100% found this document useful (1 vote)
16 views54 pages

Statistical Learning of Complex Data Francesca Greselin download

The document is a collection of selected papers from the 11th biannual meeting of the Classification and Data Analysis Group (CLADAG) held in Milan in 2017, focusing on statistical learning and data analysis. It includes contributions on various topics such as clustering, exploratory data analysis, statistical modeling, graphical models, and big data analysis. The volume aims to promote research in classification, data analysis, and multivariate statistics, featuring a range of methodologies and applications presented by international scholars.

Uploaded by

guddemankexm
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
100% found this document useful (1 vote)
16 views54 pages

Statistical Learning of Complex Data Francesca Greselin download

The document is a collection of selected papers from the 11th biannual meeting of the Classification and Data Analysis Group (CLADAG) held in Milan in 2017, focusing on statistical learning and data analysis. It includes contributions on various topics such as clustering, exploratory data analysis, statistical modeling, graphical models, and big data analysis. The volume aims to promote research in classification, data analysis, and multivariate statistics, featuring a range of methodologies and applications presented by international scholars.

Uploaded by

guddemankexm
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 54

Statistical Learning of Complex Data Francesca

Greselin download

https://textbookfull.com/product/statistical-learning-of-complex-
data-francesca-greselin/

Download more ebook from https://textbookfull.com


We believe these products will be a great fit for you. Click
the link to download now, or visit textbookfull.com
to discover even more!

Advanced linear modeling statistical learning and


dependent data 3rd Edition Christensen R

https://textbookfull.com/product/advanced-linear-modeling-
statistical-learning-and-dependent-data-3rd-edition-
christensen-r/

Statistical and Machine-Learning Data Mining, Third


Edition: Techniques for Better Predictive Modeling and
Analysis of Big Data, Third Edition Bruce Ratner

https://textbookfull.com/product/statistical-and-machine-
learning-data-mining-third-edition-techniques-for-better-
predictive-modeling-and-analysis-of-big-data-third-edition-bruce-
ratner/

Statistical Inversion of Electromagnetic Logging Data


Qiuyang Shen

https://textbookfull.com/product/statistical-inversion-of-
electromagnetic-logging-data-qiuyang-shen/

Statistical Data Analysis using SAS Intermediate


Statistical Methods Mervyn G. Marasinghe

https://textbookfull.com/product/statistical-data-analysis-using-
sas-intermediate-statistical-methods-mervyn-g-marasinghe/
Statistical Methods for Machine Learning: Discover How
to Transform Data into Knowledge with Python Jason
Brownlee

https://textbookfull.com/product/statistical-methods-for-machine-
learning-discover-how-to-transform-data-into-knowledge-with-
python-jason-brownlee/

Advanced R Statistical Programming and Data Models:


Analysis, Machine Learning, and Visualization 1st
Edition Matt Wiley

https://textbookfull.com/product/advanced-r-statistical-
programming-and-data-models-analysis-machine-learning-and-
visualization-1st-edition-matt-wiley/

Data Science and Machine Learning Mathematical and


Statistical Methods Chapman Hall Crc Machine Learning
Pattern Recognition 1st Edition Dirk P. Kroese

https://textbookfull.com/product/data-science-and-machine-
learning-mathematical-and-statistical-methods-chapman-hall-crc-
machine-learning-pattern-recognition-1st-edition-dirk-p-kroese/

Statistical Process Monitoring Using Advanced Data-


Driven and Deep Learning Approaches: Theory and
Practical Applications 1st Edition Fouzi Harrou

https://textbookfull.com/product/statistical-process-monitoring-
using-advanced-data-driven-and-deep-learning-approaches-theory-
and-practical-applications-1st-edition-fouzi-harrou/

Complex Surveys Analysis of Categorical Data 1st


Edition Parimal Mukhopadhyay (Auth.)

https://textbookfull.com/product/complex-surveys-analysis-of-
categorical-data-1st-edition-parimal-mukhopadhyay-auth/
Studies in Classification, Data Analysis,
and Knowledge Organization

Francesca Greselin
Laura Deldossi
Luca Bagnato
Maurizio Vichi Editors

Statistical
Learning of
Complex Data
Studies in Classification, Data Analysis,
and Knowledge Organization

Managing Editors Editorial Board Members

Wolfgang Gaul, Karlsruhe, Germany Daniel Baier, Bayreuth, Germany


Maurizio Vichi, Rome, Italy Frank Critchley, Milton Keynes, UK
Claus Weihs, Dortmund, Germany Reinhold Decker, Bielefeld, Germany
Edwin Diday, Paris, France
Michael Greenacre, Barcelona, Spain
Carlo Natale Lauro, Naples, Italy
Jacqueline Meulman, Leiden, The Netherlands
Paola Monari, Bologna, Italy
Shizuhiko Nishisato, Toronto, Canada
Noboru Ohsumi, Tokyo, Japan
Otto Opitz, Augsburg, Germany
Gunter Ritter, Passau, Germany
Martin Schader, Mannheim, Germany
More information about this series at http://www.springer.com/series/1564
Francesca Greselin • Laura Deldossi •
Luca Bagnato • Maurizio Vichi
Editors

Statistical Learning
of Complex Data

123
Editors
Francesca Greselin Laura Deldossi
Department of Statistics and Quantitative Department of Statistical Sciences
Methods Università Cattolica del Sacro Cuore
University of Milano-Bicocca Milan, Italy
Milan, Italy

Luca Bagnato Maurizio Vichi


Department of Economic and Social Department of Statistical Sciences
Sciences Sapienza University of Rome
Università Cattolica del Sacro Cuore Rome, Italy
Piacenza, Italy

ISSN 1431-8814 ISSN 2198-3321 (electronic)


Studies in Classification, Data Analysis, and Knowledge Organization
ISBN 978-3-030-21139-4 ISBN 978-3-030-21140-0 (eBook)
https://doi.org/10.1007/978-3-030-21140-0

Mathematics Subject Classification (2010): 62-06, 62-07, 62Fxx, 62Gxx, 62Hxx, 62Jxx, 62Kxx

© Springer Nature Switzerland AG 2019


This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

This volume collects selected papers presented at the 11th biannual meeting of
the Classification and Data Analysis Group (CLADAG) of the Società Italiana di
Statistica, held in Milan, September 13–15, 2017.
The program of the conference included 142 presentations, organized in 3
plenary talks, 21 invited sessions, 18 contributed sessions, and a poster session. We
wish to express our gratitude to the authors, whose enthusiastic participation made
the meeting possible. The conference provided a vibrant international forum for
discussion and a mutual exchange of knowledge, thanks to the 163 attendees and
authors coming from several European countries, like Austria, Denmark, France,
Germany, Great Britain, Ireland, the Netherlands, Norway, Poland, Spain, and
Switzerland, as well as from the United States and Japan. The Scientific Committee
of the Conference was chaired by Francesco Mola, and Francesca Greselin was the
Chairperson of the Local Organizing Committee.
The topics of Plenary and Invited Sessions were carefully chosen by the
Scientific Committee in view of the CLADAG mission: to promote methodological,
computational, and applied research in classification, data analysis, and multivariate
statistics. We thank all the organizers of the sessions for inviting renowned speakers.
We extend our gratitude to all the chairpersons and active participants, whose
interesting comments and suggestions made the conference a real motivating event.
The 20 manuscripts included in the present volume were selected, through a blind
review process, among the ones presented at the conference and later submitted for
the publication in the Springer book series. We are greatly indebted to the referees
(at least two scholars were involved for each paper) for the time and effort they spent
in such a careful review.
The volume is divided into five parts as follows: Clustering and Classification,
Exploratory Data Analysis, Statistical Modeling, Graphical Models, and Big Data
Analysis.
The first part, Clustering and Classification, contains methodologically oriented
papers. The paper by Fordellone and Vichi presents the combined usage of
unsupervised classification with supervised methods to enhance the assessment and
the interpretation of the obtained partition; Rainey, Tortora, and Palumbo are the

v
vi Preface

authors of the second work, that introduces a parametric version of probabilistic


distance clustering based on the Gaussian and the Student’s t multivariate density
distributions. Cappozzo and Greselin deal with an interesting application to wine
authenticity studies based on robust clustering methods, where a mixture of
Gaussian factors is used to ascertain varietal genuineness and distinguish potentially
doctored food. Simulation results are presented by Alfò, Nieddu, and Vitiello
to study the performance of the cluster-weighted beta regression in a variety of
empirical settings. The results show that the model captures individual specific
unobserved heterogeneity and its link with observed covariates and indicate some
shortcomings related to the scale of the observed quantities. The last paper of this
part, by Ranalli and Rocci, presents an overview on a recent model-based approach
to cluster ordinal variables. The aim is to extend the proposal to the case where noise
dimensions or variables are present, and to generalize the model to mixed-type data.
The second part, devoted to Exploratory Data Analysis, contains a first paper by
Bove, Ruta, and Mastandrea where multidimensional scaling and unfolding allow
to easily detect preference order, size of asymmetry, and relationships between
subjects and stimuli coming from the curvature of architectural facades. In the
second paper, De Stefano, Vitale, and Zaccarin submit exploratory results obtained
by adopting two well-known community detection methods and a new proposal,
aiming at discovering groups of scientists in the coauthorship network of Italian
academic statisticians. In their study, Okada and Tsurumi extend the asymmetric
multidimensional scaling to analyze differences among consumers and to show
how each consumer or a group of consumers relates to brands in brand switching.
In the fourth paper, Solaro faces the problem of comparing the results obtained
with different imputation methods, so critical in many fields of application, such
as cardiovascular studies in the medical context. She also assesses the quality of
imputation, through the dissimilarity profile analysis.
The third part refers to Statistical Modeling. The first paper is by Altimari,
Balzano, and Zezza who introduce an extended version of the Economic Vulner-
ability Index, adopted by the United Nations. By the partial least square approach
to structural equation model, an estimate of the extended index is obtained using
data from a panel of 98 countries over 19 years. Ascari, Migliorati, and Ongaro
present two Bayesian procedures—both based on Gibbs sampling—to estimate
the parameters of the flexible Dirichlet. This distribution is particularly suited for
compositional data, thanks to its mixture structure and to the additional parameters
that allow for a more flexible modeling of the covariance matrix. Fiori and
Motta contribute a new stochastic model of firm dynamics that leads to a Dagum
distribution for the size of business firms operating in a given industry. The model
relies on a stochastic growth process that was originally introduced in the context
of income inequality studies and sheds new light upon the connections between
growth dynamics and the meaning of parameters that appear in the steady-state
distribution of firm size. A mixture of Mincer’s models with concomitant variables
is proposed by Mazza, Battisti, Ingrassia, and Punzo. The new model provides
a flexible generalization of the Mincer model, a breakdown of the population
into several homogeneous subpopulations, and an explanation of the unobserved
Preface vii

heterogeneity. The proposal is motivated and illustrated via an application to data


provided by the Bank of Italy’s Survey of Household Income and Wealth in 2012.
In his paper, Nakai presents interesting results about the impact of women’s own
human capital on contribution to household income, by using multinomial logistic
regression upon data coming from two national social surveys conducted in 1985,
1995, 2005, and 2015 in Japan. The last paper of this part by Vernizzi and Nakai
introduces an optimization algorithm that improves the size of the final dataset after
applying the listwise deletion method. The proposed weighted optimization method
has been applied, first to some toy examples, and then to the National Longitudinal
Survey of Youth dataset.
The fourth part—devoted to Graphical Models—contains a first paper by
Marella, Vicard, Vitale, and Ababei, which proposes a procedure based on non-
parametric Bayesian networks to detect and correct measurement error. The novel
procedure is evaluated on a validation sample associated to the Bank of Italy survey
on household income and wealth. In the second paper, Musella, Vicard, and Vitale
address the issue of Bayesian network structural learning for non-paranormal data.
They propose a modified version of the Copula Grow-Shrink algorithm whose high
performance is proved through a simulation study. In addition, an application to
Italian energy market data is also provided. The third and last paper of this part is
by Nicolussi and Cazzaro; it aims to incorporate the contest-specific independence
conditions in graphical models. The authors take advantage of the Hierarchical
Multinomial Marginal parametrization to represent the dependence relationships.
The proposal is applied on the study of the trend of innovation degree for Italian
enterprises.
The last part is devoted to Big Data Analysis. The first paper, by Giuffrida,
Gozzo, Rinaldi, and Tomaselli, proposes a new approach to address, in the Big
Data world, the analysis of relational structures to improve actionable analytics-
driven decision patterns. An application to model online news is provided. In the
second paper, Pesce, Riccomagno, and Wynn discuss some issues related to the use
of experimental design to help establish causation in complex models. The question
to remove bias is also considered: various solutions are discussed, including
randomization.
We cannot conclude this brief introduction without some further thanks. We
gratefully acknowledge the Department of Statistics and Quantitative Methods
and the University of Milano Bicocca, which strongly supported the CLADAG
conference. We shared constructing ideas with some colleagues from Milan,
valuably supported by their Institutions: Piercesare Secchi from Politecnico, Laura
Deldossi from the Università Cattolica del Sacro Cuore, Pieralda Ferrari from the
Università di Milano, and Raffaella Piccarreta from the Università Bocconi. To all
of them go our most sincere appreciations. Special thanks are due to the members
of the Local Organizing Committee. They all did a great job! A particular mention
goes to Mariangela Zenga for her tireless activity and enthusiasm.
viii Preface

We are grateful to Oracle Corporation, Data Reply, and Fondazione Cariplo.


They supported the event and made possible the data challenge for Young Cladag.
They also sponsored the final concert. Maestro Alessandro Arnoldo conducted the
Milan Chamber Orchestra, and created a delightful moment for all participants.
Finally, we acknowledge Dr. Veronika Rosteck of Springer-Verlag, Heidelberg,
for her support and dedication to the making of this volume.

Milan, Italy Francesca Greselin


Milan, Italy Laura Deldossi
Piacenza, Italy Luca Bagnato
Rome, Italy Maurizio Vichi
December 2018
Contents

Part I Clustering and Classification


Cluster Weighted Beta Regression: A Simulation Study . . . . . . . . . . . . . . . . . . . . 3
Marco Alfó, Luciano Nieddu, and Cecilia Vitiello
Detecting Wine Adulterations Employing Robust Mixture
of Factor Analyzers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
Andrea Cappozzo and Francesca Greselin
Simultaneous Supervised and Unsupervised Classification
Modeling for Assessing Cluster Analysis and Improving Results
Interpretability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 23
Mario Fordellone and Maurizio Vichi
A Parametric Version of Probabilistic Distance Clustering . . . . . . . . . . . . . . . . . 33
Christopher Rainey, Cristina Tortora, and Francesco Palumbo
An Overview on the URV Model-Based Approach to Cluster
Mixed-Type Data .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
Monia Ranalli and Roberto Rocci

Part II Exploratory Data Analysis


Preference Analysis of Architectural Façades by Multidimensional
Scaling and Unfolding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57
Giuseppe Bove, Nicole Ruta, and Stefano Mastandrea
Community Structure in Co-authorship Networks: The Case
of Italian Statisticians .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 65
Domenico De Stefano, Maria Prosperina Vitale, and Susanna Zaccarin
Analyzing Consumers’ Behavior in Brand Switching . . .. . . . . . . . . . . . . . . . . . . . 73
Akinori Okada and Hiroyuki Tsurumi

ix
x Contents

Evaluating the Quality of Data Imputation in Cardiovascular Risk


Studies Through the Dissimilarity Profile Analysis . . . . . .. . . . . . . . . . . . . . . . . . . . 81
Nadia Solaro

Part III Statistical Modeling


Measuring Economic Vulnerability: A Structural
Equation Modeling Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
Ambra Altimari, Simona Balzano, and Gennaro Zezza
Bayesian Inference for a Mixture Model on the Simplex . . . . . . . . . . . . . . . . . . . . 103
Roberto Ascari, Sonia Migliorati, and Andrea Ongaro
Stochastic Models for the Size Distribution of Italian Firms:
A Proposal . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
Anna Maria Fiori and Anna Motta
Modeling Return to Education in Heterogeneous Populations:
An Application to Italy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 121
Angelo Mazza, Michele Battisti, Salvatore Ingrassia, and Antonio Punzo
Changes in Couples’ Bread-Winning Patterns and Wife’s
Economic Role in Japan from 1985 to 2015 .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 133
Miki Nakai
Weighted Optimization with Thresholding for Complete-Case
Analysis . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 143
Graziano Vernizzi and Miki Nakai

Part IV Graphical Models


Measurement Error Correction by Nonparametric Bayesian
Networks: Application and Evaluation . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 155
Daniela Marella, Paola Vicard, Vincenzina Vitale, and Dan Ababei
Copula Grow-Shrink Algorithm for Structural Learning .. . . . . . . . . . . . . . . . . . 163
Flaminia Musella, Paola Vicard, and Vincenzina Vitale
Context-Specific Independencies Embedded in Chain Graph
Models of Type I.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 173
Federica Nicolussi and Manuela Cazzaro

Part V Big Data Analysis


Big Data and Network Analysis: A Combined Approach to Model
Online News . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 183
Giovanni Giuffrida, Simona Gozzo, Francesco Mazzeo Rinaldi,
and Venera Tomaselli
Experimental Design Issues in Big Data: The Question of Bias . . . . . . . . . . . . 193
Elena Pesce, Eva Riccomagno, and Henry P. Wynn
Contributors

Dan Ababei LightTwist Software, Delft, Netherlands


Marco Alfò Dipartimento di Scienze Statistiche, Sapienza University of Rome,
Roma, Italy
Ambra Altimari Department of Economics and Law, University of Cassino and
Southern Lazio, Cassino, Italy
Roberto Ascari Department of Economics, Management and Statistics, University
of Milano-Bicocca, Milano, Italy
Simona Balzano Department of Economics and Law, University of Cassino and
Southern Lazio, Cassino, Italy
Michele Battisti Dipartimento di Scienze Giuridiche, della Società e dello Sport,
University of Palermo, Palermo, Italy
Giuseppe Bove Department of Education, Roma Tre University, Roma, Italy
Andrea Cappozzo Department of Statistics and Quantitative Methods, University
of Milano-Bicocca, Milano, Italy
Manuela Cazzaro Department of Statistics and Quantitative Methods, University
of Milano-Bicocca, Milano, Italy
Domenico De Stefano Department of Social and Political Sciences, University of
Trieste, Trieste, Italy
Anna Maria Fiori Department of Statistics and Quantitative Methods, University
of Milano-Bicocca, Milano, Italy
Mario Fordellone Dipartimento di Scienze Statistiche, Sapienza University of
Rome, Roma, Italy
Giovanni Giuffrida Department of Political and Social Sciences, University of
Catania, Catania, Italy

xi
xii Contributors

Simona Gozzo Department of Political and Social Sciences, University of Catania,


Catania, Italy
Francesca Greselin Department of Statistics and Quantitative Methods, Univer-
sity of Milano-Bicocca, Milano, Italy
Salvatore Ingrassia Dipartimento di Economia e Impresa, University of Catania,
Catania, Italy
Daniela Marella Dipartimento di Scienze della Formazione, Roma Tre University,
Roma, Italy
Stefano Mastandrea Department of Education, Roma Tre University, Roma, Italy
Angelo Mazza Dipartimento di Economia e Impresa, University of Catania,
Catania, Italy
Sonia Migliorati Department of Economics, Management and Statistics, Univer-
sity of Milano-Bicocca, Milano, Italy
Anna Motta Il Sole 24 Ore, Milano, Italy
Flaminia Musella Department of Research, Link Campus University, Roma, Italy
Miki Nakai Department of Social Sciences, College of Social Sciences, Rit-
sumeikan University, Kyoto, Japan
Federica Nicolussi Department of Economics, Management and Quantitative
Methods, University of Milano, Milano, Italy
Luciano Nieddu Università degli Studi Internazionali di Roma, Roma, Italy
Akinori Okada Research Institute, Tama University, Tokyo, Japan
Andrea Ongaro Department of Economics, Management and Statistics, Univer-
sity of Milano-Bicocca, Milano, Italy
Francesco Palumbo Department of Political Sciences, University of Naples Fed-
erico II, Napoli, Italy
Elena Pesce Department of Mathematics, University of Genoa, Genova, Italy
Antonio Punzo Dipartimento di Economia e Impresa, University of Catania,
Catania, Italy
Christopher Rainey Department of Mathematics and Statistics, San José State
University One Washington Square, San Jose, CA, USA
Monia Ranalli Department of Economics and Finance , University of Tor Vergata,
Roma, Italy
Eva Riccomagno Department of Mathematics, University of Genoa, Genova, Italy
Francesco Mazzeo Rinaldi Department of Political and Social Sciences, Univer-
sity of Catania, Catania, Italy
Contributors xiii

Roberto Rocci Department of Economics and Finance, University of Tor Vergata,


Roma, Italy
Nicole Ruta Cardiff School of Art & Design, Cardiff, UK
Nadia Solaro Department of Statistics and Quantitative Methods, University of
Milano-Bicocca, Milano, Italy
Venera Tomaselli Department of Political and Social Sciences, University of
Catania, Catania, Italy
Cristina Tortora Department of Mathematics and Statistics, San José State Uni-
versity, One Washington Square, San Jose, CA, USA
Hiroyuki Tsurumi Yokohama National University, Yokohama-shi, Japan
Graziano Vernizzi Department of Physics and Astronomy, Siena College,
Loudonville, NY, USA
Paola Vicard Dipartimento di Economia, Roma Tre University, Roma, Italy
Maurizio Vichi Department of Statistical Sciences, Sapienza University of Rome,
Rome, Italy
Maria Prosperina Vitale Department of Political and Social Studies, University
of Salerno, Fisciano, Italy
Vincenzina Vitale Dipartimento di Scienze Sociali ed Economiche, Sapienza
University of Rome, Roma, Italy
Cecilia Vitiello Dipartimento di Scienze Statistiche, Sapienza University of Rome,
Roma, Italy
Henry P. Wynn Department of Statistics, London School of Economics, London,
UK
Susanna Zaccarin Department of Business, Economic, Mathematics and Statis-
tics, University of Trieste, Trieste, Italy
Gennaro Zezza Department of Economics and Law, University of Cassino and
Southern Lazio, Cassino, Italy
Part I
Clustering and Classification
Cluster Weighted Beta Regression:
A Simulation Study

Marco Alfó, Luciano Nieddu, and Cecilia Vitiello

Abstract In several application fields, we have to model a response that takes


values in a limited range. When these values may be transformed into rates, propor-
tions, concentrations, that is to continuous values in the unit interval, beta regression
may be the appropriate choice. In the presence of unobserved heterogeneity, for
example when the population of interest is composed by different subgroups, finite
mixture of beta regression models could be useful. When conditions of exogeneity
of the covariates set are not met, extended modeling approaches should be proposed.
For this purpose, we discuss the class of cluster-weighted beta regression models.

Keywords Beta regression · Finite mixtures · Cluster-weighted regression

1 Introduction

Frequently, we are interested in describing the (conditional on covariates) distri-


bution of a continuous response variable taking values in a limited interval, which
could be mapped onto the open unit interval. Such variables can be observed in
a wide variety of empirical situations, in the form of relative frequencies of an
event (e.g., number of votes obtained by a candidate out of the total votes cast
at an election), fractions of a continuous variable (e.g., amount of GDP due to a
specific economic sector), performance ratings (e.g. student performances when
compared to the maximum performance attainable), and limited-valued indexes
(e.g., the relative Gini index). Examples of this framework can be found in medical
(see, e.g., [12, 21, 22]), education (see, e.g., [3]), and economics [2, 6] research.

M. Alfó · C. Vitiello
Dipartimento di Scienze Statistiche, Sapienza University of Rome, Rome, Italy
e-mail: [email protected]; [email protected]
L. Nieddu ()
UNINT, Rome, Italy
e-mail: [email protected]

© Springer Nature Switzerland AG 2019 3


F. Greselin et al. (eds.), Statistical Learning of Complex Data,
Studies in Classification, Data Analysis, and Knowledge Organization,
https://doi.org/10.1007/978-3-030-21140-0_1
4 M. Alfó et al.

In this context, the use of a standard linear model is not a feasible solution (see,
e.g., [17] and [14]). A naive solution is to map the response onto the real line (for
example using a probit transform), so that a standard regression model could be
used [7]. While this approach is preferable when compared to a linear model for
the original response variable, it suffers from the well-known shortcomings, see [1]
and [19]. A viable alternative is to use a regression model based on a conditional
beta distribution for the response Y given the p-dimensional vector of covariates
x, that is Y | x ∼ B(p, q), Y ∈ (0, 1), with parameters p, q > 0. This
model has been introduced by Ferrari and Cribari-Neto [8], and extended by Ospina
and Ferrari [15, 16] to account for those cases where the response is defined over
the closed interval [0, 1], via (factorizable) mixtures of discrete and continuous
distributions. For modeling purposes, Ferrari and Cribari-Neto [8] proposed the
following parameterization:

p μ(1 − μ)
E(Y ) = μ = ; V ar(Y ) = (1)
p+q 1+φ

where φ = p + q represents the precision parameter. The beta distribution may


assume different shapes according to different combinations of the (p, q), or (μ, φ),
parameters. Parameter estimates can be obtained using a maximum likelihood (ML)
approach, see [8]; for this purpose, the R library Betareg, [5], can be used. In
some cases, however, individual heterogeneity is only partially accounted for by the
observed covariates, and continuous/discrete mixed effect beta regression should be
taken into consideration. In particular, when the omitted covariates can be described
by a latent variable with a discrete distribution, or the population of interest is
composed by several subgroups, characterized by different values of regression
parameters, finite mixtures of beta regressions represent a viable option. This model
has been introduced in the literature by Grün et al. [10], who designed the R library
betamix. As in standard finite mixture models, this is based on the (sometimes
non-explicitly stated) assumption of assignment independence, see [11], which can
be also considered as a sort of exogeneity of observed covariates with respect to the
discrete latent variable. Two issues are, in this case, of interest. First, due to the wide
range of different shapes the beta distribution may take, a further parameterization
has been proposed by Chen [4], see also [1]. This is motivated by identifiability
issues and it is based on the subclass of unimodal beta densities. This prevents
the problem of a U-shaped distribution being fitted by a mixture of two J-shaped
ones. However, in the present context, we prefer an approach based on modeling the
mean of the response, rather than its mode. Second, in several empirical cases, the
omitted covariates described by the latent variables and the observed covariates may
not be independent, and this may question the reliability of parameter estimates. In
fact, if we do not account for such a dependence, the estimated impact of observed
covariates may be either due to a direct effect on the response or to the (mediated)
effect of omitted covariates. In a former paper, we introduced the class of cluster-
weighted beta regression models, see [9], to capture individual specific unobserved
heterogeneity and its link with observed covariates (see [13] and [18]).
Cluster Weighted Beta Regression: A Simulation Study 5

In the next section, after discussing the standard finite mixture approach, we
motivate the proposed approach. The EM algorithm for ML parameter estimation
is sketched in Sect. 3 while, in Sect. 4, we report the results of a simulation study.
Some concluding remarks are drawn in Sect. 5.

2 The Model

Let (Y, X) be the set including a response variable Y and a covariates vector X;
let the corresponding population be partitioned into K subpopulations, referred to
as components, and let πk (xi ) denote the prior probability that unit i belongs to
component k = 1, . . . , K. We associate to subpopulations the indicator vector
with elements zik = 1 if unit i belongs to component k. We further assume that,
conditional on being in the k-th component, k = 1, . . . , K, the following model
holds

Yi | Xi , zik = 1 ∼ B(pik , qik ), pik , qik > 0 (2)


pik    μik (1 − μik )  
μik = = ηik = h1 xi β k φik = = ξik = h2 xi γ k .
pik + qik 1 + φik

That is, each component is associated to possibly varying parameter vectors, β k


and γ k , in the models for the mean and the precision parameter of that component.
However, the component indicators are not observed and, for purpose of estimation
in an ML framework, we need to define the observed data joint density:


K
f (yi | xi ) = f (yi | xi , zik = 1)πk (xi ) . (3)
k=1

While the first term in the sum denotes the (conditional) beta density, the term
πk (xi ), k = 1, . . . , K, has been previously defined. A usual assumption is that
πk (xi ) = πk ∀k, known as assignment independence. When the independence is not
met, this assumption may lead to a severe bias in model parameter estimates. This
could be simply motivated by looking at the graph in Fig. 1; here, xi has a non-zero
impact on Yi either directly or indirectly through zik . By adopting an independence
assumption, we use a misspecified model. As a result, we remove the gray dashed
edge from xi to zik , and inflate the impact of the covariates on the response Yi .
In the case xi and zik are not independent, we need to either model the impact of
covariates on the component indicator, or go for the specification of the marginal
density of the couple (Y, X). In the former case, we cannot distinguish between the
direct and the indirect impact of X; in fact, there is no way, as opposed to the case
when repeated observations are available, to test whether the effect of the observed
covariates on Y or on Y through Z. For the latter, we may turn to consider a more
flexible family of mixture models, the cluster-weighted models (CWM), which can
6 M. Alfó et al.

Fig. 1 A simple path


xi xi
diagram. (a) True model. (b)
Misspecified model dir. dir
ind. ind.
Yi Yi
zik zik

(a) (b)

be obtained by releasing the assignment independence hypothesis. This approach


has been introduced by Gershenfeld [9] as a model based clustering approach for
the couple (Y, X). See [13] and [18] for extensions. The model is:


K
f (yi , xi ) = g(xi | zik = 1)f (yi | xi , zik = 1)πk . (4)
k=1

Here, f (yi | xi , zik = 1) is the density of the response conditional on the set of
covariates and the component the unit belongs to, and g(xi | zik = 1) denotes the
distribution of the covariates in the specific component k = 1, . . . , K.

3 ML Parameter Estimation

When the covariates are continuous, it is customary to adopt a component-specific


multivariate Gaussian density, that is X | zik = 1 ∼ MV N(ν k , Σ k ). With
non-continuous or mixed-type covariates more general models should be used
to describe component-specific covariates distribution, [13]. A (conditional) beta
model:

g1 (μi ) = ηk = xi βk g2 (φi ) = ξk = xi γk (5)

is used for the response, while the link functions g1 (·) and g2 (·) are monotone
and twice differentiable. Given these modeling assumptions, the observed data
likelihood is:


n 
K
L(ψ) = f (yi | xi , β k , γ k )g(xi | υ k , Σ k )πk
i=1 k=1

where θ k = (β k , γ k ) denotes the vector of regression parameters, while


ωk = (ν k , Σ k ), k = 1, . . . , K, represents the parameters for the condi-
tional Gaussian distribution for the set of observed covariates. Last, ψ =
(θ1 , . . . , θK , ω1 , . . . , ωK , π1 , . . . , πK−1 ) denotes the global set of model
Cluster Weighted Beta Regression: A Simulation Study 7

parameters. Let c (ψ) denote the log-likelihood function for the complete data
{yi , xi , zi }i=1,...,n . For fixed K, at the r-th iteration of the EM algorithm, r = 1, . . . ,
the E-step computes the expected value of the log-likelihood function for complete
data, conditional on the observed data and the current parameters estimates ψ (r−1) .
This is referred to as Q(ψ | ψ (r−1)). In the M-step of the algorithm, this function
is maximized with respect to ψ. The algorithm alternates the two steps until
convergence, defined in terms of the norm of the difference between two subsequent
values of model parameters:
E-step Compute Q(ψ | ψ (r−1) ) = Eψ (r−1) ( c (ψ) | X, y)
M-step maximize Q(ψ | ψ (r−1) ) w.r.t. ψ to obtain updated estimates ψ (r)
In the E-step the missing component indicator zik is replaced by the corresponding
conditional expectation

(r−1) (r−1) (r−1)


(r) f (yi | xi , θk )g(xi | ωk )πk
wik = ,
K f (y | x , θ (r−1) )g(x | ω(r−1) )π (r−1)
Σl=1 i i l i l l

that represents the posterior probability that the i-th unit comes from the k-th
component, i = 1, . . . , n, k = 1, . . . , K. ML equations for the beta model and
the Gaussian density parameters are given by the following expressions

∂ (·)  (r) ∂ log(f (yi | xi , θk )) ∂ (·)  (r) ∂ log(g(xi | ωk ))


n n
= wik =0 and = wik = 0.
∂θk ∂θk ∂ωk δωk
i=1 i=1

(r)
Both expressions are weighted score equations with weights given by wik . The
updated estimates for the Gaussian density parameters are available in closed form:
n n
− ν̂k )(xi − ν̂k )
(r) (r) (r) (r)
(r) xi wik (r) i=1 wik (xi
ν̂k = i=1 (r)
and Σ̂k = .
n n w (r)
i=1 wik Σi=1 ik

n
while π̂k(r) = (r)
i=1 wik /n, a well known result in finite mixtures.

4 Simulation Study

To study the proposed model performance in a variety of empirical settings, we


defined the following simulation experiment. The aim is to evaluate the model’s
ability to recover the true parameter values and the coverage of the corresponding
confidence intervals. Further, we evaluate whether the estimated component mem-
bership recover the true partition of the data. To make the data easy to interpret
and visualize, the covariates Xi = (Xi1 , Xi2 ) have been drawn from a multivariate
8 M. Alfó et al.

Gaussian density Xi |zik =⎧1 ∼ MV N(μk , Σ) and the number of groups has been
⎨ {0 0} k = 1
set to K = 3 where μk = {−1 +1} k = 2

{+1 +2} k = 3
For each component we have drawn nk = 700, k = 1, 2, 3 units resulting in
a total sample size n =2100. Two possible scenarios have been considered for the
covariance matrix, namely:

σ2 0
A: Σ = ; σ 2 = {0.09 + a · 0.02}, a = 1, . . . 10
0 σ2
σ12 ρσ1 σ2
B: Σ = b ; σ1 ∈ {1, 2}, σ2 ∈ {1, 2}, ρ ∈ {0, 0.2, 0.8}
ρσ1 σ2 σ22

where b is a multiplicative factor assuming values in {0.09, 0.11, 0.13}. Values for
the response variable Yi |zik = 1 have been drawn from a beta distribution with
parameters (μik , φik ) which have been chosen, conditionally on the component
membership, according to the linear predictors
∗ ∗ ∗ ∗ ∗ ∗
μik = β0k + β1k xi,1 + β2k xi,2 φik = γ0k + γ1k xi,1 + γ2k xi,2

the coefficients of the linear predictors have been set to:


⎧ ⎧
⎨ 0 −0.3 0 k=1 ⎨ 0 0 −1 k=1
{β0∗ , β1∗ , β2∗ }k = −0.5 −0.9 0.4 k=2 {γ0∗ , γ1∗ , γ2∗ }k = 0.2 0 0.5 k=2
⎩ 0.5 0.4 −0.3 k=3
⎩ 0.5 0 0.3 k=3

For each parameter configuration, 200 samples have been considered. For
estimation purpose we considered k = 1, . . . , 5 components and the best solution
was chosen according to BIC. We report in Table 1 the values of the Rand index
[20] between the original membership and the classification obtained using the
CWR approach. As expected, with increasing noise, the quality of the classification
deteriorates, decreasing from 0.98 for σ 2 = 0.09 to 0.80 for σ 2 = 0.29. As shown
in the column standard deviation in Table 1, the variability of the Rand index
distribution tends to increase as well. Overall, the coherence between the original
classification and the one obtained using the proposed approach is quite satisfactory
especially considering that our aim was to derive accurate estimates.
We have calculated the empirical coverage of the confidence intervals defined
at the nominal level 1 − α = 0.95 for parameter estimates of the beta regression
model, defined as the proportions of samples where the confidence intervals include
the true parameter values in both scenarios (Tables 2 and 3). We do not observe
a strong effect of the noise variance on the coverage, as the empirical proportions
are all very close to the nominal confidence level even for large values of the noise
variance.
Cluster Weighted Beta Regression: A Simulation Study 9

Table 1 Average Rand σ2 Average Rand Standard deviation


index and corresponding
standard deviation by values 0.09 0.98 0.0047
of σ 2 0.11 0.97 0.0064
0.13 0.95 0.0076
0.15 0.94 0.0087
0.17 0.92 0.0100
0.19 0.90 0.0111
0.21 0.88 0.0118
0.23 0.86 0.0131
0.25 0.84 0.0134
0.27 0.82 0.0143
0.29 0.80 0.0133

Table 2 Empirical coverage σ2 Component 1 Component 2 Component 3


of confidence intervals for
model parameters: Scen. A 0.09 0.95 0.93 0.95
0.11 0.97 0.94 0.94
0.13 0.97 0.96 0.96
0.15 0.97 0.94 0.95
0.17 0.95 0.93 0.97
0.19 0.94 0.95 0.96
0.21 0.95 0.95 0.94
0.23 0.95 0.94 0.94
0.25 0.93 0.96 0.94
0.27 0.95 0.95 0.95
0.29 0.95 0.93 0.94

Table 3 Empirical coverage ρ b Component 1 Component 2 Component 3


of confidence intervals for
model parameters: Scen. B 0.0 0.09 0.97 0.94 0.96
0.0 0.11 0.91 0.96 0.95
0.0 0.13 0.95 0.95 0.95
0.2 0.09 0.95 0.94 0.93
0.2 0.11 0.95 0.96 0.94
0.2 0.13 0.94 0.97 0.96
0.8 0.09 0.96 0.93 0.93
0.8 0.11 0.96 0.93 0.93
0.8 0.13 0.95 0.91 0.96

5 Concluding Remarks

We discuss cluster-weighted Beta regression to model the location and the precision
parameter for a response with a conditional beta distribution and a multivariate
Gaussian set of observed covariates. The proposal embeds the finite mixture of beta
regressions as a particular case, see [10], when g(xi | zik = 1) = g(xi ), that is
10 M. Alfó et al.

when the distribution of the observed covariates does not change across components.
In fact, in that case, the likelihood for model parameters can be factorized, and it
does not depend on the observed covariates distribution. While this is quite clear
from a theoretical point of view, there can be some shortcomings when dealing
with the implementation of the proposed method. Namely, the adopted objective
function depends on the “scale” of the observed quantities and, when no clustering
on the covariates is present, a single multivariate Gaussian component is well more
parsimonius than a finite mixture of Gaussian components which is what would
be implied by the finite mixture of the beta regression. Therefore the fact that the
proposed method suggests K = 1 may well not be a signal that the beta regression
model is homogeneous but, rather, that the Gaussian component is.

References

1. Bagnato, L., Punzo, A.: Finite mixtures of unimodal beta and gamma densities and the k-bumps
algorithm. Comput. Stat. 28, 1571–1597 (2013)
2. Buntaine, M.T.: Does the asian development bank respond to past environmental performance
when allocating environmentally risky financing. World Dev. 39, 336–350 (2011)
3. Carmichael, C.S.: Modelling student performance in a tertiary preparatory course. Master
dissertation, University of Southern Queensland (2006)
4. Chen, S.X.: Beta kernel estimators for density functions. Comput. Stat. Data Anal. 31, 131–145
(1999)
5. Cribari-Neto, F., Zeileis, A.: Beta regression in R. J. Stat. Softw. 34(2), 1–24 (2010)
6. De Paola, M., Scoppa, V., Lombardo, R.: Can gender quotas break down negative stereotypes?
Evidence from changes in electoral rules. J. Public Econ. 94, 344–353 (2010)
7. Demsetz, H., Lehn, K.: The structure of corporate ownership: causes and consequences. J.
Polit. Econ. 93(6), 1155–1177 (1985)
8. Ferrari, S., Cribari-Neto, F.: Beta regression for modelling rates and proportions. J. Appl. Stat.
31(7), 799–815 (2004)
9. Gershenfeld, N.: Nonlinear inference and cluster-wieghted modeling. Ann. N. Y. Acad. Sci.
808(1), 18–24 (1997)
10. Grün, B., Kosmidis, I., Zeileis, A.: Extended beta regression in R: shaken, stirred, mixed, and
partitioned. J. Stat. Softw. 48(11), 1–25 (2012)
11. Hennig, C.: Identifiablity of models for clusterwise linear regression. J. Classif. 17, 273–296
(2000)
12. Hunger, M., Doring, A., Holle, R.: Longitudinal beta regression models for analyzing health-
related quality of life score overtime. BMC Med. Res. Methodol. 12, 144 (2012)
13. Ingrassia, S., Punzo, A., Vittadini, G., Minotti, S.C.: The generalized linear mixed cluster-
weighted model. J. Classif. 32, 85–113 (2015)
14. Kieschnick, R., McCullough, B.D.: Regression analysis of variates observed on (0, 1):
percentages, proportions and fractions. Stat. Model. 3(3), 193–213 (2003)
15. Ospina, R., Ferrari, S.L.P.: Inflated beta distributions. Stat. Pap. 51(1), 111 (2010)
16. Ospina, R., Ferrari, S.L.P.: A general class of zero-or-one inflated beta regression models.
Comput. Stat. Data Anal. 56(6), 1609–1623 (2012)
17. Papke, L., Wooldridge, J.: Econometric methods for fractional response variables with an
application to 401(k) plan participation rates. J. Appl. Economet. 11(6), 619–632 (1996)
18. Punzo, A., Ingrassia, S.: Clustering bivariate mixed-type data via the cluster-weighted model.
Comput. Stat. 31, 989–1013 (2016)
Cluster Weighted Beta Regression: A Simulation Study 11

19. Punzo, A., Bagnato, L., Maruotti, A.: Compound unimodal distributions for insurance losses.
Insur. Math. Econ. 81, 95–107 (2018)
20. Rand, W.M.: Objective criteria for the evaluation of clustering methods. J. Am. Stat. Assoc.
66(336), 846–850 (1971)
21. Rogers, J.A., Polhamus, D., Gillespie, W.R., Ito, K., Romero, K., Qiu, R., Stephenson,
D., Gastonguay, M.R., Corrigan, B.: Combining patient-level and summary-level data for
alzheimer’s disease modeling and simulation: a beta regression meta-analysis. J. Pharma-
cokinet. Pharmacodyn. 39, 479–498 (2012)
22. Swearingen, C.J., Tilley, B.C., Adams, R.J., Rumboldt, Z., Nicholas, J.S., Bandyopadhyay, D.,
Woolson, R.F.: Application of beta regression to analyze ischemic stroke volume in NINDS
rt-PA clinical trials. Neuroepidemiology 37(2), 73–82 (2011)
Detecting Wine Adulterations Employing
Robust Mixture of Factor Analyzers

Andrea Cappozzo and Francesca Greselin

Abstract An authentic food is one that is what it claims to be. Nowadays, more
and more attention is devoted to the food market: stakeholders, throughout the
value chain, need to receive exact information about the specific product they are
commercing with. To ascertain varietal genuineness and distinguish potentially
doctored food, in this paper we propose to employ a robust mixture estimation
method. Particularly, in a wine authenticity framework with unobserved heterogene-
ity, we jointly perform genuine wine classification and contamination detection. Our
methodology models the data as arising from a mixture of Gaussian factors and
depicts the observations with the lowest contributions to the overall likelihood as
illegal samples. The advantage of using robust estimation on a real wine dataset
is shown, in comparison with many other classification approaches. Moreover, the
simulation results confirm the effectiveness of our approach in dealing with an
adulterated dataset.

Keywords Mixtures of factor analyzers · Food authenticity · Model-based


clustering · Wine adulteration · Robust estimation · Impartial trimming

1 Introduction and Motivation

The wine segment is identified as a luxury market category, with savvy as well as
non-expert customers willing to spend a premium price for a product of a specific
vintage and cultivar. Therefore, in the context of global markets, analytical methods
for wine identification are needed in order to protect wine quality and prevent its
illegal adulteration.

A. Cappozzo () · F. Greselin


Department of Statistics and Quantitative Methods, University of Milano-Bicocca, Milano, Italy
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2019 13


F. Greselin et al. (eds.), Statistical Learning of Complex Data,
Studies in Classification, Data Analysis, and Knowledge Organization,
https://doi.org/10.1007/978-3-030-21140-0_2
14 A. Cappozzo and F. Greselin

In the present work we employ an approach based on robust estimation of


mixtures of Gaussian Analyzers, for discriminating corrupted red wines samples
from their authentic variety. In a modeling context, we assume a probability distri-
bution function for the chemical and physical characteristics measured on the wines,
considering a density in the form of a mixture, whenever the dataset presents more
than a wine variety. As a consequence, the probability that a wine sample comes
from one specific grape can be estimated from the model, performing classification
through the Bayes rule. Robust estimation of the parameters in the model is adopted
to recognize the corrupted data. Particularly, we expect that adulterated observations
would be implausible under the robustly estimated model: the illegal subsample
is revealed by selecting observations with the lowest contributions to the overall
likelihood using impartial trimming, without imposing any assumption on their
underlying density.
The rest of the paper is organized as follows: in Sect. 2 the notation is introduced
and the main concepts about Gaussian Mixtures of Factor Analyzers (MFA),
trimmed MFA likelihood, and the Alternating Expectation-Conditional Maximiza-
tion (AECM) algorithm are summarized. Section 3 presents the wine dataset [7]
and classification results obtained performing a robust estimation of Gaussian
mixtures of factor analyzers. Section 4 reports a simulation study carried out
employing parameters estimated from the model in Sect. 3, in a specific framework
of contaminated dataset.
The original contribution of the present paper is given in the benchmark study on
unsupervised methods, the adaptation of the robust Bayesian Information Criterion
(BIC) introduced in [3] to MFA, and a first application of robust MFA in a somehow
realistic adulteration scenario.
An application on real data and some simulation results confirm the effectiveness
of our approach in dealing with an adulterated dataset when compared to analogous
methods, such as partition around medoids and non-robust mixtures of Gaussian
and mixtures of patterned Gaussian factors.

2 Mixtures of Gaussian Factors Analyzers

In this section we briefly recall the definition and some features of the mixture of
Gaussian Factor Analyzers (MFA) and its parameter estimation procedure. MFA
is a powerful tool for modeling unobserved heterogeneity in a population, as it
concurrently performs clustering and local dimensionality reduction, within each
cluster. Let X1 , . . . , Xn be a random sample of size n on a p-dimensional random
vector. An MFA assumes that each observation Xi is given by

Xi = μg + Λg Uig + eig (1)


Detecting Wine Adulterations Employing Robust Mixture of Factor Analyzers 15

with probability πg for g = 1, . . . , G. The total number of components in the


mixture is denoted by G, μg are p × 1 mean vectors, Λg are the p × d matrices
iid iid
of factor loadings, Uig ∼ N (0, Id ) are the factors, eig ∼ N (0, Ψ g ) are the
errors, and Ψ g are p × p diagonal matrices. Note that d < p, that is the p
observable features are supposed to be jointly explained by a smaller number of
d unobservable factors. Further, Uig and eig are independent, for i = 1, . . . , n
and g = 1, . . . G. Unconditionally, therefore, Xi has a density in the form of a
G-components multivariate normal mixture:


G
fXi (xi ; θ) = πg φp (xi ; μg , Σ g ) (2)
g=1

where φp (·; μg , Σ g ) denotes the p-multivariate normal density, whose covariance


matrix Σ g has the following decomposition Σ g = Λg Λg + Ψ g .
When estimating MFA through the usual Maximum Likelihood approach, two
issues arise. Firstly, departure from normality in the data may cause biased or
misleading inference. Some initial attempts in the literature to overcome this
issue propose to consider mixtures of t-factor analyzers [15], but the breakdown
properties of the estimators are not improved [10]. The second concern is related
to the unboundedness of the log-likelihood function [4], which leads to esti-
mation issues, like the appearance of non-interesting spurious maximizers and
degenerate solutions. To cope with this second issue, Common/Isotropic noise
matrices/patterned covariances [1] and a mild constrained estimation [9] have
been considered. The methodology considered here employs model estimation,
complemented with trimming and constrained estimation, to provide robustness,
to exclude singularities, and to reduce spurious solutions, along the lines of [8].
Therefore, with this approach, we overcome both previously mentioned issues.
A mixture of Gaussian factor components is fitted to a given dataset
x1 , x2 , . . . , xn in Rp by maximizing a trimmed mixture log-likelihood [18],
⎡ ⎤

n G
Lt rim = ζ (xi ) log ⎣ φp (xi ; μg , Λg , Ψ g )πg ⎦ (3)
i=1 g=1

where ζ (·) is a 0–1 trimming indicator function that tells us whether observation
xi is trimmed off or not. If ζ (xi )=0 xi is trimmed off, otherwise ζ(xi )=1. A
fixed
n fraction α of observations, the trimming level, is unassigned by setting
i=1 ζ (xi ) = n(1 − α), where the less plausible observations under the currently
estimated model are tentatively trimmed out at each step of the iterations that lead to
the final estimate. In the specific application to wine authenticity analysis described
in Sect. 3, they are supposed to be originated by wine adulteration.
Then, a constrained maximization of (3) is adopted, by imposing ψg,ll ≤
c ψh,mm for 1 ≤ l = m ≤ p and 1 ≤ g = h ≤ G, where {ψg,ll }l=1,...,p are
16 A. Cappozzo and F. Greselin

the diagonal element of the noise matrices Ψ g , and 1 ≤ c < +∞, to avoid the
|Σg | → 0 case. This constraint can be seen as an adaptation to MFA of those
introduced in [11]. The Maximum Likelihood estimator of Ψ g under the given
constraints leads to a well-defined maximization problem.
The Alternating Expectation-Conditional Maximization—an extension of the
Expectation-Maximization algorithm—is considered, in view of the factor structure
of the model. The M-step is replaced by some computationally simpler conditional
maximization (CM) steps, along with different specifications of missing data. The
idea is to partition the vector of parameters θ = (θ 1 , θ 2 ) , in such a way that Lt rim
is easy to be maximized for θ 1 given θ 2 and vice versa. Therefore, two cycles are
performed at each algorithm iteration:
1st cycle : we set θ 1 = {πg , μg , g = 1, . . . , G}; here, the missing data are the
unobserved group labels Z = (z1 , . . . , zn ). After applying a step of Trimming, by
assigning to the observations with lowest likelihood a null value of the “posterior
probabilities”, we get one E-step, and one CM-step for obtaining parameters in θ 1 .
2nd cycle : we set θ 2 = {Λg , Ψ g , g = 1, . . . , G}, here the missing data are
the group labels Z and the unobserved latent factors U11 , . . . , UnG . We perform a
Trimming step, then a E-step, and a constrained CM-step, i.e., a conditional exact
constrained maximization of Λg , Ψ g .
A detailed description of the algorithm is given in [8].

3 Wine Recognition Data

The wine recognition dataset, firstly analysed in [7], reports results of a chemical
and physical analysis for three different wine types, grown in the same region in
Italy. Originally, 28 attributes were recorded for 178 wine samples derived from
three different cultivars: Barolo, Grignolino, and Barbera. A reduced version of the
original dataset with only thirteen variables is publicly available in the University
of California, Irvine Machine Learning data repository, commonly used in testing
the performance of newly introduced supervised and unsupervised classifiers.
Particularly, in the unsupervised classification literature the wine recognition data
has been considered to assess cluster analysis in information-theoretic terms via
minimisation of the partition entropy [19], to prove the modelling capabilities of a
generalized Dirichlet mixture [2], to evaluate the efficacy of employing distances
based on non-Euclidean norms [5] and of Random Forest dissimilarity [20]. More
recently, also parsimonious Gaussian mixture models have been applied to the
Italian wines dataset [16].
Here our purpose is twofold: we want to explore the classification performance
of a robust estimation based on mixtures of Gaussian Factors Analyzers, and we aim
at obtaining realistic parameters for the subsequent simulation study. The dataset,
available in the pgmm R package [17], contains 27 of the 28 original variables, since
the sulphur measurements were not available. Initially, to perform model selection
and detect the most suitable values of factors d and groups G, an adaptation to the
Detecting Wine Adulterations Employing Robust Mixture of Factor Analyzers 17

Table 1 RobustBI C [3] for G


different choices of the
d 1 2 3
number of factors d and the
number of groups G for the 1 9082.58 8282.92 8223.46
robust MFA model on wine 2 8560.62 8107.62 8112.90
data, trimming level α = 0.05 3 8352.26 8042.02 8199.38
and c = 20. The smallest 4 8160.77 7969.64 8315.23
value is obtained with d = 4
and G = 2 5 8102.77 8044.03 8456.00
6 8097.06 8165.67 8735.63

Table 2 Classification table 1 2 3


for the robust MFA with
number of factors d = 4, Barolo 59 0 0
number of groups G = 3, Grignolino 0 71 0
trimming level α = 0.05 and Barbera 0 0 48
c = 20 on the wine data Trimmed observations are
classified a posteriori according
to the Bayes rule

MFA framework of the robust Bayesian Information Criterion, firstly introduced


in [3], has been considered. That is, BI C = −2Lt rim(x; θ̂) + v c log n∗ where
v c = (G − 1 + Gp + G(pd − d(d − 1)/2) + (Gp − 1)(1 − 1/c) + 1) denotes the
number of free parameters in the model (depending on the value of the constraint
c) and n∗ = n(1 − α) the number of non-trimmed observations. Robust BIC
for different choices of the number of factors d and the number of groups G are
reported in Table 1, considering a trimming level α = 0.05 and c = 20. The value
of the robust BIC is minimized for d = 4 and G = 2, suggesting a mixture with just
two components. Careful investigation on this result highlighted that robust MFA
methodology tended to cluster together Barolo and Grignolino samples as arising
from the same mixture component, while clearly separating Barbera observations. It
is worth recalling [7] that the wines in this study were collected over the time period
of 1970–1979, and the Barbera wines are predominantly from a later period than the
Barolo or Grignolino wines. Therefore, considering the nature of the phenomena
under study and the risks related to rigidly selecting the number of components in a
mixture model only on the basis of the results provided by an information criteria,
such as BIC [13], we decided to employ a robust MFA with d = 4, G = 3, and α =
0.05, leading to the classification matrix reported in Table 2. Employing a robust
MFA rather than a Gaussian mixture leads to a 60% reduction in the number of
parameters to be estimated (470 against 1217). Notice, in addition, that after robust
estimation, also the trimmed observations can be a posteriori classified according to
the Bayes rule, i.e., assigning each of them to the component g having greater value

of Dg (x, θ ) = φp (x; μg , Λg Λg + Ψ g )πg .
Results in Table 2 show that the robust MFA algorithm led to a perfect
clusterization of the samples according to their true wine type.
For completeness, the robust MFA algorithm was also applied to the more
common thirteen variable subset of the wine data and comparison with the existing
literature is reported in Table 3. The clustering performance with respect to the true
Discovering Diverse Content Through
Random Scribd Documents
Fig 26 COLT AUTOMATIC POCKET PISTOL
Seven shots; 2 inch barrel; weight, 13 oz.: 25 cal.

Fig. 27—SAVAGE AUTOMATIC POCKET PISTOL


Ten shots; 4¼ inch barrel; weight, 1 lb. 5 oz.; .32 and .38 cal.

Fig. 28.—SMITH & WESSON AUTOMATIC POCKET PISTOL


Eight shots; 3½ inch barrel; weight, 1 lb., 7¾ oz.; .35 cal.

The accuracy of the cartridge depends largely upon the uniformity exercised in the operations of
loading, the fit of the bullet, its shape, and the reliability and uniformity of the powder. The primer must
be of uniform strength also, especially in reduced charges. In ammunition for military service the shells
are crimped on the bullets to hold them in place. This does not increase the accuracy in black powder
ammunition, but it is necessary and advantageous in all smokeless ammunition including gallery
charges, in order to confine the powder and produce uniform results.
The following is a digest of the principal pistol and revolver cartridges in use at the present time.
Rim-fire Cartridges.—These are primed with a fulminate of mercury mixture around the outer edge of
the rim, or base of the shell, and are generally loaded with Lesmok, semi-smokeless, or black powder.
Figure 29.

The smallest and lightest charged ammunition in general use is the .22 caliber. In this caliber the “C. B.”
or Conical Ball Cap loaded with black powder is the smallest practicable cartridge. The charge is 1½
grains of powder and a lubricated conical bullet weighing 29 grains.

Figure 30.

An excellent cartridge in this caliber is the .22 short, (Fig. 30). This cartridge fouls very little and is
almost equal in accuracy to the .22 “long rifle” up to 50 yards. On account of its lighter report it is
preferred by many for gallery shooting.

Figure 31.

The .22 caliber “long rifle” cartridge is more extensively used for pistol shooting than any other. It is the
most accurate of the .22-caliber cartridges, being well proportioned, the bullet well lubricated, and the
shell uncrimped. In addition to this, the ammunition is inexpensive and has very clean shooting
qualities. It is, therefore, particularly well adapted for pistol shooting. This cartridge, fired from a 10-
inch barrel, will shoot regularly inside of a 2-inch circle, at 50 yards, and inside a 5-inch circle at 100
yards.
The .22-caliber Long Rifle “Armory” and the .22-caliber Smith & Wesson Long are special makes of the
long rifle cartridge that are furnished with a crimped shell, preventing the bullet from becoming
dislodged and thus adapting this popular cartridge for use in revolvers of this caliber.
In all of the foregoing cartridges only the surface of the bullet outside the shell is lubricated. Exposed in
this way, the lubricant is easily rubbed off, or melted if allowed to stand in the sunlight on a warm day.
Great care should be taken to prevent this, as, without lubrication, the bullets will lead the barrel and
cause inaccurate shooting.

Figure 32.
The .22-caliber Winchester is a cartridge with inside lubrication. It is more powerful than the .22 long
rifle, and gives good results in the pistol. The bullet has a flat point, making it suitable for game
shooting, and the lubrication being within the shell, these cartridges may be carried loose in the pocket.
All of the .22 caliber cartridges can be had with hollow-pointed bullets, which are to be preferred for
game shooting. They are also furnished loaded with smokeless powder. When this powder was first
used in .22-caliber ammunition the results were far from satisfactory, but as now manufactured the
smokeless ammunition approximates very closely in uniformity and accuracy to that loaded with black
powder.
There still remains, however, considerable difficulty with the rim-fire smokeless cartridges on account of
their liability to rust the inside of the barrel.[7] The novice is therefore cautioned not to use this
ammunition until the difficulty of rusting is overcome.
The .25-cal. Stevens is a much more powerful cartridge than any of the preceding, and gives excellent
results in the pistol. It is selected by those who wish a more powerful rim-fire cartridge than is furnished
in .22 caliber.

Figure 33.

Rim-fire cartridges in larger caliber than .25 are used for derringers (large-bore, single-shot pocket
pistols now seldom used) and inferior grades of revolvers. These cartridges sometimes lack uniformity in
caliber when made by different manufacturers, are frequently defective, and discharge occasionally in
closing the action of the arm in which they are loaded. They consequently lack the safety, reliability, and
accuracy of the corresponding calibers in central-fire ammunition. Rim-fire cartridges cannot be
reloaded.
Central-fire Cartridges.—This type of cartridge has a brass or copper primer fitted with a skeleton anvil
of brass and charged with a small quantity of priming composition containing a sensitive explosive for
igniting the powder charge. The primer fits water-tight in a socket in the center of the base of the shell.
After being discharged, the primer can be renewed and the shell reloaded.
In all the central-fire cartridges the lubrication of the bullet is inside of the shell, rendering the
ammunition much more serviceable and less liable to be damaged.
Mantled bullets designated as “metal pointed” and “full metal patched” can be supplied by the
ammunition manufacturers for all the central-fire cartridges at a cost of one dollar per thousand more
than the regular lead bullets. The mantled bullets do not deform as readily in handling, shipping, etc.,
and give slightly increased penetration in soft woods, animal tissue, etc., as compared with the plain
lead bullet with the same powder charge.

Figure 34. Figure 35.


The .32-caliber S. & W. cartridge is adapted to the Smith & Wesson, Colt, or other pocket revolvers.
Occasionally single-shot pistols are chambered for this cartridge. It is fairly accurate at ranges up to 50
yds. A gallery charge is furnished in this shell consisting of 4 grains of black powder and a spherical or
“round” bullet weighing 47 grains.
The .32-cal. S. & W. Long is more accurate and powerful than the preceding cartridge. It gives excellent
results in both the pistol and revolver. The gallery charge is the same as that of the .32 S. & W.
The .32-caliber Colt New Police is also an accurate cartridge, and was designed specially for the Colt
New Police revolver. The flat point adds to its effectiveness. A good gallery charge in this shell consists
of a powder charge of 1½ grains of Bullseye and the regular bullet.

Figure 36.

The .32-44 S. & W. and the .38-44 S. & W. were special black powder cartridges designed for the S. &
W. Russian Model revolver bored for these calibers. The shells were uncrimped and the bullets seated
inside of the shells flush with the mouth. A large variety of special bullets of varying weights were
designed for these cartridges and much experimentation was done with them. The .38-44 Caliber was
originally designed for and largely used by Chevalier Ira A. Paine, the noted pistol shot in his exhibitions.
While these cartridges proved very accurate and were popular when black powder was in general use
they are entirely unsuited for smokeless powders and consequently are now seldom used.
The .38 S. & W. is adapted to the Smith & Wesson, Colt, and other pocket revolvers. It is much more
powerful than the .32 S. & W., and is consequently more practical and better adapted for a pocket
revolver charge. When shot from a 4-inch barrel, groups of ten shots can be made in a 2-inch circle at
20 yards and in a 6-inch circle at 50 yards.
A good gallery or reduced load in this shell is Ideal Bullet No. 358242, 36072, or 360302 with 2 grains
of Bullseye powder.

Figure 37. Figure 38.

The .38 Colt New Police is almost identical with the .38 S. & W., the only difference being a slightly
heavier bullet with a flat point.

Figure 39.
The .38 Long Colt is adapted to the Colt and S. & W. Military revolvers. It was the regulation charge of
the service weapon of the U. S. Army until 1911. Under service conditions the cartridge was found to
have insufficient power, was inaccurate and on account of the deterioration of the powder with which
some of the ammunition was loaded it proved most unsatisfactory, especially in the Philippine war.

Figure 40.

The .38 Smith & Wesson Special cartridge is more powerful than the .38 Long Colt and is exceedingly
accurate. From a 6-inch barrel six shots may be placed within a 5-inch circle at 100 yards. Numerous
gallery and mid-range charges with special bullets have been designed for this cartridge. It is now the
most popular of all the revolver cartridges for target practice. Some of the special bullets are illustrated
herewith, the numbers being those used in the “Ideal Handbook”:

358242 360345 36072

125 gr. 115 gr. 110 gr.

360302 360271 360363

112 gr. 150 gr. 70 gr.

Figure 41.

A powder charge of 2¼ to 2½ grains of Bullseye will give good results with any of these bullets. Bullets
No. 360345, 360302 and 360271 cut full-size bullet holes in the targets.
The following are some of the special charges supplied by the manufacturers in this shell:

Wt. in
Weight in Powder Type
Name Manufacturer grains
Grains Charge Bullet
Bullets
Gallery U. M. C. Co. 5.2 Black 70 Spherical
Target U. M. C. Co. 2.6 Bullseye 130 R. N.
Colt Special U. M. C. Co. 3.4 Bullseye 160 F. N.
Sharp Shoulder U. M. C. Co. 2.1 Bullseye 122 F. Head
Mid Range Winchester 2.0 Bullseye 104 R. N.
Gallery Winchester 8.5 C. P. W. 70 R. N.

The .44-caliber Smith & Wesson Russian[8] was the most popular revolver cartridge for target shooting
before smokeless powder was introduced. Since smokeless charges have been adapted to it many
expert shots prefer this cartridge in the gallery contests as the large bullet hole is a decided advantage
over the smaller calibers at ranges of 20 yards and under. Nearly all the great records in revolver
shooting in the past have been made with this cartridge and many important matches have been won
with it. A great deal of experimental work has also been done with it, and many reduced charges have
been evolved. The Ideal Manufacturing Company can furnish moulds for bullets of the shapes and
weights shown in Fig. 43.

Figure 42.

429336 429251 U.M.C. U.M.C. 429106

255 gr. 256 gr. 110 gr. 130 gr. 175 gr.

429348 429106 429239 429215 429220

176 gr. 160 gr. 125 gr. 205 gr. 175 gr.

Figure 43.

Bullets No. 429336, 429348, and 429220 cut clean, full-size holes in the target. The weight of the
powder charge and bullets in grains and the accuracy of the various loads fired from 6½-inch barrel are
about as follows:

Bullseye
Bullet Diameter of Circle Enclosing Group of 10 Shots
Powder
20 yds. 30 yds. 50 yds. 100 yds. 200 yd.
4.1 256 1 in. 1½ in. 1½ in. 6 in. 15 in.
2.3 110 1 in. 2 in.
2.7 130 1¼ in. 2½ in.
3.0 160 2 in. 3 in.
2.8 176 1½ in.
3.0 175 2½ in.
2.7 125 1¼ in. 2½ in.
3.2 205 2 in. 3 in. 7 in.
2.6 175 1¼ in. 3¼ in.

These various loads adapt this shell to almost any conceivable requirement in revolver shooting.

Figure 44.

The .44 S. & W. Special is the latest and most powerful of the .44-caliber cartridges. It equals the .44 S.
& W. Russian in accuracy and is the best proportioned of the heavy revolver cartridges. The reduced
and gallery loads of the .44 S. & W. Russian will give equally good results in this shell.

Figure 45.

The .45 Colt Army is the most powerful of all the revolver cartridges. It was formerly the United States
army service ammunition. The charge was so heavy, and the recoil so excessive that it was almost
impossible to shoot it without flinching. The smokeless powder charge of 5 grains of Bullseye makes it
much more practical and very similar to the .44 S. & W. Special cartridge. Both of these are exceedingly
powerful and accurate and suitable for military service.

Figure 46. Figure 47.

The caliber of the service ammunition for the revolver of the British army is .455. This is a very accurate
cartridge, but not as powerful as the corresponding military cartridges used in this country. A special
cylindrical bullet with a deep convex hollow point is furnished in the same shell and is known as the
“man stopper.”
This form of bullet is used in the English .450 and .38 caliber cartridges also.
The .450 Welby is another English cartridge that is accurate, and pleasant to shoot. It is used largely at
Bisley in the annual revolver competitions of the National Rifle Association of Great Britain.
In order to avoid excessive fouling with black powders a self-lubricating bullet has been invented and
introduced by Smith & Wesson, which can be furnished in all calibers above .32. The bullet has a hollow
core open in the rear. Lubricant is filled into the core, after which it is closed with a lead plunger. Four
small ducts communicate from the forward end of the core to the exterior of the bullet just ahead of its
bearing on the barrel. At the moment of discharge the plunger is driven forward, forcing the lubricant
through the ducts into the barrel ahead of the bullet.
This bullet has given excellent results and will be found decidedly advantageous when black powder is
used. With it a hundred or more shots may be fired with black powder without causing sufficient fouling
to impair the accuracy.
Revolvers are sometimes chambered for the .44-40-200, the .38-40-180, and the .32-20-115 rifle
cartridges. These charges in black powder load are not as accurate as the corresponding revolver
cartridges in these calibers, but can be relied on to shoot inside a 5-inch circle at 50 yards. These
cartridges are desirable for revolvers only when it is an advantage to use the same ammunition in the
rifle and revolver, or in certain localities where only a few varieties of ammunition are to be had. The
large powder charge makes the recoil of the first two cartridges named rather unpleasant. The .32-20-
115 is the most accurate of these cartridges, and gives the best results in the pistol or revolver. All these
cartridges having flat-pointed bullets are well adapted for game shooting. None of these rifle cartridges
loaded with smokeless powder will give good results in revolvers because the brand of powder generally
used in rifle ammunition requires a long barrel to consume the charge. Fired from a short barrel only
part of the charge will be consumed and the rest will be expelled unburned, thus reducing the velocity
and power of the charge and sometimes increasing the recoil. It is of course entirely practicable to
adapt a charge of bullseye or similar smokeless powder to these shells which would make them much
more satisfactory. Another disadvantage of using the rifle cartridge in revolvers is the possibility of
inexperienced persons using the new high velocity rifle ammunition, which would prove not only most
unsatisfactory but extremely dangerous in revolvers. There are no reduced or gallery loads supplied in
these shells.

Figure 48. Figure 49.

Figure 50. Figure 51.

Figure 52. Figure 53.


Automatic Pistol Cartridges.—With the introduction of the magazine pistol special smokeless cartridges
have been devised that are rimless and have a crease around the base of the shell by which they may
be held and manipulated by the mechanism. These cartridges are exceedingly clean-shooting. Several
hundred rounds may be fired without causing more fouling than is apparent after the first few shots.
This ammunition is furnished loaded with “full-mantled” and “soft-nosed” bullets; the latter, having the
lead exposed at the point will mushroom on striking animal tissue and are sometimes referred to as
“dum dum” bullets and are intended for hunting purposes.

Figure 54. Figure 55.

Figure 57. Figure 56.

The mantled or metal cased bullet has undoubted advantages in rifle ammunition, in which low
trajectory and extreme long range are desiderata that can be obtained only by high velocities. In
ammunition for magazine pistols and revolvers, however, the prime object is to deliver the most
effective blow possible at comparatively short range.
The velocities attainable in large calibers within the permissible weight of an automatic pistol are
comparatively low. The deformation of any bullet on striking animal tissue is in direct proportion to its
velocity. It is, therefore, extremely doubtful that a metal cased bullet will ever prove as effective and
satisfactory in “stopping power” and for military service, either in the automatic pistol or the revolver, as
the large caliber lead bullet.
The .25 cal. is the smallest of the American made automatic pistol ammunition and is adapted for the
Colt and other magazine pistols. It is an accurate cartridge but the short length of barrel of the Colt
weapon makes it impossible to do accurate work with it.
Figures 49 and 50 are the well known Luger and Mauser cartridges adapted to the pistols of that name.
They are powerful charges, accurate and clean shooting. These were among the first cartridges
developed for automatic pistols and are still extensively used.
The .32 Automatic Colt is adapted to Colt and other magazine pistols of this caliber. It is an accurate,
pleasant shooting cartridge with very little recoil and excellent work can be done with it at the target.
The .35 S. & W. Automatic is adapted to the Smith & Wesson magazine pistol. It is a very accurate
cartridge, has no unpleasant recoil and like the preceding is well adapted for target shooting.
Figure 53 is the .380 Automatic Colt cartridge designed to meet the demand for a light charge in this
caliber. It is adapted to the Colt and other magazine pistols.
The .38 Automatic Colt is the best proportioned and most powerful of all automatic pistol cartridges. It
has a slightly flattened point and is extremely accurate. When fired from regulation arms this
ammunition is capable of placing ten shots inside a 3-inch circle at 50 yards and inside a 7-inch circle at
100 yards.
This was the ammunition of the first Colt automatic pistol introduced in the United States.
Figures 55 and 56 are two cartridges adapted to the .45 Automatic Colt pistol. Figure 56 is the new
service charge of the U. S. Army. They are exactly alike except that the service charge has a 230 gr.
bullet (30 gr. heavier than the other). The service charge when fired from the regulation service arm is
capable of placing 10 shots in a 3½-inch circle at 50 yards and an 8-inch circle at 100 yards.
A flat or blunt pointed bullet of about 185 gr. and a ten per cent. heavier powder charge would improve
the effectiveness and stopping power of this cartridge wonderfully without materially affecting the recoil
or the accuracy.
Figure 57 is the service charge of the regulation magazine pistol (Webley & Scott) of the British Army. It
is an accurate cartridge but it lacks sufficient power to fulfill the exacting requirements of present-day
military service.
Light or gallery charges in magazine pistol shells are impracticable on account of not having sufficient
recoil to operate the automatic mechanism. Slightly reduced loads with lead bullets may be used in
some of the arms but seldom with satisfactory results. Reduced loads can be used in most of the
weapons if the mechanisms are hand operated for each shot.
The following ballistical table gives the charges, muzzle velocities, etc., of the principal factory-loaded,
smokeless pistol and revolver cartridges. The factories aim to keep the muzzle velocities uniform for
each cartridge. To produce this result with the various brands of smokeless powder, all of which differ
more or less in strength, the weight of the powder charge necessarily varies for the different brands of
powder. Even when purchased in large quantities, different blends and packages of the same brand of
powder occasionally vary somewhat in strength. For these reasons it is impossible to designate the
exact weight or volume of any brand of powder which will in all cases produce the muzzle velocities in
the table, and the charges given must therefore be considered as approximate only.

BALLISTICAL TABLE

BULLET Muzzle
Weight in Grains Length Energy
Exact velocity Penetr
Name of and Brand Weight Round of bbl. —foot
diameter (50 ft. (inche
Cartridge of Powder in or flat in arm lbs.:
in from white
(Approximate only) grains nosed tested Wv2÷2g
inches muzzle)
Rim
Fire:
.22 Short 1.6 Lesmok .223 30 RN 6 789 41.5 3
.22 Long 2.1 " .223 35 RN 6 770 46.2 3½
.22 L. Rifle 3.4 " .223 40 RN 6 765 51.8 4
W. R.
.22 3.5 " .2275 45 FN 6 811 65.8 4
F.
Central
Fire:
Auto
.25 1.1 Bullseye .251 50 RN 2 733 59.7 3
Colt
m.m.
7.63 5.5 " .3105 86 RN 5½ 1397 373 11
Mauser
m.m.
7.65 4.1 " .3095 93 RN 4⅝ 1173.5 284.3 10
Luger
m.m.
9 4.6 " .3555 125 FN 4 1039.2 299.8 10
Luger
W. C.
.32 10.0 Sharpshooter .3125 115 FN 5½ 954 232.4 5
F.
Auto
.32 2.5 Bullseye .3125 74 RN 3¾ 938 144.8 5
Colt
S. &
.32 1.5 " .315 85 RN 4 606.7 69.5 3
W.
S. &
.32 1.5 " .315 85 RN 10 902 159 4
W.
Lg.
.32 2.0 " .313 90 RN 4 641.4 82.2 3½
Colt
Sht.
.32 1.4 " .315 80 RN 4 657.2 78.7 3½
Colt
S. &
.32 W. 2.0 " .315 98 RN 4 706.9 108.6 4
Long
Colt N.
.32 2.5 " .314 98 FN 4 706.3 108.6 4
P.
S. &
.35 W. 1.9 " .3195 76 RN 3½ 809 110.5 4
Auto
S. &
.38 2.4 " .359 145 RN 5 579.3 108.2 4½
W.
Auto
.38 4.6 " .359 130 RN 6 1175 398.0 10
Colt
Colt N.
.38 2.4 " .359 150 FN 4 579.6 111.7 4
P.
Sht.
.38 2.5 " .375 130 RN 6 608 107 4
Colt
Long
.38 3.0 " .358 148 RN 6 786 203 4½
Colt
Long Gray
.38 3.4 .358 150 RN 6 771.6 198.3 6
D. A. Walsrode
S. &
.38 3.4 Bullseye .358 158 RN 6 856.7 257.5 7
W. Spl.
Colt
.38 3.4 " .358 158 FN 6 857.6 258 7
Spl.
S. &
.38 W. Spl. 8.5 C.P.W.[9] .358 70 RN 6 1300 263 5
Gal.
S. &
W. Spl.
.38 2.1 Bullseye .358 123 RN 6 655 99 3
Mid
Range
W. C.
.38 15. Sharpshooter .400 180 FN 5 983 386.5 6
F.
Auto
.380 2.6 Bullseye .357 95 RN 3¾ 887 166 5½
Colt
Sht.
.41 2.5 " .406 160 RN 6 707 177 4
Colt
Long
.41 3.3 " .387 200 RN 6 705.6 221.2 5
Colt
S. &
.44 W. 4.1 " .431 246 RN 6 706 272 6½
Russ.
S. &
W.
.44 2.5 " .431 115 RN 6 685 118 3
Russ.
Gall.
S. &
.44 5.1 " .431 246 RN 5 755 311.5 7
W. Spl.
W. C.
.44 16.5 Sharpshooter .426 200 FN 7½ 918.8 375 6
F.
Auto
.45 4.7 Bullseye .4505 200 RN 5 910.2 368 8
Colt
Auto
.45 Colt 4.7 " .4505 230 RN 5 809 335 6
(Govt.)
Colt D.
.45 5.0 " .455 255 RN 5 770.6 336.3 5
A.
.455 Colt 4.5 " .458 265 RN 5 756.6 336.5 5
Webley
.455 7.0 Cordite .455 220 RN 5 750 280.6 4½
Auto
British
.455 5.5 " .455 265 RN 7½ 700 288 4¼
Service
CHAPTER IV

SIGHTS

T HE purpose of sights is to assist in aiming the piece. The national organizations allow only “open”
sights in pistol and revolver shooting. “Peep” or “aperture” sights are barred. The rear sight usually
consists of a notch shaped like a V or a U, the notch being as wide on top as at any part. The front
sight is a piece of thin metal set on edge. Sometimes the latter has a special shape or section
resembling a pinhead when looking at it from the breech, as in aiming.

Side View. End View. Side View. End View.


Front Sights.

Rear Sight. Appearance when aiming.


Military Sights.

Figure 58.

Military sights usually consist of a plain groove in the top of the frame for the rear sight and a tapering
front sight fixed to the barrel near the muzzle.

Appearance
Rear Sight. Side View. End View.
when aiming.
Front Sight.

“Paine” Sights.
Figure 59.

Target sights are made in endless variety to suit individual ideas. The sights most generally used for
target shooting are the “Paine” sights, named after Chevalier Ira A. Paine, who invented and was the
first to use them. The rear sight is a flat bar with a semi-circular notch, and the front sight is a “bead”
sight; that is, a sight that resembles a pinhead when aiming.

Appearance
Rear Sight. Side View. End View.
when aiming.
Front Sights.

Patridge Sights.
Figure 60.

Another sight that many of the best shots are using is the “Patridge” sight, developed by Mr. E. E.
Patridge of Boston, Mass. The rear sight has a wide rectangular notch; the front sight is plain, with a
square top, as shown.
Fig. 61 represents the “Lyman” sights as adapted to Smith & Wesson revolvers. The distinctive features
of these sights are the ivory bead of the front sight and the horizontal ivory line in the rear sight. These
sights are well adapted for hunting and shooting at objects with a dark background.
These sights have been referred to in the order in which they are most used. It is generally necessary
for individuals to try various sights before they are able to select intelligently. In target arms different-
shaped sights may be used in the same base or fitting, so that it is a comparatively easy matter to try
any or all of these sights on the same arm.
The notch of the rear sight should have a bevelled edge concave toward the front. This will secure
sharpness of outline in any light. The front sight should also be distinct and is found to be more
satisfactory when the side toward the eye is a surface at right angles to the line of sight.

Fig. 61.—Lyman Sights


Fig. 62.—THE WESPI SEARCHLIGHT SIGHT
A-Battery; B-Mercury switch; C-Electric bulb; D-E-Lenses.

Fig. 63.—The Wespi Searchlight Mounted on a Pocket Revolver.

For years means have been sought to make successful shooting at night possible. White and
phosphorescent paints have been applied to the sights and to the top of the barrel but all such methods
have proved more or less unsatisfactory even in dim light and in total darkness the target or other
object cannot be seen. A recently invented device that overcomes all these difficulties and makes
shooting at night practicable is the “Wespi” searchlight sight.[10]
This sight is a tube about 6 inches long and ¾ inches in diameter containing a miniature electric
searchlight which projects a dark spot in the center of the illuminated field. When properly mounted on
the piece the black spot indicates where the bullet will strike. This sight can be readily attached to any
pistol or revolver. As offered on the market at the present time it is adapted for short range work up to,
say, 60 feet. The illustrations show a section through the sight tube, and the sight attached to a
revolver. The weight is six ounces. (See 61 and 62 facing p. 64.)
This sight embodies the principles of the telescopic sight and can undoubtedly be modified to increase
its illuminating power and adapted so as to project well-defined dark lines similar to cross wires, on a
target; or the dark spot decreased in size to about 3 or 4 inches in diameter at 60 feet. So modified this
would be a practical sight for target shooting and would be a boon to many of the older marksmen
whose sight is failing and who find it more and more difficult to shoot in artificial light with the ordinary
sights.
Such a sight would also possess many advantages for beginners as the moving spot on the target would
indicate the unsteadiness of the holding and impress upon the marksman the importance of holding the
spot in the right position at the instant of discharge. A further improvement would be to substitute for
the dark spot, a spot of intensely bright light. This would be equally as effective as the dark spot and
would greatly increase the range at which the sight could be used, adapting it to game shooting at
night. It is hoped that the manufacturers will develop a sight as suggested for target and game
shooting.
CHAPTER V

SHOOTING POSITION

T HE position in pistol and revolver shooting is very important. In firing a long series of shots, a man
with an easy, natural position will suffer much less fatigue, and will have a decided advantage over
another whose position is straining and uncomfortable. Formerly the approved position was to stand
with the right side toward the target. This required the head to be turned ninety degrees from its
natural position, and was very uncomfortable. Undoubtedly this position is a relic of duelling days, when
it might have been argued that a smaller mark was offered to the antagonist.
The positions adopted by the leading shots vary considerably. Most of them face a trifle to the left of the
target, with the right foot 6 or 8 inches ahead of the left, and pointing directly toward the target, the
weight of the body supported equally by both legs and perfectly balanced. Others shoot with the feet
close together; some with one or both eyes open, and with the arm partly or fully extended. The
question of position depends largely upon the physique and comfort of the individual.
Mr. Winans’ position is an exceedingly strong one. His poise is very good, and he stands firmly on both
feet. The left arm falls straight down along the left side of the body. This affords rigidity when desired,
and imparts action to the figure.
Mr. Axtell has a stanch, natural position. Like Mr. Winans, he shoots with the right arm fully extended,
and he holds the weapon in the correct and most approved manner.
The position of Mr. Anderton is excellent. He enjoys perfect health, and has his large muscular
development well under control. His position is strong, natural, and comfortable.
Mr. Dietz’s position is entirely different from any of those preceding. It is tenseless and flexible
permitting him to shoot long series of shots without fatigue.
The positions of Mr. Patridge and Sergeant Petty are characteristic and typical of persons of entirely
different physique.
Mr. Gorman and Dr. Sayre are men of similar physique. Their positions which resemble each other
closely are firm, easy and natural.
Fig. 64.—Walter Winans Fig. 65.—C. S. Axtell Fig. 66.—Thomas Anderton

Mr. Lane’s position is natural and interestingly unconventional. He has perfect poise and shoots without
apparent fatigue.
Dr. Snook has a well poised and deliberate position. He shoots with his arm not fully extended and with
the feet close together.
The positions of Mr. Armstrong and Mr. Dolfen are very similar. They are men of entirely different stature
but almost identical in physique. Their positions are firm and business-like. Both shoot with the arm fully
extended.
Note.—For photographs showing shooting positions of individuals referred to in this
chapter, see illustrations facing pp. 68, 72, 76, and 78.
CHAPTER VI

TARGET-SHOOTING

I N the development of firearms and ammunition, target-shooting has always occupied an important
place. It is regularly and systematically practised in the army and navy, in order to maintain and
improve the proficiency of the men as marksmen. Target-shooting, with many different styles of
firearms, under prescribed rules and regulations, has also become extremely popular with civilians.
Target-shooting was indulged in extensively with the rifle before it became popular with the pistol and
revolver. The shorter barrel, and the greater difficulty in acquiring skill with the latter weapons, were
doubtless responsible for the mistaken idea, long prevalent, that these arms were extremely inaccurate.
When, however, a few individuals developed sufficient skill to obtain fine shooting, their performances
were considered phenomenal. Among the first to obtain a high order of skill with the muzzle-loading
pistol in the United States was Captain John Travers of Missouri. He was well known as an expert pistol
shot as early as 1860. In that year Captain Travers shot an interesting individual match in St. Louis at a
distance of 100 feet. Fifteen china plates, nine inches in diameter, were used as targets. Captain Travers
broke 11 out of 15, while his opponent broke but 9.
In 1865 Colonel William F. Cody, (Buffalo Bill) and Captain William P. Schaaf of St. Louis became
prominent as pistol shots. The latter subsequently joined Captain Travers in a three years’ tour of the
United States, giving exhibitions in nearly all the large cities.
About 1880 Ira Anson Paine, a native of Massachusetts, attracted attention by his fine marksmanship
with the pistol. In 1881 he went abroad, and for a number of years he traveled over the principal
countries of Europe, giving public exhibitions of his skill with the pistol and revolver. While in Portugal in
1882 he was knighted by the King in the presence of a notable assemblage, and made a chevalier of an
ancient military order. In his exhibitions Chevalier Paine used a Stevens Lord Model pistol and a Smith &
Wesson revolver. His skill with these arms was so far in advance of his contemporaries that he was
popularly supposed to accomplish many of his feats by trickery.
Target-shooting with the pistol and revolver, as a sport, may be said to have originated at the annual
meeting of the National Rifle Association at Creedmoor in 1886. During that meeting a revolver match
was scheduled to be shot at 25 yards on the 200-yard Standard American Rifle Target. It was a reëntry
match, with the three best scores of five shots each of any contestant to count. In this match three
scores of 48 out of 50 were made, the highest individual aggregate of three scores being 143 out of a
possible 150.
The same year a similar match was announced at the fall meeting of the Massachusetts Rifle
Association at Walnut Hill. Chevalier Paine was a competitor in this match, and made 50—49—49==148
in six entries. The next best three scores equalled 142.
These matches proved so interesting and successful that target-shooting with the pistol and revolver
became instantly popular all over the country. It was soon found that the arms possessed remarkable
accuracy, and as the skill of the shooters improved the distance was increased to 50 yards retaining the
same target.
Fig. 67.—John A. Dietz Fig. 68.—E. E. Patridge Fig. 69.—Sergt. W. E. Petty

Mr. A. C. Gould, editor of The Rifle, and Shooting and Fishing, was the first one to recognize the
possibilities of the pistol and revolver, and became greatly interested in the performances with these
arms. He assisted and encouraged the shooters, witnessed their work, and made careful and elaborate
records of all the important scores that were made in the United States from 1886 to 1900.[11] It was at
his suggestion that Chevalier Paine essayed to fire the first 100-shot score at 50 yards on the Standard
American Target, scoring 791 points. This shooting was done with a finely sighted .44-caliber Smith &
Wesson Russian Model Revolver, regulation full charge ammunition, and a 2½-pound trigger pull. A
keen rivalry for the 100-shot record soon sprang up, resulting as follows:

Oct. 15, 1886, Chevalier Ira Paine at Walnut Hill 791


March 7, 1887, Chevalier Ira Paine " " " 841
Nov. 4, 1887, F. E. Bennett " " " 857
Nov. 14, 1887, F. E. Bennett " " " 877
Dec. 5, 1887, F. E. Bennett " " " 886
Dec. 17, 1887, Chevalier Ira Paine " " " 888
Dec. 22, 1887, Chevalier Ira Paine " " " 904
Dec. 23, 1887, W. W. Bennett " " " 914

This rivalry led to a long newspaper controversy, and culminated in the famous Paine-Bennett revolver
match. The conditions were as follows: Stakes $1000.00; 100 shots per day for six consecutive days;
Smith & Wesson Russian Model Revolvers, .44 caliber; factory-loaded full charge ammunition; trigger
pull, 3 pounds; Standard American Target with 8-inch bull’s-eye; distance, 50 yards. On the fifth day of
the match, and while 9 points in the lead, Chevalier Paine entered a protest and withdrew. Mr. F. E.
Bennett continued shooting, as stipulated in the match, scoring 5093 points for the total of the six days.
The protest was referred to the National Rifle Association, which decided in favor of Mr. Bennett,
awarding him the match and the championship of America.
In practising for this match Mr. F. E. Bennett, under the same conditions, made a score of 915. This
record was not excelled until June 1, 1901, when C. S. Richmond of Savannah, Georgia, scored 918
points under substantially the same conditions.
During the summer of 1890, Mr. William E. Carlin, assisted by Mr. Hubert Reynolds, made a very
elaborate series of tests with the revolver and various kinds of ammunition, to ascertain the possibilities
of the arms, the accuracy of the ammunition, the effect of fouling, etc. About 10,000 rounds were fired,
Mr. Carlin used a butt-stock attachment, telescopic sight, and sand bag rest; and Mr. Reynolds verified
Mr. Carlin’s results from a machine rest. All the shooting was done with black powder charges in Smith &
Wesson revolvers.
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and


personal growth!

textbookfull.com

You might also like