplaced – the sentences above the
line do not have any “form” relevant to answering this question. Nonetheless,
around 35% of participants rated these cases as deductively correct, which was
well above chance. This level of error did not change when participants were
asked this question in two slightly different ways, which emphasised certainty
rather than form. However, regardless of the question posed, this level of
error seems too high to support the claim that people are sensitive to logical
form.
In Rips (2001) the critical deductively incorrect and causally consistent
cases were constructed for the other inference forms, except in the case of
conjunction elimination,5 in the same way as for modus ponens (Rips per-
sonal communication). That is, the major premise of the deductively correct
and causally consistent case was left out. Considering how these cases could
have been constructed, while retaining some logical structure in the premises,
suggests an alternative explanation for the strong differentiation between
judgements of deductive correctness and causal consistency for these cases.
Let us take conjunctive syllogism as an example. Rips’s (2001) paper gives
the impression that these cases were constructed with the same substitution
instances for p and q . However, this would have the consequence that the de-
ductively correct and causally consistent case is actually causally inconsistent:

Not(Car X10 runs into a brick wall and Car X10 stops) (5)
Car X10 runs into a brick wall
Car X10 does not stop

4 This question is also highly ambiguous as not even experts agree on the “form” of English
sentences. So exactly what participants understood from this instruction is unclear.
5 For conjunction elimination, p and q therefore q, the following argument form was used as the
logically incorrect and causally consistent case, p therefore p and q. That is, rather than omit a
major premise a conjunct was removed again, leading to an unstructured premise.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

276 M. Oaksford and U. Hahn

Consequently, Rips (2001) replaced Car X10 stops with Car X10 speeds on. The
conclusion is then that Car X10 does not speed on, which is causally consistent
with the minor premise. Constructing the deductively incorrect but causally
consistent case, while retaining some logical structure, could be achieved by
using the premises in (5) but reversing the polarity of the conclusion, that
is, Car X10 stops. The problem about (5), even with this substitution, is that
the major premise seems to express a causally inconsistent proposition. One
of the other facts about Car X10 that seems to have to be true when it hits a
wall is that it stops, but the major premise denies this (assuming that there is
only one Car X10 and that the two conjuncts are indexed to the same space-
time location). Similar arguments apply for modus ponens and disjunctive
syllogism. These considerations suggest that including some logical structure
for the deductively incorrect but casually consistent cases would appear to
have to introduce some level of causal inconsistency.
Despite the absence of logical structure for the deductively incorrect and
causally consistent cases in Rips (2001), 35% of these arguments were judged
deductively correct. We believe that a proper test of the hypothesis would have
to involve using materials like (5) but with the reversed polarity conclusion
for the deductively incorrect but causally consistent cases. We suspect that the
additional structure may lead to more judgements of deductive correctness
while the presence of a causally inconsistent major premise may lead to fewer
judgements of causal consistency. The combined effect would be to reduce
the gap between these two types of judgements found by Rips (2001, see
also Heit & Rotello, 20056 ). In sum, the way the deductively incorrect but
causally consistent cases were constructed seems to beg the question against
the unitary view. Thus, we don’t interpret the existing evidence using the Rips
(2001) paradigm as convincing evidence against a unitary view.
In this section we have concentrated on arguments that are deductively in-
correct but inductively strong and argued that these may be more closely
aligned than Rips (2001) might lead one to expect. In the next section,
we concentrate on arguments that are deductively strong but inductively
weak.

6 Heit and Rotello (2005) used similar materials to Rips (2001) but used abstract material (Ex-
periment 1) and analysed their results using signal detection theory. The results were similar
to Rips (2001). Heit and Rotello’s (2005) Experiment 2 used quantified statements where class
inclusion (all birds have property A, therefore robins have property A) was respected or violated
(premise and conclusion inverted). While these results are interesting, our feeling is that they
tell us most about the semantics/pragmatics of the words “necessary” and “plausible,” where it
has been known for some time that they don’t neatly carve up some underlying probability scale
(see Dhami & Wallsten, (2005); Rapoport, Wallsten, & Cox, 1987; Wallsten, Budescu, Rapoport,
Zwick, & Forsyth, 1986 ).
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 277

1.2 Deductively Correct and Inductively Weak Arguments

Rips (2001) also provides a priori arguments for why, when the probability
of the conclusion given the premise is 1, the argument is not necessarily
deductively correct – although if it is deductively correct this probability must
be 1. We now suggest that some deductively correct arguments may actually
have an inductive strength of 0. Moreover, we argue that this is appropriate
because they are universally agreed to be poor arguments.
Sensitivity to deductive correctness can’t really be assessed by asking partic-
ipants to evaluate just a few inferences. Deductive systems come as packages
and prescribe a whole range of inferences while proscribing others. This point
raises two questions. First, what is the deductive system that provides the stan-
dard of deductive correctness in Rips’s (2001) experiments? Second, given this
system, for how many inferences licensed by that system do deductive correct-
ness and inductive strength dissociate? Clearly if it is only for a few inferences,
then the imperialist thesis is not threatened very much.
In the last section, we assumed that the standard of deductive correctness
intended in Rips (2001) was standard propositional logic. However, Rips’s
(1994) PSYCOP system is slightly different7 but shares most properties with
standard logic. In particular, standard propositional and predicate logic and
Rips’s (1994) system are truth conditional and truth functional. That is, the
meaning of a simple proposition, p, is equivalent to the conditions that make
it true in the world. The meaning of more complex propositions, for example,
p and q , is a function of the truth values of the constituent propositions, that
is, p and q is true only when both p is true and q is true, otherwise it is false.
However, the study of deductive logic and thus what we mean by “deductive
correctness” goes beyond this truth-functional core.
The concept of a deductive logic is very general. We can regard a logic as a
pair L , |= consisting of a language L and a consequence relation |= (Scott,
1971). All pairs of formulae (or sequences of formulae) such that φ |= ϕ define
the consequence relation. Two questions arise about the resulting logic –
(1) can a procedure be defined for deriving ϕ from φ? and (2) is there
an efficient implementation of that procedure once defined? Any pair of
formulae (or sequences of formulae) such that φ |= ϕ define a deductively
correct inference with respect to that language and to that consequence relation.
Thus the concept of deductive correctness is a relative concept, although in
the psychology of reasoning only standard truth-functional logic is usually
intended when an argument is described as deductively valid or “correct.”
7 Not all inferences with conditionals licensed by standard logic are provable in Rips’s (1994)
system, but all inferences provable in Rips system are classically valid.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

278 M. Oaksford and U. Hahn

Some deductive systems preserve the truth-functional core, for example,

systems of modal logic (Hughes & Cresswell, 1968), but some systems abandon
parts of it for some connectives. In particular, in most interesting systems of
conditional logic (Nute, 1984; Bennett, 2003) the conditional is not truth
functional (Lewis, 1973; Stalnaker, 1968), and in some it is neither truth
functional nor truth conditional (Adams, 1975, 1998). Indeed, looked at
from a dynamic perspective (see next section), the Adams conditional can be
meaningfully assessed only for inductive strength. The general point is that the
question of whether people are sensitive to “deductive correctness” is much
more difficult to answer than Rips (2001) suggests. There are many possible
logics, many of which are yet to be explored. One could never exclude the
possibility that one of these will map perfectly onto people’s judgements of
deductive correctness.
Nonetheless, one thing we can be reasonably sure of is that people are
unlikely to regard deductive correctness in standard logic as a particularly
intuitive or useful way to assess arguments. Rips (2001) only used inference
patterns from the truth-functional core of standard logic and avoided any of
the more complex arguments involving conditionals, that is, only modus po-
nens was used. However, there are a range of arguments involving conditionals
that are deductively correct according to standard logic. It can only really be
argued that people are sensitive to deductive correctness in standard logic, if
they judged all of these arguments to be deductively correct. However, the
psychology of deductive reasoning has shown that human deductive reason-
ing performance rarely matches up to standard logic. In mental models and
mental logic, the deviations are then explained at the performance level, that
is, resource limitations on working memory exclude our ability to assess some
inferences. This means that the question has always been one of degree—do
people conform to standard logic enough to argue that they are sensitive to
deductive correctness in this logic? Rips (2001) uses very simple deductive
arguments to argue for an affirmative answer. However, as we now show,
there are equally simple deductively correct arguments in standard logic that
people certainly should not endorse.
For example, here are two argument forms that are simple, one-premise,
deductively correct inferences in standard logic:
¬p q
(6)
∴ p→q ∴ p→q
If the material conditional of standard logic is regarded as the proper inter-
pretation of the conditional of everyday language, then either argument form
can lead to conclusions like if the moon is made of cheese, then Tony Blair is
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 279

Prime Minister. This is because in standard logic the material conditional is

true if the antecedent is false or the conclusion is true. So in both cases, if
the premise is true – the moon is not made of cheese or Tony Blair is Prime
Minister – then the conclusion must also be true. We doubt that anyone
would judge these as deductively correct. These are the paradoxes of material
implication, and unless people endorse these inferences as deductively correct,
it can not be claimed that people are sensitive to standard deductive validity
for the conditional.
We now suggest that the reason why the arguments in (6) are paradoxical is
that, despite being deductively correct, they are inductively weak. That is, the
mismatch with argument strength explains their paradoxical nature and also
confirms that it is argument strength that is the dominant factor in evaluating
arguments. In the introduction, we proposed identifying the probability of a
conditional with the conditional probability, that is,
P ( p → q ) = P (q | p), where P ( p) > 0 (7)
Edgington (1991) refers to (7) as simply “The Equation.” P (q | p) is given
by the Ratio formula (Bennett, 2003, proposed this name; “∧” is the logical
symbol for and):
P (p ∧ q)
P (q | p) = (8)
P ( p)
The Ratio formula suggests that people should “believe p → q to the extent
that you think that p ∧ q is nearly as likely as p” (Edgington, 1991, p. 189).
So when P ( p) = 0, you should not accept p → q , as P (q | p) is 0 or undefined
(see Bennett, 2003). Thus if P ( p) = 0, then the inferences in (7) should be
regarded as maximally inductively weak, as either P (conclusion|premises) = 0
or they can’t be assigned an inductive strength at all (see Bennett, 2003). Thus,
the paradoxes of material implication are deductively correct but inductively
weak. Our guess is that for these conditional inferences, judgements of de-
ductive correctness would track inductive strength. Moreover, given their
paradoxical nature, they should track inductive strength.
In this section, we have argued that the claim that people are sensitive to
deductive correctness is difficult to assess given the generality of the concept
of a logical system. We pointed out that there were several conditional logics,
most of which, unlike standard logic, are not truth functional. These systems
define a range of inferences that are deductively correct. We argued that only
if people judged them all as deductively correct could it be argued that people
are sensitive to deductive correctness for the conditional for that logic. We
then examined the paradoxes of material implication. It seems unlikely that
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

280 M. Oaksford and U. Hahn

people would regard these as deductively correct, and that is why they are
regarded as paradoxes. However, these turned out to be inductively weak,
which explains their paradoxical status.8
In the next section, we explore psychological accounts of the conditional
more fully. In particular, we look at recent accounts of the psychological
data based on Adams’s (1975, 1998) probability conditional. Following some
recent proposals (Oaksford, 2005), we suggest that not only are conditionals
assessed for inductive strength in psychological experiments, but inductive
learning processes may also be invoked in assessing conditional inferences.
We also evaluate the inductive strength of the conditional arguments usually
investigated in the psychology of reasoning.

.  
Along with many other researchers (e.g., Anderson, 1995; Chan & Chua,
1994; George, 1997; Liu, 2003; Liu, Lo, & Wu, 1996; Politzer, 2005; Stevenson
& Over, 1995), we have recently been arguing for an account of conditional
inference based on inductive strength (Oaksford & Chater 2007; Oaksford &
Chater, 1998, 2001, 2003a–d; Oaksford, Chater, & Larkin, 2000). The main
assumption underlying this approach, as we discussed in the introduction and
in the last section, is that the probability of a conditional is the conditional
probability, that is, see (7) above. This assumption is at the heart of Adams’s
(1975, 1998) account of the probability conditional. However, in that account
the Ratio formula (8) is not definitional of conditional probability. People’s
understanding of P (q | p) is given by the subjective interpretation provided
by the Ramsey test. As Bennett (2003) says:
The best definition we have [of conditional probability] is the one provided
by the Ramsey test: your conditional probability for q given p is the prob-
ability for q that results from adding P ( p) = 1 to your belief system and
conservatively adjusting to make room for it (p. 53).
The Ratio formula indicates that in calculating the probability of the every-
day conditional, all that matters are the probabilities of the true antecedent
cases, that is, P ( p ∧ q ) and P ( p ∧ ¬q )[P ( p) = P ( p ∧ q ) + P ( p ∧ ¬q )].
This contrasts immediately with standard logic where the probability of a
conditional is given by the probability of the cases that make it true, that
is, P ( p → q ) = P (¬ p) + P ( p, q ). Evans et al. (2003) and Oberauer and
Wilhelm (2003) have confirmed that people interpret the probability of the

8 Although, as we discuss later (in the paragraph before Section 2.1), this is not the only possible
response. In some non-classical logics, the paradoxes are also not valid inferences.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 281

conditional as the conditional probability and not the probability suggested

by standard logic.
(7) has the important consequence that the probability conditional is not
truth functional and it is not truth conditional either, that is, conditionals
are not propositions that can have a truth value. If they were propositions,
then asking about their probability, P ( p → q ), would be equivalent to asking
about the probability that p → q is true. But from (7), P ( p → q ) = P (q | p),
and so the question being asked is what is the probability that q | p is true. How-
ever, “q | p” is not a claim about the world, that is, it doesn’t make sense to ask
whether this is true or false. According to the Ramsey test, this expression re-
lates to a mental process, not a fact about the world that could be true or false.9
If indicative conditionals do not have truth values, then it would seem im-
possible to define any valid inferences that follow from them. An inference is
classically valid if the truth of the premises guarantees the truth of the conclu-
sion. However, while classical validity is therefore impossible for arguments
involving conditionals, they may possess another virtue, probabilistic validity
or p-validity. Adams (1975) discovered that most classically valid arguments
also possess a property that can best be formulated using the concept of un-
certainty. The uncertainty of a proposition p, U ( p), is simply 1 − P ( p). The
property he discovered was that for many “classically valid argument[s] the
uncertainty of the conclusion cannot exceed the sum of the uncertainties of the
premises” (Bennett, 2003, p. 131). That is,

n
If p1 . . . pn entail q , then U (q ) ≤ U ( pi ) (9)
i =1

Adams (1975) calls an argument fulfilling this condition p-valid. Arguments

containing indicative conditionals, while not candidates for classical validity,
may be p-valid and so can be evaluated on that basis.

9 These considerations place constraints on the way conditionals can be embedded in truth-
functional compound sentences (Lewis, 1976). However, there is linguistic evidence that the
constraints on embeddings recommended by the probability conditional may be reflected in
natural languages (Bennett, 2003). Bennett (2003) argues that this linguistic fact seems to disarm
some of the strongest criticisms of the Equation due to Lewis (1976).
This thesis may also seem to violate intuitions that conditionals such as “if a steam roller
runs over this teacup, it will break” and “if Lincoln visited Gettysburg, then he delivered a
speech there” are determinately true. The issues highlighted by examples like these relate to how
conditionals and the predicate “true” are used in natural language. It is possible that the teacup
does not break (as it is made of titanium, i.e., it just bends a bit). Moreover, only someone who
knew Lincoln gave the Gettysburg address would be in a position to affirm this conditional,
but, given this knowledge, why would they ever have grounds to assert this conditional? “True”
might also be used in natural language to mean a probability infinitesimally different, but still
different, from 1 (Pearl, 1988).
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

282 M. Oaksford and U. Hahn

All p-valid arguments are classically valid, but not all classically valid ar-
guments are p-valid. For example, as we have seen, the paradoxes of material
implication (6) are not p-valid, that is, they can be inductively weak. Strength-
ening of the antecedent (3) also turns out not to be p-valid. Although it makes
sense to assign a high probability to if Tweety is a bird, then Tweety can fly,
because most birds fly, anyone would assign a very low probability, that is, 0,
to if Tweety is a bird and Tweety is one second old, then Tweety can fly (how
old Tweety is can be adjusted to produce probabilities for the conclusion
that are greater than 0). Thus the uncertainty of the conclusion, that is, 1, is
greater than the uncertainty of the premise – it is an inductively weak argu-
ment. So the probability conditional is defeasible or non-monotonic at its very
foundation.
P -validity holds for an argument form or schemata. For example, if it is
p-valid, there can be no instances where (9) is violated. Inductive strength
applies to particular arguments, that is, content matters. So for example, if
“is one second old” were replaced with “is green,” then this instance of strength-
ening the antecedent might be viewed as inductively strong.10 Strengthening
of the antecedent nonetheless is not p-valid because (9) does not hold for
all instances of this inference, that is, (9) does not hold in all models of the
premises. An instance of a p-valid inference may also be inductively weak. For
example, in a modus ponens inference, someone may be maximally uncertain
about whether birds fly and whether Tweety is a bird, that is, the probability
of both premises is .5. As we will see later, the probability of the conclusion is
therefore .25. That is, although (9) is not violated, this is an inductively weak
argument, that is, the probability of the conclusion given the premises is low.
However, when it comes to conditionals, p-validity is perhaps not that
interesting a concept, that is, it is useful only in the static case of fixed prob-
ability distributions. Conditionals in probability logic can also be looked at
dynamically, that is, they allow people to update their beliefs (prior probabil-
ity distributions) given new information. So if a high probability is assigned
to if x is a bird, x flys, then on acquiring the new information that Tweety is
a bird, one’s degree of belief in Tweety flys should be revised to one’s degree
of belief in Tweety flys given Tweety is a bird, that is, my degree of belief in the
conditional. So using P0 to indicate prior degree of belief and P1 to indicate
posterior degree of belief, then:

P1 (q ) = P0 (q | p), when P1 ( p) = 1. (10)

10 Although as “is green” seems totally irrelevant to the whether birds fly, this argument feels
circular, that is, the conclusion is just a restatement of the premises. We look at argumentative
fallacies, like circularity, in relation to an account of argument strength in Section 4.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 283

This is exactly the formulation we used in the introduction to the chapter to

illustrate evaluating an argument for inductive strength.
One possible consequence of moving to an account of the conditional that
avoids the problems for standard logic, that is, the paradoxes, is an account
of conditional inference based on inductive strength. This is not a necessary
consequence of avoiding the paradoxes, as there other systems of condi-
tional logic. For example, the Stalnaker conditional (Stalnaker, 1968), which
is based on possible worlds, avoids the paradoxes but is not always consistent
with the equation. Nonetheless, as we now show, an account based on induc-
tive strength seems to resolve some fundamental issues in the psychology of
conditional reasoning. Moreover, such an account may incorporate inductive
processes (Oaksford & Chater 2007). Finally, there are relevant data that allow
us to briefly compare two different accounts of inductive strength that appear
in Rips (2001) and to which we have recently made appeal.

2.1 Inferential Asymmetries and Inductive Processes

in “Deductive” Reasoning
Oaksford and Chater (2007) recently showed how the account of conditional
inference based on inductive strength (Oaksford, Chater, & Larkin, 2000) was
compatible with the dynamic view of inference in probability logic (Adams,
1998). Along with modus ponens (MP) and modus tollens (MT), psychologists
have also investigated two other inference forms that are truth-functional
fallacies, denying the antecedent (DA) and affirming the consequent (AC):
p → q, ¬p p → q, q
(DA) (AC)
¬q p
The problem with applying the scheme for dynamic updating by conditional-
ization to inferences other than modus ponens is that the premises all involve
the conditional if p then q. However, the appropriate conditional probability
for, say, the categorical premise of a modus tollens inference, P (¬q ) = 1, to
conditionalize on is P (¬ p|¬q), but the premises of this inference do not
entail a value for this conditional probability.
Oaksford et al.’s (2000) solution to this problem was to assume that people’s
prior world knowledge included estimates of the prior marginal probabilities,
P0 ( p) and P0 (q ). Oaksford & Chater (2007) then showed how with P0 (q|p)
= a, P0 ( p) = b and P0 (q ) = c , these probabilities entail relevant conditional
probabilities for all four inferences:
MP P1 (q ) = P0 (q | p) = a (11)
1 − c − (1 − a)b
DA P1 (¬q ) = P0 (¬q |¬ p) = (12)
1−b
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

284 M. Oaksford and U. Hahn

ab
AC P1 ( p) = P0 ( p|q ) = (13)
c
1 − c − (1 − a)b
MT P1 (¬ p) = P0 (¬ p|¬q ) = (14)
1−c
Equations (11) to (14) show the posterior probabilities of the conclusion of
each inference, assuming the posterior probability of the categorical premise
is 1. By using Jeffrey conditionalization, these cases can be readily generalized
to when the probability of the categorical premise is less than 1. For example,
for MP, assuming P1 ( p) = d:
MP P1 (q ) = P0 (q | p)P1 ( p) = ad (15)
(15) actually represents a lower bound on the posterior probability, with the
upper bound at ad + 1 – d.
In studies investigating all four inferences, a significant asymmetry has been
observed between modus ponens (1) and modus tollens (2), with (1) being
endorsed considerably more than (2). Moreover, affirming the consequent
is endorsed more than denying the antecedent. In most other psychological
accounts, this latter asymmetry is explained as an instance of the asymmetry
between (1) and (2) on the assumption that some people believe that if p
then q pragmatically implicates its converse, if q then p. Oaksford, Chater, and
Larkin (2000; Oaksford & Chater, 2003a, b) showed that as long P0 (q | p) is
less than 1, then (11) to (14) can generate these asymmetries. The precise
magnitude of the asymmetries depend on the particular values of P0 (q|p),
P0 ( p), and P0 (q ).
However, no values of these probabilities are capable of generating the em-
pirically observed magnitudes of the MP–MT and AC–DA asymmetries –
the best fit values underestimate the former and overestimate the latter.
Oaksford & Chater (2007) argued that this is related to perceived failures
of what is called the rigidity assumption for conditionalization, which is that
P1 (q | p) = P0 (q | p) (16)
This assumption is implicit in (15). This means that learning that the cate-
gorical premise has a particular posterior value does not alter the conditional
probability. However, consider the following example for MT, where the con-
ditional premise is if you turn the key the car starts and the categorical premise
is the car does not start. What is odd about the assertion of the categorical
premise is that it seems informative only in a context where the car was ex-
pected to start, and the only way one would normally expect a car to start
was if the key had been turned. Consequently, the assertion of the categorical
premise raises the possibility of a counterexample to the conditional.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 285

In several recent approaches to conditional inference (e.g., Schroyens,

Schaeken, & Handley, 2003; Quinn & Markovits, 2002), after reaching a
tentative conclusion based on a mental-models representation, people search
for a possible counterexample in long-term memory. If they find one, then
they do not draw the inference. Oaksford & Chater (2007) proposed a prob-
abilistic version of this account for modus tollens, where an initially high
estimate of P0 (q | p) is revised down by Bayesian revision on a single coun-
terexample to derive a revised lower estimate for P0 (q | p), P0R (q | p). It is this
value that is then used in (14) to calculate the relevant conditional probabil-
ity to conditionalize on. A similar revision is not hypothesised to occur for
MP, as this inference directly provides the appropriate conditional premise to
conditionalize on.
How should the probability of the conditional be updated? Oaksford and
Chater (2007) assumed that people’s degree of belief in a conditional is deter-
mined by the evidence that they possess for a dependency between p and q
and the evidence that they have that p and q are independent. The equation
states that the degree of belief in the conditional, that is, the hypothesis that
there is a dependency between p and q , should be equal to P0 (q | p). Hence
the degree of belief in the independence hypothesis is 1 –P0 (q | p). In the
independence hypothesis P0 (q | p) = P0 (q ). Someone’s new degree of belief
in the conditional and consequently their revised estimate of P0 (q | p), that
is, P0R (q | p), is the probability of the dependence model given the occurrence
of a p, ¬q case. This can be calculated by simple Bayesian revision. For ex-
ample, assume that P0 (q | p) = .97 (as in a recent meta-analysis [Schroyens &
Schaeken, 2003]), which is also the prior degree of belief assigned to the hy-
pothesis of a dependency between p and q (by equation 7). A prior degree
of belief of .03 is therefore assigned to the independence hypothesis. Then,
assuming that P0 ( p) = .5 and P0 (q ) = .6, so in the independence hypothe-
sis P0 (q | p) = .6, on updating,P0R (q | p) = .71. The revision is quite marked
because a counterexample is so much more likely in the independence than
in the dependence model. Moreover, using this value in (14), the conditional
probability for MT, and so P1 (¬p), is .64. This revision process considerably
increases the magnitude of the MP–MT asymmetry. Without the revision pro-
cess, MP endorsement should be 97% and MT, 96.25%, that is, they should be
endorsed in almost equal proportion. However, after revision the asymmetry
is considerably greater, with MT endorsements falling to 64%.
In Schroyens and Schaeken’s (2003) account, the retrieval of a possible
counterexample falsifies the conditional and so means that the inference can
not be valid. In the present inductive account the possibility of a counterex-
ample revises one’s degree of belief in the conditional by Bayesian updating,
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

286 M. Oaksford and U. Hahn

which in turn reduces the posterior degree of belief in the conclusion. Oaks-
ford and Chater (2007) generalise this account to DA and AC and show how
it provides a much better account of the MP–MT and AC–DA asymmetries.
For example, using the same figure as in the above example, AC should be en-
dorsed at 59% and DA at 51%. In sum, evaluating conditional inferences for
inductive strength rather than deductive correctness gets round paradoxical
problems for the standard logic of the conditional. Moreover, it provides a bet-
ter account of the data by invoking inductive processes of Bayesian updating
to cope with rigidity violations.

2.2 Measures of Argument Strength Applied to Conditional Inference

Rips (2001) actually introduces two different concepts of inductive strength,
both of which Hahn, Oaksford, and Corner (2005) looked at independently as
measures of argument strength to apply to reasoning fallacies more generally
(see Section 3). The measure that features in the bulk of Rips’s (2001) paper is
the probability of the conclusion given the premises. However, in a footnote
(Rips, 2001, p. 128, footnote 1), he considers another measure of argument
strength:
P (conclusion|premises) − P (conclusion)
Argument Strength = (17)
1 − P (conclusion)
This measure, which is called conditional agreement (Bishop, Feinberg, &
Holland, 1975), has the advantage that it looks at the change in the probability
of the conclusion brought about by the premises and scales this by how far
someone needs to go to be convinced of the conclusion. As Rips (2001) ob-
serves, it is important to have a measure like this because otherwise arguments
where the premises are irrelevant to the conclusion would not be assigned
0 argument strength but would be as strong as the prior probability of the
conclusion.
In this section we apply different measures of argument strength to some
data from a meta-analysis of conditional inference (Schroyens & Schaeken,
2003). This analysis consists of the data from sixty experiments on conditional
inference conducted over the last thirty years and reports the frequency of
endorsement of the MP, DA, AC, and MT inferences.11 Recently, Oaksford and

11 These studies all involved abstract material. Consequently, it could be argued that people have
no prior beliefs to bring to bear. However, the conversational pragmatics of the situation, where
the conditional is asserted, immediately suggest, via the Ratio formula, that P0 (q | p) is high and
so P0 ( p) is non-zero. In modelling these data Oaksford & Chater (2007) only assumed that
P0 ( p) is maximally uncertain, that is, .5. Moreover, in general we believe that prior knowledge is
always bought to bear analogically even on abstract materials, that is, perhaps only a small subset
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 287

Chater (2007) fitted the probabilistic model based on inductive strength,

outlined in the last section, to the data from this meta-analysis. The observed
probability of drawing each of the four inferences provides the measure of
P (conclusion|premises). The best-fit model parameters provide estimates of
P (conclusion), that is, the prior marginal probabilities, P0 ( p) and P0 (q ).
One theme that has emerged so far is the theoretical possibility that people
may regard a particular instance of an inferential fallacy to be nonetheless
inductively strong. In particular, they may regard them as being as inductively
strong as some valid inferences. Rips’s (2001) data does not address this
issue because, as we have seen, his deductively incorrect arguments were not
instances of classical logical reasoning fallacies, that is, the premises contained
no relevant logical structure. By applying these measures of argument strength
to these data, we can assess whether this possibility is observed empirically.
But first, we introduce a problem for the measure of argument strength in
(17) and introduce a measure that may get round it.
As a measure of argument strength, (17) has some counterintuitive prop-
erties. Consider an argument that moves someone from having a very low
degree of belief in a conclusion, P (conclusion) = .01, to complete uncertainty,
P (conclusion) = .5. Intuitively, this argument is stronger than one that moves
someone from a high degree of belief in a conclusion, P (conclusion) = .9, to
an even higher degree of belief in a conclusion, P (conclusion) = .99. However,
according to (17), the argument strength of the former is .495, while that of
the latter is .9. That is, counterintuitively, the second argument is by far the
stronger of the two.
There are many different measures that might be useful as an index of
argument strength, probably as many as there are indices of deviations from
independence, that is, P (conclusion|premises)= P (conclusion), of which Hat-
tori and Oaksford (in press) recently counted at least forty. Hahn et al. (2005)
suggested that the likelihood ratio may be a better measure of argument
strength. The likelihood ratio maps the prior odds onto the posterior odds
and so can also be calculated from P (conclusion|premises) andP (conclusion):
P (conclusion|premises)(1 − P (conclusion))
Likelihood Ratio = (18)
P (conclusion)(1 − P (conclusion|premises))
Using (18) with the previous example, the argument strength of the first
argument, taking someone from .01 to .5, is 93, whereas the argument strength
of the second argument, taking someone from .9 to .99, is 11. So the likelihood
ratio matches intuition in judging the first argument to be stronger than the

of exemplars of conditional relationships are accessed to provide rough and ready estimates of
some of relevant parameters.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

288 M. Oaksford and U. Hahn

second. However, as the degree of conviction gets closer to 1 in the second

argument, then this measure may reverse this pattern. Nonetheless, as long
as non-extreme values are assumed, this measure seems to capture intuition.
Hahn et al. (2005) mention some other advantages of the likelihood ratio
but, like Rips (2001), do not attempt to decide the issue of which is the best
measure, although they suggest that argument strength may be better viewed
as a bi-dimensional concept, with P (conclusion|premises) as one dimension
and change, perhaps measured by the likelihood ratio, as the second.
What happens when these measures are used to estimate argument strength
for the data from Schroyens and Schaeken’s (2003) meta-analysis? In these
data, the three measures, P (conclusion|premises), conditional agreement (17),
and the likelihood ratio (18), diverged on the ordering of argument strength
over the four inferences. P (conclusion|premises) is simply the endorsement
rate of each inference in the data. Estimates of the prior probability of the
conclusion, P (conclusion), were derived from fitting the model in (11) to
(14) to the data. For P (conclusion|premises) the average ordering was MP
(.97 [SD = .03]) > MT (.72[.14]) > AC (.60[.19]) > DA (.54[.17]), which
agreed with the likelihood ratio, MP (2572 [3822]) > MT (3.69[2.78]) > AC
(2.48[2.56]) > DA (2.26[1.17]). The extreme values of the likelihood ratio
for MP occurred because for several experiments all participants endorsed
this inference, which meant P (conclusion|premises) was in the extreme range.
Conditional agreement led to DA being ranked as a stronger argument than
AC: MP (.92 [.13]) > MT (.44[.28]) > DA (.26[.23]) > AC (.21[.38]). That
there is some disagreement about the strengths of the different arguments
suggests that there is scope for experimental investigations discriminating
between different accounts of argument strength. Participants would need to
be asked not just how much they endorse the conclusion but by how far the
premises had altered their prior beliefs, alternatively their prior beliefs, given
the pragmatics of the task (see footnote 2) could be measured.
Are there experiments where the fallacies, AC or DA, are arguments as
strong or stronger than the logically valid inferences? We concentrate on the
relationship between the fallacies and MT because of the revision strategy
we discussed above, where MP is evaluated with respect to P0 (q | p), whereas
MT and the fallacies are evaluated with respect toP0R (q | p). According to
the likelihood ratio, for 25% of the experiments, AC or DA was as strong
or stronger than MT. The other two measures agreed on this analysis for
all but one study. It could be argued that this pattern could arise because
these are studies for which Oaksford and Chater’s (2007) probabilistic model
provides poor fits, and so the estimates of the prior marginal probabilities are
inaccurate. This argument has some initial plausibility. However, two of the
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 289

experiments had the poorest fits to the data and were technically outliers. With
these experiments removed, the fit, measured by the co-efficient of variation
R 2 , did not differ significantly between the studies where AC or DA were
stronger than MT (mean R 2 = .92, SD = .07) and the rest (mean R 2 = .93,
SD = .08), t(56) = .57, p = .57.
It could also be argued that these results are not surprising because in other
theories, like mental models (Johnson-Laird & Byrne, 2002) or versions of
mental logic (Rips, 1994), AC can be drawn more often than MT. In these
theories, this happens if the task rule is interpreted as a bi-conditional, in
which case all four inferences are valid. However, if p then q does not logically
entail if and only if p then q. Moreover, the simple alphanumeric stimuli
used in the experiments in Schroyens and Schaeken’s (2003) meta-analysis
do not provide an appropriate context for the bi-conditional interpretation.
Consequently, if people do interpret the rules in this way, they are still in
error. There are some studies for which DA is drawn at least as often, if not
more often, than MT, although, within a study, we doubt whether this effect
was ever significant. This data can be interpreted as showing that some logical
reasoning fallacies are regarded as at least as strong as deductively correct
inferences (although the data is also consistent with the interpretation that
people erroneously assume that the conditional implicates its converse). In the
probabilistic account we introduced in the last section, DA can be endorsed
more than MT when P0 ( p) is greater than P0 (q ), which is probabilistically
consistent as long as P0 (q | p) is not too high.
In this section, we have discussed alternative accounts of argument strength
to which we think people may be sensitive. However, the appropriate exper-
iment to judge between appropriate measures has not been conducted. This
is because in standard conditional reasoning experiments, people are asked
to judge only their absolute rate of endorsement of the conclusion, which
is a function of both the prior, P (conclusion), and the amount of change,
P (conclusion|premises) – P (conclusion), brought about by an argument. How-
ever, the analysis in this section did not show much discrimination between a
measure of change, the likelihood ratio, and P (conclusion|premises). This will
require experiments where change is explicitly addressed and where people
are not just asked how much do you endorse a conclusion? but also how good
an argument is it for you?
We also observed that there are experiments where the argument strength
associated with a reasoning fallacy was as high as that associated with a de-
ductively correct inference. This is at least suggestive of a rather extreme
dissociation between deductive correctness and argument strength that goes
beyond that observed by Rips (2001). In Rips’s data, no deductively incorrect
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

290 M. Oaksford and U. Hahn

argument was evaluated as of similar inductive strength to a deductively cor-

rect argument. Moreover, Rips’s (2001) deductively incorrect arguments were
not instances of logical reasoning fallacies. If people are genuinely sensitive to
deductive correctness, it might have been hoped that they would have a facil-
ity to detect reasoning fallacies and so always evaluate them as significantly
weaker than a deductively correct argument (although see Massey, 1981). In
the next section, we show how we have recently used the concept of argument
strength to provide a new analysis of a range of what have been called informal
reasoning fallacies. These analyses extend our proposal that while people may
well be capable of assessing deductive correctness when explicitly asked to,
this is rarely, if ever, their focus of interest in evaluating an argument.

.     


Fallacies, or arguments that seem correct but aren’t, for example, DA and
AC, have been a longstanding focus of debate. Catalogues of reasoning and
argumentation fallacies originate with Aristotle and populate books on logic
and informal reasoning to this day. The classic tool brought to the analysis of
fallacies such as the argument from ignorance is formal logic, and it is widely
acknowledged to have failed in providing a satisfactory account. Testament
to this is the fact that fallacies figure in logic textbooks under the header
of “informal reasoning fallacies” (see, e.g., Hamblin, 1970) – an acknowl-
edgement of the inability to provide a sufficient formal logical treatment. In
particular, logical accounts have proved unable to capture the seeming ex-
ceptions to fallacies that arise with simple changes in content that leave the
structure of the argument unaffected. This suggests that either it is not formal
aspects of fallacies that make them fallacious, or else that the relevant formal
aspects are not being tapped into by classical logics.
Oaksford and Hahn (2004; see also Hahn & Oaksford, 2005; Hahn, Oaks-
ford, & Bayindir, 2005; Hahn, Oaksford, & Corner, 2005) provide evidence of
such variation and put forward an alternative, Bayesian account: individual
arguments are composed of a conclusion and premises expressing evidence
for that conclusion. Both conclusion and premises have associated probabil-
ities which are viewed as expressions of subjective degrees of belief. Bayes’s
theorem then provides an update rule for the degree of belief associated with
the conclusion in the light of the evidence. Argument strength, then, on this
account is a function of the degree of prior conviction, the probability of
evidence, and the relationship between the claim and the evidence, in partic-
ular how much more likely the evidence would be if the claim were true. That
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 291

is, as in discussing conditional inference, different instances of argumentative

fallacies may vary in argument strength, conceived of as the probability of the
conclusion given the premises.
We now illustrate this approach by appeal to a particular informal reasoning
fallacy: the argument from ignorance.

3.1 The Argument from Ignorance

A classic informal argument fallacy, which dates back to John Locke, is the
so-called argument from ignorance, or argumentum ad ignorantiam.

Ghosts exist, because nobody has proven that they don’t. (19)

This argument does indeed seem weak, and one would want to hesitate in
positing the existence of all manner of things whose non-existence simply had
not been proven, whether these be UFO’s or flying pigs with purple stripes.
However, is it really the general structure of this argument that makes it weak,
and if so what aspect of it is responsible? Other arguments from negative
evidence are routine in scientific and everyday discourse and seem perfectly
acceptable:

This drug is safe, because no one has found any side effects. (20)

Should all arguments from negative evidence be avoided, or can a systematic

difference between the two examples be recognized and explained?
A Bayesian account can capture the difference between (19) and (20) as
we show below. Moreover, it can capture the difference between positive and
negative evidence which allows one to capture the intuition that the positive
argument (21) is stronger than the negative argument (22):

Drug A is toxic because a toxic effect was observed (positive argu-

ment). (21)

Drug A is not toxic because no toxic effects were observed (negative argu-
ment, i.e., the argument from ignorance). (22)

Though (22) too can be acceptable where a legitimate test has been performed,
that is,

If drug A were toxic, it would produce toxic effects in legitimate tests.

Drug A has not produced toxic effects in such tests.
Therefore, A is not toxic.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

292 M. Oaksford and U. Hahn

Demonstrating the relevance of Bayesian inference for negative versus pos-

itive arguments involves defining the conditions for a legitimate test. Let e
stand for an experiment where a toxic effect is observed and ¬e stands for an
experiment where a toxic effect is not observed; likewise let T stand for the
hypothesis that the drug produces a toxic effect and ¬T stand for the alterna-
tive hypothesis that the drug does not produce toxic effects. The strength of
the argument from ignorance is given by the conditional probability that the
hypothesis, T , is false given that a negative test result, ¬e, is found, P (¬T|¬e).
This probability is referred to as negative test validity. The strength of the ar-
gument we wish to compare with the argument from ignorance is given by
positive test validity, that is, the probability that the hypothesis, T , is true
given that a positive test result, e, is found, P (T | e). These probabilities can
be calculated from the sensitivity (P (e | T)) and the selectivity (P (¬e | ¬T))
of the test and the prior belief that T is true (P (T )) using Bayes’s theorem.
Let n denote sensitivity, that is, n = P(e | T), l denote selectivity, that is, l =
P(¬e | ¬T), and h denote the prior probability of drug A being toxic, that is,
h = P (T ), then,

nh
P (T | e)= (23)
nh + (1 − l )(1 − h)
l (1 − h)
P (¬T|¬e)= (24)
l (1 − h) + (1 − n)h

Sensitivity corresponds to the “hit rate” of the test and 1 minus the selectivity
corresponds to the “false positive rate.” There is a trade-off between sensitivity
and selectivity which is captured in the receiver-operating characteristic curve
(Green & Swets, 1966) and which plots sensitivity against the false positive
rate (1 – selectivity). Where the criterion is set along this curve will determine
the sensitivity and selectivity of the test.
Positive test validity is greater than negative test validity as long as the
following inequality holds:

h 2 (n − n2 ) > (1 − h)2 (l − l 2 ) (25)

Assuming maximal uncertainty about the toxicity of drug A, that is, P (T ) =

.5 = h, this means that positive test validity, P (T | e), is greater than negative
test validity, P (¬T | ¬e), when selectivity (l ) is higher than sensitivity. As
Oaksford and Hahn (2004) argue, this is often a condition met in practice for
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 293

a variety of clinical and psychological tests. Therefore, in a variety of settings,

positive arguments are stronger than negative arguments.12
Oaksford and Hahn (2004) also provide experimental evidence to the effect
that positive arguments such as (21) are indeed viewed as more convincing
than their negative counterparts under the conditions just described. The
evidence from their experiment further shows that people are sensitive to
manipulations in the amount of evidence (one versus fifty studies or tests) as
predicted by the account. Finally, participants in their experiment displayed
sensitivity to the degree of prior belief a character in a dialogue initially
displayed toward the conclusion, as the Bayesian account predicts. This find-
ing captures the “audience dependence” of argumentation assumed in the
rhetorical research tradition (e.g., Perelman & Olbrechts-Tyteca, 1969).
Hahn, Oaksford, and Bayindir (2005) generalised this account to other
versions of the argument from ignorance. One might still query whether the
account just outlined amounts to a satisfactory treatment of the argument
from ignorance. This is because the textbook example of the ghosts (19)
differs from Oaksford and Hahn’s (2004) experimental materials in one,
possibly important way. The argument for ghosts involves not only negative
evidence, but also a flip in polarity between evidence and conclusion: negative
evidence is provided to support the positive existence of something. In other
words the inference is of the form
not proven (not exist) → exist (26)
as opposed to merely
not proven (exist) → not exist (27)
The examples in Oaksford and Hahn (2004) arguably have the structure in
(27), not the structure in (26). But it may be the opposite polarity case (26)
that constitutes the true fallacy of the argument from ignorance.
Classical logic licenses an inference from not(not p) to p, but not the
inference underlying (26) which might be rendered as
not says (not p) → p (28)
This is because when one has not said “not p,” one can either have said “ p” or
not spoken about “ p” at all. For example, in an argument one might defend
oneself with the claim “I didn’t say you were rude,” which could be true either
because one had specifically claimed the opposite or because one had not

12 This means that people’s criterion is set such that positive evidence counts more for a hypothesis
than negative evidence counts against it.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

294 M. Oaksford and U. Hahn

mentioned rudeness at all. So maybe nothing at all can be inferred in such

cases?
Hahn, Oaksford, and Bayindir (2005) established that (28) can be regarded
as a strong argument by using a form of the argument from ignorance based
on epistemic closure which is related to the negation-as-failure procedure in
artificial intelligence (Clark, 1978). The case can be made with an informal
example: imagine your colleagues at work are gathering for a staff picnic. You
ask the person organizing the picnic whether your colleague Smith is coming,
to which you receive the reply that “Smith hasn’t said that he’s not coming.”
Should this allow you to infer that he is in fact coming, or has he simply
failed to send the required reply by e-mail? Your confidence that Smith will
be attending will vary depending on the number of people that have replied.
If you are told that no one has replied so far, assuming Smith’s attendance
seems premature; if by contrast you are told that everyone has replied, you
would be assured of his presence. In between these two extremes your degree
of confidence will be scaled: the more people have replied the more confident
you will be. In other words, the epistemic closure of the database in question
(the e-mail inbox of the organizer) can vary from no closure whatsoever to
complete closure, giving rise to corresponding changes in the probability that
not says (not p) does in fact suggest that p.
Hahn, Oaksford, and Bayindir’s (2005) experiments confirmed that people
are sensitive to variations in the epistemic closure of a database and that this
affects their willingness to endorse argument like (28). Moreover, in their
data, it was found that arguments like (28) can be regarded as even stronger
than the standard negative evidence case (27). Therefore, as our example
suggested, there would seem to be nothing in the structure of arguments like
the Ghosts example that make them inherently unacceptable.
The real reason we consider negative evidence on ghosts to be weak, that is,
why (19) is a weaker argument than (20), is because of the lack of sensitivity
(ability to detect ghosts) we attribute to our tests as well as our low prior
belief in their existence. That is, the weakness of this argument is due to
the probabilistic factors that affect the strength of the argument, that is,
P (conclusion|premises). Hahn and Oaksford (2006) have shown how this
account generalises to other inferential fallacies, such as circularity and the
slippery slope argument.
In summary, in this section we have shown how the concept of argument
strength can resolve the problem of why some instances of informal argument
fallacies nonetheless seem like perfectly acceptable arguments that should
rationally persuade an audience of the conclusion. This is the same problem
that we investigated in microcosm in the section on conditional inference.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 295

. 
Rips (2001, 2002) suggested that recent proposals in the psychology of de-
ductive reasoning seem to imply reasoning imperialism. That is, proponents
of the view that people evaluate even deductive arguments for inductive or
argument strength are interpreted as seeing no role for deductive correct-
ness (e.g., Oaksford & Chater, 1998). Moreover, some researchers seem to be
committed to the opposite point of view that deductive correctness can pro-
vide an account of inductive strength (Johnson-Laird, 1994; Johnson-Laird,
Legrenzi, Girotto, Legrenzi, & Caverni, 1999). While we can not reply for the
latter imperialists, we have attempted to reply for the former. However, as
we have pointed out, we do not view ourselves as arguing for an imperialist
position, and in concluding we will point out why. But first we summarise
what we hope has been learnt from this discussion.
We first addressed Rips’s (2001) paper. We showed that the data are perhaps
not as convincing as they may first seem, because the examples of deductively
incorrect and inductively strong arguments had no logical structure and yet
up to 35% of participants identified them as deductively correct. Moreover, we
argued that establishing the claim of sensitivity to deductive correctness can be
established only for a whole logic and not by examining isolated inferences. We
therefore looked at some other simple inferences that are valid in propositional
logic, the implicit standard in Rips (2001), and showed that some are para-
doxical, that is, they lead to intuitively unacceptable conclusions. That is, one
would not want to reason according to this standard. We diagnosed the prob-
lem as a mismatch between argument strength and deductive correctness –
if argument strength can be low for a deductively correct argument, it may be
viewed as paradoxical. This is prima facie strong evidence for the primacy of
argument strength.
We then looked at conditional inference in particular. We showed that
probability logic (Adams, 1998) uses an account of inductive strength which
resolves many of the paradoxes observed for standard deductive validity.
Moreover, we showed how a dynamic view of inference in probability logic
is consistent with Oaksford et al’s. (2000) probabilistic account. This account
also provides a better account of the data by invoking inductive processes
of updating beliefs in the light of possible counterexamples. The possibil-
ity of counterexamples arises because of possible violations of the rigidity
condition for conditionalization. We also evaluated the inductive strength of
some conditional argument forms and showed that people judge some con-
ditional reasoning fallacies to be as strong as a deductively correct inference.
This should not be possible according to the account Rips defends based on
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

296 M. Oaksford and U. Hahn

standard logic where all deductively correct inferences must have maximal
argument strength, that is, P (conclusion|premises) = 1.13
We finally showed that the concept of argument strength can be applied
much more generally to the informal argument fallacies, such as the argument
from ignorance, to resolve the paradox of why some instances of these fallacies
seem perfectly acceptable and capable of rationally persuading someone of
a conclusion. This again provides a strong argument for the centrality of
argument strength in human reasoning and argumentation.
However, as we said in the introduction to the chapter, the upshot of
all these arguments is not to deny that people possess some understand-
ing of deductive validity, but rather to establish that in determining hu-
man inferential behaviour, argument strength is the more central con-
cept. In AI knowledge representation, technical problems in developing
tractable non-monotonic logics that adequately capture human inferen-
tial intuitions has provided the impetus for the development of alterna-
tive approaches that explicitly include some measure of the strength of an
argument (Gabbay, 1996; Fox, 2003; Fox & Parsons, 1998; Pollock, 2001;
Prakken & Vreeswijk, 2002). It is this inclusion that provides these sys-
tems with their nice default properties. These systems often address un-
certainty and argument strength qualitatively rather than use the probabil-
ity calculus. In some accounts (e.g., Pollock, 2001) argument strengths are
assumed to range from 0 to ∞, whereas in others they are treated as qual-
itative categories with varying numbers of bins, for example, ++ , +, –, – –
(Fox & Parsons, 1998; see also some of the work reviewed in Prakken &
Vreeswijk, 2002). In the latter case, argument strengths can combine like
multivalued truth tables. Some of these rules for combining qualitative argu-
ment strengths are consistent with the probability calculus (Fox & Parsons,
1998). One important aspect of these approaches is that most use logical
entailment to determine whether a proposition is relevant to another propo-
sition and so can transmit an element of argument strength from one to the
other. These accounts illustrate an important fact that to build practical sys-
tems that track argument strength across complex arguments requires some
account of structure.
These systems also raise other interesting questions that reasoning re-
searchers probably need to address. For example, some of these systems are
not consistent with the probability calculus. The reason is the apparently

13 The reason this is not the case in probability logic is that the moment the Equation is accepted,
one is outside the realm of classical or standard logic, in which the probability of a conditional
is P(¬p) + P(p, q).
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 297

commonsense principle embodied in Theophrastus’s rule: the strength that a

chain of deductively linked arguments confers on the conclusion can not be
weaker than the weakest link in the chain (Walton, 2004). This is a condition
that can not be guaranteed by the probability calculus. Examples that seem
to conform to Theophrastus’ rule but not to the probability calculus have
persuaded Walton (2004) and, for example, Pollock (2001) that a third form
of reasoning should be countenanced in addition to deductive and induc-
tive/statistical reasoning, that is, plausibilist reasoning.
In argumentation theory, it is well recognized that many fallacies are fal-
lacies of relevance. For example, the argumentum ad misericordiam, that is,
the appeal to pity, may be relevant to arguments about, for example, whether
to make charitable donations despite the lack of a logical connection. This
suggests that accounts of argumentative relevance that rely purely on logical
consequence are unlikely to do justice to the concept of relevance required
by a theory of informal argument. Similar arguments have been made with
respect to relevance logics (Anderson & Belnap, 1970). The concept of rele-
vance goes beyond logical entailment, even relevant entailment. For example,
it is relevant to whether an animal has palpitations that it has a heart, but
this is due to the causal structure of the world, not the logical relations
between propositions (Oaksford & Chater, 1991; Veltman, 1986). In sum,
existing theories of argumentation that transmit plausibilities via logical con-
sequence relations may not capture all the ways in which relevance relations
are established in argument. Possible alternatives include informational rele-
vance in Bayesian networks (Pearl, 1988). However, such networks also rely
on structure to get around the unconstrained nature of general probabilistic
reasoning.
The relevance of these AI approaches to human cognition could be ques-
tioned. However, these attempts to provide computational accounts of reason-
ing processes do place constraints on the computations that must underpin
human reasoning in the brain. The constraint on which they are unambiguous
is that basing an account of inference on argument strength, probabilistic or
not, is going to require an account of structure, logical or not. Consequently,
given that most practical systems currently use logical entailment to transmit
argument strengths, it seems best to view these notions as cognitively com-
plementary. This view has the consequence that while most judgements of the
acceptability of an argument are based on argument strength, people must
also be sensitive to the deductive relations between propositions in order to
appropriately transmit them in a passage of reasoning. Thus even if there are
dissociable brain regions dedicated to deduction and induction (Goel, Gold,
Kapur, & Houle, 1997; Osherson, Perani, Cappa, Schnur, Grassi, & Fazio,
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

298 M. Oaksford and U. Hahn

1998), deductive correctness and inductive strength should generally be seen

as working together rather than as independent systems.

References
Adams, E. W. (1975). The logic of conditionals. Dordrecht: Reidel.
Adams, E. W. (1998). A primer of probability logic. Stanford, CA: CSLI Publications.
Anderson, A. R., & Belnap, N. D. (1975). Entailment: The logic of relevance and necessity,
Vol. 1, Princeton: Princeton University Press.
Anderson, J. R. (1995). Cognitive psychology and its implications. New York: W. H.
Freeman.
Bennett, J. (2003). A philosophical guide to conditionals. Oxford: Oxford University Press.
Bishop, Y. M. M., Feinberg, S. E., & Holland, P. W. (1975). Discrete multivariate analysis.
Cambridge, MA: MIT Press.
Chan, D., & Chua, F. (1994). Suppression of valid inferences: Syntactic views, mental
models, and relative salience. Cognition, 53, 217–238.
Chater, N., & Oaksford, M. (1996). The falsity of folk theories: Implications for psy-
chology and philosophy. In W. O’Donohue & R. F. Kitchener (Eds.), The philosophy
of psychology (pp. 244–256). London: Sage Publications.
Chater, N., & Oaksford, M. (1999). The probability heuristics model of syllogistic rea-
soning. Cognitive Psychology, 38, 191–258.
Clark, K. L. (1978). Negation as failure. In H. Gallaire & J. Minker (Eds.), Logic and
databases (pp. 293–322). New York: Plenum Press.
Copi, I. M., & Cohen, C. (1990). Introduction to logic (8th Ed.). New York: Macmillan
Press.
Dhami, M. K., & Wallsten T. S. (2005). Interpersonal comparison of subjective proba-
bilities: Towards translating linguistic probabilities. Memory and Cognition, 33, 1057–
1068.
Edgington, D. (1991). The matter of the missing matter of fact. Proceedings of the
Aristotelian Society, Suppl. Vol. 65, 185–209.
Evans, J. St. B. T., & Over, D. E. (2004). If. Oxford: Oxford University Press.
Evans, J. St. B. T., Handley, S. J., & Over, D. E. (2003). Conditionals and conditional
probability. Journal of Experimental Psychology: Learning, Memory, & Cognition, 29,
321–335.
Enderton, H. (1972). A mathematical introduction to logic. New York: Academic
Press.
Fox, J. (2003). Probability, logic and the cognitive foundations of rational belief. Journal
of Applied Logic, 1, 197–224.
Fox, J., & Parsons, S. (1998). Arguing about beliefs and actions. In A. Hunter and
S. Parsons (Eds.), Applications of uncertainty formalisms (Lecture notes in artificial
intelligence 1455) (pp. 266–302). Berlin: Springer Verlag.
Gabbay, D. (1996). Labelled deduction systems. Oxford: Oxford University Press.
George, C. (1997). Reasoning from uncertain premises. Thinking and Reasoning, 3,
161–190.
Green, D. M., & Swets, J. A., (1966). Signal detection theory and psychophysics. New York:
Wiley.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 299

Hahn, U., & Oaksford, M. (2006). A Bayesian approach to informal argument fallacies.
Synthese, 152, 207–236.
Hahn, U., Oaksford, M., & Bayindir, H. (2005). How convinced should we be by neg-
ative evidence? In B. Bara, L. Barsalou, and M. Bucciarelli (Eds.), Proceedings of the
27th Annual Conference of the Cognitive Science Society (pp. 887–892), Mahwah, NJ:
Lawrence Erlbaum.
Hahn, U., Oaksford, M., & Corner, A. (2005). Circular arguments, begging the question
and the formalization of argument strength. In A. Russell, T. Honkela, K. Lagus, & M.
Pöllä, (Eds.), Proceedings of AMKLC’05, International Symposium on Adaptive Models
of Knowledge, Language and Cognition (pp. 34–40), Espoo, Finland, June 2005.
Hamblin, C. L. (1970). Fallacies. London: Methuen.
Hattori, M., & Oaksford, M. (in press). Adaptive non-interventional heuristics for co-
variation detection in causal induction: Model comparison and rational analysis.
Cognitive Science.
Heit, E., & Rotello, C. M. (2005). Are there two kinds of reasoning? In B. Bara, L.
Barsalou, & M. Bucciarelli (Eds.), Proceedings of the 27th Annual Conference of the
Cognitive Science Society (pp. 923–928). Mahwah, NJ: Lawrence Erlbaum.
Hughes, G. E., & Cresswell, M. J. (1968). An introduction to modal logic. London:
Methuen.
Johnson-Laird, P. N., & Byrne, R. M. J. (1991). Deduction. Hove, Sussex: Lawrence
Erlbaum.
Johnson-Laird, P. N., & Byrne, R. M. J. (2002). Conditionals: A theory of meaning,
pragmatics, and inference. Psychological Review, 109, 646–678.
Johnson-Laird, P. N., Legrenzi, P., Girotto, V., Legrenzi, M. S., & Caverni, J. P. (1999).
Naive probability: A mental model theory of extensional reasoning. Psychological
Review, 106(1), 62–88.
Lewis, D. (1973). Counterfactuals. Oxford: Blackwell.
Lewis, D. (1976). Probabilities of conditionals and conditional probabilities. Philosoph-
ical Review, 85, 297–315.
Liu, I. M. (2003). Conditional reasoning and conditionalisation. Journal of Experimental
Psychology: Learning, Memory, & Cognition, 29, 694–709.
Liu, I. M., Lo, K., & Wu, J. (1996). A probabilistic interpretation of “If-then.” Quarterly
Journal of Experimental Psychology, 49A, 828–844.
Massey, G. J. (1981). The fallacy behind fallacies. Midwest Studies in Philosophy, 6, 489–
500.
Nute, D. (1984). Conditional logic. In D. Gabbay & F. Guenthner (Eds.), Handbook of
philosophical logic, Vol. 2 (pp. 387–439). Dordrecht: Reidel.
Oaksford, M., & Chater, N. (1991). Against logicist cognitive science. Mind & Language,
6, 1–38.
Oaksford, M., & Chater, N. (1994). A rational analysis of the selection task as optimal
data selection. Psychological Review, 101, 608–631.
Oaksford, M., & Chater, N. (1996). Rational explanation of the selection task. Psycho-
logical Review, 103, 381–391.
Oaksford, M., & Chater, N. (1998). Rationality in an uncertain world: Essays on the
cognitive science of human reasoning. Hove, Sussex: Psychology Press.
Oaksford, M., & Chater, N. (2001). The probabilistic approach to human reasoning.
Trends in Cognitive Sciences, 5, 349–357.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

300 M. Oaksford and U. Hahn

Oaksford, M. & Chater, N. (2002). Commonsense reasoning, logic and human rational-
ity. In R. Elio (Ed.), Common sense, reasoning and rationality (pp. 174–214). Oxford:
Oxford University Press.
Oaksford, M., & Chater, N. (2003a). Computational levels and conditional reasoning:
Reply to Schroyens and Schaeken (2003). Journal of Experimental Psychology: Learning,
Memory, & Cognition, 29, 150–156.
Oaksford, M., & Chater, N. (2003b). Conditional probability and the cognitive science
of conditional reasoning. Mind & Language, 18, 359–379.
Oaksford, M., & Chater, N. (2003c). Modeling probabilistic effects in conditional infer-
ence: Validating search or conditional probability? Revista Psychologica, 32, 217–242.
Oaksford, M., & Chater, N. (2003d). Probabilities and pragmatics in conditional infer-
ence: Suppression and order effects. In D. Hardman & L. Macchi (Eds.), Thinking:
Psychological perspectives on reasoning, judgment and decision making (pp. 95–122).
Chichester, UK: John Wiley & Sons.
Oaksford, M., & Chater, N. (2003e). Optimal data selection: Revision, review and re-
evaluation. Psychonomic Bulletin & Review, 10, 289–318.
Oaksford, M., & Chater, N. (2007). Bayesian rationality: The probabilistic approach to
human reasoning. Oxford: Oxford University Press.
Oaksford, M., Chater, N., & Larkin, J. (2000). Probabilities and polarity biases in condi-
tional inference. Journal of Experimental Psychology: Learning, Memory & Cognition,
26, 883–899.
Oaksford, M., & Hahn, U. (2004). A Bayesian approach to the argument from ignorance.
Canadian Journal of Experimental Psychology, 58, 75–85.
Oberauer, K., & Wilhelm, O. (2003). The meaning of conditionals: Conditional prob-
abilities, mental models, and personal utilities. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 29, 321–335.
Osherson, D., Perani, D., Cappa, S., Schnur, T., Grassi, F., & Fazio, F. (1998). Distinct
brain loci in deductive vs. probabilistic reasoning. Neuropsychologica, 36, 369–376.
Pearl, J. (1988). Probabilistic reasoning in intelligent systems: Networks of plausible infer-
ence. San Mateo, CA: Morgan Kaufman.
Perelman, C., & Olbrechts-Tyteca, L. (1969). The new rhetoric: A treatise on argumenta-
tion. Notre Dame, IN: University of Notre Dame Press.
Politzer, G. (2005). Uncertainty and the suppression of inferences. Thinking and Rea-
soning, 11, 5–34.
Pollock, J. L. (2001). Defeasible reasoning with variable degrees of justification. Artificial
Intelligence, 133, 233–282.
Prakken, H., & Vreeswijk, G. A.W. (2002). Logics for defeasible argumentation. In D.
M. Gabbay & F. Guenthner (Eds.), Handbook of philosophical logic, 2nd ed., Vol. 4
(pp. 219–318). Dordrecht: Kluwer Academic Publishers.
Quinn, S., & Markovits, H. (2002). Conditional reasoning with causal premises: Evidence
for a retrieval model. Thinking and Reasoning, 8, 179–19.
Rapoport, A., Wallsten, T. S., & Cox, J. A. (1987). Direct and indirect scaling of mem-
bership functions of probability phrases. Mathematical Modelling, 9, 397–417.
Rips, L. J. (1994). The psychology of proof. Cambridge, MA: MIT Press.
Rips, L. J. (2001a). Two kinds of reasoning. Psychological Science, 12, 129–134.
Rips, L. J. (2001b). Reasoning imperialism. In R. Elio (Ed.), Common sense, reasoning,
and rationality (pp. 215–235). Oxford University Press.
P1: JZP
05216772443c11 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:43

Induction, Deduction, and Strength 301

Rips, L. J. (2002). Reasoning. In H. F. Pashler (Series Ed.) and D. L. Medin (Vol. Ed.),
Stevens’ handbook of experimental psychology: Vol. 2. Cognition (3rd Ed.). New York:
Wiley.
Schroyens, W., & Schaeken, W. (2003). A critique of Oaksford, Chater, and Larkin’s
(2000) conditional probability model of conditional reasoning. Journal of Experimen-
tal Psychology: Learning, Memory, & Cognition, 29, 140–149.
Schroyens, W. J., Schaeken, W., & Handley, S. (2003). In search of counter-examples: De-
ductive rationality in human reasoning. Quarterly Journal of Experimental Psychology,
56A, 1129–1145.
Scott, D. (1971). On engendering an illusion of understanding. The Journal of Philosophy,
68, 787–807.
Skyrms, B. (1975). Choice and chance (2nd Ed.). Belmont, CA: Wadsworth.
Stalnaker, R. C. (1968). A theory of conditionals. In N. Rescher (Ed.), Studies in Logical
Theory (pp. 98–113). Oxford: Blackwell.
Stevenson, R. J., & Over, D. E. (1995). Deduction from uncertain premises. Quarterly
Journal of Experimental Psychology, 48A, 613–643.
Veltman, F. (1985). Logics for conditionals. PhD. Thesis, Faculteit der Wiskunde en
Natuurwetenschappen, University of Amsterdam.
Wallsten, T. S., Budescu, D. V., Rapoport, A., Zwick, R., & Forsyth, B. (1986). Measuring
the vague meanings of probability terms. Journal of Experimental Psychology: General,
115, 348–365.
Walton, D. N. (2004). Relevance in argumentation. Mahwah, NJ: Lawrence Erlbaum.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

12

Individual Differences, Dual Processes, and Induction

Aidan Feeney

In this chapter I hope to demonstrate that answering the question “who does
what in reasoning experiments?” can also help us answer fundamental ques-
tions about the nature of thought. Because it is nearly always the case that
some experimental participants display phenomena of interest and others
don’t, almost every experiment run by cognitive psychologists produces data
on individual differences. But, as most psychologists do not wish to take dif-
ferences between their participants as the starting point of their investigations,
individual differences tend to be ignored. Thus, the means that experimental
psychologists report when they describe their data abstract across differences
between individuals, with only the standard deviation or standard error in-
dicating the extent to which participants varied in their responses. In this
chapter I will address what the study of differences between individuals might
tell us about a range of issues in inductive reasoning, including what consti-
tutes a good inference, what processes underlie induction, and how induction
differs from deduction.
My primary focus will be on the general question of how individual-
differences data support dual-process theories of thinking (Evans & Over,
1996; Stanovich, 1999; Sloman, 1996). Dual-process theories have been ap-
plied to deduction, decision making, and induction. I will outline the dual-
process approach and describe evidence in its favour, paying particular atten-
tion to individual differences. I will also describe how individual-differences
methodology has been used to arbitrate between different normative ac-
counts of specific thinking tasks (see Stanovich & West, 1998b, c). With this
background covered I will consider what individual differences might tell us
about category-based induction. In particular, what can the method tell us
above and beyond more standard experimental methods about people’s in-
ductive reasoning abilities? Does the method have anything to say about how
a dual-process approach might be applied to induction across categories? And
302
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 303

can it inform the debate about the normative status of inductive phenom-
ena? I will describe the results of two studies designed with these questions
in mind.
I will conclude with a consideration of other questions about induction
to which the individual-differences methodology might help provide some
answers. I will briefly describe a recent study designed to investigate who
is susceptible to inductive reasoning fallacies, that is, judgements about the
strength of inductive arguments that depart from the prescriptions of some
normative theory of induction. In addition, I will consider recent claims
about dissociations between induction and deduction (Rips, 2001; Rips and
Asmuth, Chapter 10, this volume; Oaksford and Hahn, Chapter 11, this
volume; Heit & Rotello, 2005). I will outline a dual-process perspective on
this dissociation and argue that an individual-differences approach allows
one to test the dual-process account. First, however, as dual-process theories
of thinking have motivated much of the individual differences work I will
describe, it is important to review them in some detail here.

-   

It is now a relatively common claim that there are two types of thinking.
Roughly speaking, these types correspond to an intuitive mode of thought
and a more analytic way of thinking. Recently, there have been at least three
attempts to put this claim on a sound theoretical footing (Evans & Over,
1996; Sloman, 1996; Stanovich, 1999; for recent reviews see Evans, 2003, and
Osman, 2005). All versions of the theory agree that one type of thinking
is fast, associative, and automatic while the other is slow, rule-based, and,
to some degree at least, controlled. I will refer to the first of these types of
thinking as Type 1 and to the second as Type 2. All three theories claim
that the two types of thinking overlap and that they are not confined to
any particular thinking task. For this reason the distinction between types of
reasoning is orthogonal to the philosopher’s distinction between induction
and deduction (Sloman, 1996; but see Oaksford and Hahn, Chapter 11, this
volume; and Rips and Asmuth, Chapter 10, this volume). They also agree that
when engaged in the same thinking task, the systems can produce conflicting
responses. Sloman enumerates a range of thinking tasks where he claims that
conflicting responses are available. These range from Tversky and Kahneman’s
(1983) conjunction fallacy to the case of belief bias in syllogistic reasoning (see
Evans, Barston, & Pollard, 1983). Evans and Over (1996) and Stanovich (1999)
explicitly claim that the second type of thinking calls on working memory
resources and hence is likely to be predicted by measures of cognitive ability.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

304 A. Feeney

As Sloman equates Type-2 thinking with symbol manipulation, this claim is

implicit in his account.
Given that claims about the existence of dual processes are almost always
contentious, I will briefly review the evidence most often cited to support
dual-process claims about thinking. Conveniently for the purposes of this
overview, the task used to measure belief bias in syllogistic reasoning (Evans,
Barston, & Pollard, 1983) has been explored using a variety of methodologies
and participants. In the syllogisms used in belief bias tasks, participants are
presented with two quantified premises and are asked to verify a conclusion
provided by the experimenter. For example:
All beekeepers are artists Argument 1
All artists are archers
Therefore, all beekeepers are archers
Participants are presented with syllogisms whose conclusions either neces-
sarily follow or do not follow from the premises and which are believable or
unbelievable given background knowledge. A believable and logically valid
problem taken from Evans et al. (1983) is as follows:
No police dogs are vicious Argument 2
Some highly trained dogs are vicious
Therefore, some highly trained dogs are not police dogs
Problem 3 is an unbelievable invalid problem taken from the same experiment.
No millionaires are hard workers Argument 3
Some rich people are hard workers
Therefore, some millionaires are not rich people
Evans and colleagues found main effects of logic and of conclusion believabil-
ity in their study. That is, participants were more likely to accept conclusions
that were either believable or logically necessary given the premises. Further-
more, they showed that the results were not due to the existence of separate
groups of logical and belief-based responders in their sample. Sometimes
individual participants responded on the basis of belief and sometimes on
the basis of what was necessarily true. This is one of the best examples of
thinking tasks that give rise to conflicting responses (see Sloman, 1996). That
is, Type-1 and Type-2 processes produce different answers in competition
with one another. Sometimes the Type-1 answer is accepted whereas on other
occasions the Type-2 response overcomes the Type-1 response.
There are a number of lines of investigation that support the dual-process
interpretation of Evans et al.’s results. For example, Gilinsky and Judd (1994)
showed that susceptibility to effects of belief on conflict problems increases
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 305

with age. This makes sense if logical or Type-2 thinking is associated with
working memory and IQ, both of which decline as we get older (e.g.,
Hasher & Zacks, 1988; Horn & Cattell, 1967). In another line of investigation
it has been shown that susceptibility to belief bias on conflict problems may be
reduced by stressing the concept of logical necessity in the instructions (Evans,
Allen, Newstead, & Pollard, 1997). This result is predicted because of the con-
trolled nature of Type-2 thinking processes. Further evidence for the dual-
process interpretation comes from a recent study by Evans and Curtis-Holmes
(2005) using a speeded response paradigm (for a related use of this paradigm
see Shafto, Colony, and Vitkin, Chapter 5 this volume). When participants
were required to make a speeded response, they were more susceptible to the
effects of belief in conflict problems. To summarise then, responses based on
Type-1 processes are primary, but responses based on Type-2 processing can
dominate if participants are instructed to suppress Type-1 outputs, if they have
sufficient time for the slower Type-2 processes to operate, or if they are young.
Before proceeding it is worthwhile to consider two other recent sources
of particularly compelling evidence. Goel and Dolan (2003) have reported
the results of an MRI study of belief bias in which they found very strong
evidence for separate types of thinking. When participants gave belief-based
responses on conflict problems, researchers observed increased activation in
ventral medial prefrontal cortex (VMPFC). When these same participants
gave logical responses on conflict problems, the researchers observed in-
creased activation in right inferior prefrontal cortex. Goel and Dolan argue
that VMPFC is associated with the effects of beliefs in reasoning (see also
Adolphs, Tranel, Bechara, Damasio, & Damasio, 1996), whereas activation in
inferior prefrontal cortex is due to the inhibition of a response based on Type-1
thinking processes in favour of a Type-2 or logical response. The notion that
inhibition is important to logical reasoning is further supported by some
data on children’s reasoning reported by Handley, Capon, Beveridge, Dennis,
and Evans (2004). These authors found that scores on the conflict index in
children aged ten years were correlated with performance on a task designed
to measure ability to inhibit a prepotent response. Children who were more
likely to respond logically had higher inhibitory control. Both of these studies
suggest the existence of two systems for reasoning, one of which is dominant
and whose effects must be suppressed by the other.
Handley et al.’s study is interesting in the current context because it
employed an individual-differences methodology. However, the individual-
differences variable measured was response inhibition and, although there is
currently considerable interest in the relationship between measures of inhi-
bition and working memory (see Kane & Engle, 2003), there has been little
other work on inhibition and thinking. By far the most popular individual
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

306 A. Feeney

differences variable to have been employed by researchers interested in dual-

process theories of thinking is general cognitive ability. It is to a review of the
literature on cognitive ability and thinking that I will now turn.

    

 
In a series of papers, Keith Stanovich and his colleagues have described the re-
sults of a large-scale research program designed to investigate the relationship
between cognitive ability and thinking. For the most part using SAT scores
as an indicator of ability, these researchers have investigated which tasks are
predicted by ability and which are not. This program of research is notable
because of the range of deductive, decision-making, and (to some extent)
inductive tasks it has investigated. It is motivated by two arguments, the first
of which is that individual-differences studies allow one to test dual-process
theories of thinking. Explicit in Evans and Over’s and Stanovich’s formulation
of the theory (and implicit in Sloman’s) is the claim that Type-2 processes
are subject to capacity limitations whereas Type-1 processes are not. Thus, on
tasks where Type-2 processes are required in order for the normatively correct
answer to be produced, participants who are lower in cognitive resources will
tend to perform more poorly than those who are high in cognitive resources.
There are a number of lines of work in the literature linking IQ measures
to general cognitive constraints. First, performance on intelligence tests is
held to be constrained by basic cognitive abilities. For example, Carpenter,
Just, and Shell (1990) gave an account of individual differences in the series-
completion problems used in Raven’s Progressive Matrices test. Their account
implicated variations in the ability to induce abstract relations and the ability
to coordinate goals in working memory as the basis for individual differences.
In support of this account, Unsworth and Engle (2005) have recently shown
that individual differences in working memory predict performance on Raven.
Although other evidence (e.g., Fry & Hale, 1996) suggests that processing
speed mediates the relationship between memory and IQ, nonetheless there
is very good evidence for relationships between constraints on basic cognitive
abilities and performance on tests of ability. As SAT scores are known to
be very closely related to IQ (Frey & Detterman, 2004), it is reasonable to
conclude that the same constraints limit SAT performance.
According to Stanovich (1999), the finding that correct performance on
certain tasks, but not others, is associated with SAT scores lends support to
the suggestion that there are at least two types of thinking and that each is
involved to different extents in the accomplishment of a variety of thinking
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 307

tasks. Stanovich (1999, pg. 143) claims that large differences in performance
due to cognitive ability will be found only for thinking tasks that engage both
systems. When a thinking task engages the associative system only, then no
difference will be observed. In addition, individual differences are more likely
to be found when the systems produce conflicting responses.
Stanovich illustrates this argument with reference to performance on de-
ontic and indicative versions of Wason’s selection task (see Stanovich & West,
1998b). In the selection task participants are presented with a conditional
rule that is said to govern a set of instances. In standard indicative versions
of the task, the rule governs what is printed on each side of sets of cards. For
example, “If there is an A on one side of the card (p), then there is a 3 on
the other side (q).” Participants are shown four cards. On the visible sides are
printed an A (p), an L (not-p), a 3 q), and a 7 (not-q). Participants are asked
which card or cards need to be turned over to test the rule. Deontic versions of
the selection task concern what should or ought to be done. Most often, they
concern social laws, obligations, or permissions. For example, “If someone is
drinking alcohol in a bar (p) then they must be over 18 (q)”. Printed on the
visible side of the cards in this example might by “Beer” (p), “Coke” (not-p),
“21” (q), and “16” (not-q).
One of the earliest findings in the literature on the indicative selection
task (see Wason, 1966) was that people tended not to select the cards that
could falsify the rule (the p and the not-q cards in the example above).
However, when deontic content is included (see Johnson-Laird, Legrenzi, &
Legrenzi, 1972), people regularly select the falsifying cases. There are many
explanations of why people behave differently on the indicative and deontic
tasks (see Cheng & Holyoak, 1985; Cosmides, 1986; Manktelow & Over,
1991; Oaksford & Chater, 1994). Stanovich argues that most explanations
suggest that people’s selections on the deontic task are not determined by
conscious, analytic reasoning but by largely unconscious, heuristic processes.
Correct responding on the indicative task, on the other hand, requires that
the analytic system overcome the unconscious associative system which, in
most participants, is responsible for the selection of those cards that match
the items in the rule.
If Stanovich is right, then correct performance on the indicative task should
be associated with cognitive ability whereas correct performance on the de-
ontic task should not. Stanovich and West (1998b) confirmed this predic-
tion. They found clear evidence that people who reported higher SAT results
were more likely to select the potentially falsifying cards on the indicative
selection task. There was much less evidence for such an association on
the deontic task. Broadly similar results have been reported by Newstead,
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

308 A. Feeney

Handley, Harley, Wright, and Farrelly (2004), who also found that whether
associations between ability and performance on deontic and indicative tasks
are found depend on the ability range of the participants sampled.
The second argument that Stanovich uses to motivate the individual-
differences approach relates to what counts as the correct normative response
on any particular task. There has been substantial debate about which is
the correct normative account of tasks as varied as Kahneman and Tversky’s
(1983) Linda problem and Wason’s selection task (see Cohen, 1981; Oaksford
& Chater, 1994). Stanovich and West (1998b, c) argued that an association
between cognitive ability and certain patterns of performance on these tasks
might favour one normative account over another (see also Cohen, 1981).
For example, Oaksford and Chater (1994) have recast Wason’s selection task
as one of deciding which card or cards is likely to be most informative. From
a Bayesian perspective, given certain assumptions about the probability of
items mentioned in the conditional rule, it turns out that the most informa-
tive cards on the indicative task are the p and q cards. Oaksford and Chater
argue that their Bayesian reanalysis of the task is closer to a correct normative
analysis than is the Popperian notion of falsification, which claims that the
“correct” solution is to select the p and not-q cards.
To arbitrate between these competing normative accounts, Stanovich and
West looked for associations between cognitive ability and the tendency to
select cards in accordance with each of these distinct normative approaches.
They found that more able participants were more likely to select the falsifying
cards than were the less able participants. Less able participants, on the other
hand, were more likely to select the p and q cards. Stanovich concludes that
the individual-differences data suggest that from a normative point of view,
it is more appropriate to think about the indicative selection task as one
requiring falsification rather than decisions about expected information gain.
Of course this argument may not always hold. For example, if one happens
to sample from the bottom of the range of ability scores, then the most able
participants in a particular sample may very well perform in non-normative
ways on a particular thinking task (see Newstead et al., 2004). However, if
one consistently finds, across a range of samples and abilities, that the most
able participants consistently favour one response over another, then this
lends support to the claim that the response favoured by these participants is
the one that is normatively correct.
Unfortunately, although Stanovich and colleagues have investigated a num-
ber of phenomena (including belief bias), very little data has been pub-
lished on the relationship between inductive reasoning and general cog-
nitive ability. One exception is work by Stanovich and West (1998a) on a
task where participants make a decision in the face of conflicting evidence
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 309

comprised of a vivid testimonial from a friend and a statistical summary of the

experiences of a number of people (see also Jepson, Krantz, & Nisbett, 1983;
Fong, Krantz, & Nisbett, 1986). This task satisfies Stanovich’s (1999) condi-
tions for the observation of an association between intelligence and norma-
tively correct responses. This is because in order to resist a response based
on the vivid single case, Type-2 processes are required to decontextualise the
problem. A response based on the vivid non-statistical information is likely
to be delivered by context-sensitive Type-1 processes. Stanovich and West
(1998a) report correlations of .33 and .38 between the tendency to resist the
effects of the salient single case in favour of the statistical evidence and a com-
posite measure of thinking ability. This association suggests that sensitivity
to evidence quality in inductive reasoning is related to cognitive ability and
hence reflects the operation of Type-2 thinking processes.

   :  

 
What might an individual-differences approach tell us about category-based
induction where people are asked about their willingness to project a (usually)
blank property from members of one or more premise categories to members
of another conclusion category? One potential use of the methodology is to
explore the diversity effect (see Heit, Chapter 1, this volume; Shafto, Coley, &
Vitkin Chapter 5, this volume). Consider Arguments 4 and 5 below.
Tigers have property X Argument 4
Cows have property X
All mammals have property X
Tigers have property X Argument 5
Lions have property X
All mammals have property X
Participants are said to be sensitive to evidential diversity if they prefer argu-
ments based on dissimilar rather than similar premise categories. Because the
premises in Argument 4 concern more diverse categories than do the premises
in Argument 5, Argument 4 is said to be stronger than Argument 5. As we
will see, an individual-differences approach may be useful in approaching
normative and psychological questions about diversity.

The Normative Argument for Diversity

Although there is general agreement that, all else being equal, diverse evidence
is strong evidence, it has been difficult to formally prove this intuition to be
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

310 A. Feeney

true. As a consequence there has been disagreement in the literature about

whether diversity is a sound normative principle for inductive reasoning. Sen-
sitivity to diversity is advocated by a range of philosophers of science (Nagel,
1939; Carnap, 1950; Hempel, 1966), and in the psychological literature the
soundness of diversity is assumed in a variety of models (see Osherson et al.,
1990; Sloman, 1993; Heit, 1998). In addition, a variety of Bayesian justifica-
tions for its soundness have been offered (see Howson & Urbach, 1989).
However, Bayesian accounts of diversity have been called into question
(Wayne, 1995; for a reply see Myrvold, 1995), and arguments against the
generality of a diversity principle have been made by Lo, Sides, Rozelle, and
Osherson (2002) and by Medin, Coley, Storms, and Hayes (2003). Medin et al.
have demonstrated a number of exceptions to diversity (for an experimental
reappraisal of one of these exceptions see Heit & Feeney, 2005). Other ex-
ceptions to diversity include those found in cross-cultural work comparing
Itzaj-Mayan Indians to North American students. Although the students dis-
played sensitivity to diversity (see López, Atran, Coley, Medin, and Smith,
1997), members of the Itzaj tribe did not. Proffitt, Coley, and Medin (2000)
showed that some tree experts do not show strong diversity effects when
reasoning about trees. Reviewing all of this work, Heit, Hahn, and Feeney
(2005) have argued that these exceptions to diversity do not invalidate the
diversity principle. Instead, they may involve the sensible use of background
knowledge beyond simple similarity information when evaluating arguments.
Knowledge about diversity is but one of many sources of knowledge that might
be used to evaluate inductive inferences.
Heit et al.’s attempt at reconciliation notwithstanding, there is clearly some
disagreement about the normative status of the diversity principle. In some
ways (but not in others), this disagreement parallels the normative debates we
encountered when considering Stanovich’s individual-differences approach to
thinking. Below I will describe two individual-difference studies that apply
Stanovich’s logic to the diversity effect. These studies examined whether there
is a positive correlation between sensitivity to evidential diversity and cog-
nitive ability. If such a correlation were to be observed, then, by Stanovich’s
argument, we would have additional reason to believe that, all things being
equal, sensitivity to diversity is a normatively sensible strategy.

Dual-Process Predictions
Using an individual-differences approach to study diversity will also enable us
to give preliminary consideration to how a dual-process view might be applied
to category-based induction. Certainly, at least one theory of induction makes
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 311

clear dual-process predictions. Sloman – a dual-process theorist – has also put

forward an important account of category-based induction called the feature-
based model (FBM) (Sloman, 1993). According to this model, category-
based induction is explicable without the need to assume a stable category
structure. In the FBM the premises of an argument are encoded by training
a connectionist network to associate the property to be projected with the
features of the categories in the premises. Argument strength is determined
by the amount of activation of the target property in the presence of the
conclusion category. The greater the featural overlap between premise and
conclusion categories, the stronger the argument. In this approach, diverse
premises lead to associations with a wider variety of features and hence are
more likely to lead to strong activation of the target property in the presence of
the conclusion category. This similarity-based model is associative in nature.
It claims that category-based induction is achieved by Type-1 processes and,
therefore, that the tendency to be sensitive to evidential diversity should not
be associated with a measure of cognitive ability. However, Sloman (1996) has
made the point that although a feature-based process may be pre-eminent
in category-based induction, rule-based or Type-2 processes may also play a
role. Furthermore, under certain conditions, relationships between categories
do affect judgements of inductive strength (see Experiment 4, Sloman, 1998).
Individual-differences studies allow us to test the strong predictions of the
FBM. Any indication of a significant association between ability and sensitivity
to diversity would suggest that, contrary to the predictions of the FBM, Type-
1 and Type-2 processes may be involved in the evaluation of category-based
inductive arguments.
It is somewhat more difficult to see how dual-process theory might map
onto other theories of category-based induction. For example, according to
Osherson and colleagues’ Similarity-Coverage Model (SCM), the strength of
a category-based inductive argument is determined by the similarity of the
categories in the premises to the category in the conclusion and the degree
to which the categories in the premises cover the superordinate category that
includes all categories mentioned in the argument (Osherson, Smith, Wilkie,
López, & Shafir, 1990). It is the coverage component of the model that accounts
for diversity effects. The coverage measure is the average maximum similarity
between the categories in the premises and instances of the superordinate
category that come to mind when assessing the argument. As diverse categories
in the premises are likely to be similar to a wider variety of instances from the
superordinate, the coverage measure in the model tends to be elevated by the
presence of diverse premise categories. This model is different from the FBM
because it assumes the existence of a stable category structure. It also predicts
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

312 A. Feeney

that for arguments like 6 and 7, where the conclusion category is specific
rather than general, in order to display sensitivity to diversity, people must
generate the lowest level superordinate category that contains the categories
in the premises and the conclusion.
Bears have property X Argument 6
Gorillas have property X
Tigers have property X
When the conclusion in the argument is general, however, as it is in Argument
4, then people do not have to generate a covering category.
Bears have property X Argument 7
Mice have property X
Tigers have property X
The need to generate a covering category has been implicated by López,
Gelman, Gutheil, and Smith (1992) as an important factor in the develop-
ment of category-based induction. Lopez et al. found developmental trends
where sensitivity to diversity emerged for general arguments before specific
arguments. An individual-differences study will allow for an alternative test
of Lopez’s argument. If generating a covering category is an additional source
of difficulty, then sensitivity to diversity for specific arguments should be
related to cognitive ability. In addition, we should observe less sensitivity to
diversity when the conclusion category is specific rather than general. Perhaps
we might also expect to observe a weaker relationship between sensitivity to
diversity and ability with specific conclusions, as only the most able people in
any sample may display such sensitivity at above chance levels.
The relevance account of category-based induction (Medin, Coley, Storms,
& Hayes, 2003) attributes diversity effects to participants’ hypotheses about
the nature of the blank feature. Although they focus on the roles played by
causal and environmental knowledge in producing exceptions to diversity,
when no such knowledge is available, participants will use the degree of
similarity between the categories in the premises to guide their hypothesis
formation. Dissimilar categories lead to hypotheses that the feature is general
and hence highly projectible, whereas similar categories lead to hypotheses
about specific and hence non-projectible features (see also Heit, 1998).
The relevance account appears to give no grounds for dual-process predic-
tions in its current form. However, the relevance theory of linguistic prag-
matics (Sperber & Wilson, 1986/1995), upon which the relevance account is
based, uses logical rules for non-demonstrative inference (see Sperber &
Wilson, 1995, chapter 2). Thus, people work out the implications of an
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 313

utterance by deriving deductive conclusions from assumptions made avail-

able by the context and what is explicated by the utterance. For such an
approach to be applied to category-based induction, it is likely that people
would have to explicitly represent principles of inductive inference and draw
inferences about what follows from those rules and information in the ut-
terance. For example, given dissimilar premise categories, people might infer
that the blank property is general and, based on this interim conclusion and
the further premise that general properties are highly projectible, might con-
clude that members of the category in the conclusion are also likely to possess
the property. Certainly the relevance account of induction is not committed
to such a thorough application of relevance theory. Nonetheless, under such
an application category-based induction might come to have a rather explicit
or Type-2 character, and one might predict that sensitivity to an inductive
principle such as diversity would be associated with cognitive ability.
Finally, Heit (1998) has offered a Bayesian account of induction pitched at
a computational rather than an algorithmic level. As a computational-level
account specifies what the goals of people’s cognition should be (see Marr,
1982), the Bayesian account might make the simple prediction that if there are
differences in people’s sensitivity to diversity, then they should be predicted by
ability. If there is variability in the extent to which people attain their cognitive
goals, then people higher in cognitive capacity should be more likely to attain
them.

Some Recent Evidence

In both of the studies to be described here (for a more complete description
see Feeney, in press), I presented participants with a series of category-based
inductive arguments and the AH4 (Heims, 1968), a test of cognitive ability
previously used to study individual differences in induction (see Newstead
et al., 2004). The test has verbal/numerical and spatial components, each
of which has sixty-five items and is administered separately under timed
conditions.
In the first study 100 students at the University of Durham were presented
with 48 inductive arguments. Half of these had general conclusions and half
had specific conclusions. Half of the items were designed to measure sensi-
tivity to diversity. The other items were control items designed to examine
monotonicity effects and will not be described here. There were twelve pairs
of diversity items. One item in each pair concerned similar premise categories,
while the other concerned dissimilar premise categories. One of the categories
in the pair always remained the same. For example, in one pair the diverse
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

314 A. Feeney

argument concerned hippos and hamsters, while the non-diverse argument

concerned hippos and rhinos. All of the arguments concerned blank premises
such as Property X or Property B. Participants’ task was to indicate on a per-
centage scale the extent to which they thought the conclusion was supported
by the premises. For the purposes of statistical analyses I counted the number
of pairs where the strength rating for the diverse argument was stronger than
for the non-diverse argument.
There are two analyses of interest here. In the first I examined whether
overall, participants were sensitive to diversity at a rate greater than chance.
I found significant diversity effects for arguments with general conclusions
(sensitivity in 58% of trials) but not for arguments with specific conclu-
sions (sensitivity in 46% of trials). In general, people do not appear to be
highly sensitive to premise diversity when evaluating arguments. The fact that
they were significantly less sensitive for arguments with specific conclusions,
t(99) = 4.05, p < .001, lends some support to the claim made by Lopez et al.
that the generation of a superordinate category for arguments with a specific
conclusion is an extra source of difficulty.
The mean AH4 score for this study was 95.6 (S.D. = 14.4). This is close
to the norm for this kind of sample (mean 96.4, S.D. = 15.01). When the
reasoning data were collapsed across conclusion type, there was a significant
association between AH4 score and sensitivity to diversity (r = .27, p < .01).
This association was significant when items with general conclusions were
considered on their own (r = .27, p < .01). However, when items with spe-
cific conclusions only were analysed, the association between AH4 scores and
sensitivity to diversity was no longer significant (r = .16, p > .1). Closer
analysis of response patterns revealed that not one participant displayed di-
versity for all six pairs of arguments with specific conclusions, whereas ten
participants displayed maximum sensitivity in the general condition. The
requirement to generate a superordinate covering category seems to have ren-
dered sensitivity to diversity very difficult for the vast majority of participants
in this sample
In a second study I achieved greater control over our experimental materi-
als than I did in the first. In the first study I used different materials for the
diversity-specific and diversity-general conditions. In the second study I used
the same materials for each condition. One-hundred-and-fifteen students
at the University of Durham completed the AH4 and thirty-six reasoning
problems. Eighteen of these problems concerned monotonicity and will not
be discussed here. The set of thirty-six problems was composed of six sets
of six. Each set of six employed the same five mammal categories, three of
which were premise categories (mice, dolphins, squirrels), one of which was
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 315

a specific conclusion category (dogs), and one a general conclusion category

(I used mammal for all six problem sets). For the specific and general diverse
items, two taxonomically dissimilar mammal categories were presented in the
premises (mice, dolphins), whereas for the non-diverse items two taxonomi-
cally similar items were presented (mice, squirrels). Problems were presented
in one of four different random orders.
Once again I found that sensitivity to diversity was significantly greater than
chance for arguments with general conclusions (59%) but not for arguments
with specific conclusions (52%). The significant difference in sensitivity to
diversity due to conclusion type, t(114) = 2.96, p < .005, is further support
for the claim that the requirement to generate a covering category interferes
with diversity-based reasoning.
Although the raw diversity findings from this study are similar to those
obtained in Study 1, there are differences in the individual-differences data.
Mean performance on the AH4 was 106.3 (S.D. = 10.65). There are two things
to note about this result. First, it is significantly greater than the norm for a
sample of university students (mean = 96.4; S.D. = 15.01). As different degree
courses have different academic entry requirements, the difference between
the mean AH4 scores observed in these studies is most likely due to the degree
course from which the sample for each study was drawn. Second, the standard
deviation of the mean for this sample is considerably smaller than the normal
standard deviation. Accordingly, all of the correlation coefficients reported
below will be adjusted for truncated variance.
When the reasoning data were collapsed across conclusion type, I ob-
served a significant association between sensitivity to diversity and AH4 scores
(r(adj) = .28, t < .005). The association between reasoning performance and
ability is significant when arguments with general conclusions are consid-
ered on their own (r(adj) = .23, p < .02) and when arguments with specific
conclusions are considered on their own (r(adj) = .25, p < .01).

Interpreting the Data

These two studies have produced evidence that bears on theories of induc-
tive reasoning, specific claims about the normative basis for the diversity
effect, and on the general claim that there are two types of thinking. First,
the finding that people tend not to be sensitive to premise diversity when the
conclusion is specific suggests that Lopez et al. may be correct in their pre-
diction that having to generate a covering category is a source of difficulty in
induction. However, whereas Lopez et al. made their prediction for children’s
reasoning, our finding concerns the reasoning of university undergraduates.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

316 A. Feeney

An examination of overall sensitivity to diversity across both studies might

tempt one to argue that people never reliably generate the covering category
and, therefore, to conclude that evidence for the diversity effect with specific
conclusions has been overstated. However, the individual-differences analysis
of Study 2 allowed us to observe a significant association between cognitive
ability and sensitivity to diversity in specific arguments. Although we did not
find evidence for such an association in Study 1, we appear to have sampled
a higher part of the distribution of ability scores in Study 2. As Newstead
et al. (2004) have demonstrated in the case of the selection task, whether an
association is observed between ability and performance on difficult thinking
tasks often depends on the nature of the sample. In a high ability sample, those
who are highest in ability tend to be sensitive to premise diversity with specific
conclusions. This appears to support the prediction that the requirement to
generate a covering category is an additional source of difficulty in induction.
Sensitivity to diversity under these circumstances appears to be relatively rare.
The association that we observed between sensitivity to diversity and in-
telligence, by Stanovich’s normative argument, provides some evidence that,
all things being equal, sensitivity to diversity is normatively sound. One con-
cern about this argument relates to the point in the previous paragraph: in
any particular study whether certain effects are found to be associated with
cognitive ability will depend on the make-up of the sample. In Study 1 we
found an association between sensitivity to diversity and ability. However, the
sample in Study 1 was of average ability. It was possible that, for the higher
ability sample in Study 2, no association would be found. Although we did
find a relationship in Study 2, a failure to do so would have been very prob-
lematic for the normative argument. In principle, it is always possible that an
association will not be found with a sample that is of higher average ability
than the current sample. For this reason, the normative argument is strongest
for those phenomena that have been tested by a number of studies across a
range of abilities.
Our data also have implications for dual-process theories. In particular, they
have implications for the prediction made by the FBM that category-based
induction is achieved as a result of Type-1 processes and may be modelled
wholly in associationist terms (Sloman, 1993). We have seen here that individ-
ual differences in cognitive ability predict sensitivity to diversity. Associations
such as this one are taken by dual-process theorists to indicate that Type-2
processes are at play. In this case, such a conclusion is problematic for the
FBM. Type-2 processes might be involved in the diversity effect because peo-
ple may explicitly represent principles of inductive reasoning. Upon noticing
that the categories in the premises are dissimilar, people may bring to mind
the principle that diverse evidence leads to strong inductive inferences. This
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 317

possibility has also been suggested by McDonald, Samuels, and Rispoli (1996)
and, as we noted earlier, is consistent with certain applications of relevance
theory to category-based induction.
Finally, there are two ways in which category-based inductive phenom-
ena generally, and the diversity effect specifically, differ from other reasoning
and decision-making phenomena studied by dual-process theorists. First,
category-based phenomena require people to draw on their extensive knowl-
edge of categories in the world. It is possible, therefore, that our individual-
differences findings may be due to people high in cognitive ability having
richer representations of categories than people lower in ability. This would
mean that their associative systems had more effective knowledge bases to
make inferences over, leading to greater sensitivity to diversity. Whilst this
counter-explanation does not apply to other dissociations between Type-1
and Type-2 processes, it is an important one in this context. However, in
the case of diversity, this counter-explanation seems to boil down to the un-
likely possibility that some undergraduate students have insufficient featural
knowledge to distinguish similar from dissimilar mammals.
Second, Stanovich (1999) argues that associations between cognitive abil-
ity and performance on particular tasks will be strongest when different
responses are cued by Type-1 and Type-2 processes. As sensitivity to diversity
is predicted by the associative account (Sloman, 1993) and a range of other
more rule-based accounts (e.g., Osherson et al., 1990; Heit, 1998), the task
we have employed here is, in this respect, unlike other tasks previously used
to dissociate types of thinking. Nonetheless, we have observed significant
associations between sensitivity to diversity and cognitive ability. This obser-
vation suggests that sensitivity to diversity is more likely when there is greater
involvement of Type-2 processes.

    

As we have seen, the individual-differences paradigm may be very useful in
helping us to answer questions about diversity. There are a range of other
questions raised in other chapters of this book to which the approach might
usefully be applied. Some of these questions concern normative issues, whilst
others are more psychological.

Normative Issues
There exist several cases in the literature on induction where the findings
predicted by a particular normative account are not observed, or, alternatively,
phenomena are observed that are not predicted by the normative account.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

318 A. Feeney

Consider Murphy and Ross’s (Chapter 8, this volume) description of data

from inductive tasks involving uncertain categorisation that appear to be
inconsistent with a rational analysis. For example, Murphy, Malt, and Ross
(1995) asked participants to say how likely an individual described as walking
towards a house would be to check the sturdiness of the doors of the house.
In one condition the individual was either an estate agent or a cable TV
repairman. In another he was either an estate agent or a burglar. In both
conditions the individual was most likely to be an estate agent. Because
pretests had shown that people believe that burglars would be more likely to
check the doors than a TV repairman, participants should have rated more
highly the likelihood that the uncategorized individual will check the windows
when there is a possibility that he is a burglar. However, in this experiment and
many others involving abstract as well as concrete materials, the alternative
category made no difference to people’s likelihood judgements.
A related demonstration concerns people’s inductive inferences based on
cross-classifiable categories such as “bagel” which may be classified as a bread
or a breakfast food. Here, Ross and Murphy (1999) find that although a
majority of participants rely on just one of the categories when evaluating
the likelihood that a target will possess some property, a minority take both
possible classifications into account.
These findings pose a problem for Bayesian accounts of induction because
they suggest that people are, in certain respects, poor Bayesians. However, as
Murphy and Ross remark in their chapter, their problems resemble some of
the problems, such as the taxi-cab problem, used by Kahneman and Tversky
to study decision making. Just as Ross and Murphy find that increasing the
relevance of the alternative category increases people’s tendency to take it
into consideration, so Kahneman and Tversky observed that making the base
rate causally relevant increased the rate at which it played a role in people’s
judgements. Stanovich and West (1998a) have demonstrated an association
between cognitive ability and base rate neglect so that the most able partic-
ipants in a sample are less likely to ignore the base rate. A similar question
may be asked about induction under uncertainty. Will the tendency to ignore
the alternative category be predicted by ability? Perhaps Murphy and Ross’s
means hide predominantly high ability participants who take the alternative
into account. This may be especially true of the cross-classification studies
where a minority take both classifications into account. Perhaps this minority
tend to be highest in ability. If so, then we would have a situation, similar
to that which pertains in several other areas of high-level cognition (for a
review see Stanovich, 1999), where only a minority of participants behave in
accordance with the norm for the task. Because those participants tend to
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 319

be the most able, the observation of non-normative behaviour overall is less

problematic.
Another example of an experimental effect that is not predicted by Bayesian
accounts of induction is the conjunction fallacy (Medin et al., 2003), where
adding categories to the conclusion causes the argument to become stronger.
Medin et al. observed two types of conjunction fallacy. The conjunction fallacy
by property reinforcement is observed when the addition of a second category
to the conclusion of an argument causes the feature shared by the categories
in the argument to become more relevant. For example, Medin et al. observed
that the argument from chickens to cows and pigs was stronger than that
from chickens to cows or from chickens to pigs. Medin et al. account for this
finding by suggesting that the environment shared by members of all three
categories becomes more available when all three are present in the argument.
The causal conjunction fallacy occurs when the addition of an extra category
to the conclusion causes the causal relationship between the categories in
the argument to become relevant. Thus, the argument from grain to mice
and owls is stronger than that from grain to mice or from grain to owls. Of
course, in probabilistic terms this is non-normative because the probability
of a conjunction of events can never be higher than the probability of either
one of the conjuncts.
Again, it may be the case that these fallacies are resisted by participants
who are highest in ability. Indeed, Stanovich and West found an association
between ability and resistance to the conjunction fallacy in Kahneman and
Tversky’s famous Linda problem, where many participants estimate it to be
more likely that Linda is a feminist bankteller than that she is a bankteller.
As has been the case with other types of thinking, an individual-differences
approach may very well resolve apparent discrepancies between average be-
haviour and the predictions of normative or computational-level analyses.
A recent study of category-based induction by Feeney, Shafto, and Dun-
ning (in press) applied the individual-differences approach to both types
of conjunction fallacy. One-hundred-and-thirty participants completed the
AH4 and thirty-six category-based arguments. Half of these arguments were
designed to test susceptibility to the conjunction fallacy by property reinforce-
ment and the other half to test susceptibility to the causal conjunction fallacy.
The eighteen property reinforcement arguments consisted of six items, each
comprised of three arguments. One of the arguments in each item concerned
all three categories, with one category in the premise and the other two in
the conclusion. The conclusion of the other two arguments concerned one or
the other of the conclusion categories from the first argument. These single-
conclusion category arguments were randomly labelled A or B. Each of the
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

320 A. Feeney

corresponding six causal items also contained a three-category argument and

a pair of two-category arguments. Following Medin et al. the categories in
the three-category arguments always made up a causal chain. For example,
grain, mice, and owls may share a property because mice eat grain and owls
eat mice. Accordingly, because grain is closer to mice in this chain than it is
to owls, we labelled one type of single-conclusion argument Close (in this
example, grain to mice) and the other (grain to owls) as Distant.
Our data contained clear evidence for the existence of a conjunction fallacy
in category-based induction. Preliminary analysis revealed that participants
were susceptible to the fallacy by property reinforcement on 47% of trials and
to the causal version of the fallacy on 53% of trials. We split our participants
into low and high ability groups. The mean overall score on the AH4 was
94.57 (S.D. = 14.32). The mean score for the low ability group was 83.09
(S.D. = 11.00), whilst the mean score for the high ability group was 105.70
(S.D. = 5.88). For the property reinforcement materials, an ANOVA on rat-
ings of inductive strength revealed a significant interaction between ability
and number of conclusions in the argument. However, the equivalent interac-
tion from the analysis of strength ratings for the causal materials did not ap-
proach significance. The means in these interactions are shown in Figure 12.1,
where it may be seen that the high ability participants did not tend to
distinguish between two-conclusion category and one-conclusion category
property reinforcement arguments, whereas participants lower in ability did.
Figure 12.1 also shows that all participants, regardless of ability, tended to
respect causal distance in their ratings of argument strength. So, for exam-
ple, the argument from grain to mice was rated stronger than that from
grain to owls. In fact, participants were more likely with these materials to
respect causal distance than they were to commit the causal conjunction
fallacy.
The results are intriguing. First of all, because the conjunction fallacy by
property reinforcement is not observed in more able participants, our results
suggest that it may be relatively unproblematic for Bayesian accounts (see
Heit, 1998) of category-based induction. However, both high and low ability
participants were equally likely to commit the causal conjunction fallacy and
to respect causal distance in their ratings of argument strength. The causal
conjunction fallacy remains, therefore, problematic for Bayesian accounts of
induction. Causal knowledge appears to be very compelling when people
are evaluating inductive arguments. In this respect our results are similar
to those reported by other researchers who have shown the pre-eminence
of causal information in induction (see Rehder & Hastie, 2001; Rehder, this
volume).
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 321

(a)

5.5

5
Mean Rating

Two conclusion
4.5 One conclusion A
One conclusion B
4

3.5
LA HA
Ability
(b)

5.5

5 Two conclusion
Mean Rating

One conclusion
4.5
Close
One conclusion
4
Distant

3.5
LA HA
Ability
figure 12.1. Interaction between Ability and Conclusion Type from Feeney, Shafto, and
Dunning (in press).

The Relationship between Induction and Deduction

The final issue to which we will briefly apply the individual-differences method
is the distinction made by philosophers between deductive or necessary in-
ferences and inductive or plausible inferences. One important question about
this distinction is whether it corresponds to a psychological dissociation be-
tween the two types of reasoning. There are a variety of views on this issue.
One view is that induction and deduction are achieved by the same processes
(see Johnson-Laird, Legrenzi, Girotto, Legrenzi, and Caverni, 1999). Another
view is that there is a clear dissociation between induction and deduction (see
Rips, 2001). Support for this view comes from an experiment by Rips (2001;
see also Heit & Rotello, 2005) where participants were asked to evaluate the
same set of inferences under inductive or deductive instructions. There were
four types of inference: valid and causally strong; valid but causally weak;
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

322 A. Feeney

table 12.1. Examples of the materials used by Rips (2001) to demonstrate a

dissociation between induction and deduction

Valid Strong Valid Weak Invalid Strong Invalid Weak

If Car #10 runs into If car #10 runs into
a brick wall it will a brick wall it will
stop. speed up.
Car #10 runs into a Car #10 runs into a Car #10 runs into a Car #10 runs into a
brick wall. brick wall. brick wall. brick wall.
------------------- ------------------- ------------------- -------------------
Car #10 will stop. Car #10 will speed Car #10 will stop. Car #10 will speed
up. up.

invalid but causally strong; invalid and causally weak. An example of each of
these inferences is to be seen in Table 12.1. For valid but causally weak and
invalid but causally strong problems, Rips observed an interaction such that
the former were rated stronger than the latter under deductive instructions,
whereas under inductive instructions this pattern was reversed. Rips argued
that this finding constitutes a psychological dissociation between induction
and deduction.
A dual-process view, on the other hand, is that although there are distinct
psychological processes for thinking, these processes do not map onto in-
duction and deduction (see Sloman, 1996). According to a dual-process view,
both Type-1 and Type-2 processes may be involved to a greater or lesser extent
in any thinking task, inductive or deductive. When participants distinguish
between deductive and inductive reasoning instructions, they do so because
they have explicit knowledge about the difference between concepts such as
necessity and plausibility, and, given different instructions, Type-2 processes
allow them to apply different standards for evaluating arguments.
Support for the dual-process view comes from an experiment by Feeney,
Dunning, and Over (in preparation) where eighty-one students from the Uni-
versity of Durham completed a subset of Rips’s problems under induction
conditions. Their task was to decide whether the conclusion of each argu-
ment was plausible given the information in the premises. Causal strength
and validity were independently manipulated, and participants saw four argu-
ments of each type. They circled “strong” if they thought that the argument
was plausible and “not strong” if they thought that it was implausible. As
with other studies described in this chapter, participants in this study com-
pleted the AH4. Their mean score was 94.40 (S.D. = 12.94). We carried out
a median split on our data by cognitive ability. The mean score of the forty-
one participants in the low ability group was 84.24 (S.D. = 8.20), whilst
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 323

0.9
Proportion 'Strong' responses

0.8

0.7
High Ability
0.6
Low Ability
0.5

0.4

0.3
Valid Invalid
figure 12.2. Means involved in the interaction between Ability and Validity from Feeney,
Dunning, and Over (in submission).

the mean score of the forty participants in the high ability group was 104.8
(S.D. = 7.42).
A 2(Ability) × 2 (Validity) × 2(Strength) mixed design ANOVA on the
proportion of plausible judgements in each condition revealed significant
main effects of Validity, F(1, 79) = 114.59, MSe = .07, p < .001, and of
Strength, F(1, 79) = 172.45, MSE = .06, p < .001. Whilst the interaction
between Ability and Strength was non-significant, F(1, 79) = 1.05, MSE =
.06, p > .3, Ability did interact significantly with Validity, F(1, 79) = 5.14,
MSE = .07, p < .03. The means involved in this interaction are presented in
Figure 12.2, where it may be seen that high ability participants judge a greater
proportion of valid arguments to be strong than do low ability participants,
and they judge a greater proportion of invalid arguments to be “not strong”
than do low ability participants.
It is clear from Figure 12.2 that overall, logical validity plays an important
role in people’s judgements of inductive strength. However, what is most
striking is that validity plays a greater role in the judgements of participants
high in cognitive ability than in the decisions made by participants lower in
ability. These results are consistent with the claim that inductive judgements
are mediated by two types of process. One type of process is sensitive to
content and context, whereas the other is sensitive to structure. The first type
of process underlies effects of belief in reasoning, but because it does not
depend on cognitive capacity, effects of belief do not interact with cognitive
ability. The second type of process underlies our sensitivity to logical validity.
Because it is dependent on cognitive capacity, sensitivity to logical validity in
this study interacts with cognitive ability.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

324 A. Feeney


In this chapter I have described how results from studies using the individual-
difference method have informed recent dual-process theorising about think-
ing. I have shown how the method may provide valuable insights into the pro-
cesses involved in induction. For example, sensitivity to evidential diversity is
associated with cognitive ability and, in the presence of a specific conclusion,
is almost impossible for all but the most able participants. These results sug-
gest that the diversity effect is not explicable in wholly associationist terms
and that the generation of a covering category is, as appears to be predicted by
the SCM, an extra source of difficulty in induction. The method also appears
useful in testing Bayesian accounts of inductive reasoning and dual-process
accounts of the relationship between induction and deduction.
Individual differences are but one of several methods for testing models of
induction, and as a consequence, individual-differences studies are unlikely
to lead to the formulation of an entirely novel theory of induction. However,
as the range of approaches described in this volume attests, there is currently
a broad range of theories of induction formulated at a number of levels of
explanation. One might argue that currently the field does not need new
theories so much as it does evidence with which to discriminate between
existing theories and approaches. Individual-differences studies are likely to
be a rich source of discriminating evidence.

 
Please address correspondence to Aidan Feeney, Applied Psychology, Univer-
sity of Durham Queen’s Campus, Thornaby, Stockton-on-Tess TS17 6BH,
United Kingdom, Email: [email protected]

References
Adolphs, R., Tranel, D., Bechara, A., Damasio, H., & Damasio, A. R. (1996). Neuropsy-
chological approaches to reasoning and decision-making. In A. R. Damasio (Ed.),
Neurobiology of decision-making. Berlin: Springer-Verlag.
Carnap, R. (1950). Logical foundations of probability. Chicago: University of Chicago
Press.
Carpenter, P. A., Just, M. A., & Shell, P. (1990). What one intelligence test measures: A
theoretical account of processing in the Raven Progressive Matrices test. Psychological
Review, 97, 404–431.
Cheng, P. W., & Holyoak, K. J. (1985). Pragmatic reasoning schemas. Cognitive Psychology,
17, 391–416.
Cohen, L. J. (1981). Can human irrationality be experimentally demonstrated? Behav-
ioral & Brain Sciences, 4, 317–370.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 325

Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how
humans reason? Studies with the Wason selection task. Cognition, 31, 187–276.
Evans, J. St. B. T. (2003). In two minds: Dual process accounts of reasoning. Trends in
Cognitive Sciences, 7, 454–459.
Evans, J. St. B. T., Barston, J. L., & Pollard, P. (1983). On the conflict between logic and
belief in syllogistic reasoning. Memory & Cognition, 11, 295–306.
Evans, J. St. B. T., & Curtis-Holmes, J. (2005). Rapid responding increases belief bias:
Evidence for the dual process theory of reasoning. Thinking and Reasoning, 11, 382–
389.
Evans, J. St. B. T., Newstead, S. E., Allen, J. L., & Pollard, P. (1994). Debiasing by
instruction: The case of belief bias. European Journal of Cognitive Psychology, 6, 263–
285.
Evans, J. St. B. T., & Over, D. E. (1996). Rationality and reasoning. Hove, UK: Psychology
Press.
Feeney, A. (in press). How many processes underlie category-based induction? Effects
of conclusion specificity and cognitive ability. Memory & Cognition.
Feeney, A., Dunning, D. & Over, D. E. (in preparation). Reason distinctions: An indi-
vidual differences study.
Feeney, A., Shafto, P., & Dunning, D. (in press). Who is susceptible to conjunction
fallacies in category-based induction? Psychonomic Bulletin & Review.
Fong, G. T., Krantz, D. H., & Nisbett, R. E. (1986). The effects of statistical training on
thinking about everyday problems. Cognitive Psychology, 18, 253–292.
Frey, M. C., & Detterman, D. K. (2004). Scholastic assessment or g? The relationship
between the Scholastic Assessment Test and general cognitive ability. Psychological
Science, 15, 373–378.
Fry, A. F., & Hale, S. (1996). Processing speed, working memory, and fluid intelligence:
Evidence for a developmental cascade. Psychological Science, 7, 237–241.
Gilinsky, A. S., & Judd, B. B. (1994). Working memory and bias in reasoning across the
life-span. Psychology and Aging, 9, 356–371.
Goel, V., & Dolan, R. J. (2003). Explaining modulation of reasoning by belief. Cognition,
87, B11–B22.
Handley, S. J., Capon, A., Beveridge, M., Dennis, I., & Evans, J. St. B. T. (2004). Working
memory and inhibitory control in the development of children’s reasoning. Thinking
and Reasoning, 10, 175–195.
Hasher, L., & Zacks, R. T. (1988). Working memory, comprehension, and aging: A review
and a new view. The Psychology of Learning and Motivation, 22, 193–225.
Heim, A. W. (1967). AH4 group test of intelligence [Manual]. London: National Founda-
tion for Educational Research.
Heit, E. (1998). A Bayesian analysis of some forms of inductive reasoning. In M. Oaksford
& N. Chater (Eds.), Rational models of cognition (pp. 248–274). Oxford: Oxford
University Press.
Heit, E., & Feeney, A. (2005). Relations between premise similarity and inductive
strength. Psychonomic Bulletin and Review, 12, 340–344.
Heit, E., Hahn, U., & Feeney, A. (2005). Defending diversity. In W. Ahn, R. Goldstone,
B. Love, A. Markman, & P. Wolff (Eds.), Categorization inside and outside of the
laboratory: Essays in honor of Douglas L. Medin (pp. 87–99). Washington DC: American
Psychological Association.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

326 A. Feeney

Heit, E., & Rotello, C. (2005). Are there two kinds of reasoning? Proceedings of the Twenty-
Seventh Annual Conference of the Cognitive Science Society (pp. 923–928). Mahwah,
NJ: Erlbaum.
Hempel, C. G. (1966). Philosophy of natural science. Englewood Cliffs, NJ: Prentice Hall.
Horn, J. L., & Cattell, R. B. (1967). Age differences in fluid and crystallized intelligence.
Acta Psychologica, 26, 107–129.
Howson, C., & Urbach, P. (1993). Scientific reasoning: The Bayesian approach. Chicago:
Open Court.
Jepson, C., Krantz, D. H., & Nisbett, R. E. (1983). Inductive reasoning: Competence or
skill? Behavioral and Brain Sciences, 6, 494–501.
Johnson-Laird, P. N., Legrenzi, P., Girotto, V., Legrenzi, M., & Caverni, J.-P. (1999). Naive
probability: A mental model theory of extensional reasoning. Psychological Review,
106, 62–88.
Johnson-Laird, P., Legrenzi, P., & Legrenzi, S. (1972). Reasoning and a sense of reality.
British Journal of Psychology, 63, 395–400.
Kane, M. J., & Engle, R. W. (2003). Working memory capacity and the control of
attention: The contributions of goal neglect, response competition, and task set to
Stroop interference. Journal of Experimental Psychology: General, 132, 47–70.
Lo, Y., Sides, A., Rozelle, J., & Osherson, D. (2002). Evidential diversity and premise
probability in young children’s inductive judgment. Cognitive Science, 26, 181–
206.
López, A., Atran, S., Coley, J. D., Medin, D. L., & Smith, E. E. (1997). The tree of life:
Universal and cultural features of folkbiological taxonomies and inductions. Cognitive
Psychology, 32, 251–295.
López, A., Gelman, S. A., Gutheil, G., & Smith, E. E. (1992). The development of
category-based induction. Child Development, 63, 1070–1090.
McDonald, J., Samuels, M., & Rispoli, J. (1996). A hypothesis assessment model of
categorical argument strength. Cognition, 59, 199–217.
Malt, B. C., Ross, B. H., & Murphy, G. L. (1995). Predicting features for members of nat-
ural categories when categorization is uncertain. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 21, 646–661.
Manktelow, K. I., & Over, D. E. (1991). Social roles and utilities in reasoning with deontic
conditions. Cognition, 39, 85–105.
Marr, D. (1982). Vision. San Francisco: W. H. Freeman & Co.
Medin, D. L., Coley, J., Storms, G., & Hayes, B. K. (2003). A relevance theory of induction.
Psychonomic Bulletin & Review, 10, 517–532.
Murphy, G. L., & Ross, B. H. (1999). Induction with cross-classified categories. Memory
& Cognition, 27, 1024–1041.
Myrvold, W. C. (1995). Bayesianism and diverse evidence: A reply to Andrew Wayne.
Philosophy of Science, 63, 661–665.
Nagel, E. (1939). Principles of the theory of probability. Chicago: University of Chicago
Press.
Newstead, S. E., Handley, S. J., Harley, C., Wright, H., & Farrelly, D. (2004). Individual
differences in deductive reasoning. Quarterly Journal of Experimental Psychology, 57,
33–60.
Oaksford, M., & Chater, N. (1994). A rational analysis of the selection task as optimal
data selection. Psychological Review, 101, 608–631.
P1: JZP
0521672443c12 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:45

Individual Differences, Dual Processes, and Induction 327

Osherson, D. N., Smith, E. E., Wilkie, O., López, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Osman, M. (2004). An evaluation of dual-process theories of reasoning. Psychonomic
Bulletin & Review, 11, 988–1010.
Proffitt, J. B., Coley, J. D., & Medin, D. L. (2000). Expertise and category-based induction.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 811–828.
Rehder, B., & Hastie, R. (2001). Causal knowledge and categories: The effects of causal
beliefs on categorization, induction, and similarity. Journal of Experimental Psychology:
General, 130, 323–360.
Rips, L. J. (2001). Two kinds of reasoning. Psychological Science, 12, 129–134.
Ross, B. H., & Murphy, G. L. (1999). Food for thought: Cross-classification and category
organization on a complex real-world domain. Cognitive Psychology, 38, 495–553.
Sloman, S. A. (1993). Feature-based induction. Cognitive Psychology, 25, 231–280.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psychological
Bulletin, 119, 3–22.
Sperber, D., & Wilson, D. (1986/1995). Relevance: Communication and cognition. Oxford:
Basil Blackwell.
Stanovich, K. E. (1999). Who is rational? Studies of individual differences in reasoning.
Mahwah, NJ: Erlbaum.
Stanovich, K. E., & West, R. F. (1998a). Individual differences in rational thought. Journal
of Experimental Psychology: General, 127, 161–188.
Stanovich, K. E., & West, R. F. (1998b). Cognitive ability and variation in selection task
performance. Thinking and Reasoning, 4, 193–230.
Stanovich, K. E. & West, R. F. (1998c). Individual differences in framing and conjunction
effects. Thinking and Reasoning, 4, 289–317.
Tversky, A., & Kahneman, D. (1983) Extensional vs. intuitive reasoning: The conjunction
fallacy in probability judgment. Psychological Review, 90, 293–315.
Unsworth, N., & Engle, R. W. (2005). Working memory capacity and fluid abilities:
Examining the correlation between Operation Span and Raven. Intelligence, 33, 67–
81.
Wason, P. (1966) Reasoning. In B. Foss (Ed.), New horizons in Psychology (pp. 135–151).
Harmondsworth: Penguin.
Wayne, A. (1995). Bayesianism and diverse evidence. Philosophy of Science, 62, 111–121.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

13

Taxonomizing Induction
Steven A. Sloman

The chapters in this book offer a variety of perspectives on induction. This

is not shocking, as inductive problems are both fundamental and unsolv-
able. They are fundamental in that organisms must solve them in order to
know most of the things about their environment that they cannot observe
directly, like what’s around the next corner or what to expect from that woolly
mammoth over there. They are unsolvable in that uncertainty cannot be
eliminated. However we define induction, as discussed in Heit’s Chapter 1, it
involves conclusions that we cannot be sure about. There seems to be general
agreement with Hume’s (1748) conclusion that no logic of justification for
inductive inference is possible.
Despite the variety of perspectives, the chapters revolve around three central
themes concerning, respectively, the prevalence of induction, how to carve
up the space of inductive inferences, and the role of causality. I will briefly
discuss each in turn.

.    

The consensus in these essays is that most of the conclusions that people come
to have an inductive basis and that little human reasoning is well-described
as deductive. Indeed, the brunt of Oaksford and Hahn’s Chapter 11 is that
probabilistic inference is the dominant mode of human reasoning, a claim
echoed by Heit and by Thagard (Chapter 9).
Even the Rips and Asmuth Chapter 10 leaves the reader with the impres-
sion that deductive reasoning is exceptional. Rips and Asmuth make a strong
argument that inductive proofs in math have the strength of deductive infer-
ence, that they are deductive proofs in disguise. The existence of deductive
proofs entails the existence of deductive reasoning, so the existence of math-
ematicians and logicians proves that human beings are capable of deductive
328
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 329

inference. But in the course of their argument, Rips and Asmuth review a
lot of evidence that students find deductive proofs hard to generate and that
they confuse them with demonstrations via examples. Their data give the
impression that nonexperts rarely rely on deductive inference.
The authors seem to agree that reasoning is either inductive or deductive.
Heit, Oaksford and Hahn, and Rips and Asmuth all make this dichotomy
explicit. So does Thagard, who views abduction as a species of induction
because both involve uncertain conclusions. On this view, if little reasoning
is deductive, then almost all reasoning is inductive.
The field of psychology at large seems to agree with this conclusion. Less
and less work is being done on how people make deductive inferences. One
of the central players in that line of work has concluded that logic is rarely
a reasonable standard to evaluate human reasoning and that most reasoning
is pragmatic and probabilistic (Evans, 2002). With the exception of Johnson-
Laird (2005), this seems to be a widely held view.
Of course, the claim that most reasoning is inductive is not terribly strong.
If we follow Skyrms (2000) in holding that an inductive argument is merely
one in which the conclusion is not necessary given the premises (see Heit’s
chapter), we are left with a vast number of argument forms and types. The
set of inductive arguments includes those that seem almost deductively valid
(if Houdini could hold his breath for fifteen minutes, then he could hold his
breath for fifteen minutes and one second), arguments that are exceptionally
weak (lobster is expensive today, therefore lobster was expensive in the 17th
century), causal arguments (A causes B and A occurred, therefore B probably
occurred), abductions (see Thagard’s chapter), and many more. Is there one
big homogeneous space of inductive arguments, or can they be grouped in
meaningful ways? This is the second theme of the book.

.       

This question can refer to a normative issue, whether there is more than one
system for justifying inductive leaps, or a descriptive issue, whether there are
natural fault lines that distinguish different modes or systems that people use
to make inductive inferences. I address the normative issue very briefly and
then focus on the major topic of this book, the descriptive issue.

Normative Criteria
The implicit claim in Tenenbaum, Kemp, and Shafto (Chapter 7), made ex-
plicitly by Oaksford and Hahn, is that there is a single normative model of
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

330 S. A. Sloman

induction, a criterion for justification defined in Bayesian terms as proba-

bilistic support for a conclusion. Bayesian theory provides a way to maintain
a coherent set of beliefs by providing a rule to update the probability that is
assigned to beliefs in the face of new data. This rule has the virtue that, in the
case of repeated events, the updated probabilities converge to the long-run
relative frequency of the event.
I will not take issue with Bayesianism, though others have (see Howson &
Urbach, 1993, for a discussion). But I will point out that even if this view is
right, the conclusion that induction has a single set of normative criteria
does not immediately follow, because probabilistic support comes in varieties.
Different senses of probability can be distinguished. One common distinction
is between an extensional type that measures the inherent unpredictability
of events such as the randomness of a coin toss, referred to as “aleatory
uncertainty,” and the type of probability that refers to degree of ignorance or
“epistemic uncertainty,” like the degree of confidence that you assign to your
answer to a trivial pursuit question (before looking it up). Randomness in the
world and lack of certainty in a judgment have different sources. Randomness
is a property of events in the world independent of the judge and uncertainty
a property of a mental state of the judge. They also present different obstacles
to prediction. Randomness generally refers to a principled and fixed limit on
the accuracy of prediction, whereas uncertainty can be reduced with more
knowledge or expertise.
Having such different sources and prospects, these two forms of probability
appear to provide quite different forms of justification. The fact that they both
can be updated via a common equational form (Bayes’s rule) is not a very deep
commonality. Bayes’s rule requires models of its terms, both its prior belief
and likelihood terms (see Tenenbaum et al., Chapter 7 for an explanation of
what these terms refer to). Neither of these is trivial. Tenenbaum et al. offer a
sophisticated theory of the source of prior beliefs, that is, how people generate
a space of hypotheses and assign probability to them. Likelihoods require a
model that spells out how each hypothesis expects data to be generated in order
to provide, for every hypothesis, a probability assignment for every possible
pattern of data. These models can have very different forms and make different
assumptions depending on the kind of probability at issue. Thus the criteria
for inductive justification that they provide are fundamentally different for
sufficiently distinct problems.
Moreover, the aleatory/epistemic distinction is not the only important one.
Tenenbaum et al. discuss the variety of different relational structures that
are necessary for the range of problems that people encounter. To illustrate,
they propose one type of relational structure for the biological domain (a
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 331

hierarchical one). A different type would be required to count days of the

week (a periodic one). That choice and all others influence the normative
model used in the final analysis. So, again, the argument for a single normative
criterion for inductive inference still needs to be made.
That question is not at the heart of this book; this book is really concerned
with the psychological/cognitive scientific question of how people actually
reason inductively, not how they should or whether they are justified. Al-
though descriptive and normative models are related conceptually in many
ways, they must be evaluated in entirely different ways. Descriptive models
must be evaluated according to their ability to describe behavior; normative
models must be evaluated according to their ability to accurately describe
what is right and what is wrong. The distinction has been muddied how-
ever by the development of the “rational analysis” school of cognitive science
(Oaksford & Chater, 1998), as exemplified by the Tenenbaum et al. chapter.

On Computational Models
Tenenbaum, Kemp, and Shafto offer a valuable perspective on induction,
one that they claim is a computational-level model of human induction. Is
their model normative, identifying whether behaviors are well designed to
achieve their users’ goals, or descriptive, identifying which behaviors will be
produced? When Marr (1982) introduced the concept of a computational
model, he intended it as a functional description of what an organism is
trying to accomplish, the problems it faces, as opposed to the strategies it
uses to accomplish them (an algorithmic model) or how those strategies are
realized in a physical system (an implementational model).
A computational model in Marr’s sense is neither normative nor descriptive.
It is not normative in that it merely characterizes a problem without specifying
whether or not that is the right problem to solve. The characterization might
be correct or incorrect, but either way nothing is implied about whether the
organism’s thought or behavior is effective or not. A computational theory is
not descriptive because it is a representation of a problem in the world, not
necessarily anything that an organism does or even knows about at any level.
A computational theory of nest building by an ant colony describes a problem
that no single ant is trying to solve. The solution is an emergent property of
the behavior of a large number of ants blind to any larger purpose. Similarly, a
computational model of human induction, like the Bayesian model suggested
by Tenenbaum et al., is an abstract representation of a problem. Solving
that problem, or just trying to, does not require that people have access to
that computational representation. People do not need to know Bayes’s rule
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

332 S. A. Sloman

in order to make inductions in a way that approximates or is consistent

with it.
The model is not descriptive unless it takes a stand on how the computation
is performed or how well people conform to the constraints it imposes. In
other words, the normative/descriptive distinction becomes relevant when we
assess how well the organism is doing at achieving its goals as represented by
the computational model. If it wants gummy bears and it gets gummy bears,
then its behavior is normatively appropriate. How it goes about getting gummy
bears is a descriptive question. In the domain of induction, a normative theory
would prescribe how to obtain an optimal solution to the problem. The only
reason to compare that optimal solution to what people do, a comparison
important to Tenenbaum et al., is to see if the normative theory is also
descriptive. If it is, then people are rational and we have a rational model of
behavior. If it is not, then we remain without a descriptive theory.
Positing that a normative theory is also descriptively adequate is certainly
reasonable, but it is known to be false in general. Dozens of demonstrations of
fallacious reasoning, including the conjunction fallacy (Tversky & Kahneman,
1983), some cases of base-rate neglect (Barbey & Sloman, in Press), and many
others (Gilovich, Griffin, & Kahneman, 2002), all show that people, for good
or for bad, do not always behave according to normative principles when
reasoning inductively.
A lot can be and has been learned about how the mind works by examining
the hypothesis that people are normative. However, this research strategy can
also lead to confusion on both sides of the normative/descriptive divide. One
kind of error occurs when a normative theorist thinks his or her theory is
descriptively adequate when it is not. A complementary error occurs when
a descriptive theorist fails to consider ways in which people make mistakes.
Worse, empirical methodologies are sometimes used that are designed to hide
human error. This can only interfere with the development of both descriptive
and prescriptive theory.

Descriptive Models
The most pertinent question for this book is whether human inductive rea-
soning occurs in one or more fundamentally different ways. The chapters
offer several ways of cutting the pie:

no cut. Oaksford and Hahn argue for a single method of making inductive
inferences, one that conforms to Bayesian inference. They show that a range
of judgments about the strength of an inductive inference is consistent with
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 333

Bayesian tenets. This broadly supports the stance of Tenenbaum, Kemp, and
Shafto.
The strength of this proposal is its breadth. It offers a conception that covers
induction in all its forms and, unlike any other model, proposes a method that
is both mathematically well-defined and highly flexible in its ability to adjust
to all forms of relational knowledge. A danger of this approach is that with
great flexibility comes great freedom to fit all patterns of data, both real and
imagined. Theoretical power is a curse when it comes to demonstrating the
empirical value of a model because it allows the model to wrap itself around
any potential result, even those not found.

by domain. Tenenbaum et al.’s suggestion that different relational structures

are used with different kinds of arguments motivates their effort to spell
out how different relational structures are chosen and applied to particular
arguments based both on prior domain knowledge and specific facts presented
within an argument. So, although they propose an overarching formal relation
that applies to all inductive inference, they believe that the way the relation
is applied depends on domain-specific relational structure. Inductions about
biological relations are supported by fuzzy hierarchical knowledge about
biological taxonomy. Inductions about the spread of diseases are supported by
knowledge of the causal relations that govern interactions between predators
and their prey. Tenenbaum et al. propose that there are also other relational
structures for induction.
The proposal is impressive in its coverage and potential for generating
domain-specific theories. The danger of this approach is a particular case of
the concern about Bayesianism stated earlier: It opens the possibility of mul-
tiplying relational structures, each shaped to fit a specific study. Presumably,
relational structures are not independent entities but are generated on the
basis of basic principles. For instance, biological hierarchies can be grounded
in one or another causal model: The hierarchy might emerge from a causal
analysis of how biological entities function, top-down from causal beliefs
about evolutionary processes, or in some other way. In fact, it might be that
once the causal model that grounds the hierarchy is spelled out, the hierarchy
itself will turn out to be superfluous. Inductions might be mediated directly
by causal knowledge (see Hayes, Chapter 2, for a related point).
The fact that some relational structure can be used along with Bayesian
principles to successfully fit data on induction does not provide indisputable
evidence that people use that relational structure to reason. The relational
structure may serve as a proxy for some other kind of knowledge that might
not be domain specific, like causal knowledge.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

334 S. A. Sloman

similarity versus causal knowledge. Several chapters argue that people

use similarity relations to make inductions when no other useful informa-
tion is available but use causal knowledge when they can (Shafto, Coley, &
Vitkin; Rehder; Blok, Osherson, & Medin). This idea can be traced to Quine
(1970), who argued that natural kinds are classified using an innate sense of
similarity of objects to one another until a conceptual theory becomes avail-
able to replace the unanalyzed similarity construct. Theoretical knowledge
often becomes available to the child through learning; it becomes available to
scientists as their field of study matures.
Hayes points out that distinctions among similarity, causal, and categorical
reasoning tend to be ecologically invalid. Similarity is generally an indicator of
categorical relations, and categories cohere only by virtue of causal structure.
So the norm in the real world is that cues tied to the various strategies are
highly intercorrelated and mutually supportive. Experimenters have managed
to construct situations that disentangle the cues, but such situations might
eliminate people’s ability to do what they are especially good at, namely, to
combine different kinds of convergent information. Nevertheless, people are
at times faced with objects and events that they cannot analyze in any depth.
Sometimes, an object or event is so new that one is limited to a superficial
analysis that affords comparison only by similarity and not by unobservable
attributes like causal structure. I had this experience recently in Portugal when
a barnacle was put on my lunch plate, apparently a delicacy. All I could do
was remark on its striking superficial properties (properties quite unrelated
to eating lunch). Admittedly, even in this case of ignorance and surprise I was
able to infer that what I was looking at was in some way related to organic
matter. Hayes is fundamentally correct that induction is guided not by one
sort of cue or another, but by a desire to explain why categories have properties
(Sloman, 1994), and this gives the various reasoning strategies the common
goal of making inferences in a way that respects the latent structure of an
object.

associative versus deliberative reasoning. Feeney argues in his Chap-

ter 12 that induction is mediated by two different systems of reasoning (Evans
& Over, 1996; Stanovich, 1999; Sloman, 1996), one governed by associative
principles and the other more deliberative, rule-based, and dependent on
working memory capacity. His evidence comes from a couple of studies of
individual differences showing that participants who score higher on a test
of intelligence demonstrate more sensitivity to the diversity principle, the
principle that multiple pieces of evidence for a conclusion are strong to the
degree that the pieces are diverse (dissimilar). On the assumption that more
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 335

intelligent people can allocate deliberative reasoning resources more effec-

tively than less intelligent people but that everyone is equally capable of
associative reasoning, this suggests that sensitivity to diversity emerges from
deliberative effort, not associative reasoning, an idea previously championed
by McDonald, Samuels, and Rispoli (1996).
Sloman (1998) also argued for a dual-system model of induction but on an
entirely different basis. In that paper, I report several experiments showing an
inclusion similarity phenomenon, namely, that people will make inductions
proportional to similarity even when doing so neglects transparent categorical
relations between premise and conclusion. For example, people did not always
judge
Every individual body of water has a high number of seiches.
Therefore, every individual lake has a high number of seiches.
to be perfectly strong even when they agreed that a lake is a body of water.
Moreover, they judged
Every individual body of water has a high number of seiches.
Therefore, every individual reservoir has a high number of seiches.
to be even weaker (again even when they agreed that the premise category in-
cludes the conclusion category), presumably because reservoirs are less typical
bodies of water than lakes. These kinds of judgments fall out of an associative
model of reasoning: People made their judgments based on similarity or some
kind of feature-based analysis, not on the basis of categorical knowledge.
But if the latent categorical premise was made explicit,
All reservoirs are bodies of water.
All bodies of water have a high number of seiches.
All reservoirs have a high number of seiches.
then everyone judged the argument perfectly strong. By turning the argument
into an explicit syllogism, people were induced to use a strategy that is highly
sensitive to categorical relations. My argument in 1998 was that this strategy is
deliberative and not merely associative because, unlike associative knowledge,
it depends on an explicit categorical relation and leads to certainty in the
conclusion.

intended meaning versus knowledge-based inference. Most of the

discussion in this book assumes that inductions take place through a cogni-
tive process mediated by knowledge. But several authors note that the com-
municative context of induction might also influence people’s inductions.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

336 S. A. Sloman

Communicative context affects induction in a couple of ways. For one, in-

duction can be influenced by the semantics of a language. Medin and Waxman
point out that asymmetries in children’s inferences first pointed out by Carey
(1985) can be explained this way. Carey found that young children were more
willing to project an unfamiliar part from humans to dogs than from dogs
to humans. One interpretation of Medin and Waxman is that “human” is an
ambiguous term, implying on one hand “animal of human form” and, on the
other hand, “non-animal entity.” They suggest that arguments in the human
to dog direction could suggest the more inclusive animal sense of “human,”
whereas arguments in the dog to human direction could suggest the less in-
clusive non-animal sense. This would explain the observed asymmetry by
appealing to different intended meanings of a word.
Induction can also be influenced by pragmatic considerations. Feeney and
Shafto, Coley, and Vitkin mention that many of the conclusions we draw
are given to us or at least guided by people around us. We seek expertise
everywhere. Even experimenters might be deemed experts so that the format
of the arguments they present might be treated as suggesting appropriate
conclusions. For instance, Medin, Storms, Coley, and Hayes (2003) suggest
that some monotonicities and nonmonotonicities have a pragmatic basis. If
told that grizzly bears have a property, you might be willing to conclude that
it is more likely that squirrels have the property than you thought beforehand.
But if told that three different kinds of bear have the property (grizzlies, polar
bears, and black bears say), then you might conclude that the experimenter
is trying to tell you something about what kind of thing has the property
and therefore that it seems more likely that only bears do. This is a plausible
inference on the assumption that the experimenter is picking categories so as
to be informative, that is, that the choice of premises is pragmatically driven.
However, it is also a reasonable inference from the assumption that cate-
gories are being randomly chosen from the set of things that have the property.
If a wide range of mammals had the property, then it is very unlikely that
three bears would be randomly chosen; they would be much more likely to be
chosen if only bears had the property. In other words, some pragmatic effects
might reflect sampling assumptions that people make about how premises
and conclusions are chosen (Tenenbaum & Griffiths, 2001). However, there is
evidence that people find arguments with more premises stronger regardless
of what they are told about sampling procedures (Fernbach, 2006). This fact
contradicts both the sampling and the pragmatic inference theories.

categorical versus non-categorical inference. Murphy and Ross

(Chapter 8) demonstrate convincingly that in their experimental paradigm
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 337

people tend to focus on a single category, neglecting alternative categories. The

observation that people focus on only a single possibility and neglect alterna-
tives converges with other category research (Lagnado & Shanks, 2003), as well
as research showing that people tend to focus on only a single causal model
when troubleshooting (Fischhoff, Slovic, & Lichtenstein, 1978), when mak-
ing probability judgments (Dougherty, Gettys, & Thomas, 1997), and when
engaged in various forms of deductive inference (Johnson-Laird, 2005).
Tenenbaum (1999) found that a model that posits that people make in-
ductions by weighting each category-hypothesis in a Bayesian manner fits
human data fairly well. Such a model assumes that people consider multiple
categories in induction. To account for this inconsistency with their find-
ings, Murphy and Ross suggest that people are not even using categories in
Tenenbaum’s task but rather relying on associations between attributes and
the new property. This is precisely the theoretical issue that separates Osher-
son, Smith, Wilkie, & Shafir (1990) similarity-coverage model of induction
from Sloman’s (1993) feature-based model. The notion of coverage in the
similarity-coverage model is category coverage, the degree to which premise
categories cover the lowest-level category that includes all premise and con-
clusion categories. This notion requires that people have a category hierarchy
available and that the determination of category membership is an explicit
part of the inductive process. The feature-based model also has a notion of
coverage – feature coverage – the extent to which the features (properties
or attributes) of the premise categories cover the features of the conclusion
category. No hierarchical category knowledge is assumed.
Murphy and Ross are right that the question of whether people use cat-
egories is prior to the question of how many categories people consider. It
may be that relatively little induction is mediated by categorical knowledge
at all. Indeed, Sloman’s (1998) data suggest that hierarchical knowledge is
considered only when it is made explicit. After all, so many other kinds of
knowledge are available to mediate induction such as similarity comparisons
and causal knowledge. This observation supports Tenenbaum et al.’s insight
that categorical knowledge is only one kind of relational structure and that
one of the key problems of induction is how to choose from and deploy a
variety of different relational structures.

abduction versus other forms of induction. Thagard observes the

generality of the process of abduction and notes that, while it can be con-
strued as a form of induction in that it involves uncertain conclusions, it is
distinguished in the process of reasoning that it entails. The critical element
of abduction according to Thagard is the process of explanation that it entails.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

338 S. A. Sloman

Belief changes following an abductive process pertain to explanations of ob-

servations. Explanations earn belief according to the quality of their account
of observations.
Thagard is surely right that abductive processes have been neglected in
cognitive theory (though see Lombrozo, in press). But the importance of
explanation is not limited to abduction. It is known to be just as central to
other inductive forms. Explanations frequently mediate induction by selecting
relevant features (Sloman, 1994; Sloman & Wisniewski, 1992). To decide if
property A should be projected from category X to category Y, people explain
why category X has property A and evaluate whether that explanation would
pertain to category Y. Should the property of being hard to see at night be
projected from Black Labradors to Blond Labradors? Presumably not as Black
Labs are invisible because of their color and that color is not shared by Blond
Labs. Explaining why Black Labs have the property is a critical part of the
inductive inference.

. 
Although the chapters in this collection offer a variety of perspectives on
how to distinguish inductive forms, a theme common to many chapters is
the central role played by causal knowledge in inductive inference. Hayes
focuses on the role of causal relations in guiding people to make inferences
based on causally relevant properties rather than causally irrelevant ones.
Shafto et al. and Feeney provide strong evidence that people reason with
causal relations when they have sufficient knowledge. Rehder offers a causal
model of inductive inference. Tenenbaum et al. claim that causal structure is
an important type of relational structure for inference. Thagard argues that
causality is the core of explanation, and that explanation is the heart of the
abductive inferences that he discusses.
These authors are all in one way or another echoing Hume (1748), who
argued that inductive inferences are parasitic on causal beliefs (cf. Sloman &
Lagnado, 2005). As regards the psychology of induction, the jury is out about
whether all inductive inferences have a causal basis or whether only some do.
As many of the chapters attest, the evidence is overwhelming that other prin-
ciples also apply to particular inductive inferences. Some inferences clearly
appeal to similarity and others to categorical relations (Murphy & Ross). But it
might be that these other relations are just proxies for causal structure. Objects
that are more similar tend to share more causal roles, so similarity provides a
heuristic basis for making causally appropriate inferences. Hadjichristidis et
al. (2004) show that people’s inferences are proportional to a property’s causal
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 339

centrality but only to the extent that the objects are similar. This suggests that
people treat similarity as a cue for the extent to which objects share causal
structure. And when similarity conflicts with causal relations, causality wins
(Hayes).
Category knowledge serves as a proxy for causal structure in a couple of
senses. First, some categories are defined by their role in a causal relation. The
category of “brown things” is the set of things that cause light to reflect off
of them such that visual systems perceive brown. The category of “goals in a
hockey game” is the set of events that consist of an outcome that gives a hockey
team a point. Second, more complex categories are also associated with causal
roles, frequently multiple causal roles. For instance, the set of objects that we
call “automobile” tend to participate in causal roles involving transportation
of people, systems of steering and acceleration, expensive repairs, and so forth.
The most direct explanation for how an induction is made is not necessarily
causal. But there might be a causal explanation at some level operating in the
background to motivate the relational structure used to make the inductive
inference in every case.

.   
The range of topics covered by the chapters in this book exposes the point
that Shafto et al. make explicitly: People use a variety of strategies and frames
for assessing inductive strength in everyday life. Strategies seem to be arrayed
along a series of levels that require progressively more conceptual effort and
provide progressively more justifiable responses. The hierarchy conforms to
the trade-off between cognitive effort and cognitive effect that determines
relevance (Wilson & Sperber, 2004). At the bottom of the ladder, induction
makes use of rough guesses and, at the top, involves deliberation including
construction and calculation:
Most effort, most justifiable: Set theoretic frames (including
categorical)
Some effort, moderate justifiability: Causal models
Least effort, least justifiable: Associative strategies (correlational;
similarity)
The bottom level involves pattern recognition (Sloutsky & Fisher, 2004). Ob-
jects that have a particular pattern have some property; therefore objects with
similar patterns have that property. If I know a particular kind of mushroom
is poisonous, then I avoid all mushrooms that look like it unless I have more
specific information to help me along. But if I have causal knowledge to
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

340 S. A. Sloman

deploy, and I have the time and motivation to deploy it, then I can do better.
For instance, if I can identify the poison-relevant attributes of the mushroom,
or if I can figure out from the first mushroom where the poison resides and
how to get rid of it, or if I can use the first mushroom to obtain some kind
of chemical imprint of the poison, then I can use that new information to
generalize to new mushrooms. I can use a causal model to draw a more spe-
cific and more certain conclusion about a novel object (Sloman, 2005). This
requires more effort, but the greater specificity and certainty offers a bigger
payoff.
However, causal knowledge is not perfect and causal relations might all be
probabilistic (Pearl, 2000), so in some cases one can do better than by using
a causal model. For instance, if in addition to causal knowledge, I know that
99% of the mushrooms in the area are poisonous, I would be unlikely to risk
eating the mushroom under any circumstances. This third kind of inference
is set theoretic in the sense that it involves a representation or a calculation
over a set of objects (all the mushrooms in the area). Lagnado and Sloman
(2004) argue that thinking about sets involves an outside view of a category
in terms of its extension as a set of instances, and that such thinking requires
special effort. People find it easier to think about categories from the inside, in
terms of their properties and relations among those properties. In the context
of induction, the distinction can be seen most clearly in the inclusion fallacy
of Shafir, Smith, and Osherson (1990):

Robins have sesamoid bones.

Therefore, all birds have sesamoid bones.

versus

Robins have sesamoid bones.

Therefore, all penguins have sesamoid bones.

Most people judge the first argument stronger, although it cannot be if you
believe that all penguins are birds. The first argument entails that penguins
as well as all other birds have the property; the second that only penguins
do. So the second argument must be stronger, and this is obvious once you
see the category inclusion relation between penguins and birds – a relation
among sets. But most people do not consider that relation until it is pointed
out to them or made transparent in some other way. This kind of thinking
generally involves the greatest effort, especially without cues to aid it along,
but it offers the biggest payoff. In this case, it affords certainty about the
stronger argument.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 341

The first level of the hierarchy is clearly associative and the top level is
clearly not. This is consistent with Feeney’s evidence that some inferences
are made intuitively and others involve deliberative processing. Some
dual-process models (Sloman, 1996) imply that we always engage in asso-
ciative reasoning at some level, that is, that the similarity-based response is
always available. In order to engage in deliberative processing, we need to
take whatever cues we can from language, perception, or elsewhere in order
to construct an effective organizational frame.
Beyond dual processes, the proposed framework is consistent with other
ideas suggested in this book. It is consistent with the idea that people use
similarity if a causal model is not available as long as no set-theoretic alterna-
tive is. One set-theoretic alternative is a category hierarchy, and Murphy and
Ross’s evidence that people rely on a single category when making inductions
may reflect a constraint on how people think about sets. The proposal is also
consistent with Tenenbaum et al.’s proposal that people use a variety of rela-
tional schemes that are individually appropriate for different inferences. But
the proposal does entail that we cannot reduce every inductive inference that
people make to a single inferential model, unless we choose a model – like
unconstrained Bayesian inference – that is so general that it does not have pre-
dictive power. To successfully predict human induction, the next generation
of models will have to respect this variety in inductive procedures.


This work was funded by NSF Award 0518147. It has benefited greatly from
the comments of Aidan Feeney, Evan Heit, and especially Phil Fernbach.

References
Barbey, A. K., & Sloman, S. A. (2006). Base-rate respect: From ecological rationality to
dual processes. Behavioral and Brain Sciences.
Dougherty, M. R. P., Gettys, C. F., & Thomas, R. P. (1997). The role of mental simulation
in judgments of likelihood. Organizational Behavior and Human Decision Processes,
70, 135–148.
Evans, J. St. B. T. (2002). Logic and human reasoning: An assessment of the deduction
paradigm. Psychological Bulletin, 128, 978–996.
Evans, J. St. B. T., & Over, D. E. (1996). Rationality and reasoning. Hove, UK: Psychology
Press.
Fernbach, P. M. (2006). Sampling assumptions and the size principle in property in-
duction. Proceedings of the Twenty-Eigth Annual Conference of the Cognitive Science
Society, Vancouver, Canada.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

342 S. A. Sloman

Fischhoff, B., Slovic, P., & Lichtenstein, S. (1978). Fault trees: Sensitivity of estimated
failure probabilities to problem representations. Journal of Experimental Psychology:
Human Perception and Performance, 4, 330–344.
Gilovich, T., Griffin, D., & Kahneman, D. (2002). Heuristics and biases: The psychology
of intuitive judgment. Cambridge: Cambridge University Press.
Hadjichristidis, C., Sloman, S. A., Stevenson, R. J., Over, D. E. (2004). Feature centrality
and property induction. Cognitive Science, 28, 45–74.
Howson, C., & Urbach, P. (1993). Scientific reasoning: The Bayesian approach. Second
edition. Open Court, Chicago.
Hume, D. (1748). An enquiry concerning human understanding. Oxford: Clarendon.
Johnson-Laird, P. (2005). Mental models and thought. In K. Holyoak & R. Morrison
(Eds.), The Cambridge handbook of thinking & reasoning, pp. 185–208. New York:
Cambridge University Press.
Lagnado, D. A., & Shanks, D. R. (2003). The influence of hierarchy on probability
judgment. Cognition, 89, 157–178.
Lagnado, D., & Sloman, S. A. (2004). Inside and outside probability judgment. D. J.
Koehler & N. Harvey (Eds.), Blackwell handbook of judgment and decision making,
pp. 157–176. Oxford, UK: Blackwell Publishing.
Lombrozo, T. (in press). The structure and function of explanations. Trends in Cognitive
Science.
Marr, D. (1982). Vision: A computational investigation into the human representation
and processing of visual information. New York: W. H. Freeman and Company.
McDonald, J., Samuels, M., & Rispoli, J. (1996). A hypothesis assessment model of
categorical argument strength. Cognition, 59, 199–217.
Oaksford, M., & Chater, N. (Eds.) (1998). Rational models of cognition. Oxford, UK:
Oxford University Press.
Osherson, D.N., Smith, E. E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based
induction. Psychological Review, 97, 185–200.
Quine, W. V. (1970). Natural kinds. In N. Rescher (Ed.), Essays in honor of Carl G.
Hempel, pp. 5–23. Dordrecht: D. Reidel.
Shafir, E., Smith, E., & Osherson, D. (1990). Typicality and reasoning fallacies. Memory
and Cognition, 18, 229–239.
Skyrms, B. (2000). Choice and chance: An introduction to inductive logic. (Fourth edition).
Belmont, CA: Wadsworth.
Sloman, S. A. (1994). When explanations compete: The role of explanatory coherence
on judgments of likelihood. Cognition, 52, 1–21.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psychological
Bulletin, 119, 3–22.
Sloman, S. A. (1998). Categorical inference is not a tree: The myth of inheritance
hierarchies. Cognitive Psychology, 35, 1–33.
Sloman, S. A., & Lagnado, D. (2005). The problem of induction. In K. Holyoak & R.
Morrison (Eds.), The Cambridge handbook of thinking & reasoning, pp. 95–116. New
York: Cambridge University Press.
Sloman, S. A., & Wisniewski, E. (1992). Extending the domain of a feature-based model
of property induction. Proceedings of the Fourteenth Annual Conference of the Cognitive
Science Society, Bloomington, IN.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47

Taxonomizing Induction 343

Sloutsky, V. M., & Fisher, A. V. (2004). Induction and categorization in young children:
A similarity-based model. Journal of Experimental Psychology: General, 133, 166–188.
Stanovich, K. E. (1999). Who is rational? Studies of individual differences in reasoning.
Mahwah, NJ: Erlbaum.
Tenenbaum, J. B. (1999). Bayesian modeling of human concept learning. In M. Kearns,
S. Solla, & D. Cohn (Eds.), Advances in neural information processing systems 11,
pp. 59–65. Cambridge: MIT Press.
Tenenbaum, J. B., & Griffiths, T. L. (2001). Generalization, similarity and Bayesian
inference. Behavioral and Brain Sciences, 24, 629–640.
Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning: The con-
junction fallacy in probability judgment. Psychological Review, 90, 293–315.
Wilson, D., & Sperber, D. (2004). Relevance theory. In L.R. Horn & G. Ward (Eds.), The
handbook of pragmatics, pp. 607–632. Oxford, UK: Blackwell Publishing.
P1: IBE
0521672443c13 CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:47
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

Index

Abrahamsen, A.A., 242, 244 138, 139, 140, 164, 165, 179, 180, 183,
Adams, E.W., 272, 278, 280, 281, 283, 295, 201, 310, 326
298 Attribution, 231
Adolphs, R., 305, 324 Availability, 114–134
Aguilar, C.M., 141, 163
Ahn, W., 86, 111, 113, 139, 163, 241, 244 Bacon, F., 12, 22
Alfonso-Reese, L.A., 173, 201 Baer, L., 55, 80
Algebra and mathematical induction, 259–60 Bailenson, J.N., 96, 111
Allen, J.L., 305, 325 Baillargeon, R., 80, 215, 224, 235, 244
Analogical reasoning, 50, 230, 240–241, 243 Baker, A., 122, 134, 135
Anderson, A.R., 297, 298 Baldwin, D.A., 26, 42, 51, 115, 131, 132, 135
Anderson, J.R., 134, 169, 197, 198, 201, 206, Baraff, E., 115, 124, 126, 135, 136, 194, 203
207, 212, 219, 220, 222, 224, 261, 262, Barbey, A.K., 332, 341
264, 268, 280, 298 Barsalou, L.W., 59, 78, 131, 132, 135, 138, 163
Andrews, G., 50, 52 Barston, J.L., 303, 304, 325
Anggoro, F., 64, 78 Base rate neglect, 211, 332
Argument from ignorance, 272, 290–294, 296 Basic level categories, 29, 103
Argument strength, Batsell, R.L., 154, 163
and AI knowledge representation, 297 Bauer, P., 29, 42, 51
and certainty, 5 Bay, S., 230, 246
and informal reasoning fallacies, 290–294 Bayesian approach,
and likelihood ratio, 287–288 and normative justification, 329–331
measures of, 5, 286–290 to belief updating in conditional reasoning,
Aristotle, 273, 290 285–286
Artificial categories, 208–210 to deductive reasoning, 8
Artificial expert systems, 152, 161–163 to evidential diversity, 14, 21
Asgharbeygi, N., 230, 246 to inductive reasoning, 21–22, 45–46,
Ashby, F.G., 173, 201 110–111, 167–200, 206–207, 218–223,
Asher, S.J., 170, 173, 193, 202 313, 318–321
Asmuth, J., 22, 266, 268, 328, 329 to informal reasoning fallacies, 290–294
Asymmetries in conditional inference, 283–286 Bayesian networks, 187, 230
Asymmetries in similarity judgment, 141 Bayes’ rule, 177
Asymmetries of projection , 55–78, 141, 172, Bayindir, H., 290, 293, 294, 299
184 Bechara, A., 305, 324
Atran, S., 16, 24, 32, 51, 55, 57, 60, 61, 66, 71, Bechtel, W., 242, 244
78, 79, 96, 103, 111, 112, 123, 124, 135, Belief bias, 303–306

345
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

346 Index

Belnap, N.D., 297, 298 effects of,

Bennett, J., 278, 279, 280, 281, 298 in adults, 38–40, 69, 196–197, 223,
Beveridge, M., 305, 325 319–321, 333–334, 338–339
Bishop, C.M., 233, 246 in children, 26, 38–40, 43, 66–67,
Bishop, Y.M.M., 286, 298 experimentally provided, 86–95, 97–107,
Blake, R.M., 229, 245 pre-existing, 84–85, 96–97, 102–103
Blanchette, I., 240, 245 Causal reasoning,
Blok, S.V., 22, 63, 78, 164, 334 and abduction, 228, 230, 234–237, 241
Bloom, P., 55, 78, 107, 112, and induction, 81–111, 127,
Bloomfield, A., 266, 268 Causal structures, 104–107, 110
Blum, B., 197, 204 chain network structure, 105–107
Bodenhausen, G.V., 215, 224 common cause structure, 104–107
Bonawitz, E.B., 194, 201, 203 common effect structure, 105–107
Bonini, N., 138, 145, 165 see also, causal transmission model
Booth, A.E., 30, 51 Causal transmission model, 184–185, 194–197
Bott, L., 132, 135 Caverni, J.P., 8, 24, 270, 294, 298, 321, 325
Bottom-up approaches to development of Chan, D., 280, 298
induction, 48–49 Charniak, E., 229, 245
Boyer, 27, 49, 51 Chater, N., 6, 8, 22, 23, 24, 52, 169, 197, 201,
Boyes-Braem, P., 31, 53, 103, 113 203, 269, 270, 271, 273, 274, 280, 283,
Bradshaw, G., 230, 246 284, 285, 286, 287, 288, 295, 297, 298,
Brenner, D., 154, 163 299, 300, 301, 307, 308, 325, 326, 331,
Brooks, L.R., 222, 224 342
Brown, A.L., 37, 43, 44, 52 Cheng, P.W., 89, 112, 230, 246, 307, 325
Bruner, J.S., 26, 51 Children’s mathematical induction, 258
Budescu, D.V., 276, 301 Children’s reasoning
Burnett, R.C., 230, 241, 246 about artefacts, 29, 216
Byrne, R.M.J., 261, 267, 269, 289, 299 about cross-classified categories, 118
about multiple versus single premise
Callanan, M., 33, 37, 43, 53 arguments, 26, 32–35
Cannon, I., 151, 164 about social categories, 29
Capon, A., 305, 325 biasing effects of belief on, 305
Cappa, S., 10, 24, 248, 268, 297, 300 effects of knowledge on, 31–32, 49–50, 59,
Caramazza, S., 192, 202 61, 69, 118–119, 127–128
Carey, S., 17, 18, 22, 25, 28, 30, 31, 33, 42, 43, effects of labels on, 27–30, 42
51, 55, 56, 57, 58, 60, 61, 67, 68, 69, 70, effects of properties on, 35–38, 118–119
78, 79, 151, 164, 168, 170, 187, 192, 201 role of category coherence in, 102
Carnap, R., 13, 22, 310, 324 role of hierarchical structure in, 26, 31–32,
Carpenter, P.A., 305, 324 45
Casey, K.L., 37, 52, 55, 78 role of perceptual similarity in, 26–27
Casler, K., 37, 43, 44 role of taxonomic relations in, 27–30,
Category-based conjunction fallacy, 120–121, 118–119
318–321 with diverse evidence, 17–19
Category coherence, 84, 102–107 Chin-Parker, S., 215, 225
in children, 102 Choi, I., 118, 124, 125, 127, 135
relation to taxonomic hierarchy, 103 Chua, F., 280, 293
Cattell, R.B., 305, 326 Churchland, P, M., 238, 245
Causal knowledge, Circularity, 272,
and induction, 38–40, 81–111, 139, Claire, T., 227, 245
151–152, 172, 174, 184–185, 194–197, Clark, K.L., 294, 298
338–341 Cohen, C., 273, 298
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

Index 347

Cohen, L.J., 163, 164, 308, 325 Deak, G.O., 42, 51

Coleman, D., 163, 164 Dedekind, R., 252, 254, 255, 267
Coley, J.D., 16, 22, 24, 28, 29, 31, 32, 38, 42, 43, Dedekind-Peano axioms, 252
45, 46, 51, 55, 57, 60, 66, 67, 68, 69, 70, Deductive correctness, 277–280
71, 79, 96, 103, 109, 110, 111, 112, 113, Deductive logics, 277–278
115, 118, 119, 120, 121, 122, 123, 124, Deductive reasoning, 226, 240
125, 126, 127, 197, 130, 131, 132, 133, its relation to inductive reasoning, 1, 2–11,
134, 135, 136, 138, 139, 140, 164, 165, 263–267, 269–298, 321–323, 328–329
172, 194, 197, 200, 201, 202, 203, 216, the problem view, 2–7
222, 223, 225, 305, 309, 310, 312, 319, the process view, 2, 7–11
320, 326, 327, 336, 338, 339 Deductive validity, 4, 8
Collin, A., 63, 78 and certainty, 4–5
Comparison processes, 60, 63, 66, 73–78, 109 Dennis, M.J., 86, 111
Computational-level accounts of induction, Dennis, I., 305, 325
169, 187, 331–332 Detterman, D.K., 306, 325
Confirmation, 145 De Ruyter van Steveninick., 233, 246
Connectionist models, 174–175 Dhami, M.K., 276, 293, 298
see also Feature-based model Diagnostic reasoning, 88–89
Conclusion homogeneity effect, 42 Diagrams and mathematical induction,
Conceptual development, 57, 168 261–263
Conditional probability and inductive Diesendruck, G., 29, 51
reasoning, 137–163 Discovery-based learning, 230
Conditionals and conditional probability, Diversity effects
269–270, 278–290 in adults, 16, 32, 96, 98–100, 171–172, 180,
Conjunction fallacy, 137, 303, 319–321, 332 191–192, 309–317
see also category-based conjunction fallacy in children, 17–19, 32–35, 42–43, 311, 315
Counterexamples and mathematical Diversity principle, 11–22, 139, 143
induction, 264–265 eliminative justification for, 14
Copi, I.M., 273, 298 normative considerations, 13–15, 310
Corballis, P.M., 236, 245 Dolan, R.J., 240, 245, 305, 325
Corner, A., 286, 287, 288, 290, 299 Domain specificity,
Cosmides, L., 307, 325 emergence of, 37–38, 57, 60
Cox, J.A., 276, 300 in theories of induction, 110–111, 115,
Craver, C.F., 242, 246 167–170, 330–331, 333, 341
Cresswell, M.J., 278, 299 Donahue, E., 160, 164
Criterion-shift account of reasoning, 9 Dougherty, M. R. P., 337, 341
Cross-classification and induction, 118, Dual-process theories, 7–11, 303–308,
215–218, 318 310–317, 321–324, 334–335, 341
Cross-cultural studies of induction, 16, 56, 64, Ducasse, C.J., 229, 245
69–75, 96, 103, 120, 123–125, 127–128, Dunbar, K.N., 230, 236, 240, 245
310 Dunning, D., 319, 321, 322, 323, 325
Crivello, F., 240, 245
Crupi, V., 145, 165 Earman., 14, 22
Curtis-Holmes, J., 305, 325 ECHO model, 238
Ecological reasoning, 38, 69, 71–72, 77,
Damasio, A.R., 240, 245, 305, 325 110–111, 114, 116, 118–119, 122,
Damasio, H., 305, 325 125–128, 131–134, 168, 172
Darden, L., 229, 242, 245, 246 Edgington, D., 279, 298
Davidson, N.S., 29, 42, 51 Eells, E., 145, 164
Dawes, R., 160, 164 Eimas, P.D., 28, 52
Day, R.S., 115, 134, 135 Eliaser, N.M., 259, 260, 267
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

348 Index

Eliasmith, C., 233, 239, 244, 245 Fry, A.F., 306, 325
Emotion and reasoning, 227–228, 233–234, Fugelsang, J.A., 236, 245
240–241
Enderton, H., 274, 298 Gabbay, D., 296, 298, 299, 300
Engle, R.W., 305, 306, 326, 327 Gap model, 139, 142
Essentialism, 48–49, 103, 107, 111 Gap2 model, 109, 142–152, 163
Estin, P., 103, 111 Gati, I., 137, 166
Evans, J. St. B.T., 9, 23, 261, 267, 270, 280, 298, Gazzaniga, M. S., 236, 245
302, 303, 304, 305, 306, 325, 329, 334, Gelman S.A., 18, 19, 23, 24, 28, 29, 30, 31, 33,
341 34, 36, 37, 42, 43, 44, 45, 46, 48, 49, 51,
Evolutionary model, 181–183, 188–197 52, 53, 55, 57, 62, 78, 80, 85, 100, 102,
Example-based strategies in maths induction, 103, 107, 112, 138, 139, 163, 165, 168,
258–263, 266 172, 186, 187, 201, 204, 216, 224, 241,
Exemplars and inductive inference, 221–223 244, 312, 314, 315, 326
Expert systems, Gentner, D., 41, 50, 51, 66, 74, 79, 90, 113, 137,
see Artificial expert systems 141, 164, 165
Expertise, 16, 38, 96, 116, 124–127, 129, 310 Geometry and mathematical induction,
Explanation, 260–263
accounts of, 234–237 George, C., 280, 298
and abductive reasoning, 337–338 Gettys, C.E., 337, 341
and causality, 83, 95–102,109, 234–237 Gigerenzer, G., 138, 164
and induction, 41, 85, 338 Gilinsky, A.S., 304, 325
Extensional reasoning, 90–95 Gilovich, T., 230, 245, 332, 342
Girotto, V., 8, 24, 271, 295, 299, 321, 326
Fallis, D., 265, 267 Gleitman, H., 60, 78
Farrar., 27, 49, 51 Gleitman, L.R., 60, 78
Farrelly, D., 307, 308, 313, 316, 325 Goel, V., 10, 23, 240, 245, 248, 267, 305, 325
Fazio, F., 10, 24, 248, 268, 297, 300 Gold, B., 10, 23, 248, 267
Feature alignment, 141 Gold, E., 160, 164
Feature-based model, 20–21, 44–45, 65, 76, Goldfarb, W., 253, 267
130, 174–175, 310–311, 316–317, 337 Goldstone, R.L., 41, 51, 66, 79, 137, 141, 164,
Feature centrality, 86–87 165
Feeney, A., 6, 15, 17, 22, 23, 38, 52, 120, 135, Gooding, D.C., 13, 23
307, 310, 313, 319, 321, 322, 323, 324, Goodman, N., 15, 23, 62, 78, 137, 164, 174,
325, 334, 336, 338, 341 201
Feinberg, S.E., 286, 296 Gopnik, A., 168, 201, 204, 228, 245
Feldman, J., 236, 245 Gordon, M., 200, 203
Fernbach, P.M., 336, 341 Gorman, M.E., 13, 23
Fischhoff, B., 337, 342 Goswami, U., 50, 51
Fisher A.V., 28, 42, 45, 48, 49, 52, 339, 342 Graham, S.A., 27, 28, 30, 42, 51, 52, 54
Fitelson, B., 145, 164 Grassi, F., 10, 24, 248, 267, 297, 300
Flach, P.A., 230, 245 Grattan-Guinness, I., 252, 267
Florian, J., 29, 42, 51 Gray, W.D., 31, 53, 103, 113
Folkbiology, 55, 70, Green, B., 192, 202
Fong, G.T., 309, 325 Green, D.M., 292, 298
Foodwebs, 184–186, 194–195 Greenberg, M.J., 258, 263, 267
Forsyth, B., 276, 301 Greenfield, P.M., 26, 51
Fox, J., 296, 298 Griffin, D., 230, 245, 332, 342
Freedman, R., 55, 78 Griffiths, T.L., 14, 24, 168, 177, 182, 186, 194,
Frege, G., 248, 253, 254, 255, 265, 267 197, 201, 202, 204, 330, 331, 332, 333,
Frey, M.C., 306, 325 336, 337, 338, 341, 343
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

Index 349

Gutheil, G., 18, 23, 24, 30, 33, 34, 42, 45, 49, Houde, O., 240, 245
52, 55, 57, 78, 100, 112, 138, 165, 312, Houle, S., 10, 23, 248, 268
314, 315, 326 Howson, C., 14, 23, 310, 326, 330, 342
Hu, R., 119, 127, 136
Hadjichristidis, C., 86, 87, 88, 101, 108, 110, Huang, Y.T., 30, 51
112, 338, 342 Hughes, G.E., 278, 299
Hahn, E., 27, 35, 36, 43, 48, 53 Hume, D., 13, 23, 328, 338, 342
Hahn, U., 15, 18, 19, 22, 23, 34, 42, 52, 138, Hypothesis-based model, 66
164, 286, 287, 288, 290, 292, 293, 294,
299, 300, 310, 326, 328, 329, 331 Inagaki, K., 31, 52, 55, 61, 65, 66, 67, 68, 78
Hale, S., 306, 326 Inclusion similarity effect, 335
Halford, G., 50, 52 Incoherence, 154, 156, 160
Hallett, M., 253, 267 Individual differences,
Hamblin, C.L., 290, 299 and cognitive constraints, 306
Hampton, J., 151, 164 and conjunction fallacy, 319–321
Handley, S.J., 6, 23, 280, 285, 298, 301, 305, and diversity, 309–317
307, 308, 313, 316, 325, 326 and dual process theories, 10, 306–309
Hanson, N.R., 229, 245 and induction, 309–324
Harel, G., 260, 268 and normative theories, 308, 309–310
Harley, C., 307, 308, 313, 316, 326 and Wason’s selection task, 307–308
Harman, G., 7, 23, 228, 245 and uncertain classification, 318
Hasher, L., 305, 325 Inductive bias, 168
Hastie, R., 103, 104, 106, 107, 111, 113, 320, Inductive selectivity, 117, 125–127, 132–134
327 Infancy, reasoning in, 27–28, 30, 35–36,
Hatano, G., 31, 52, 55, 61, 65, 67, 68, 78 Informal reasoning fallacies, 272–273, 290–294
Hattori, M., 287, 299 Inhelder, B., 26, 27, 52
Hayes, B.K., 1, 16, 19, 22, 23, 31, 32, 34, 38, 39, Intuitive theories, 168, 187
40, 43, 44, 45, 46, 52, 53, 54, 60, 67, 78, and cognitive development, 186
97, 109, 110, 111, 112, 119, 120, 121,
131, 132, 135, 138, 165, 197, 202, 310, Jaswal, V., 29, 42, 52
312, 319, 320, 326, 333, 334, 336, 338, Jeffrey, R.C., 153, 164
339 Jepson, C., 35, 53, 309, 326
Heim, A. W., 313, 325 Johnson, D.M., 31, 52, 103, 113
Heit, E., 1, 8, 11, 14, 15, 16, 17, 18, 19, 21, 22, Johnson, S., 55, 79
24, 34, 35, 38, 42, 44, 45, 46, 51, 84, 85, Johnson-Laird, P.N., 7, 8, 23, 24, 147, 164, 269,
100, 109, 110, 112, 117, 120, 130, 132, 271, 273, 289, 295, 299, 307, 321, 326,
135, 138, 139, 164, 170, 172, 175, 188, 329, 337, 342
189, 201, 202, 207, 216, 218, 224, 276, Jones, S.S., 28, 52
299, 303, 309, 310, 312, 313, 317, 320, Josephson, J.R., 229, 230, 231, 232, 233, 234,
321, 325, 326, 328 235, 236, 237, 238, 239, 240, 245
Hempel, C.G., 13, 23, 310, 326 Josephson, S.G., 229, 230, 231, 232, 233, 234,
Henderson, A.M., 30, 51 235, 236, 237, 238, 239, 240, 245
Heyman, G.D., 29, 37, 42, 43, 51 Judd, B.B., 304, 325
Hintzman, D.L., 170, 173, 193, 202 Juslin, P., 142, 164
Holland, J.H., 13, 23 Just, M.A., 306, 324
Holland, P.W., 286, 299
Holyoak, K.J., 13, 23, 243, 245, 246, 247, 307, Kahn, T., 34, 43, 52
324 Kahneman, D., 69, 79, 114, 115, 136, 138, 164,
Horn, J.L., 305, 326 166, 211, 225, 230, 245, 257, 267, 303,
Horowitz, L.M., 115, 134, 135 308, 318, 319, 327, 332, 342, 343
Horwich, P., 14, 23 Kakas, A.C., 230, 245
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

350 Index

Kalish, C.W., 37, 43, 52, 216, 224, 241, 244 Loose, J.J., 48, 52
Kalish, W.B., 139, 163 Lopez, A., 8, 16, 18, 19, 20, 21, 22, 24, 27, 32,
Kane, M.J., 305, 326 33, 34, 44, 49, 52, 58, 59, 76, 79, 121,
Kane, R., 118, 136 123, 130, 135, 138, 139, 151, 165, 170,
Kapur, S., 10, 23, 248, 267 171, 173, 174, 180, 188, 192, 193, 203,
Kaye, R., 255, 267 216, 225, 310, 311, 312, 314, 315, 317,
Keeble, S., 235, 246 326, 337
Keil, F.C., 48, 52, 55, 57, 78, 85, 113, 137, 164 Love, B.C., 82, 96, 100, 112
Kelemen, D., 37, 43, 44, 52 Luenberger, D.G., 155, 165
Kemp, C., 110, 111, 115, 122, 135, 177, 182, Lynch, E.B., 55, 57, 61, 66, 67, 68, 69, 70, 71,
183, 186, 187, 193, 194, 200, 202, 203, 72, 73, 74, 78, 79, 80, 124, 135
204, 329, 330, 331, 332, 333, 337, 338,
341 Maass, W., 233, 246
Kilbreath, C.S., 27, 28, 42, 51 Machamer, P., 233, 242, 246
Kim, N.S., 86, 111 Machine-learning algorithms, 168
Kim, S.W., 86, 113 Macrae, C.N., 215, 224
Kincannon, A.P., 13, 23 Madden, E.H., 229, 245
Klahr, D., 230, 245 Madole, K.L., 41, 53
Kline, M., 252, 268 Magnani, L., 229, 231, 246
Knuth, D.E., 255, 268 Malt, B.C., 207, 208, 224, 318, 326
Koedinger, K.R., 261, 262, 264, 268 Mandler, J.M., 35, 36, 43, 53, 236, 246
Kotovsky, L., 235, 244 Manktelow, K.I., 307, 326
Krantz, D.H., 35, 53, 309, 325, 326 Mansinghka, V., 200, 203
Kruschke, J.K., 173, 193, 202 Marcus-Newhall, A., 231, 246
Kunda, Z., 35, 53, 227, 231, 245 Mareschal, D., 48, 52
Markman A.B., 30, 53, 55, 74, 78, 79, 141,
Lagnado, D.A., 213, 214, 224, 230, 246, 337, 164
338, 139, 165, 340, 342 Markman, E.M., 26, 28, 29, 32, 33, 36, 42, 45,
Lakoff, G., 236, 246 51, 53, 85, 112, 172, 179, 201, 202
Langley, P., 230, 246 Markov mutation model, 182
Larkin, J., 270, 283, 295, 300 Markovits, H., 285, 300
Lassaline, M.E., 38, 41, 52, 86, 90, 101, 102, Marr, D., 169, 187, 202, 313, 326, 331, 342
103, 104, 105, 106, 107, 108, 109, 110, Mathematical induction, 248–267, 328–329
111, 112, 139, 164 Martin, W.G., 260, 268
Lawson, C., 31, 32, 45, 51 Massey, G.J., 290, 299
Legrenzi, P., 8, 24, 271, 295, 299, 307, 321, 326 Maximal similarity, 33, 311
Legrenzi, M.S., 8, 24, 271, 295, 299, 307, 321, as an approximation for evolutionary
326 model, 194, 196
Lemmon, E.J., 256, 257, 268 versus summed similarity, 173–174, 188,
Leslie A.M., 57, 79, 235, 246 192–194
Levin, D.T., 137, 164 Mazoyer, B., 240, 245
Lewis, D., 278, 281, 299 McCarrel, N.S., 37, 43, 53
Lichtenstein, S., 337, 342 McClelland, J.L., 35, 48, 53, 174, 175, 203
Lien, Y., 230, 239, 246 McCloskey, M., 192, 202
Lipton, P., 228, 246 McDermott, D., 229, 245
Litt, A., 237, 243, 244, 247 McDonald, J., 46, 53, 66, 79, 317, 326, 335, 342
Liu, I.M., 280, 299 McDonough, L., 35, 36, 43, 53
Lo, Y., 14, 15, 19, 30, 33, 34, 42, 43, 52, 138, McKenzie, B.R.M., 197, 202
139, 164, 280, 299, 310, 326 Mechanisms
Logical fallacies, 272, and abduction, 241–243
Lombrozo, T., 338, 342 causal, 83, 85
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

Index 351

Medin, D.L., 16, 22, 23, 24, 28, 32, 38, 46, 51, Neural mechanisms,
53, 55, 57, 59, 60, 61, 63, 64, 66, 67, 68, and inference, 237–239
69, 70, 71, 72, 73, 74, 77, 78, 79, 96, as dynamical systems, 232
103, 109, 110, 111, 112, 113, 119, 120, Neuropsychological studies of reasoning, 10,
121, 123, 124, 131, 132, 135, 137, 138, 240, 248–249, 297–298, 305
139, 140, 141, 142, 163, 164, 165, 168, Newstead, S.E., 261, 267, 305, 307, 308, 313,
174, 192, 197, 202, 216, 222, 223, 224, 316, 325, 326
225, 241, 244, 310, 312, 319, 320, 325, Nguyen, S.P., 118, 135, 216, 225
326, 327, 334, 336 Nilsson, N., 153, 165
Melartin, R.L., 26, 32, 45, 51 Nisbett, R.E., 13, 23, 35, 53, 124, 125, 135, 246,
Mellers, B., 240, 246 231, 309, 325, 326
Meltzoff, A., 168, 201 Niyogi, S., 168, 186, 204
Mental model theory, 7–8, 273–274, 285–286 Non-diversity via property reinforcement
Mervis, C.B., 31, 53, 59, 79, 103, 113 effect, 16–17, 120
Michalski, R., 63, 78, Non-diversity via causality effect, 97
Michotte, A., 235, 246 Non-monontonicity effects
Mikkelsen, L.A., 197, 202 in adults, 121,
Miller, C., 60, 78 in children, 33, 43, 46
Miller, D.T., 215, 225, 227, 245 Non-monotonicity via property reinforcement
Milne, A.B., 215, 224 effect, 121, 129
Milson, R., 134 Non-monotonicity and deduction, 270–271,
Mirels, H.L., 160, 164 282,
Misinterpreted necessity model, 261 Norenzayan, A., 124, 135
Mitchell, T.M., 168, 178, 202 Norman, S.A., 115, 134, 135
Modus ponens, 229, 269 Normative accounts of induction, 13–15,
and mathematical induction, 252–254, 329–332
265–267 Norvig, P., 152, 165
Moloney, M., 31, 32, 45, 51 Nosofsky, R.M., 170, 173, 193, 203
Monotonicity effects, Nute, D., 278, 299
in adults, 32,
in children, 32–35, 43 O’Reilly, A.W., 31, 51, 102, 112
Monotonicity principle, 140, 144, 148–149 Oakes, L: 41, 53
Moutier, F., 240, 245 Oaksford, M., 6, 8, 22, 23, 24, 169, 197, 201,
Multidimensional scaling, 157 203, 269, 270, 271, 273, 274, 280, 283,
Multimodal representations, 232–234, 242 284, 285, 286, 287, 288, 289, 290, 292,
Murphy, G.L., 41, 53, 118, 133, 135, 168, 174, 293, 294, 295, 297, 298, 299, 300, 301,
192, 197, 198, 202, 203, 207, 208, 209, 303, 307, 308, 325, 326, 328, 329, 331,
210, 211, 212, 213, 214, 216, 217, 218, 332, 342
219, 220, 221, 222, 223, 224, 225, 318, Oberauer, K., 280, 300
326, 327, 336, 337, 338, 341 Olbrechts-Tyteca, L., 293, 300
Mutation history, 181–183 Olver R.R., 26, 51
Mutation principle, 179, 181–183 One-process accounts of reasoning, 7–11,
Myers, T.S., 86, 112 332–333
Myrvold, W.C., 310, 326 Ortiz, C.L., 139, 165
Ortony, A., 103, 112
Nagel, E., 13, 24, 310, 326 Osherson, D.N., 8, 10, 14, 15, 16, 19, 20, 21, 22,
Nair, M., 91, 93, 94, 95, 112 24, 27, 30, 32, 33, 34, 42, 43, 44, 52, 58,
Narayan, S., 236, 245 59, 63, 76, 78, 79, 82, 85, 86, 109, 112,
Neapolitan, R., 160, 165 113, 121, 123, 130, 135, 136, 137, 138,
Needham, A., 235, 244 139, 140, 142, 145, 151, 154, 163, 164,
Nelson, L.J., 215, 225 165, 170, 171, 172, 173, 174, 180, 188,
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

352 Index

Osherson (cont.) functional, 36, 43

189, 191, 192, 193, 197, 200, 201, 203, hidden, 18–19, 28, 34, 114
216, 222, 225, 248, 267, 297, 300, 310, knowledge of, 21, 35–38, 117–119, 172
311, 317, 326, 327, 334, 337, 340, 342 and surprise, 62–63
Osman, M., 303, 327 their interaction with categories, 35–38, 43,
Ostrin, R., 60, 78 84–85, 117–119, 125–128, 132–133,
Over, D.E., 9, 23, 86, 87, 88, 101, 108, 110, 112, 216,
270, 280, 298, 301, 302, 303, 306, 307, Prospective reasoning, 89–90
322, 323, 325, 326, 334, 338, 341, 342 Prototypes, 28, 55, 58–61, 72–73
Psillos, S., 229, 246
Palmer S.E., 79
Paradoxes of material implication, 272, QPf algorithm, 157–161, 163
278–280, 282, 295 Quine, W.V. O., 165, 254, 268, 334, 342
Parsons, S., 296, 298 Quinn, P.C., 28, 52
Patalano, A.L., 215, 225 Quinn, S., 285, 300
Pearl, J., 152, 161, 165, 235, 246, 281, 297, 300
Pearlstone, Z., 115, 135 Rakison, D.H., 27, 35, 36, 43, 48, 52
Peirce, C.S., 226, 227, 246 Ramsey test, 280–281
Peng, K., 124, 135 Raney, 27, 49, 51
Perani, D., 10, 24, 248, 268, 297, 300 Rapoport, A., 276, 300, 301
Perceptual similarity, 26–30, 42, 57, 114 Ratio formula, 279–280, 286
Perelman, C., 293, 300 Rational analysis, 169, 186, 193–194, 198
Perfors, A., 200, 202 Read, S., 231, 246
Persson, M., 142, 164 Rehder, B., 22, 86, 87, 89, 90, 91, 97, 98, 99, 100,
Piaget, J., 26, 27, 52 103, 104, 106, 107, 109, 111, 112, 113,
Pineau, S., 240, 245 230, 236, 241, 246, 320, 327, 334, 338
Plausibilist reasoning, 297 Relational shift, 50
Pohorille, A., 230, 245 Relevance theory, 46–47, 60, 66, 109, 119,
Poincare, H., 253, 268 131–132, 312–313, 339
Poletiek, F., 6, 24 Rieke, F., 233, 246
Politzer, G., 280, 300 Rips, L.J., 9, 10, 11, 22, 24, 59, 79, 84, 113, 138,
Pollard, P., 303, 304, 305, 325 139, 157, 158, 160, 170, 173, 203, 216,
Pollock, J.L., 296, 297, 300 222, 225, 248, 266, 268, 269, 270, 271,
Polya, G., 251, 259, 268 273, 274, 275, 276, 277, 278, 283, 286,
Popper., 15, 24 287, 288, 289, 290, 295, 300, 301, 303,
Posner, T., 37, 43, 44, 52 321, 322, 327, 328, 329
Pragmatics, 336 Rispoli, J., 46, 53, 66, 79, 317, 327, 335, 342
Prakken, H., 296, 300 Ritov, I., 240, 246
Predicate stability, 139–140, 163 Rogers, T.T., 35, 48, 53, 174, 175, 203
Probabilistic accounts of reasoning, Rosch, E.H., 31, 53, 59, 79, 103, 113, 179, 203
see Bayesian approach Roser, M.E., 236, 245
Probability logic, 272, 283, 295 Ross, N., 55, 57, 66, 67, 68, 69, 70, 71, 72, 73,
Probabilistic taxonomic theories, 183 74, 79
Probabilistic validity, 281–282 Ross, B.H., 30, 53, 118, 133, 135, 138, 139, 163,
Problem of induction, 13, 167, 328 197, 198, 203, 207, 208, 209, 210, 211,
Proffitt, J.B., 16, 24, 96, 111, 112, 124, 135, 216, 212, 213, 214, 215, 216, 217, 218, 219,
222, 223, 225, 310, 327 220, 221, 222, 223, 224, 225, 318, 326,
Properties 336, 337, 338, 341
context dependence and independence, 131 Ross, L., 231, 233, 246
distinctive, 21, 64–66, 73–78 Rotello, C.M., 11, 23, 276, 299, 303, 321, 326
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

Index 353

Rozelle, J., 14, 15, 19, 23, 24, 30, 33, 34, 35, 42, Similarity effects
43, 52, 84, 85, 109, 110, 112, 117, 130, in adults, 83, 171
135, 138, 139, 164, 172, 202, 216, 218, in children, 42, 55–56,
224, 310, 326 Similarity Coverage model, 19–21, 33, 44, 59,
Rubin, D.C., 250, 268 76, 82–85, 95–100, 108, 130, 138–139,
Russell, B., 251, 253, 268 151, 170, 173–174, 188–194, 311–312,
Russell, S.J., 152, 165 324, 337
Simon, H., 230, 246
Salmon, W., 12, 24 Simons, D.J., 137, 164
Samuels, M., 46, 53, 66, 79, 317, 326, 335, SINC model, 45
342 Skyrms, B., 4, 24, 153, 165, 270, 301, 329,
Schaffer, M.M., 143, 165 342
Schaeken, W., 270, 285, 286, 288, 289, 300, Slippery slope argument, 272
301 Sloman, S.A., 8, 9, 20, 21, 22, 24, 27, 44, 65, 75,
Schnur, T., 10, 24, 248, 268 297, 300 76, 79, 83, 85, 86, 87, 88, 101, 108, 110,
Schroyens, W.J., 270, 285, 286, 288, 289, 300, 112, 113, 130, 136, 138, 139, 165, 170,
301 171, 174, 175, 203, 230, 246, 302, 303,
Schulz, L., 168, 201, 203, 204 304, 306, 310, 311, 316, 317,322, 327,
Schwartz, A., 240, 246 332, 334, 335, 337, 338, 340, 341, 342
Scientific reasoning, 12–13, 15, 186–187 Sloutsky, V.M., 28, 42, 45, 48, 49, 53, 339,
Scott, D., 277, 301 342
Scripts, 118, 128 Slovic, P., 337, 342
Seaton, C.E., 118, 135 Smith, E.E., 8, 16, 18, 19, 20, 21, 22, 24, 27, 30,
Selection task. 32, 33, 34, 42, 43, 44, 49, 53, 55, 58, 59,
see Wason’s selection task 76, 79, 82, 85, 86, 96, 100, 112, 113,
Shafir, E., 8, 16, 19, 20, 21, 22, 24, 27, 32, 33, 121, 123, 124, 125, 130, 135, 136, 137,
44, 53, 58, 59,76, 79, 82, 85, 112, 113, 138, 139, 142, 151, 157, 158, 160, 164,
121, 123, 130, 135, 136, 137, 138, 139, 165, 170, 171, 172, 173, 174, 180, 188,
142, 151, 165, 170, 171, 172, 173, 174, 192, 193, 197, 200, 201, 203, 216, 225,
180, 188, 192, 193, 197, 200, 203, 216, 310, 311, 312, 314, 315, 317, 326, 327,
225, 310, 311, 317, 327, 337, 340, 342 337, 340, 342
Shafto, P., 22, 38, 53, 96, 109, 110, 111, 113, Smith, L.S., 28, 52, 256, 258, 268
115, 124, 125, 126, 127, 130, 131, 132, Smith, W.C., 137, 164
133, 135, 136, 172, 194, 197, 200, 203, Smolensky, P., 239, 246
305, 309, 319, 321, 325, 329, 330, 331, Social reasoning, 11, 124–125
332, 333, 337, 338, 339, 341 Sousa, P., 55, 57, 61, 67, 68, 69, 70, 71, 72, 73,
Shanks, D.R., 213, 214, 224, 337, 342 74, 78
Shrager, J., 230, 246 Speeded response paradigm, 305
Shell, P., 306, 324 Spelke, E.S., 57, 79
Shelley, C.P., 229, 230, 231, 240, 246, Sperber, D., 312, 327, 339, 343
247 Springer, K., 36, 53, 85, 113
Shepard, R.N., 15, 24, 197, 203 Stalnaker, R.C., 278, 283, 301
Shipley, E.F., 62, 79, 102, 113 Stalnaker conditional, 283
Shoben, E.J., 59, 79, 160, 157, 158, 165 Stanovich, K.E., 9, 10, 24, 302, 303, 306, 307,
Shum, M.S., 96, 111 308, 309, 310, 316, 317, 318, 319, 327,
Sides, A., 14, 15, 19, 24, 30, 33, 34, 42, 43, 52, 334, 343
138, 139, 164, 307, 310, 326 Stepanova, O., 124, 126, 135, 136
Silverman, B., 173, 203 Stern, L.D., 170, 173, 193, 202
Similarity, 1, 8, 100–102, 109–110, 133, Stevenson, R.J., 86, 87, 88, 101, 108, 110, 112,
137–163, 170, 334, 341 280, 301, 338, 342
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

354 Index

Stewart, I., 251, 252, 253, 254, 268 Two process accounts of reasoning,
Steyvers, M., 197, 204 see dual process theories
Stob, M., 86, 112 Typicality effect
Stochastic processes, 184–185 in adults, 58, 82, 97–98, 171
Storms, G., 38, 46, 53, 60, 67, 79, 96, 109, 110, in children, 29–30, 42, 58–61
111, 112, 119, 120, 121, 131, 132, 135,
138, 165, 197, 202, 310, 312, 319, 320, Ucan E.K., 55, 57, 61, 67, 68, 69, 70, 71, 72, 73,
326, 336 74, 78
Stromsten, S., 182, 194, 201, 202 Uncertain categorisation and induction,
Structured representations, 169, 184–185, 197–199, 205–215, 218–224, 318
192 Uncertainty, aleatory, vs. epistemic, 330
Subordinate level categories, 31–32 Universal generalization, 250, 256–258
Superordinate level categories, 19–20, 29, Unsworth, D., 306, 327
31–32, 35–36 Urbach, P., 14, 23, 310, 326, 330, 342
Sum formula (for mathematical induction),
249–250, 252, Valderrama, N., 55, 78,
Surprise, effects of Van Fraassen, B., 229, 247
in inductive reasoning 62, Vapnarsky, V., 55, 57, 61, 67, 68, 69, 70, 71, 72,
in abductive reasoning, 227 73, 74, 78
Swets, J.A., 292, 298 Vardi, M., 154, 163
Veltman, F., 297, 301
Tall, D., 251, 252, 253, 254, 268 Vera, A., 55, 57, 78
Taxonomic principle, 179–180 Verde, M.F., 219, 220, 225
Taxonomic relations, 57, 114, 116, 118–119, Vitkin, A.Z., 118, 119, 127, 135, 136, 338, 339
122, 125–128, 131–134, 168, 185–186 Vreeswijk, G.A.W., 296, 300
Taxonomic trees, 180 Vygotsky, L.S., 26, 54
Tenenbaum, J.B., 14, 22, 24, 110, 111, 115,
136, 168, 177, 182, 183, 186, 193, 194, Wagar, B.M., 244, 247
197,199, 200, 201, 202, 203, 204, 207, Wagenmakers, E.J., 197, 204
218, 219, 220, 221, 222, 225, 329, 330, Wallsten, T.S., 276, 298, 300
331, 332, 333, 336, 337, 338, 341, Walton, D.N., 297, 301
343 Warland, D., 233, 246
Tentori, K., 138, 145, 165 Wason, P., 307, 308, 325
Testa, T.J., 60, 63, 79 Wason’s selection task, 6, 307–308
Thagard, P., 13, 22, 89, 113, 228, 229, 230, 231, Waxman, S.R., 22, 28, 30, 51, 55, 64, 77, 78, 79,
233, 234, 237, 238, 240, 242, 243, 244, 336
245, 246, 247, 328, 329, 337, 338 Wayne, A.P., 14, 15, 24, 310, 327
Theophrastus’ rule, 297 Weisberg, R. W., 13, 24
Theory-based priors for reasoning, 178–187 Welder, A.N., 27, 28, 42, 51
Thomas, R.P., 337, 341 Wellman, H.M., 57, 80, 168, 186, 187, 204
Thompson, S., 38, 39, 40, 43, 44, 50, Wenzel, A.E., 250, 268
54 West, R.F., 302, 307, 308, 309, 318, 319, 327
Tranel, D., 305, 324 Widdowson, D., 37, 43, 44, 52
Tsvachidis, S., 154, 163 Wilhelm, O., 280, 300
Tulving, E., 115, 135 Wilkie, O., 16, 19, 20, 21, 22, 24, 27, 32, 33, 44,
Turner, H., 139, 166 53, 58, 59, 76, 79, 82, 112, 121, 123,
Tversky, A., 65, 69, 79, 114, 115, 136, 137, 138, 130, 135, 138, 151, 165, 170, 171, 173,
141, 164, 166, 211, 225, 257, 267, 303, 174, 180, 188, 192, 193, 203, 216, 225,
308, 318, 319, 327, 332, 343 310, 311, 317, 327, 337
Tweney, R.D., 13, 23 Wilson, D., 312, 327, 339, 343
P1: PJU
0521672443ind CUFX144-Feeney 0 521 85648 5 July 20, 2007 7:50

Index 355

Wisniewski, E.J., 66, 80, 338, 342 Yopchick, J.E., 118, 135
Wright, H., 307, 308, 313, 316, 326
Wu, J., 280, 299 Zacks, R.T., 305, 325
Wu, M.L., 90, 113 Zago, L., 204, 245
Zwick, R., 276, 301
Xu, F., 177, 179, 197, 204 Zytkow, J., 230, 246