ADVANCED ROBUST

AND NONPARAMETRIC METHODS

IN EFFICIENCY ANALYSIS
METHODOLOGY AND APPLICATIONS
Studies in Productivity and Efficiency

Series Editors:

Rolf Färe
Shawna Grosskopf
Oregon State University
R. Robert Russell
University of California, Riverside

Books in the series:

Fox, Kevin J.:

Efficiency in the Public Sector
Ball, V. Eldon and Norton, George W.:
Agricultural Productivity:
Measurement and Sources of Growth
Färe, Rolf and Grosskopf, Shawna:
New Directions:
Efficiency and Productivity
Daraio, C. and Simar, L.:
Advanced Robust and Nonparametric Methods in Efficiency Analysis:
Methodology and Applications
ADVANCED ROBUST
AND NONPARAMETRIC METHODS
IN EFFICIENCY ANALYSIS
METHODOLOGY AND APPLICATIONS

by

Cinzia Daraio
Institute of Informatics and Telematics (CNR), Pisa
and
Department of Electrical Systems and Automation,
School of Engineering, University of Pisa

Léopold Simar
Institute of Statistics
Université Catholique de Lovain, Louvain la Neuve, Belgium

123
Library of Congress Control Number: 2006927067
ISBN-10: 0-387-35155-8 e-ISBN 0-387-35231-7
ISBN-13: 978-0387-35155-1 e-ISBN-13: 978-0387-35231-1

Printed on acid-free paper.

© 2007 Springer Science+Business Media, LLC

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC., 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they
are not identified as such, is not to be taken as an expression of opinion as to whether or not they are
subject to proprietary rights.

Printed in the United States of America.

987654321

springer.com
 Ai miei genitori
Luisa e Innocenzo Daraio

À ma famille
Contents

v
Dedication
xi
List of Figures
xv
List of Tables
Preface xvii
Acknowledgments xxi
1. INTRODUCTION 1
1.1 What this work is about 1
1.2 Improving the nonparametric approach in frontier analysis 4
1.3 An outline of the work 8

Part I Methodology
2. THE MEASUREMENT OF EFFICIENCY 13
2.1 Productivity and Efficiency 13
2.2 A short history of thought 16
2.3 The economic model 19
2.4 A taxonomy of efficient frontier models 25
2.5 The nonparametric frontier approach 30
2.5.1 Data Envelopment Analysis (DEA) 31
2.5.2 Free Disposal Hull (FDH) 33
2.6 Recent developments in nonparametric efficiency analysis 39
3. STATISTICAL INFERENCE IN NONPARAMETRIC
FRONTIER ESTIMATION 43
3.1 Statistical foundation 43
3.2 Introducing stochastic noise in the model 45
viii Contents

3.3 Asymptotic results 47

3.3.1 Consistency 47
3.3.2 Sampling distributions 49
3.4 Bootstrap techniques and applications 50
3.4.1 Bootstrap in frontier models 52
3.4.2 Correcting the bias 54
3.4.3 Bootstrap confidence intervals 55
3.4.4 Is the bootstrap consistent? 56
3.4.5 Applications of the bootstrap 63
3.4.6 Bootstrapping FDH estimators 64
4. NONPARAMETRIC ROBUST ESTIMATORS:
PARTIAL FRONTIERS 65
4.1 A re-formulation based on the probability of being dominated 66
4.2 Order-m frontiers and efficiency scores 68
4.3 Order-α quantile-type frontiers 72
4.4 Properties of partial frontier estimators 77
4.4.1 Statistical properties 77
4.4.2 Robust estimators of the full frontier 77
4.4.3 Advantages of using partial frontiers 78
4.4.4 Detection of outliers 79
4.5 Summary of the results for the output oriented case 81
4.6 Parametric approximations of robust nonparametric frontiers 85
4.6.1 Two stage methods 87
4.6.2 The bootstrap algorithms 89
4.7 Multivariate parametric approximations 90
4.7.1 Generalized Cobb-Douglas parametric model 91
4.7.2 Translog parametric model 93
5. CONDITIONAL MEASURES OF EFFICIENCY 95
5.1 Explaining efficiency in the literature 96
5.2 Introducing external-environmental variables 100
5.2.1 Conditional full frontier measures 100
5.2.2 Conditional order-m measures 101
5.2.3 Conditional order-α measures 103
5.2.4 Summary for the output oriented case 105
5.3 Bandwidth selection 108
5.3.1 Univariate case 109
5.3.2 Multivariate case 110
Contents ix

5.4 An econometric methodology 113

5.4.1 Global effect of Z on the production process 113
5.4.2 A decomposition of conditional efficiency 119
5.5 Simulated illustrations 121
5.5.1 Univariate Z 122
5.5.2 Multivariate Z 124

Part II Applications
6. ECONOMIES OF SCALE, SCOPE AND EXPERIENCE
IN THE ITALIAN MOTOR-VEHICLE SECTOR 135
6.1 Introduction 135
6.2 Data description 139
6.2.1 Definition of outputs and inputs 142
6.2.2 An exploratory investigation 145
6.2.3 Aggregation of inputs and outputs 148
6.3 Testing returns to scale and bootstrapping efficiency scores 151
6.4 Economies of scale 157
6.5 Economies of scope 160
6.6 Economies of experience 163
6.7 Conclusions 164
7. AGE, SCALE AND CONCENTRATION EFFECTS
IN A PUBLIC RESEARCH SYSTEM 167
7.1 Introduction 167
7.2 Data description 176
7.3 Scale and concentration effects 178
7.4 Age effects on CNR scientific productivity 181
7.5 Robust parametric approximation of multioutput
distance function 186
7.6 Conclusions 191
8. EXPLORING THE EFFECTS OF MANAGER TENURE, FUND
AGE AND THEIR INTERACTION 193
8.1 Introduction 193
8.2 Data description 196
8.3 Impact of mutual fund manager tenure on performance 199
8.4 Interaction between manager tenure and fund age 208
x Contents

8.5 Conclusions 216

9. CONCLUSIONS 217
References 221
Topic Index 243
Author Index 245
List of Figures

2.1 An overview of the literature on efficient frontier estimation. 27

2.2 Input and Output slacks. 33
2.3 FDH and DEA estimation of C(y) and ∂C(y). 36
2.4 FDH and DEA estimation of P (x) and ∂P (x). 36
2.5 Influence of outliers. 38
4.1 Input order-m frontier in the bivariate case. 69
4.2 Input order-α frontier in the bivariate case. 73
4.3 An illustration of α input measure of efficiency in the
bivariate case. 76
4.4 Illustration of full and partial frontiers in the input orientation. 76
4.5 Illustration of full and partial frontiers in the output orientation. 85
5.1 The Boiler case. Z favorable. 116
5.2 The Fridge case. Z unfavorable. 118
5.3 The Fridge case with outliers. Full frontier case 119
5.4 The Fridge case with outliers. Partial frontiers case 119
5.5 Simulated example with univariate Z. Likelihood CV
criterion. 123
5.6 Simulated example with univariate Z. Plots of the percentages. 123
5.7 Simulated example with univariate Z. Smoothed non-
parametric plots. 125
5.8 Simulated example with multivariate Z. Likelihood
cross validation plot. 126
5.9 Density of Z and its Contour plot. 127
5.10 Simulated example with multivariate Z. Plots of the
percentage of points outside partial frontiers. 128
xii List of Figures

5.11 Simulated example with multivariate Z. Surface. 128

5.12 Simulated example with multivariate Z. Plots. 129
5.13 Density of Z and its Contour plot. Z1 and Z2 are correlated. 129
5.14 Simulated example with multivariate Z. Z1 and Z2 are
correlated. 130
5.15 Simulated example with multivariate Z and 5 outliers.
Full frontier. 131
5.16 Simulated example with multivariate Z and 5 outliers.
α-frontier. 132
6.1 Nominal rate of growth by line of business. 140
6.2 Real rate of growth by line of business. 140
6.3 Projection of the variables for the Italian Motor-vehicle
insurers dataset. 147
6.4 Projection of individual insurers. 147
6.5 Output factor versus Input factor and DEA frontier for
the Italian Motor-vehicle dataset, year 2000. 152
6.6 Output factor versus Input factor and DEA frontier for
the Italian Motor-vehicle dataset, year 2000: a zoom. 153
6.7 Boxplots of the input Shephard efficiency scores. 156
6.8 Economies of scale in the Italian motor-vehicle insur-
ance business. Full frontier case 159
6.9 Economies of scale in the Italian motor-vehicle insur-
ance business. Partial Frontiers case 159
6.10 Economies of experience in the Italian motor-vehicle
insurance business. 164
7.1 Estimation of the density of Z, CNR institutes. 180
7.2 Scale and concentration effects on CNR institutes. Surface. 182
7.3 Scale and concentration effects on CNR institutes. Plots
of Qz . 183
7.4 Scale and concentration effects on CNR institutes. Plots
of Qzα . 183
7.5 Scale and concentration effects on CNR institutes. Plots
of Qzm . 184
7.6 Age effects on CNR institutes: FDH, order-α and order-
m estimates. 185
8.1 Scatterplot matrix of inputs (X) and output (Y) for Ag-
gressive Growth Mutual Funds. 200
List of Figures xiii

8.2 Scatterplot matrix of output and external factors for Ag-

gressive Growth Mutual Funds. Z1 is manager tenure
and Z2 is Fund age. 200
8.3 Scatterplot matrix of external factors and Inputs for Ag-
gressive Growth Mutual Funds. 201
8.4 Influence of manager tenure on the performance of Ag-
gressive Growth Mutual Funds. 201
8.5 Influence of manager tenure on the performance of Ag-
gressive Growth Mutual Funds. A zoom. 201
8.6 Influence of fund age on the performance of Aggressive
Growth Mutual Funds. 208
8.7 Influence of fund age on the performance of Aggressive
Growth Mutual Funds. A zoom. 209
8.8 Influence of manager tenure and fund age on the per-
formance of Aggressive Growth Mutual Funds. Surface. 210
8.9 Influence of manager tenure (Z1 ) and fund age (Z2 ) on
the performance of Aggressive Growth Mutual Funds. Plots. 210
8.10 Influence of manager tenure (Z1 ) and fund age (Z2 ) on
the performance of Aggressive Growth Mutual Funds.
Density of Z and its contour plot. 211
List of Tables

1.1 Limitations and Advancements of nonparametric methods

in frontier analysis. 7
3.1 Summary of the bootstrap principle. 54
5.1 Summary of efficiency measures with references. 106
6.1 Definition of Nominal and Real rate of growth. 141
6.2 Definition of rate of growth by line of business. 141
6.3 Definition of Inputs, Outputs and External Factors. 144
6.4 Descriptive statistics on inputs, outputs and external
factor considered in the analysis. Italian motor vehicle
insurers. 144
6.5 Correlations matrix. Italian motor vehicle insurers. 145
6.6 Eigenvalues and percentages of variances explained.
Italian motor vehicle insurers. 146
6.7 Correlations of the first two pc’s with the original variables
(Factors loadings). Italian motor vehicle insurers. 146
6.8 Input Factor inertia. Italian motor vehicle insurers. 149
6.9 Eigenvectors of the matrix X
X. Italian motor vehicle
insurers. 149
6.10 Correlations between the input factor (Fin ) and inputs.
Italian motor vehicle insurers. 149
6.11 Output Factor inertia. Italian motor vehicle insurers. 150
6.12 Eigenvectors of the matrix Y
Y . Italian motor vehicle
insurers. 150
6.13 Correlations between the output factor (Fout ) and outputs.
Italian motor vehicle insurers. 150
6.14 Bootstrap results: Italian motor-vehicle insurers. 154
xvi List of Tables

6.14 (Continued) 155

6.15 Some descriptive statistics on Age and size (Mktshare)
of Italian motor-vehicle insurers. 162
7.1 Research areas. 176
7.2 Definition of variables. 177
7.3 Definition of Inputs and Outputs. 178
7.4 Average values of Inputs and Outputs by research area. 179
7.5 CNR institutes - Descriptive statistics on inputs, outputs
and external factors considered in the analysis. 179
7.6 Definition of inputs and outputs used for the translog
estimation of CNR output distance function. 187
7.7 Output distance function parameter estimates for CNR
research institutes. 188
7.8 Bootstrapped confidence intervals for order-α frontier
translog approximation. 189
7.9 Bootstrapped confidence intervals for order-m frontier
translog approximation. 190
8.1 Descriptive statistics on inputs, output, external factors
for Aggressive Growth mutual funds. 198
8.2 Unconditional measures of efficiency, and various indicators,
Aggressive Growth Mutual Funds. 202
8.3 Conditional measures of efficiency and various indicators
(Z is manager tenure), by groups of Aggressive Growth
Mutual Funds. 204
8.4 Some descriptive statistics by group of funds with different
impact of manager tenure. Z manager tenure. 205
8.5 Rank of AG mutual funds with Qzm > 1 ordered by
INCR. Z is manager tenure. 206
8.6 Rank of AG mutual funds with Qzm < 1 ordered by
DECR. Z is manager tenure. 207
8.7 Some descriptive statistics on efficiency measures and
indicators by group of funds with different impact of
manager tenure. Multivariate Z. 212
8.8 Some descriptive statistics by group of funds with different
impact of manager tenure. Bivariate Z. 213
8.9 Rank of AG mutual funds with Qzm < 1 ordered by
DECR. Bivariate Z. 214
8.10 Rank of AG mutual funds with Qzm > 1 ordered by
INCR. Bivariate Z. 215
Preface

The topic of production efficiency has attracted attention since Adam Smith’s
pin factory and even before. However, a rigorous analytical approach to the
measurement of efficiency in production originated with the work of Koopmans
(1951) and Debreu (1951), empirically applied by Farrell (1957). Farrell’s
seminal work gave rise to a considerable amount of studies.
The basic idea of efficiency analysis is to make a comparison among a group
of firms or branches or among Decision Making Units (DMUs), in order to
evaluate how the resources (or inputs) are used to obtain (produce) the products
(services or outputs). This evaluation process is based on the estimation of a
benchmark frontier against which the DMUs are assessed, using DMUs’ inputs
and outputs. The level of efficiency of each DMU is gauged as the distance
from the estimated (‘efficient’) frontier.
In literature on efficiency analysis, the nonparametric approach has received
a considerable amount of interest, both from a theoretical and an applied per-
spective. This mainly because it does not require many assumptions and par-
ticularly because it does not need the specification of a functional form for the
frontier. Hence, the parameters of the functional form of the frontier do not
have to be estimated in this approach, from which the name ‘nonparametric’
approach derives, whereas in the parametric approach, the parameters of the
efficient frontier must be estimated. Data Envelopment Analysis (DEA) and
Free Disposal Hull (FDH) are among the most known and applied nonparamet-
ric techniques for the measurement of the efficiency in production and service
activities (see e.g., Cooper, Seiford and Tone, 2000, for about 1,500 references
of their applications). Nevertheless, this traditional nonparametric approach
(DEA/FDH based) presents some severe limitations that are not always taken
into account by researchers who apply it in empirical works. These limits
should be carefully considered in order to provide a correct interpretation of the
obtained results.
xviii Preface

To summarise, the main drawbacks of the traditional nonparametric approach

are known to be:

Deterministic and non-statistical nature;

Influence of outliers and extreme values;

Lack of parameters for the economic interpretation;

Unsatisfactory techniques for the introduction of environmental or exter-

nal variables in the measurement (estimation) of the efficiency.

The main objective of this book is to provide a systematic and comprehensive

treatment of recent developments in efficiency analysis in order to overcome
these drawbacks.
In Part I of the book (Methodology), we introduce a complete set of tools
for measuring the efficiency of firms (or production units, or Decision Making
Units) and for explaining the observed efficiency differentials. This is a general
and flexible toolbox that does not make any assumption on the behaviour of the
units under analysis. Therefore, it can be used in theories that generalise the
neoclassical theory of production as well as alternative ones (as is the case of
the evolutionary theory of production).
The Methodology presents, in an intuitive, rigorous and self-contained way,
the state-of-the-art on ‘advanced’ frontier models based on techniques that do
not impose any functional specification of the frontier (nonparametric methods)
and are not affected by extremes and outliers in the data (robust methods).
The following issues are dealt with in detail and further developed:

Statistical inference in nonparametric frontier estimation. Here, we intro-

duce stochastic elements in nonparametric frontier models, and we present
the application of the bootstrap in efficiency analysis (Simar and Wilson,
1998, 2000b, 2006a), including: estimation of bias and confidence in-
tervals of efficiency estimates; hypothesis testing; comparison of groups’
efficiency.

Robust estimators of frontiers. Order−m frontiers (Cazals, Florens and

Simar, 2002; Daraio and Simar, 2005a) and α-quantile frontiers (Daouia
and Simar, 2004), do not envelop all data points and for that are not influ-
enced by extreme values. They are also characterised by useful statistical
properties and interesting economic interpretations.

Parametric approximation of nonparametric and robust frontiers. These

approximations, introduced by Florens and Simar (2005) and Daouia,
Florens and Simar (2005), provide ‘robust’ estimates of parameters usable
Preface xix

for economic interpretation. These techniques are extended, in this book,

to the full multi-input multi-output setup.
Nonparametric and robust conditional frontier models. These models,
recently introduced by Daraio and Simar (2005a, 2005b), are able to cap-
ture and measure the effect of external environmental variables on the
efficiency, in a way that overcomes most drawbacks of previous methods.

This book has been specifically designed for applied economists who have
an interest in the advantages of traditional nonparametric methods (DEA/FDH)
for efficiency analysis, but are sceptical about adopting them because of the
drawbacks they present.
In Part II of the book (Applications), we propose three empirical illustrations
taken from different economic fields: insurance sector, scientific research and
mutual funds industry. These applications perfectly illustrate how the tools
we propose can be used to analyse economies of scale, economies of scope,
dynamics of age and agglomeration effects, trade-offs in production and service
activities, groups comparison as well as help explain efficiency differentials.
These extensively treated empirical applications, based on real data, show the
usefulness of our approach in applied economics. Through these applications
we illustrate how various statistical tools can be combined to shed light on the
key features of the studied production process.
Moreover, this book has also been written for researchers with a background
in Operations Research (OR) and/or Management Science (MS), who would
like to deepen their knowledge of these new robust and nonparametric tech-
niques, which have been recently presented at specialised conferences and have
appeared on the scientific journals in recent years. In this book they will find
a readable, synthetic but also accurate presentation of these recent advances -
without the burden of technicalities and formal demonstrations - together with
an extensive illustration of their use in empirical works.
Acknowledgments

This book is the result of a scientific collaboration between the two au-
thors started in 2000 during the Ph.D. of Cinzia Daraio at the Scuola Superiore
Sant’Anna of Pisa, Italy, and reinforced by her Master of Arts in Statistics at-
tended at the Université Catholique de Louvain, Belgium, in 2001-2002. Within
this research collaboration a series of papers, published or forthcoming, as well
as several joint research projects have been realized or are in progress. This
book can be viewed as a milestone of this still continuing and active synergy.

The research activity from which this book originated has been funded by
various national and international research funds. In particular the financial
support of the following projects is gratefully acknowledged:

The AQUAMETH (Advanced Quantitative Methods for the Evaluation of

the Performance of Public Sector Research) Project within the European
Network of Excellence PRIME;

The Interuniversity Attraction Pole, Phase V (No. P5/24) from the Belgian
Government (Belgian Science Policy);

The Italian national project (iRis) Reorganizing the Italian Public Research
system to foster technology transfer: governance, tools and implementa-
tion;

The Italian national project (MIUR 40% - prot. 2004132703) System

spillovers on the competitiveness of Italian economy: quantitative analysis
for sectoral policies.
xxii Acknowledgments

We would like to thank all our colleagues and friends who encouraged us
in the writing of this book. In particular we are grateful to Carlo Bianchi, An-
drea Bonaccorsi, Hal Fried, Giampiero M. Gallo, Shawna Grosskopf, Jacques
Mairesse, Henk Moed, Paula Stephan, Robert Russell and Paul Wilson.
Finally, special thanks go to Sean Lorre, Springer editor, and to Deborah
Doherty, Springer author support, for their kind cooperation.
Chapter 1

INTRODUCTION

1.1 What this work is about

Theoretical mainstream production analysis has always focused on produc-
tion activity as an optimization process. In conventional microeconomic theory,
it is assumed, in fact, that producers optimize by not wasting resources in a sys-
tematic way: producers operate somewhere on the boundary of their production
possibility sets. However, consistent empirical evidence shows that not all pro-
ducers succeed in all circumstances. Hence, it is important to analyse the degree
to which producers fail to optimize and the extent of departures from technical
and economic efficiency.
On the other hand, empirical production analysis has concentrated on central
tendency, or “average” or “most likely” relationship constructed by intersecting
data with a function rather than surrounding data with a frontier. There have
been attempts to reconcile the two fields by developing quantitative techniques
for surrounding rather than intersecting data.
The approach of production frontiers is an effort to define empirically an
envelopment of production data. This approach combines the construction
of production frontiers with the measurement and interpretation of efficiency
relative to the constructed frontiers.
The structure of economic frontiers is of interest in itself: it is important
to know whether, for example, technologically efficient production is charac-
terised by economies of scale or scope. The structure of production frontiers
can differ from the structure of average production functions constructed from
the same data. Best practice is not just better than average practice, it may
also be structurally different, and it is important to know whether the structure
of efficient production differs from the structure of average production. Best
practices (captured by the frontier approach) may be better than average prac-
2 Introduction

tices (measured in a regression-based framework) in the sense that they exploit

available substitution possibilities or scale opportunities that average practices
do not. Public policy based on the structure of best practice frontiers may be
very different from policy based on the structure of average practice functions.
The distance between production frontiers and observed producers is of obvi-
ous policy interest. It is important to know how inefficient observed production
is, what types of producers are most and least efficient, and what type of in-
efficiency is most prevalent. It is also important to know whether inefficient
firms (or Decision Making Units) are working in presence of economies (dis-
economies) of scale or scope and the extent of trade-offs in the production or
service activities.
Interest in the general topic of productivity and efficiency has grown greatly
in recent years. There are in fact “efficiency” studies in virtually every country
and about virtually every market or non-market production or service activity.
The general area of production economics is not the only discipline inter-
ested in frontiers and efficiency. Attention from economists in frontiers was
reawakened in the 1970s when the works of Debreu (1951), Koopmans (1951,
1957) and Farrell (1957) were rediscovered.
The approach developed by Shephard (1953, 1970, 1974) in production
analysis represents an important contribution, mainly because it is based on
distance functions that are intimately related to the measurement of the effi-
ciency. As a matter of fact, a lot of subsequent works are based on Shephard’s
distance functions.
A valuable body of works has emerged in the Operations Research (OR)
and in the Management Science (MS) fields. The OR/MS approach has its
own orientation, and relies heavily on linear programming techniques. Extend-
ing ideas of Farrell (1957), the OR discipline has developed the popular Data
Envelopment Analysis approach (DEA, Charnes, Cooper and Rhodes, 1978)
to the estimation of production frontiers and efficiency measurement which
employs linear programming techniques. On the other hand, in a more eco-
nomic framework, the Free Disposal Hull estimator (FDH, Deprins, Simar and
Tulkens, 1984) of frontier was introduced. Nevertheless, the linear program-
ming approach came to be accepted as a computational method for measuring
efficiency in different kinds of economic decision-making problems since the
work by Dorfman, Samuelson and Solow (1958).
FDH and DEA are the most known nonparametric approaches in efficiency
analysis. FDH can be seen as a more general nonparametric estimator of fron-
tiers than DEA, or as a non-convex version of DEA. In fact, FDH relies on
the free disposability of inputs and outputs (i.e., the possibility of not using or
destroying goods -inputs, outputs- without costs) which define the production
possibility set; whilst, DEA relies on the free disposability and on the convexity
What this work is about 3

of the production set (convexity implies that if two observations are possible,
then all the linear combinations that lie between them are also possible).
The preference of the nonparametric approach over the parametric approach
(based on the functional specification of the frontier), is due to the small amount
of assumptions required and mainly to the fact that we do not have to specify
the functional form of the relation inputs-outputs and we do not need to specify
a distributional form for the inefficiency term.
Nonetheless, traditional nonparametric estimators based on envelopment
techniques (i.e. DEA /FDH types) were for a long time limited by several draw-
backs: deterministic (meaning that all deviations from the efficient frontier are
considered as inefficiency, and no noise is allowed) and non statistical nature;
influence of outliers and extreme values; lack of parameters for the economic
interpretation; unsatisfactory techniques for the introduction of environmental
or external variables in the measurement of the efficiency.
Our work treats at length recently introduced robust and nonparametric ap-
proaches in efficiency analysis which overcome most traditional limits of the
nonparametric approach listed above. In doing so, we provide computationally
feasible methods of calculation (both of the efficient frontier and of the distance
from it) and explanation of efficiency differentials.
We believe that the robust and nonparametric approach in frontier analysis
has reached a level of generality and has overcome most of its limits, so that it
can be considered as being more flexible and more suitable for the evaluation
of complex production and service activities, with respect to other approaches,
like the parametric approach.
The economic model underlying our robust and nonparametric frontier ap-
proach is very general: it does not make any assumptions about the behaviour
of the firms (or DMUs) and does not introduce prices of factors which are
considered as the link of DEA-based models with the neoclassical theory of
production (Ray, 2004). Moreover, in a lot of empirical applications prices are
not available (as is the case for scientific research, several no-profit services,
and so on).
Our book is designed to fill a gap in the literature by systematically propos-
ing the recent developments of the nonparametric approach and illustrating its
usefulness for empirical research through three full economic applications. We
propose an intuitive and readable, but in the meantime rigorous presentation
of advanced nonparametric and robust methods in efficiency analysis, without
the burden of technicalities and demonstrations. This methodology does not
impose any assumption on the behavior of firms and therefore, it is a general
and flexible tool suitable for applications both in theories of production that
generalize the neoclassical theory, and in alternative approaches.
The material contained in this work offers a background for researchers of
different disciplines.
4 Introduction

It can be used following different reading paths:

Applied economists may be interested in the whole book, both the Method-
ology and the Applications parts.

Researchers with OR background could be more interested in the Method-

ological part, and may perhaps skip Part II (Applications).

Researchers in MS may well start their reading with the Applications, and
then go back to the Methodology, for a better understanding of the applied
techniques.

This book could also be adopted for specialised courses in efficiency analysis,
for graduate students or undergraduate students of the last years.

1.2 Improving the nonparametric approach in frontier

analysis
The main purpose of this book is to propose a general and comprehensive
approach for the measurement and explanation of firms’ (or DMUs’) efficiency
differentials by using advanced robust and nonparametric frontier models. In
fact, some recent developments in efficiency analysis are able to overcome most
drawbacks of the traditional nonparametric approach (DEA/FDH based).
One traditional limitation of the nonparametric approach is its deterministic
nature and the difficulty in making statistical inference. In this setting, it is
assumed that all deviations from the efficient frontier are due to inefficiencies.
The problem of handling noise in this framework is due to the fact that the
model is not identified, unless some restrictions are assumed. Aigner, Lovell
and Schmidt (1977) consider approaches that assume a parametric function for
the frontier; Kneip and Simar (1996) analyse the case of panel data. More
general results for handling noise in nonparametric frontier models have been
introduced by Hall and Simar (2002) and by Simar (2003b). In Section 3.2 we
comment this literature as well as recently introduced approaches based on local
maximum likelihood techniques (Kumbhakar, Park, Simar and Tsionas, 2004).
In the nonparametric approach, moreover, statistical inference is not easy, due
to the complex nature of the estimation which is based on very few assumptions.
Nevertheless, advances have been made, and statistical inference is available
by using asymptotic results (Simar and Wilson, 2000a) or by applying the
bootstrap (Simar and Wilson, 1998, 2000, Kneip, Simar and Wilson, 2003).
We present this literature in Chapter 3, together with the available hypothesis
testing procedures.
Improving the nonparametric approach in frontier analysis 5

The problem of extremes or outliers can be treated by applying methods to

detect them, and using estimators which are robust to them. This is achieved
by partial frontiers of order-m or α which are developed in Chapter 4. These
frontiers do not envelop all data points and for that are less affected, and hence
more robust, to extreme values and outliers in the data.
These frontiers do not also suffer of the so called curse of dimensionality,
typical of nonparametric methods (DEA/FDH), which requires a large amount
of observations in order to avoid large bias and imprecise estimation (i.e., huge
confidence intervals). Partial frontiers are estimators that approach the “true”
frontier as fast as parametric estimators (i.e., the speed of convergence of order-
m and α frontiers is root−n where n is the number of firms or DMUs analysed).
Whereas DEA estimators have a speed of convergence of n2/(p+q+1) - where
p is the number of the inputs and q is the number of the outputs used in the
analysis. The rate of convergence of FDH is n1/(p+q) . This indicates for the
DEA/FDH estimators the necessity of increasing the number of observations
when the dimension of the input-output space increases to achieve the same
level of statistical precision. The dimensionality problem does not apply for
these robust frontier estimators: this is an important property because most
of the empirical applications relies on small samples and the collection of an
increasing number of data is difficult and sometimes almost impossible.
Several authors show formally how DEA efficiency scores are affected by
sample size; they demonstrate that comparing measures of structural ineffi-
ciency between samples of different sizes leads to biased results. This sample
size problem can be easily overcome using the robust nonparametric approach
based on partial frontiers. In this setting, the parameters m or α may be chosen
equal for two or more groups of observations, and their average performance
can then be compared.
Another limitation of traditional nonparametric techniques is the more diffi-
cult economic interpretation of the production process in terms of elasticities of
substitution, marginal products, partial elasticities, and so on, due to the lack of
parameters. To overcome this drawback Florens and Simar (2005) and Daouia,
Florens and Simar (2005) propose the full theory for parametric approximation
of nonparametric frontiers, which applies also for robust order-m and order-α
frontiers. This topic will be treated and developed in Chapter 4.
Finally, the conventional techniques for explaining efficiency in the nonpara-
metric frontier setting are unsatisfactory. In fact, two main approaches have
been proposed in literature but both are flawed by restrictive prior assumptions
on the Data Generating Process (DGP) and/or on the role of these external fac-
tors on the production process. The first family of models is based on a one-stage
approach (see e.g. Färe, Grosskopf and Lovell, 1994, p. 223-226), where these
factors are considered as free disposal inputs and /or outputs which contribute
to define the attainable set but which are not active in the optimization process
6 Introduction

defining the efficiency scores. An external-environmental factor is considered

as an input if it is favorable to efficiency and as an output if it is detrimental to
efficiency. The drawback of this approach is twofold: first we have to know
a priori what is the role of the factor on the production process, and secondly
we assume the free disposability (and eventually convexity, if DEA is used)
of the corresponding attainable extended production set. The second family
of models is based on a two-stage approach. Here the estimated efficiency
scores are regressed, in an appropriate limited dependent variable parametric
regression model (like truncated normal regression models) on the environmen-
tal factors. Recently, some models in this family propose also three-stage and
four-stage analysis as extension of the two-stage approach (Fried, Schmidt, and
Yaisawarng, 1999; Fried, Lovell, Schmidt, and Yaisawarng, 2002).
However, as pointed out by Simar and Wilson (2003), also these models,
like the others in the two-stage family, are flawed in that usual inference on
the obtained estimates of the regression coefficient is not available. They state:
“none of the studies that employ two-stage approaches have described the un-
derlying data-generating process. Since the DGP has not been described, there
is some doubt about what is being estimated in the two-stage approaches”, and
“A more serious problem in all of the two-stage studies that we have found
arises from the fact that DEA efficiency estimates are serially correlated; con-
sequently, standard approaches to inference [...] are invalid”. Simar and Wilson
(2003) then propose two bootstrap-based algorithms to obtain a more accurate
inference. However, even this bootstrap-based approach, through more cor-
rect, shares some of the inconveniences of the two-stage based approaches.
Firstly, it relies on a separability condition between the input × output space
and the space of values of the external-environmental variables: i.e., the value
of external-environmental variables does not influence the position of the fron-
tier of the attainable set. Secondly, the regression in the second stage relies
on some parametric assumptions (like linear model and truncated normal error
term).
Daraio and Simar (2005a, b) propose a full nonparametric and robust ap-
proach to evaluate the influence of external-environmental variables which
overcomes most of the drawbacks of earlier approaches. This topic will be
treated in detail in Chapter 5.
In this work, we propose a readable and rigorous description of the most re-
cent developments in robust and nonparametric efficiency models that overcome
the main traditional limitations of the nonparametric approach (see Table 1.1)
and an extensive illustration of its usefulness for applied economics, through
several applications with real data, in different economic fields (scientific re-
search, insurance industry, mutual funds).
The combination of the advancements in statistical inference, robust methods
and improvements in the explanation of efficiency differentials is used in the
Improving the nonparametric approach in frontier analysis 7

Table 1.1. Limitations and Advancements of nonparametric methods in frontier analysis.

Limitations of Proposed
Nonparametric methods Advancements
Deterministic nature and no easy Noise - Hall and Simar, (2002); Simar (2003b).
inference
Statistical properties - Simar and Wilson, (2000a);
Kneip, Simar and Wilson, (2003).

Bootstrapping - Simar and Wilson,

(1998, 2000b, 2006a);
Kneip, Simar and Wilson, (2003).

Hypothesis testing - Simar and Wilson, (2001, 2002).

No parameters for economic Parametric approximation of nonparametric

interpretation frontiers - Florens and Simar (2005);

Daouia, Florens and Simar (2005) and this book.

Sensitive to extremes / outliers Robust order-m frontiers

Curse of dimensionality Cazals, Florens and Simar (2002);

Sample size bias in comparison Daraio and Simar (2005a);

Robust order-α frontiers

Daouia and Simar (2004)

Unsatisfactory methods for explaining Problems of two-stage approaches

efficiency Simar and Wilson (2003).

Probabilistic approach
Daraio and Simar (2005a, b) and this book.

Applications Part of this book to analyse confidence intervals of efficiency

scores as well as efficiency distributions, economies of scale and scope, trade-
offs in production and service activities together with comparison of groups’
efficiency in the fields of scientific research, mutual funds industry and insurance
sector.
8 Introduction

1.3 An outline of the work

The book is structured in two parts.
Part I describes the state-of-the-art of robust and nonparametric methods in
efficient frontiers estimation.
Part II applies the comprehensive approach proposed in Part I to three dif-
ferent topics: insurance industry, scientific research and mutual funds.
In particular, Part I is organized as follows.
Chapter 2 introduces the measurement of efficiency. In Sections 2.1 and
2.2 some basic definitions and a short history of thought are reported, while in
Section 2.3 the economic model underlying the frontier estimation is described.
Section 2.4 is dedicated to make an introduction on the several efficient fron-
tier models available and, finally, Section 2.5 introduces the nonparametric
approach (DEA/FDH based) in efficiency analysis.
Chapter 3 focuses on the last developments of the statistical inference in
the nonparametric setting. Section 3.1 illustrates the statistical foundation of
nonparametric frontier estimation. In Section 3.2 some stochastic elements are
introduced, while Section 3.3 summarizes the main asymptotic properties of
nonparametric estimators. Finally, Section 3.4 introduces the available appli-
cations of the bootstrap in nonparametric frontier models.
Chapter 4 introduces robust nonparametric measures based on partial fron-
tiers. Section 4.1 presents a probabilistic formulation of the frontier estimation
setting. Section 4.2 is devoted to robust order-m frontiers (input oriented case),
while Section 4.3 introduces robust order-α frontiers (input oriented case) as
well as some measures of efficiency. The main properties of partial frontiers are
described in Section 4.4. In Section 4.5 the various output oriented cases are
outlined. The general framework of the parametric approximation of nonpara-
metric and robust frontiers is offered in Section 4.6, while the generalization
to the multi-input multi-output parametric approximation is reported in Sec-
tion 4.7 where the cases of the Generalized Cobb-Douglas and of the Translog
functions are detailed.
Chapter 5 is devoted to conditional measures of efficiency and the explanation
of efficiency differentials. In Section 5.2 we propose a unifying description of all
conditional efficiency measures, and introduce a new probabilistic conditional
indicator, while Section 5.3 illustrates the practical data-driven selection of
the bandwidth in this framework. The econometric methodology to evaluate
the impact of external factors on the performance is presented in Section 5.4
where also a decomposition of the conditional measure is offered, aiming at
disentangling the effect of the external factors on the performance. Finally,
Section 5.5 offers several simulated examples, with univariate and multivariate
external factors, which show the power of conditional measures and how to
practically implement the econometric methodology and interpret the obtained
results.
An outline of the work 9

In Part II, the three applications have a similar structure. The introductory
section presents the relevant literature on the topic treated and states the main
research questions addressed in the chapter. Then, the description of the used
dataset is presented in a second section, where some descriptive statistics are
reported. In particular, in Section 6.2 also an exploratory Principal Component
Analysis is reported as well as a procedure to aggregate inputs and outputs. After
that, the various methods described in Part I are applied in the different fields and
the empirical results are commented. Finally, the concluding section reported
at the end of each chapter summarizes and formulate policy implications on the
main results obtained.
Specifically, Chapter 6 deals with the Italian insurance industry. It focuses
on the motor vehicle sector. It provides tests on returns to scale, bootstrapped
confidence intervals for the efficiency estimates, and empirical evidence on
economies of scale, scope and experience.
Chapter 7 analyses a public research system: the research institutes of the
Italian National Research Council (CNR). Economies of scale, agglomeration
and age effects on scientific productivity are investigated, evaluating in partic-
ular the interaction between scale and agglomeration effects. A robust scale
elasticity is also estimated using the newly introduced multi-output parametric
approximation of robust nonparametric frontiers.
Finally, Chapter 8 focuses on US Aggressive Growth mutual funds. It exam-
ines how manager tenure and funds age affect the performance of mutual funds.
The interaction of manager tenure and age of funds is also assessed providing
detailed results on groups of best and worst performers.
The last chapter sums up the main points and concludes the book.
Chapter 2

THE MEASUREMENT OF EFFICIENCY

This chapter is about the measurement of efficiency of production units. It

opens with a section concerning basic definitions of productivity and efficiency.
After that, an historical and background section follows, reporting some of the
most important contributions until around ’90s. Then, the axiomatic under-
pinning of the Activity Analysis framework used to represent the production
process is described in the economic model section. Afterwards, efficient fron-
tier models are classified according to three main criteria: specification (or not)
of the form of the frontier; presence of noise in the estimation procedure; type
of data analyzed (cross-section or panel data). The presentation of the most
known nonparametric estimators of frontiers (i.e., Data Envelopment Analysis
(DEA) and Free Disposal Hull (FDH)) is subsequently. Finally, a section sum-
marizing recent developments in nonparametric efficiency analysis concludes
the chapter.

2.1 Productivity and Efficiency

According to a classic definition (see e.g. Vincent 1968) productivity is the
ratio between an output and the factors that made it possible. In the same way,
Lovell (1993) defines the productivity of a production unit as the ratio of its
output to its input.
This ratio is easy to compute if the unit uses a single input to produce a
single output. On the contrary, if the production unit uses several inputs to
produce several outputs, then the inputs and outputs have to be aggregated so
that productivity remains the ratio of two scalars.
We can distinguish between a partial productivity, when it concerns a sole
production factor, and a total factor (or global) productivity, when referred to
all (every) factors.
14 The measurement of efficiency

Similar, but not equal, is the concept of efficiency. Even though, in the effi-
ciency literature many authors do not make any difference between productivity
and efficiency. For instance, Sengupta (1995) and Cooper, Seiford and Tone
(2000) define both productivity and efficiency as the ratio between output and
input.
Instead of defining the efficiency as the ratio between outputs and inputs, we
can describe it as a distance between the quantity of input and output, and the
quantity of input and output that defines a frontier, the best possible frontier for
a firm in its cluster (industry).
Efficiency and productivity, anyway, are two cooperating concepts. The
measures of efficiency are more accurate than those of productivity in the sense
that they involve a comparison with the most efficient frontier, and for that they
can complete those of productivity, based on the ratio of outputs on inputs.
Lovell (1993) defines the efficiency of a production unit in terms of a com-
parison between observed and optimal values of its output and input. The
comparison can take the form of the ratio of observed to maximum potential
output obtainable from the given input, or the ratio of minimum potential to
observed input required to produce the given output. In these two comparisons
the optimum is defined in terms of production possibilities, and efficiency is
technical.
Koopmans (1951; p. 60) provided a definition of what we refer to as tech-
nical efficiency: an input-output vector is technically efficient if, and only if,
increasing any output or decreasing any input is possible only by decreasing
some other output or increasing some other input.
Farrell (1957; p. 255) and much later Charnes and Cooper (1985; p. 72) go
back over the empirical necessity of treating Koopmans’ definition of technical
efficiency as a relative notion, a notion that is relative to best observed practice
in the reference set or comparison group. This provides a way of differentiating
efficient from inefficient production units, but it offers no guidance concerning
either the degree of inefficiency of an inefficient vector or the identification of
an efficient vector or combination of efficient vectors against which comparing
an inefficient vector.
Debreu (1951) offered the first measure of productive efficiency with his coef-
ficient of resource utilization. Debreu’s measure is a radial measure of technical
efficiency. Radial measures focus on the maximum feasible equiproportionate
reduction in all variable inputs, or the maximum feasible equiproportionate
expansion of all outputs. They are independent of unit of measurement.
Applying radial measures the achievement of the maximum feasible input
contraction or output expansion suggests technical efficiency, even though there
may remain slacks in inputs or surpluses in output. In economics the notion of
efficiency is related to the concept of Pareto optimality. An input-output bundle
is not Pareto optimal if there remains the opportunity of any net increase in
Productivity and Efficiency 15

outputs or decrease in inputs. Pareto-Koopmans measures of efficiency (i.e.,

measures which call a vector efficient if and only if it satisfies the Koopmans
definition reported above, coherent with the Pareto optimality concept) have
been analysed in literature. See e.g., Färe (1975), Färe and Lovell (1978) and
Russell (1985, 1988, 1990) among others.
Farrell (1957) extended the work initiated by Koopmans and Debreu by
noting that production efficiency has a second component reflecting the ability of
producers to select the “right” technically efficient input-output vector in light of
prevailing input and output prices. This led Farrell to define overall productive
efficiency as the product of technical and allocative efficiency. Implicit in
the notion of allocative efficiency is a specific behavioral assumption about the
goal of the producer; Farrell considered cost-minimization in competitive inputs
markets, although all the behavioral assumptions can be considered. Although
the natural focus of most economists is on markets and their prices and thus on
allocative rather than technical efficiency and its measurement, he expressed
a concern about human ability to measure prices accurately enough to make
good use of allocative efficiency measurement, and hence of overall economic
efficiency measurement. This worry expressed by Farrell (1957; p. 261) has
greatly influenced the OR/MS work on efficiency measurement. Charnes and
Cooper (1985; p. 94) cite Farrell concern as one of several motivations for the
typical OR/MS emphasis on the measurement of technical efficiency.
It is possible to distinguish different kind of efficiency, such as scale, alloca-
tive and structural efficiency.
The scale efficiency has been developed in three different ways. Farrell
(1957) used the most restrictive technology having constant returns to scale
(CRS) and exhibiting strong disposability of inputs. This model has been de-
veloped in a linear programming framework by Charnes, Cooper and Rhodes
(1978). Banker, Charnes and Cooper (1984) have shown that the CRS measure
of efficiency can be expressed as the product of a technical efficiency measure
and a scale efficiency measure. A third method of scale uses nonlinear specifi-
cation of the production function such as Cobb-Douglas or a translog function,
from which the scale measure can be directly computed (see Sengupta, 1994
for more details).
The allocative efficiency in economic theory measures a firm’s success in
choosing an optimal set of inputs with a given set of input prices; this is dis-
tinguished from the technical efficiency concept associated with the production
frontier, which measures the firm’s success in producing maximum output from
a given set of inputs.
The concept of structural efficiency is an industry level concept due to Farrell
(1957), which broadly measures in what extent an industry keeps up with the
performance of its own best practice firms; thus it is a measure at the industry
level of the extent to which its firms are of optimum size i.e. the extent to which
16 The measurement of efficiency

the industry production level is optimally allocated between the firms in the
short run. A broad interpretation of Farrell’s notion of structural efficiency can
be stated as follows: industry or cluster A is more efficient structurally than in-
dustry B, if the distribution of its best firms is more concentrated near its efficient
frontier for industry A than for B. In their empirical study, Bjurek, Hjalmarsson
and Forsund (1990) compute structural efficiency by simply constructing an
average unit for the whole cluster and then estimating the individual measure of
technical efficiency for this average unit. On more general aggregation issues,
see Färe and Zelenyuk (2003) and Färe and Grosskopf (2004, p. 94 ff).

2.2 A short history of thought

The theme of productive efficiency has been analysed since Adam Smith’s
pin factory and before1 . However, as we have seen in the previous section,
a rigorous analytical approach to the measurement of efficiency in production
originated only with the work of Koopmans (1951) and Debreu (1951), empir-
ically applied by Farrell (1957).
An important contribution to the development of efficiency and productivity
analysis has been done by Shephard’s models of technology and his distance
functions (Shephard 1953, 1970, 1974). In contrast to the traditional produc-
tion function, direct input and output correspondences admit multiple outputs
and multiple inputs. They are thus able to characterize all kinds of technologies
without unwarranted output aggregation prior to analysis. The Shephard direct
input distance function treats multiple outputs as given and contracts inputs vec-
tors as much as possible consistent with technological feasibility of contracted
input vector. Among its several useful properties, one of the most important is
the fact that the reciprocal of the direct input distance function has been pro-
posed by Debreu (1951) as a coefficient of resource utilization, and by Farrell
(1957) as a measure of technical efficiency. This property has both a theoretical
and a practical significance. It allows the direct input distance function to serve
two important roles, simultaneously. It provides a complete characterization
of the structure of multi-input, multi-output efficient production technology,
and a reciprocal measure of the distance from each producer to that efficient
technology.
The main role played by the direct input distance function is to gauge tech-
nical efficiency. Nevertheless, it can also be used to construct input quantity
indexes (Tornqvist, 1936; Malmquist, 1953) and productivity indexes (Caves,
Christensen, and Diewert, 1982). Similarly, the direct output distance func-
tion introduced by Shephard (1970) and the two indirect distance functions of
Shephard (1974) can be used to characterize the structure of efficient production

1 This section is based on Färe, Grosskopf and Lovell (1994), pp. 1-23; and Kumbhakar and Lovell (2000),

pp. 5-7.
A short history of thought 17

technology in the multi-product case, to measure efficiency to that technology,

and to construct output quantity indexes (Bergson,1961; Moorsteen, 1961) and
productivity indexes (Färe, Grosskopf, and Lovell, 1992).
Linear programming theory is a milestone of efficiency analysis. The work
of Dantzig (1963) is closely associated with linear programming since he con-
tributed to the basic computational algorithm (the simplex method) used to
solve this problem. Charnes and Cooper (1961) made considerable contribu-
tions to both theory and application in the development of linear programming,
and popularize its application in DEA in the late 70s (see Charnes, Cooper and
Rhodes, 1978). Forsund and Sarafoglou (2002) offer an interesting historical
reconstruction of the literature developments subsequent to Farrell’s seminal
paper that lead to the introduction of the DEA methodology.
The use of linear programming and activity analysis can be found in the work
of Leontief (1941, 1953) who developed a special case of activity analysis which
has come to be known as input-output analysis. Whereas Leontief’s work was
directed toward constructing a workable model of general equilibrium, effi-
ciency and productivity analysis is more closely related to the microeconomic
production programming models developed by Shephard (1953, 1970, 1974),
Koopmans (1951, 1957) and Afriat (1972). In these models observed activities,
such as the inputs and outputs of some production units, serve as coefficients
of activity or intensity variables forming a series of linear inequalities, yielding
a piecewise linear frontier technology.
The work of Koopmans and Shephard imposes convexity on the reference
technology, therefore, the DEA estimator relies on the convexity assumption.
The Free Disposal Hull (FDH) estimator, that maintains free disposability while
relaxes convexity, was introduced by Deprins, Simar and Tulkens (1984).
By enveloping data points with linear segments, the programming approach
reveals the structure of frontier technology without imposing a specific func-
tional form on either technology or deviations from it.
Frontier technology provides a simple means of computing the distance to
the frontier - as a maximum feasible radial contraction or expansion of an ob-
served activity. This means of measuring the distance to the frontier yields an
interpretation of performance or efficiency as maximal-minimal proportionate
feasible changes in an activity given technology. This explanation is consis-
tent with Debreu’s (1951) coefficient of resource utilization and with Farrell’s
(1957) efficiency measures. However, neither Debreu nor Farrell formulated
the efficiency measurement problem as a linear programming problem, even
though Farrell and Fieldhouse (1962) envisaged the role of linear program-
ming. The full development of linear programming techniques took place later.
Boles (1966), Bressler (1966), Seitz (1966) and Sitorius (1966) developed the
piecewise linear case, and Timmer (1971) extended the piecewise log-linear
case.
18 The measurement of efficiency

Linear programming techniques are also used in production analysis for non-
parametric ‘tests’2 on regularity conditions and behavioral objectives. Afriat
(1972) developed a series of consistency ‘tests’ on production data by assuming
an increasing number of more restrictive regularity hypotheses on production
technology. In so doing he expanded his previous work on utility functions
(Afriat 1967) based on the revealed preference analysis (Samuelson, 1948).
These ‘tests’ of consistency, as well as similar ‘tests’ of hypotheses proposed
by Hanoch and Rothschild (1972), are all based on linear programming formu-
lations. Diewert and Parkan (1983) suggested that this battery of tools could be
used as a screening device to construct frontiers and measure efficiency of data
relative to the constructed frontiers. Varian (1984, 1985, 1990) and Banker and
Maindiratta (1988) extended the Diewert and Parkan approach. In particular,
Varian seeks to reduce the “all-or-nothing” nature of the tests - either data pass
a test or they do not - by developing a framework for allowing small failures to
be attributed to measurement in the data rather than to failure of the hypothesis
under investigation.
All these studies use nonparametric linear programming models to explore
the consistency of a dataset, or a subset of a dataset, with a structural (e.g.
constant return to scale) or parametric (e.g. Cobb-Douglas) or behavioral (e.g.
cost minimization) hypothesis. These tools, originally proposed as screening
devices to check for data accuracy, provide also guidance in the selection of
parametric functional forms as well as procedures useful to construct frontiers
and measure efficiency. The problem of nonparametric exploration of regularity
conditions and behavioral objectives has been treated also by Chavas and Cox
(1988, 1990), Ray (1991), and Ray and Bhadra (1993).
Some works have indirectly influenced the development of the efficiency
and productivity analysis. Hicks (1935, p.8) states his “easy life” hypothesis
as follows: “people in monopolistic positions [...] are likely to exploit their
advantage much more by not bothering to get very near the position of maximum
profit, than by straining themselves to get very close to it. The best of all
monopoly profits is a quite life”. The suggestion of Hicks, i.e. the fact that
the absence of competitive pressure might allow producers the freedom to not
fully optimize conventional objectives, and, by implication, that the presence
of competitive pressure might force producers to do so, has been adopted by
many authors (see e.g. Alchian and Kessel, 1962, and Williamson, 1964).
Another field of work, related to efficiency literature, is the property rights
field of research, which asserts that public production is inherently less effi-
cient than private production. This argument, due originally to Alchian (1965),
states that concentration and transferability of private ownership shares create

2 Here and below when we use the word test between quotation mark we mean qualitative indicators that are

not real statistical test procedures.

The economic model 19

an incentive for private owners to monitor managerial performance, and that

this incentive is diminished for public owners, who are dispersed and whose
ownership is not transferable. Consequently, public managers have wider free-
dom to pursue their owns at the expense of conventional goals. Thus Niskanen
(1971) argued that public managers are budget maximizers, de Alessi (1974)
argued that public managers exhibit a bias toward capital-intensive budgets,
and Lindsay (1976) argued that public managers exhibit a bias toward “visi-
ble” inputs. However, ownership forms are more varied than just private or
public. Hansmann (1988), in facts, identifies investor-owned firms, customer-
owned firms, worker-owned firms, as well as firms without owners (nonprofit
enterprisers). Each of them deals in a different way with problems associated
with hierarchy, coordination, incomplete contracts and monitoring and agency
costs. This leads to the expectation that different ownership forms will generate
differences in performance.3
As a more micro level is concerned, Simon (1955, 1957) analyzed the per-
formance of producers in the presence of bounded rationality and satisfying
behavior. Later Leibenstein (1966, 1975, 1976, 1978, 1987) argued that pro-
duction is bound to be inefficient as a result of motivation, information, mon-
itoring, and agency problems within the firm. This type of inefficiency, the
so called “X-inefficiency” has been criticized by Stigler (1976) and de Alessi
(1983) among others since it reflects an incompletely specified model rather
than a failure to optimize.
The problem of model specification - including a complete list of inputs
and outputs, and perhaps conditioning variables as well, a list of constraints,
technological, and other (e.g. regulatory) is a difficult issue to face. Among
others, Banker, Chang and Cooper (1996) analyse the effects of misspecified
variables in DEA. Simar and Wilson (2001) propose a statistical procedure to
test for the relevance of inputs/outputs in DEA models.
This literature suggests that the development of efficiency analysis is par-
ticularly useful if and when it could be used to shed empirical light on the
theoretical issues outlined above.

2.3 The economic model

In this paragraph we describe the main axioms on which the economic model
underlined the measurement of efficiency is based on.4
Much empirical evidence suggests that although producers may indeed at-
tempt to optimize, they do not always succeed. Not all producers are always so
successful in solving their optimization problems. Not all producers succeed

3 This expectation is based on a rich theoretical literature. See e.g. the “classical” survey by Holmstrom and

Tirole (1989).
4 See also Färe and Grosskopf (2004), pp.151-161.
20 The measurement of efficiency

in utilizing the minimum inputs required to produce the outputs they choose
to produce, given the technology at their disposal. In light of the evident fail-
ure of at least some producers to optimize, it is desirable to recast the analysis
of production away from the traditional production function approach toward
a frontier based approach. Hence we are concerned with the estimation of
frontiers, which envelop data, rather than with functions, which intersect data.
In this setting, the main purpose of productivity analysis studies is to evaluate
numerically the performance of a certain number of firms (or business units or
Decision Making Units, DMU) from the point of view of technical efficiency,
i.e. their ability to operate close to, or on the boundary of their production set.
The problem to be analyzed is thus set in terms of physical input and output
quantities.
We assume to have data in cross-sectional form, and for each firm we have
the value of its inputs and outputs used in the production process. Measuring
efficiency for any data set of this kind requires first to determine what the
boundary of the production set can be; and then to measure the distance between
any observed point and the boundary of the production set.
Given a list of p inputs and q outputs, in economic analysis the operations of
any productive organization can be defined by means of a set of points, Ψ, the
production set, defined as follows in the Euclidean space Rp+q + :

Ψ = {(x, y) | x ∈ Rp+ , y ∈ Rq+ , (x, y) is feasible}, (2.1)

where x is the input vector, y is the output vector and “feasibility” of the vec-
tor (x, y) means that, within the organization under consideration, it is physi-
cally possible to obtain the output quantities y1 , ..., yq when the input quantities
x1 , ..., xp are being used (all quantities being measured per unit of time). It is
useful to define the set Ψ in terms of its sections, defined as the images of a
relation between the input and the output vectors that are the elements of Ψ.
We can define then the input requirement set (for all y ∈ Ψ) as:

C(y) = {x ∈ Rp+ |(x, y) ∈ Ψ}. (2.2)

An input requirement set C(y) consists of all input vectors that can produce the
output vector y ∈ Rq+ .
The output correspondence set (for all x ∈ Ψ) can be defined as:

P (x) = {y ∈ Rq+ |(x, y) ∈ Ψ}. (2.3)

P (x) consists of all output vectors that can be produced by a given input vector
x ∈ Rp+ .
The production set Ψ can also be retrieved from the inputs sets, specifically:

Ψ = {(x, y) | x ∈ C(y), y ∈ Rq+ }. (2.4)

The economic model 21

Furthermore, it holds that:

(x, y) ∈ Ψ ⇔ x ∈ C(y), y ∈ P (x), (2.5)
which tells us that the output and input sets are equivalent representations of
the technology, as is Ψ.
The isoquants or efficient boundaries of the sections of Ψ can be defined in
radial terms (Farrell, 1957) as follows. In the input space:
∂C(y) = {x|x ∈ C(y), θx ∈ C(y), ∀θ, 0 < θ < 1)} (2.6)
and in the output space:
∂P (x) = {y|y ∈ P (x), λy ∈ P (x), ∀λ > 1}. (2.7)

The axiomatic approach to production theory (Activity Analysis framework)

assumes that the technology (production model) satisfies certain properties or
axioms. These properties can be equivalently stated on Ψ, P (x), x ∈ Rp+ ,
C(y), y ∈ Rq+ .
Some economic axioms (EA) are usually done in this framework (on these
concepts see also Shephard, 1970).

EA1: No free lunch. (x, y) ∈ Ψ if x = 0, y ≥ 0, y = 0.5

This axiom states that inactivity is always possible, i.e., zero output can be
produced by any input vector x ∈ Rp+ , but it is impossible to produce output
without any inputs.

EA2: Free disposability. Let x
∈ Rp+ and y
∈ Rq+ , with x
≥ x and y
≤

y, if (x, y) ∈ Ψ then (x
, y) ∈ Ψ and (x, y
) ∈ Ψ.
This is the free disposability assumption, named also the ‘possibility of destroy-
ing goods without costs’, on the production set Ψ.
The free disposability (also called strong disposability) of outputs can be
stated as follows: y1 ∈ P (x), y2 ≤ y1 then y2 ∈ P (x) or equivalently y1 ≤ y2
then C(y2 ) ⊆ C(y1 ). The free disposability of inputs can be defined as below:
x1 ∈ C(y), x2 ≥ x1 then x2 ∈ C(y) or equivalently x1 ≤ x2 then P (x1 ) ⊆
P (x2 ).
The free disposability of both inputs and outputs is as follows:
∀(x, y) ∈ Ψ, if x
≥ x and y
≤ y then (x
, y
) ∈ Ψ.
We have also a weak disposability of inputs and outputs:

5 Here and throughout inequalities involving vectors are defined componentwise, i.e. on an element-by-

element basis.
22 The measurement of efficiency

Weak disposability of inputs:

x ∈ C(y) ⇒ ∀α ≥ 1, αx ∈ C(y) or P (x) ⊆ P (αx);

Weak disposability of outputs:

y ∈ P (x) ⇒ ∀α ∈ [0, 1], αy ∈ P (x) or C(αy) ⊆ C(y).

The weak disposability property allows us to model congestion and overuti-

lization of inputs/outputs.

EA3: Bounded. P (x) is bounded ∀x ∈ Rp+ .

EA4: Closeness. Ψ is closed, P (x) is closed, ∀x ∈ Rp+ , C(y) is closed,

∀y ∈ Rq+ .

EA5: Convexity. Ψ is convex. The convexity of Ψ can be stated as follows:

If (x1 , y1 ), (x2 , y2 ) ∈ Ψ, then ∀α ∈ [0, 1] we have :

(x, y) = α(x1 , y1 ) + (1 − α)(x2 , y2 ) ∈ Ψ.

EA6: Convexity of the requirement sets. P (x) is convex ∀x ∈ Rp+

and C(y) is convex ∀y ∈ Rq+ .
If Ψ is convex, then the inputs and outputs sets are also convex, i.e. EA5 implies
EA6.
A further characterization of the shape of the frontier relates to returns to
scale (RTS). According to a standard definition in economics, RTS express the
relation between a proportional change in inputs to a productive process and
the resulting proportional change in output. If an n per cent rise in all inputs
produces an n per cent increase in output, there are constant returns to scale
(CRS). If output rises by a larger percentage than inputs, there are increasing
returns to scale (IRS). If output rises by a smaller percentage than inputs, there
are decreasing returns to scale (DRS). Returns to scale can be described as
properties of the correspondence sets C(y) and/or P (x). We follow here the
presentation of Simar and Wilson (2002, 2006b). The frontier exhibits constant
returns to scale (CRS) everywhere if and only if:

∀(x, y) s.t. x ∈ ∂C(y) then αx ∈ ∂C(αy), ∀α > 0

or equivalently6 ,
∀α > 0, C(αy) = αC(y).

6 Analogous expressions hold in terms of P (x): ∀α > 0, P (αx) = αP (x).

The economic model 23

Constant Returns to Scale in the neighborhood of a point (x, y) s.t. x ∈

∂C(y) are characterized by C(αy) = αC(y) for some α > 0.
Increasing Returns to Scale in the neighborhood of a point (x, y) s.t. x ∈
∂C(y) implies that (αx, αy) ∈ Ψ for α < 1.
Decreasing Returns to Scale in the neighborhood of a point (x, y) s.t. x ∈
∂C(y) implies that (αx, αy) ∈ Ψ for α > 1.

A frontier that exhibits increasing, constant and decreasing returns to scale

in different regions is a Variable Returns to Scale (VRS) frontier.
The assumptions we have introduced here are intended to provide enough
structure to create meaningful and useful technologies. Generally speaking, we
will not impose all of these axioms on a particular technology, rather we will
select subsets of these assumptions that are suitable for the particular problem
under study.
***
Turning back to the production set itself, the above definitions allow us to
characterize any point (x, y) in Ψ as:

input efficient if x ∈ ∂C(y)

input inefficient if x ∈ ∂C(y)
output efficient if y ∈ ∂P (x)
output inefficient if y ∈ ∂P (x).

From what stated above, DMUs are efficient, e.g. in an input-oriented frame-
work, if they are on the boundary of the input requirement set (or, for the output
oriented case, on the boundary of the output correspondence set). In some cases,
however, these efficient firms may not be using the fewest possible inputs to
produce their outputs. This is the case where we have slacks. This is due to the
fact that the Pareto-Koopmans efficient subsets of the boundaries of C(y) and
P (x), i.e. eff C(y) and eff P (x), may not coincide with the Farrell-Debreu
boundaries ∂C(y) and ∂P (x), i.e.7 :
 
eff C(y) = x | x ∈ C(y), x
∈ C(y) ∀x
≤ x, x
= x ⊆ ∂C(y), (2.8)

 
eff P (x) = y | y ∈ P (x), y
∈ P (x) ∀y
≥ y, y
= y ⊆ ∂P (x). (2.9)

7 We give an illustration in Section 2.5 in Figure 2.2 where we describe DEA estimators of efficient frontier.
24 The measurement of efficiency

Once the efficient subsets of Ψ have been defined, we may define the efficiency
measure of a firm operating at the level (x0 , y0 ) by considering the distance
from this point to the frontier. There are several ways to achieve this but a
simple way suggested by Farrell (1957), in the lines of Debreu (1951), is to use
a radial distance from the point to its corresponding frontier. In the following
we will concentrate our attention on radial measures of efficiency. Of course, we
may look at the efficient frontier in two directions: either in the input direction
(where the efficient subset is characterized by ∂C(y)) or in the output direction
(where the efficient subset is characterized by ∂P (x)).
The Farrell input measure of efficiency for a firm operating at level (x0 , y0 )
is defined as:
θ(x0 , y0 ) = inf{θ|θx0 ∈ C(y0 )} = inf{θ|(θx0 , y0 ) ∈ Ψ}, (2.10)
and its Farrell output measure of efficiency is defined as:
λ(x0 , y0 ) = sup{λ|λy0 ∈ P (x0 )} = sup{λ|(x0 , λy0 ) ∈ Ψ}. (2.11)
So, θ(x0 , y0 ) ≤ 1 is the radial contraction of inputs the firm should achieve
to be considered as being input-efficient in the sense that (θ(x0 , y0 )x0 , y0 ) is a
frontier point. In the same way λ(x0 , y0 ) ≥ 1 is the proportionate increase of
output the firm should achieve to be considered as being output efficient in the
sense that (x0 , λ(x0 , y0 )y0 ) is on the frontier.
It is interesting to note that the efficient frontier of Ψ, in the radial sense, can
be characterized as the units (x, y) such that θ(x, y) = 1, in the input direction
(belonging to ∂C(y)) and by the (x, y) such that λ(x, y) = 1, in the output
direction (belonging to ∂P (x)). If the frontier is continuous, frontier points are
such that θ(x, y) = λ(x, y) = 1. The efficient frontier is unique but we have
two ways to characterize it.
It is sometimes easier to measure these radial distances by their inverse,
known as Shephard distance functions (Shephard, 1970). The Shephard input
distance function provides a normalized measure of Euclidean distance from a
point (x, y) ∈ Rp+q+ to the boundary of Ψ in a radial direction orthogonal to y
and is defined as:
δ in (x, y) = sup{θ > 0|(θ−1 x, y) ∈ Ψ} ≡ (θ(x, y))−1 , (2.12)
with δ in (x, y) ≥ 1, ∀(x, y) ∈ Ψ. Similarly, the Shephard output distance
function provides a normalized measure of Euclidean distance from a point
(x, y) ∈ Rp+q
+ to the boundary of Ψ in a radial direction orthogonal to x:

δ out (x, y) = inf{λ > 0|(x, λ−1 y) ∈ Ψ} ≡ (λ(x, y))−1 . (2.13)

For all (x, y) ∈ Ψ, δ out (x, y) ≤ 1. If either δ in (x, y) = 1 or δ out (x, y) = 1
then (x, y) belongs to the frontier of Ψ and the firm is technically efficient.
A taxonomy of efficient frontier models 25

As pointed out in Simar and Wilson (2001), no behavioral assumptions are

necessary for measuring technical efficiency. From a purely technical view-
point, either the input or the output distance function can be used to measure
technical efficiency - the only difference is in the direction in which distance to
the technology is measured. The way of looking at the frontier will typically
depend on the context of the application. For instance, if the outputs are ex-
ogenous and not under the control of the Decision Makers (e.g. as in most of
the public services), input efficiency will be of main interest, since the inputs
are the only elements under the control of the managers. But even in this case,
both measures are available.

2.4 A taxonomy of efficient frontier models

The analysis of the existent literature is a necessary step for the advancement
of a discipline. This is particularly true for the field of efficiency and produc-
tivity research that in the last decades has known an exponential increasing
in the number of methodological and applied works. For a DEA bibliogra-
phy over 1978-1992, see Seiford (1994, 1996) and for an extension till 2001
see Gattoufi, Oral and Reisman (2004). In Cooper, Seiford and Tone (2000)
about 1,500 DEA references are reported. Other bibliographic studies include:
Emrouznejad (2001) and Taveres (2002).

As a consequence, a comprehensive review of the overall literature would

require another whole work. Therefore, the aim of this section is to propose
a general taxonomy of efficient frontier models that gives an overview on the
different approaches presented in literature for estimating the efficient frontier
of a production possibility set. Here the review could be biased toward the
nonparametric approach, due to our commitment and involvement with non-
parametric methods most. Anyway, we give several references also on the
parametric approach that could be useful for those interested in it.

In the previous section we described the economic model underlying the

frontier analysis framework based on the Activity Analysis Model. This model
is based on some representations of the production set Ψ on which we can
impose different axioms. Nevertheless, the production set Ψ, the boundary of
the input requirement set ∂C(y) and of the output correspondence set ∂P (x),
together with the efficiency scores in the input and output space, θ(x, y) and
λ(x, y), are unknown.

The econometric problem is thus how to estimate Ψ, and then ∂C(y), ∂P (x),
θ(x, y), λ(x, y), from a random sample of production units X = {(Xi , Yi ) | i =
1, ..., n}.
26 The measurement of efficiency

Starting from the first empirical application of Farrell (1957) several different
approaches for efficient frontier estimation and efficiency score calculation have
been developed.8
In Figure 2.1 we propose an outline of what we believe have been the most
influential works in productivity and efficiency analysis, starting from the pi-
oneering work by Farrell (1957). Of course, our outline is far from being
complete and all-inclusive. Figure 2.1 shows some of the articles, books and
special issues of journals (i.e. Journal of Econometrics JE, Journal of Produc-
tivity Analysis JPA, European Journal of Operational Research, EJOR) that
have mainly influenced the writing of this work, trying to balance them accord-
ing to the adopted approach.
As it is evident from Figure 2.1 we have taken into consideration mainly the
nonparametric approach as we believe that thanks to its last developments, it
can be considered as being very flexible and very useful for modeling purpose.
We may classify efficient frontier models according to the following criteria:9
1 The specification of the (functional) form for the frontier function;
2 The presence of noise in the sample data;
3 The type of data analyzed.
Based on the first criterium (functional form of the frontier) is the classifi-
cation in:
Parametric Models. In these models, the attainable set Ψ is defined trough
a production frontier function, g(x, β), which is a known mathematical
function depending on some k unknown parameters, i.e. β ∈ Rk , where
generally y is univariate, i.e. y ∈ R+ . The main advantages of this
approach are the economic interpretation of parameters and the statistical
properties of estimators; more critical are the choice of the function g(x, β)
and the handling of multiple inputs, multiple outputs cases (for more on
this latter aspect see Section 4.7 below where we introduce multivariate
parametric approximations of nonparametric and robust frontiers).
Nonparametric Models. These models do not assume any particular func-
tional form for the frontier function g(x). The main pros of this approach
are the robustness to model choice and the easy handling of multiple in-
puts, multiple outputs case; their main limitations are the estimation of
unknown functional and the curse of dimensionality10 , typical of nonpara-
metric methods.

8 Foran introduction see e.g., Coelli, Rao and Battese (1998) and Thanassoulis (2001).
9 These criteria follow Simar and Wilson (2006b), where a comprehensive statistical approach is described.
10 The curse of dimensionality, shared by many nonparametric methods, means that to avoid large variances

and wide confidence interval estimates a large quantity of data is needed.

 Nonparametric
approach
1978 1984 1990 1992 1993 1994 1995 1996 1998 1999 2000 2001 2002 2003 2004

Charnes Banker Special Simar Banker Fare Korostelev Banker Kneip Cooper Park Simar Cazals Simar Fare
Cooper Charnes issue Grosskopf Simar Park Seiford Simar Wilson Florens Grosskopf
Florens 2005
Rhodes Cooper JE Tulkens Lovell Tsybakov Grosskopf Simar Tone Weiner Simar
Simar
Ray
Special Simar Simar Gijbels Simar Hall
Deprins Sickles
issues Wilson MammenWilson Simar
Simar JPA / Kneip Park
Tulkens Daraio
EJOR Simar Simar Simar
Simar
Wilson 2006
Simar Fried
Wilson Lovell
Debreu 1951 Schmidt
A taxonomy of efficient frontier models

(eds.)
Koopmans 1951
Kumbhakar
1994 1998 2003 a,b Semiparametric Park
Farrell 1957 Simar
Park and Simar Park, Sickles and Simar Park, Sickles and Simar approach
Tsionas
Shephard 1953
Simar
Shephard 1970 Wilson

Meeusen
van den Broeck
van den Broeck
Aigner Koop
Aigner COLS Lovell Osiewalski Ritter Kumbhakar
Chu Afriat Richmond Schmidt Stevenson Greene Steel Simar Lovell

1968 1972 1974 1977 1980 1990 1994 1997 2000

Figure 2.1. An overview of the literature on efficient frontier estimation.

Parametric
approach
27
28 The measurement of efficiency

Based on the second criterium (presence of noise) is the classification in:

Deterministic Models, which assume that all observations (Xi , Yi ) belong
to the production set, i.e.

P rob{(Xi , Yi ) ∈ Ψ} = 1

for all i = 1, ..., n. The main weakness of this approach is the sensitivity
to “super-efficient” outliers. Robust estimators are able to overcome this
drawback.
Stochastic Models, in which there might be noise in the data, i.e. some
observations might lie outside Ψ. The main problem of this approach is
the identification of noise from inefficiency.
Based on the third criterium (type of data analyzed) is the classification in:
Cross-sectional Models, in which the data sample is done by observations
on n firms or DMUs (Decision Making Units):

X = {(Xi , Yi )|i = 1, ..., n}

Panel Data Models, in which the observations on the n firms are available
over T periods of time:

X = {(Xit , Yit ) | i = 1, ..., n; t = 1, ..., T }.

Panel data allow the measurement of productivity change as well as the

estimation of technical progress or regress.
Generally speaking, productivity change occurs when an index of outputs
changes at a different rate than an index of inputs does. Productivity change
can be calculated using index number techniques to construct a Fisher (1922)
or Tornqvist (1936) productivity index. Both these indices require quantity and
price information, as well as assumptions concerning the structure of technology
and the behavior of producers. Productivity change can also be calculated using
nonparametric techniques to construct a Malmquist (1953) productivity index.
These latter techniques do not require price information or technological and
behavioral assumptions, but they require the estimation of a representation of
production technology. Nonparametric techniques are able not only to calculate
productivity change, but also to identify the sources of measured productivity
change.
A survey of the theoretical and empirical work on Malmquist productivity
indices can be found in Färe, Grosskopf and Russell (1998). On the theo-
retical side the survey includes a number of issues that have arisen since the
Malmquist productivity index was proposed by Caves, Christensen and Diewert
A taxonomy of efficient frontier models 29

(1982). These issues include the definition of the Malmquist productivity in-
dex; although all are based on the distance functions that Malmquist employed
to formulate his original quantity index, variations include the geometric mean
form used by Färe, Grosskopf, Lindgren and Roos (1989) and the quantity in-
dex form by Diewert (1992). The survey of the empirical literature presents
studies on the public sector, banking, agriculture, countries and international
comparisons, electric utilities, transportation, and insurance. See also Lovell
(2003), and Grosskopf (2003) for an historical perspective and an outline of the
state of the art in this area.
Although productivity change is not the main focus of FDH, it can be inferred
from information on efficiency change and technical change that is revealed by
FDH. The technique was developed by Tulkens that named it “sequential FDH”.
For an illustration of the sequential FDH see Lovell (1993, pp. 48-49). On this
topic see also Tulkens and Vanden Eeckaut (1995a, 1995b).
By combining the three criteria mentioned above, several models have been
studied in literature:
Parametric Deterministic Models, see e.g. Aigner and Chu (1968), Afriat
(1972), Richmond (1974), Schmidt (1976) and Greene (1980) for cross-
sectional and panel data;
Parametric Stochastic Models, most of these techniques are based on the
maximum likelihood principle, following the pioneering works of Aigner,
Lovell and Schmidt (1977) and Meeusen and van den Broeck (1977). For
a recent review see Kumbhakar and Lovell (2000). In the context of
panel data, stochastic models (see Schmidt and Sickles, 1984, and Corn-
well, Schmidt, and Sickles, 1990) have semiparametric generalizations,
in which a part of the model is parametric and the rest is nonparamet-
ric (see Park and Simar, 1994; Park, Sickles and Simar, 1998; and Park,
Sickles and Simar, 2003a, b).
Nonparametric Deterministic Models for cross-sectional and panel data.
Traditional references on these models include: Färe, Grosskopf and
Lovell (1985, 1994), Fried, Lovell and Schmidt (1993), and Charnes,
Cooper, Lewin and Seiford, 1994. Recent and updated references are
Cooper, Seiford and Tone (2000), Ray (2004) and Färe and Grosskopf
(2004).
Nonparametric Stochastic Models for cross-sectional data (see Hall and
Simar, 2002; Simar, 2003b; Kumbhakar, Park, Simar and Tsionas, 2004)
and panel data (see Kneip and Simar, 1996; and Henderson and Simar,
2005).
The mainly used approaches in empirical works are the nonparametric (de-
terministic) frontier approach and the (parametric) stochastic frontier approach.
30 The measurement of efficiency

In the following, when we refer to nonparametric frontier approach we indicate

the deterministic version of it; when we talk about stochastic frontier approach
we refer to its parametric version.
The nonparametric frontier approach, based on envelopment techniques
(DEA FDH), has been extensively used for estimating efficiency of firms as
it relays only on very few assumptions for Ψ. On the contrary, the stochastic
frontier approach (SFA) allows the presence of noise but it demands parametric
restrictions on the shape of the frontier and on the Data Generating Process
(DGP) in order to permit the identification of noise from inefficiency and the
estimation of the frontier. Fried, Lovell and Schmidt (2006) offer an updated
presentation of both approaches. A statistical approach which unifies paramet-
ric and nonparametric approaches can be found in Simar and Wilson (2006b).

2.5 The nonparametric frontier approach

In this section we introduce the most known nonparametric estimators of
efficient frontiers.
As we have seen in Section 2.3 devoted to the presentation of the economic
model, we can equivalently look at the efficient boundary of Ψ from the input
space or from the output space.
The input oriented framework, based on the input requirement set and its
efficient boundary, aims at reducing the input amounts by as much as possible
while keeping at least the present output levels. This is also called “input-
saving” approach to stress the fact that the outputs level remains unchanged
and input quantities are reduced proportionately till the frontier is reached.
This is a framework generally adopted when the decision maker can control the
inputs but has not the control of the outputs. For instance, this is the case of
public enterprises which are committed to offer some public services and are
interested in the management of the inputs, in the sense of their minimization.
Alternatively, we can take into account the output space and look at the output
correspondence set and its efficient boundary. The output oriented framework
looks at maximize output levels under at most the present input consumption.
This approach is also known as “output-augmenting” approach, because it holds
the input bundle unchanged and expand the output level till the frontier is
reached. In practice, whether the input or output-oriented measure is more
appropriate would depend on whether input conservation is more important
than output augmentation.
For the relation existent among input and output efficiency measures, see
Deprins and Simar (1983).
The main nonparametric estimators available are the Data Envelopment
Analysis (DEA) and the Free Disposal Hull (FDH) which we describe in the
subsections that follow.
The nonparametric frontier approach 31

2.5.1 Data Envelopment Analysis (DEA)

The DEA estimator of the production set, initiated by Farrell (1957) and op-
erationalized as linear programming estimators by Charnes, Cooper and Rhodes
(1978), assumes the free disposability and the convexity of the production set
Ψ. It involves measurement of efficiency for a given unit (x, y) relative to the
boundary of the convex hull of X = {(Xi , Yi ), i = 1, ...., n}:
 
n 
n
 DEA = (x, y) ∈ Rp+q
Ψ + |y ≤ γi Yi ; x ≥ γi Xi , for (γ1 , ..., γn )
i=1 i=1

n 
s.t. γi = 1; γi ≥ 0, i = 1, ...., n (2.14)
i=1

 DEA is thus the smallest free disposal convex set covering all the data.
Ψ
The Ψ DEA in (2.14) allows for Variable Returns to Scale (VRS) and is often
 DEA−V RS (see Banker, Charnes and Cooper, 1984). It may be
referred as Ψ
adapted to other returns to scale situations. It allows for:
n
Constant Returns to Scale (CRS) if the equality constrained i=1 γi =1
in (2.14) is dropped;
n
Non Increasing Returns to Scale

(NIRS) if the equality constrained i=1 γi
= 1 in (2.14) is changed in ni=1 γi ≤ 1;
Non
n
Decreasing Returns to Scale (NDRS)
n
if the equality constrained
i=1 γi = 1 in (2.14) is modified in i=1 γi ≥ 1.

The estimation of the input requirement set is given for all y by: C(y) =
p  
{x ∈ R+ |(x, y) ∈ ΨDEA } and ∂ C(y) denotes the estimator of the input
frontier boundary for y.
For a firm operating at level (x0 , y0 ) the estimation of the input efficiency
score θ(x0 , y0 ) is obtained by solving the following linear program (here and
hereafter we consider the VRS case):
 
θDEA (x0 , y0 ) = inf θ | (θx0 , y0 ) ∈ Ψ
 DEA (2.15)

 
n 
n

θDEA (x0 , y0 ) = min θ | y0 ≤ γi Yi ; θx0 ≥ γi Xi ; θ > 0;
i=1 i=1

n 
γi = 1; γi ≥ 0; i = 1, ...., n . (2.16)
i=1

 0 , y0 ) measures the radial distance between (x0 , y0 ) and (x

θ(x ∂ (x0 |y0 ), y0 )
where x (x0 |y0 ) is the level of the inputs the unit should reach in order to
∂
32 The measurement of efficiency

be on the “efficient boundary” of Ψ  DEA with the same level of output, y0 ,

and the same proportion of inputs; i.e. moving from x0 to x ∂ (x0 |y0 ) along
the ray θx0 . The projection of x0 on the efficient frontier is thus equal to
 0 , y0 )x0 .
∂ (x0 |y0 ) = θ(x
x
For the output oriented case, the estimation is done, mutatis mutandis, fol-
lowing the previous steps. The output correspondence set is estimated by:
P (x) = {y ∈ Rq+ |(x, y) ∈ Ψ  DEA } and ∂ P (x) denotes the estimator of the
output frontier boundary for x.
The estimator of the output efficiency score for a given (x0 , y0 ) is obtained
by solving the following linear program:
 
 DEA (x0 , y0 ) = sup λ | (x0 , λy0 ) ∈ Ψ
λ  DEA , (2.17)

 
n 
n
 DEA (x0 , y0 ) = max λ | λy0 ≤
λ γi Yi ; x0 ≥ γi Xi ; λ > 0;
i=1 i=1
n 
γi = 1; γi ≥ 0; i = 1, ...., n . (2.18)
i=1

In Figure 2.2 we display the DEA estimator and illustrate the concept of slacks
through an example. If we look at the left panel assuming that all firms produce
the same level of output, we can see that the DMU E could actually produce 1
unit of y with less input x1 , i.e., it could reduce x1 by one unit (from 4 to 3)
moving from E to D. This is referred to as input slack: although the DMU is
technical efficient, there is a surplus of input x1 .11 In general, we say that there
is slack in input j of DMU i, i.e., xji , if:

n
 i , yi )
γi xi < xji θ(x (2.19)
i=1

is true for some solution value of γi , i = 1, ..., n (see Färe, Grosskopf and
Lovell, 1994, for more details).
The same kind of reasoning can be done for the output oriented case, i.e. the
DMU L could increase the production of y1 moving from L to M. See Figure
2.2, right panel for a graphical illustration.
Slacks may happen for DEA estimates (as shown in Figure 2.2), as well as
for FDH estimates (presented in the next section). It is interesting to note that
if the true production set Ψ has no slacks, than slacks are only a small sample
problem. Nevertheless, it is always useful to report slacks whenever they are

11 Remember the “possibility of destroying goods without costs” underlying the frontier representation of
the economic model.
The nonparametric frontier approach 33

Slacks − DEA input oriented Slacks − DEA output oriented

X2 Y2 M
L

P(x)

C(y)

D E

X1 Y1
O 3 4 O

Figure 2.2. Input and Output slacks.

there. It is left to the analyst to decide if it is better to correct for the slacks or
just point them.
Once the efficiency measures have been computed, several interesting analy-
sis could be done, such as the inspection of the distribution of efficiency scores
and the analysis of the “best performers” or efficient facet of the frontier closer
to the analysed DMU, generally called peer-analysis, to study the technical
efficient units and try to learn from them.

2.5.2 Free Disposal Hull (FDH)

The FDH estimator, proposed by Deprins, Simar and Tulkens (1984), is a
more general version of the DEA estimator as it relies only on the free disposabil-
ity assumption for Ψ, and hence does not restrict itself to convex technologies.
This seems an attractive property of FDH since it is frequently difficult to find
a good theoretical or empirical justification for postulating convex production
sets in efficiency analysis. At this purpose, Farrell (1959) indicates indivisibil-
ity of inputs and outputs and economies of scale and specialization as possible
violations of convexity. It is important to note also that if the true production
set is convex then the DEA and FDH are both consistent estimators; however,
as pointed later in this section, FDH shows a lower rate of convergence (due to
the less assumptions it requires) with respect to DEA. On the contrary, if the
true production set is not convex, than DEA is not a consistent estimator of the
production set, while FDH is consistent.
The FDH estimator measures the efficiency for a given point (x0 , y0 ) relative
to the boundary of the Free Disposal Hull of the sample X = {(Xi , Yi ), i =
1, ...., n}. The Free Disposal Hull of the set of observations (i.e. the FDH
34 The measurement of efficiency

estimator of Ψ) is defined as:

 
 F DH = (x, y) ∈ Rp+q | y ≤ Yi ; x ≥ Xi , (Xi , Yi ) ∈ X .
Ψ (2.20)
+

It is the union of the all positive orthants in the inputs and of the negative orthants
in the outputs whose origin coincides with the observed points (Xi , Yi ) ∈
X (Deprins, Simar and Tulkens, 1984). See Figures 2.3 and 2.4 where the
FDH estimator is compared with the DEA estimator of the input and output
requirement sets, respectively.
The efficiency estimators, in this framework, are obtained (as for the DEA
case) using a “plug-in principle”, i.e., by substituting the unknown quantities (in
this case Ψ) by their estimated values (here Ψ  F DH , for the DEA case Ψ  DEA ).
The estimated input requirement set and the output correspondence set are
the following:

C(y)  F DH },
= {x ∈ Rp+ |(x, y) ∈ Ψ

P (x) = {y ∈ Rq+ |(x, y) ∈ Ψ

 F DH }.

Their respective efficient boundaries are:


∂ C(y) 
= {x|x ∈ C(y), 
θx ∈ C(y)∀0 < θ < 1},

∂ P (x) = {y|y ∈ P (x), λy ∈ P (x)∀λ > 1}.

Hence, the estimated input efficiency score for a given point (x0 , y0 ) ∈ Ψ is:
 
θF DH (x0 , y0 ) = inf θ | θx0 ∈ C(y
 0)
 
 F DH ,
= inf θ | (θx0 , y0 ) ∈ Ψ (2.21)

and the estimated output efficiency score of (x0 , y0 ) is given by:

 
 F DH (x0 , y0 ) = sup λ | λy0 ∈ P (x0 )
λ
 
 F DH .
= sup λ | (x0 , λy0 ) ∈ Ψ (2.22)

It is clear that for a particular point (x0 , y0 ), the estimated distance to the
frontiers are evaluated by means of the distance, in the input space (“input
oriented”) from this point to the estimated frontier of the input requirement

set (∂ C(y)), and in the output space (“output oriented”) by the distance from
(x0 , y0 ) to the estimated frontier of the output correspondence set (∂ P (x)).
The nonparametric frontier approach 35

It is worthwhile to note that the FDH attainable set in (2.20) can also be
characterized as the following set:

 
n 
n 
n
 F DH = (x, y) ∈ Rp+q | y ≤
Ψ γi Yi ; x ≥ γi Xi , γi = 1;
+
i=1 i=1 i=1

γi ∈ {0, 1}, i = 1, ..., n . (2.23)

Therefore the efficiencies can be estimated by solving the following integer

linear programs; for the input-oriented case we have:

 
n 
n 
n

θF DH (x0 , y0 ) = min θ | y0 ≤ γi Yi ; θx0 ≥ γi Xi , γi = 1;
i=1 i=1 i=1

γi ∈ {0, 1}, i = 1, ..., n , (2.24)

and for the output-oriented case:

 
n 
n 
n
 F DH (x0 , y0 ) = max λ | λy0 ≤
λ γi Yi ; x0 ≥ γi Xi , γi = 1;
i=1 i=1 i=1

γi ∈ {0, 1}, i = 1, ..., n .(2.25)

The latter expressions allow to make the comparison easier between the FDH
and the DEA estimators (compare for instance (2.23) with (2.14)).
Figure 2.3 illustrates the estimation of the input requirement set C(y) and
of its boundary ∂C(y) through FDH and DEA methods. The dashed line rep-
resents the FDH estimation of ∂C(y), while the solid line shows the DEA
estimation of it. The squares are the observations. The DEA and FDH esti-
mates of efficiency score of production unit B, in Figure 2.3, are respectively:
θDEA (x0 , y0 ) = |OB

|/|OB| ≤ 1, θF DH (x0 , y0 ) = |OB
|/|OB| ≤ 1.
In Figure 2.4 we show the FDH and DEA estimation of the output corre-
spondence set P (x) and its boundary ∂P (x). The dash-dotted line represents
the FDH estimator of ∂P (x), while the solid line the DEA estimator of it.
The black squares, as before, represent the DMUs. For firm B, the estimates
of its efficiency score, in output oriented framework, are: λ  F DH (x0 , y0 ) =




|OB |/|OB| ≥ 1, λDEA (x0 , y0 ) = |OB |/|OB| ≥ 1.

Practical computation of the FDH

In practice, the FDH estimator is computed by a simple vector compari-
son procedure that amounts to a complete enumeration algorithm proposed in
Tulkens (1993), which is now explained.
36 The measurement of efficiency

X2 Input Oriented Framework

B=(Xo,Yo)

A .
C(y)
B’

.
B”
∂ C(y)

O X1

Figure 2.3. FDH and DEA estimation of C(y) and ∂C(y).

Y2 Output oriented framework

B”

B’

B = (Xo,Yo)

P(x) ∂ P(x)

O Y1

Figure 2.4. FDH and DEA estimation of P (x) and ∂P (x).

The nonparametric frontier approach 37

For a DMU (x0 , y0 ), in a first step, the set of observations which domi-
nates it is determined, and then the estimate of its efficiency score, relative to
the dominating facet of Ψ  is computed. In the simplest case, with a technol-
ogy characterized by one input and one output, the set of observations which
dominate (x0 , y0 ) is defined as:
 
D0 = i|(Xi , Yi ) ∈ X , Xi ≤ x0 , Yi ≥ y0 . (2.26)

The “input oriented” efficiency estimate is done through:

X 
θF DH (x0 , y0 ) = min i∈D0
i
, (2.27)
x0
and the “output oriented” efficiency is computed via:
Y 
 F DH (x0 , y0 ) = max i∈D
λ
i
. (2.28)
0
y0

It has to be noted that as Xi ≤ x0 then θF DH ≤ 1. As for the input-oriented

case, from the fact that Yi ≥ y0 follows that λ F DH ≥ 1.
In a multivariate setting, the expression (2.21) can be computed through:
  X i,j 
θF DH (x0 , y0 ) = min i∈D0 max j=1,...,p , (2.29)
xj0

where X i,j is the j th component of X i ∈ Rp+ and xj0 is the j th component of

x0 ∈ Rp+ .
It is a maximin procedure (for the “input oriented” framework): the “max”
part of the algorithm identifies the most dominant DMUs relative to which a
given DMU is evaluated. Once the most dominant DMUs are identified, slacks
are calculated from the “min” part of the algorithm.
The multivariate computation of expression (2.22) is done by:
  Y i,j 
 F DH (x0 , y0 ) = max i∈D min j=1,...,q
λ (2.30)
0
y0j

where Y i,j is the j th component of Y i ∈ Rq+ and y0j is the j th component of

y0 ∈ Rq+ .
The FDH estimator has been applied in several contexts. For a detailed
presentation of FDH concepts see Vanden Eeckaut (1997).
Recently, some authors have raised explicit doubts about the economic mean-
ing of FDH, but from the exchange between Thrall (1999) and Cherchye, Ku-
osmanen and Post (2000), published on the Journal of Productivity Analysis, it
38 The measurement of efficiency

emerged that FDH can be economically more meaningful than convex monotone
hull, also under non-trivial alternative economic conditions.
Hence, FDH technical efficiency measures remain meaningful for theories
of the firm that do allow for imperfect competition or uncertainty (see e.g.
Kuosmanen and Post, 2001, and Cherchye, Kuosmanen and Post, 2001).
One of the main drawbacks of deterministic frontier models (DEA /FDH
based) is the influence of “super-efficient” outliers.
This is a consequence of the fact that the efficient frontier is determined by
sample observations which are extreme points. Simar (1996) points out the need
for identifying and eliminating outliers when using deterministic models. If they
cannot be identified, the use of stochastic frontier models is recommended.
See Figure 2.5 for an illustration of the influence of outliers in case of FDH
estimation. The same is valid for the DEA case. If point A is an extreme point,
outlying the cloud of other points, the estimated efficient frontier is strongly
influenced by it. In fact, in Figure 2.5, the solid line is the frontier that envelops
point A, while the dash-dotted line does not envelop point A.

FDH frontier of ψ
y
FDH frontier without A

ψFDH

Free disposability

Figure 2.5. Influence of outliers on the FDH estimation of the production set Ψ.

We will come back on this problem in Chapter 4 where we propose robust

nonparametric approaches based on various nonparametric measures less influ-
enced by extreme values and outliers, which have also nice statistical properties.
Recent developments in nonparametric efficiency analysis 39

2.6 Recent developments in nonparametric efficiency

analysis
In the following, we recall briefly some stream of works that have contributed
to the latest advancement of the nonparametric efficiency literature.12

Sensitivity of results to data variation and discrimination (in a DEA

framework)
The focus of studies on sensitivity and stability is the reliability of classifica-
tion of DMUs into efficient and inefficient performers. Most analytical methods
for studying the sensitivity of results to variations in data have been developed
in a DEA framework.
After a first stream of works concentrated on developing solution methods
and algorithms for conducting sensitivity analysis in linear programming, a
second current of studies analysed data variations in only one input or one
output for one unit at a time. A recent stream of works makes it possible to
determine ranges within which all data may be varied for any unit before a
reclassification from efficient to inefficient status (or vice versa) occurs, and
for determining ranges of data variation that can be allowed when all data are
varied simultaneously for all DMUs. For a review and some references see
Cooper, Li, Seiford, Tone, Thrall and Zhu (2001).
As we have seen above, DEA models have a deterministic nature, meaning
that they do not account for statistical noise. Some authors (e.g., Land, Lovell
and Thore, 1993; Olesen and Petersen, 1995) have proposed the application
of the chance-constrained programming to the DEA problem in order to over-
come its deterministic nature. The basic idea is that of make DEA stochastic
by introducing a chance that the constraints on either the envelopment problem
or the multiplier problem may be violated with some probability. However,
the chance-constrained efficiency measurement requires a large amount of data
in addition to inputs and outputs. Moreover, it is based on a strong distribu-
tional assumption on the process determining the chance of a constrained to
be violated. The analyst in fact has to provide also information on expected
values of all variables for all DMUs, and variance-covariance matrices for each
variable across all DMUs. An alternative to this approach is given by a fuzzy
programming approach to DEA and FDH efficiency measurement.
There is an increasing number of studies that apply the fuzzy set theory
in productivity and efficiency contexts. In some production studies, the data
that describe the production process cannot be collected accurately due to the
fact that measurement systems have not been originally designated for the pur-

12 Seealso Lovell (2001) and Fried, Lovell and Schmidt (2006) for a presentation of some recent fruitful
research areas introduced in parametric and nonparametric approaches to efficiency analysis.
40 The measurement of efficiency

pose of collecting data and information that are useful for production studies.
Sengupta (1992) was the first to introduce a fuzzy mathematical programming
approach where the constraints and objective function are not satisfied crisply.
Seaver and Triantis (1992) proposed a fuzzy clustering approach for identify
unusual or extreme efficient behavior. Girod and Triantis (1999) implemented
a fuzzy linear programming approach, whilst Triantis and Girod (1998), and
Kao and Liu (1999) used fuzzy set theory, to let the traditional DEA and FDH
account for inaccuracies associated with the production plans. A fuzzy pair-
wise dominance approach can be found in Triantis and Vanden Eeckaut (2000)
where, a classification scheme that explicitly accounts for the degree of fuzzi-
ness (plausibility) of dominating units is reported.
According to a classification proposed by Angulo-Meza and Pereira Estellita
Lins (2002), the methods for increasing discrimination within efficient DMUs
in a DEA setting can be classified into two groups:
Methods with a priori information. In these methods, the information pro-
vided by a decision-maker or an expert about the importance of the variables
can be introduced into the DEA models. There are three main methods devoted
to incorporating a priori information or value judgments in DEA:

Weight restrictions. The main objective of the weight restrictions methods

is to establish bounds within which the weights can vary, preserving some
flexibility/ uncertainty about the real value of the weights.13

Preference structure models. These models have been introduced by Zhu

(1996) within a framework of non-radial efficiency measures. In this
approach, the target for inefficient DMUs is given by a preference structure
(represented through some weights) expressed by the decision-maker.

Value efficiency analysis. This method, introduced by Halme, Joro, Ko-

rhonen, Salo and Wallenius (2000), aims at incorporate the decision-
maker’s value judgements and preferences into the analysis, using a two
stage procedure. The first stage identifies the decision maker’s most pre-
ferred solutions through a multiple objective model. The second stage
consists in the determination of the frontier based on the most preferred
solutions chosen.

Methods that do not require a priori information. These family of models

aims at increase discrimination in DEA without the subjectivity, the possibility
of biased or wrong judgements, typical of the methods that introduce a priori

13 See Allen, Athanassopoulos, Dyson and Thanassoulis (1997), and Pedraja-Chaparro, Salinas-Jimenes,
Smith and Smith (1997) for a review of some methods within this approach, including direct weight restric-
tions, cone ratio models, assurance region and virtual inputs and outputs restrictions.
Recent developments in nonparametric efficiency analysis 41

information. The main methods that minimize the intervention of the experts
are:

Super efficiency. Andersen and Petersen (1993) proposed this method to

rank efficient DMUs.
Cross-evaluation. The main idea of this method is to use DEA in a “peer-
evaluation” instead of a classical “self evaluation” evaluated by the clas-
sical DEA models.
Multiple objective approach. A Multiple Criteria Data Envelopment
Analysis has been proposed by Li and Reeves (1999) to solve the prob-
lems of lack of discrimination and inappropriate weighting schemes in
traditional DEA.

Extensions to the basic DEA Models

Directional distance functions have been introduced by Chambers, Chung
and Färe, (1996) and are based on Luenberger (1992) benefit functions. These
functions represent a kind of generalization of the traditional distance functions.
Their application leads to measures of technical efficiency from the potential for
increasing outputs while reducing inputs at the same time. In order to provide
a measure of “directional” efficiency, a direction, along which the observed
DMU is projected onto the efficient frontier of the production set, has to be
chosen. This choice is arbitrary and of course affects the resulting efficiency
measures. In addition, those measures are no more scale-invariant. See Färe
and Grosskopf (2004) for more details on these “new directions” in efficiency
analysis.
Examples of the literature that try to link DEA with a theoretical foundation
or that try to overcome and generalize the economic assumptions underlying
DEA include: Bogetoft (2000) which links the theoretically oriented agency,
incentives and contracts literature with the more practical oriented efficiency
measurement literature; and Briec, Kerstens and Vanden Eeckaut (2004a, b)
which extend the duality properties to non-convex technologies and propose
congestion-based measures in this framework.
Producers face uncertainty about technology reliability and performance.
The structure of technology and the existence and magnitude of inefficiency are
sensitive to the treatment of risk and uncertainty. On productivity measurement
under uncertainty see Chambers and Quiggin (2000) and Chambers (2004).

Statistical inference in efficiency analysis

All what we have seen in the previous description of recent developments
does not allow for a statistical sensitivity analysis, neither for rigorous statistical
testing procedures. This is because the previous literature does not relies on a
42 The measurement of efficiency

statistical model; there is not, in fact, a definition of the Data Generating Process
(DGP) and there is no room for statistical inference based on the construction
of confidence intervals, estimation of the bias, statistical tests of hypothesis and
so on.
There is instead a new approach, recently developed, which aims exactly at
the analysis of the statistical properties of the nonparametric estimators, trying
to overcome most limitations of traditional nonparametric methods and allow-
ing for statistical inference and rigorous testing procedures. This literature is
the main focus of this book. To the review of the statistical properties of non-
parametric frontier estimators we devote the following Chapter 3. Chapter 4
deals in detail with a family of robust nonparametric measures of efficiency,
which are more resistent to the influence of outliers and errors in data while
having good statistical properties which let inference feasible in this complex
framework. Finally, Chapter 5 illustrates and develop further the topic of con-
ditional and robust measures of efficiency and an alternative way to evaluate
the impact of external-environmental variables based on conditional measures
of efficiency.
Chapter 3

STATISTICAL INFERENCE IN
NONPARAMETRIC FRONTIER
ESTIMATION

Since efficiency is measured relative to an estimate of the frontier, estimates

of efficiency from nonparametric models are subject to uncertainty due to sam-
pling variation. Bootstrap methods hence, may be used to assess this uncertainty
by estimating bias, confidence intervals, testing hypothesis and so on. In the
following, we summarize the main results available in literature for statistical
inference based on nonparametric estimators of frontiers and efficiency scores
(DEA/FDH). In so doing we briefly define the Data Generating Process (DGP)
and introduce the statistical foundation of nonparametric frontier models. Then
we report the main asymptotic results available and the most important to know
properties of these estimators. Finally we briefly describe the bootstrap proce-
dure in this complex case as well as its useful applications in the nonparametric
frontier estimation context.

3.1 Statistical foundation

The statistical underpinning of nonparametric frontier models aims at defin-
ing a statistical model which allows to analyze the statistical properties of the
nonparametric estimators. The knowledge of these properties would, in princi-
ple, allow for statistical inference (consistency analysis, bias correction, confi-
dence intervals, test of hypothesis, . . . ) in this framework14 .
As pointed out in Simar (1996), even in a nonparametric setup, no statistical
inference can be conducted without a clear definition of a statistical model that
describes the DGP.

14 For
a selective survey on statistical inference in nonparametric frontier estimation see Grosskopf (1996).
Recent reviews are Simar and Wilson (2000a and 2006a).
44 Statistical inference in nonparametric frontier estimation

Let us start with a simple case where we only have a univariate output and
we want to carry out an output oriented efficiency analysis. By considering this
simple case, the reader will more easily understand the analogy with standard
regression models, parametric or nonparametric ones. The nonparametric fron-
tier can be defined as some unknown function ψ(x) sharing some properties
(monotonicity, possibly concavity, . . . ) that can be expressed as follows:

yi = ψ(xi ) − ui , ui ≥ 0 (3.1)

where yi ∈ R+ is the output and xi ∈ Rp is the vector of inputs. The random

term ui ≥ 0 represents a combination of random elements which explains why
an observed firm is below the frontier and so is (output-)inefficient. The statis-
tical model is complete when we define the conditional probability distribution
function (pdf) of u, f (u|x). This pdf can be very general, just satisfying some
regularity conditions, the conditioning on x indicates that we may consider het-
eroscedastic models where the distribution of the inefficiencies may depend on
the level of the input x. Combined with the pdf on x, this induces a joint pdf
on (x, y), f (x, y).
Under such statistical model, the observation of the sample X = {(xi , yi ),
i = 1, . . . , n} allows one to make inferences on the unknown functional ψ and
the unknown f (x, y) or on any of its component, generally, f (u|x) will have
some interest for the researcher. The DGP is, hence, completely characterized
by f (x, y) which has as support Ψ = {(x, y) ∈ Rp+ × R+ | y ≤ ψ(x)}. If
restrictive parametric assumptions are done on ψ and/or on f (x, y) (typically
through parametric assumptions on f (u|x)), there exists a lot of tools for doing
inference by using appropriate ordinary least squares (OLS) procedures or even
maximum likelihood techniques. We will come back to this kind of parametric
models in Sections 4.6 and 4.7 where more advanced techniques of estimation
will be proposed. In a nonparametric setup, we have to use other tools for doing
inference.
In this section we describe a more general statistical model that is useful to
derive asymptotic properties of our nonparametric estimators. As we will see
below, it also provides the appropriate bootstrap algorithms for doing inference
in practice. In fact very few hypothesis are required to define this DGP. We
follow here and summarize the presentation as in Simar and Wilson (2006a).
These Statistical Assumptions (SA) complement the economic assumptions EA
described in Section 2.3.

SA1: Random Sample. The sample observations (xi , yi ) in X are realiza-

tions of identically and independently distributed random variables (X, Y ) with
probability density function f (x, y) with support Ψ ⊂ Rp+ ×Rq+ : Prob((X, Y )
∈ Ψ) = 1.
Introducing stochastic noise in the model 45

This assumption is common in most empirical studies and just states that the
observations are considered as random draws from a population of firms (this is
typically what is done in the simple model described above, even in a parametric
approach).

SA2: Positiveness. The density f (x, y) is strictly positive on the boundary

of Ψ and is continuous in any direction toward the interior of Ψ.
This is a sufficient (but not necessary) condition for proving the consistency of
nonparametric estimators. It says that the probability of observing firms in an
open neighborhood of the frontier is strictly positive. Note that assumption SA2
could be relaxed but at a cost of losing some performance of the nonparametric
estimators (rate of convergence).
The next assumption is not strictly speaking a statistical assumption but is a
technical condition which insures that the true boundary is sufficiently smooth.

SA3: Smoothness. For all (x, y) in the interior of Ψ, the functions θ(x, y)
and λ(x, y) are differentiable in both arguments.
It is a sufficient condition used by Kneip, Simar and Wilson (2003) to derive
the asymptotic distribution of the DEA estimator, for the FDH estimator (where
Ψ is not assumed to be convex), only Lipschitz continuity of the functions is
required in Park, Simar and Weiner (2000).
Summing up, the DGP P is completely characterized by the knowledge of
f (x, y) and of its support Ψ with the regularity conditions (SA1–SA3) described
above. Hence, we can write P = P(Ψ, f (·, ·)).

3.2 Introducing stochastic noise in the model

The statistical model defined above introduces the stochastic elements which
allow to describe how a random observation (Xi , Yi ) is generated on the attain-
able set Ψ. The DGP is characterized by the joint probability density function
f (x, y). We will see below that this allows to make inference on the quantities
of interest in efficiency analysis, even if the complexity of the problem will
force the practitioner to use bootstrap algorithms for practical purposes. We
will also see in Chapter 4, that this statistical model can also be reformulated
in terms of nonstandard cumulative distribution functions, where the Farrell-
Debreu efficiency scores can be shown to be defined from their properties. This
will provide a natural way for introducing the robust estimators of the frontier
and of the efficiencies based on partial-frontier concepts.
All these approaches contain stochastic elements in a so-called deterministic
frontier framework, where it is supposed that all the observations (Xi , Yi ), i =
1 . . . , n belong to the attainable set Ψ. Mathematically we write Prob{(Xi , Yi )
∈ Ψ} = 1. This seems to be a quite natural assumption in most applications but
this implies that no stochastic noise is allowed in the model. In econometrics,
46 Statistical inference in nonparametric frontier estimation

stochastic noise is generally introduced in regression models, for instance, for

allowing error in measurements, random shocks, chance,. . . In a frontier setup,
we refer then to stochastic frontier models where, till now, most of the work in
the literature has been done in a complete parametric setup. Pioneering work
refer to Aigner, Lovell and Schmidt (1977) and Meeusen and van den Broeck
(1977), with other developments in Greene (1990, 1993) and Stevenson (1980).
A recent comprehensive reference is Kumbhakar and Lovell (2000).
All these approaches rely on very restrictive parametric assumption about the
shape of the frontier and about the distribution of the efficiencies (some non-
negative random variable like half-normal, truncated normal, exponential, or
gamma densities) and about the distribution of the noise (usually a normal vari-
ate). Usually maximum likelihood estimators (MLE) are derived and statistical
inference is rather straightforward and valid if the model is correctly specified.
Parametric estimators incur the risk of misspecification which generally results
in inconsistency. Beside the restrictive parametric assumptions and the risk of
misspecification, parametric stochastic frontier models suffers from other prob-
lems in finite sample situation (a common situation for the practitioner). We
only briefly mention here the identification problems as described in Ritter and
Simar (1997) and the MLE computations themselves along with the difficulty
of inference on the efficiency scores as described in Simar and Wilson (2005).
Nonparametric estimators avoid the risk of misspecification but at a cost
of complexity. As pointed in Simar and Wilson (2006a), introducing noise
in DEA/FDH framework is still a challenge and an open issue. The problem
is difficult because without some restrictions on the model, a stochastic non-
parametric model is not identified. Hall and Simar (2002) discuss this issue in
details. However, Hall and Simar show that if the noise is not too large (in terms
of noise to signal ratio) a reasonable estimator of the boundary can be found by
identifying the point near the deterministic boundary where a nonparametric
estimator of the density of the noisy data presents a maximum of the absolute
value of its gradient. These ideas have been extended to the multivariate setup
in Simar (2003b) where stochastic versions of DEA/FDH estimators have been
developed when the noise is of moderate size. In particular the method provides
versions of the DEA/FDH very robust to outliers and extreme values.
In a different setup, Kumbhakar, Park, Simar and Tsionas (2004) propose
a general approach for nonparametric stochastic frontier model using local
maximum likelihood methods. The idea is to start with an anchorage parametric
model, in the line of Aigner, Lovell and Schmidt (1977) and then by localizing,
generate a very flexible model approximating a much wider class of models
than the chosen anchorage model.
In the case of a panel of data, much more information is available and the
identification problem can more easily be handled. This was done in Kneip and
Simar (1996) where a general nonparametric model is proposed in this particular
Asymptotic results 47

setup. The drawback is that in practice, a large number of time periods is needed
for getting sensible results. New directions in this area have been proposed by
Henderson and Simar (2005).
In this book we focus the presentation on deterministic frontier models, so
that the popular nonparametric envelopment estimators can be considered. But
since we know these estimators are sensible to extreme value and outliers (due
to the absence of noise in the model), we pay special attention to develop
estimators which are robust to these extreme points: this will be the major topic
of Chapter 4. In the next sections we summarize the main known statistical
properties of the DEA/FDH estimators and we indicate how the bootstrap can
be implemented to solve practical inferential problems.

3.3 Asymptotic results

When estimating a statistical model from a “partial” information coming
from a sample of size n, a natural question which should be raised is the follow-
ing: “Has the estimate I obtain anything to see with the unknown characteristic
of the model it is supposed to estimate?”.

3.3.1 Consistency
The first minimal property one would like to achieve is consistency. Roughly
speaking consistency means that if the sample size increases, an estimator θ will
converge to the true but unknown value θ it is supposed to estimate. Mathe-
matically, we will say that θ → θ as n → ∞, meaning that as the sample size
p

increases to infinity, the probability of the error |θ − θ| being greater than any
positive value ε > 0 converges to zero. This is a minimal property that an
estimator should have to be reliable. Another important issue is then the rate of
convergence of the consistent estimator. It indicates the possibility of getting
sensible results with finite samples estimators. In classical
√ parametric statistics
(like linear regression models), estimators achieve n-consistency, meaning
that the order of the error of estimation is decreasing to zero like n−1/2 when
n → ∞. We write:

θ − θ = Op (n−1/2 ). (3.2)

In nonparametric frontier estimation, for decades nobody worried about these

issues, neither for the DEA nor for the FDH estimators. The first result which
appeared was due to Banker (1993), who proved the consistency of the DEA
efficiency scores in the very particular univariate case (one input for input
orientation, or one output in the output orientation). For instance, in the input
orientation (one input) the obtained result can be written as:

θDEA (x, y) → θ(x, y),

p
(3.3)
48 Statistical inference in nonparametric frontier estimation

where no indication was given about the rates of convergence. These rates
where obtained for the DEA (where convexity of Ψ is required) and for the
FDH case (where convexity of Ψ is not required) in Korostelev, Simar and
Tsybakov (1995). For instance in the output oriented case (one output) they
obtained:
1
 F DH , Ψ) = Op (n− p+1 ),
d (Ψ
2
 DEA , Ψ) = Op (n − p+2
d (Ψ ),
where d (Ψ, Ψ) is the Lebesgue measure of the difference between the two
sets and where p is the number of inputs (similar rates are obtained for the
corresponding efficiency measures). The rates of convergence reflect the curse
of dimensionality typical of many nonparametric statistical techniques; if p is
large, the estimators exhibit very low rates of convergence, and much larger
quantity of data is needed to get sensible estimates (i.e. to avoid large variances
and very wide confidence interval estimates) than in the case of small number
of inputs p. Note that for p = 1 we obtain a better rate n−2/3 than the standard
parametric rate n−1/2 .
Much later, Kneip, Park and Simar (1998) for the DEA case and Park, Simar
and Weiner (2000) for the FDH case obtained the proof of the consistency of
the estimated efficiency scores in the full multivariate setup (p, q > 1) along
with their rates of convergence. The difficulty here was to handle the radial
nature of the difference between the efficiency scores. Formally they obtain:
2
− p+q+1
θDEA (x, y) − θ(x, y) = Op (n ), (3.4)
1
− p+q
θF DH (x, y) − θ(x, y) = Op (n ). (3.5)
These results again reflect the curse of dimensionality which is even worse for
the multivariate case since the convergence rates are affected by p + q rather
than merely by p (or q), as for the former univariate case.
These results are encouraging: the methods used by researchers since decades
where indeed consistent! But these results are of little practical importance for
doing inference. To achieve this we need the sampling distributions of the
estimators in order to derive the eventual bias, or to compute its standard devia-
tion or even better to build confidence intervals for individual efficiency scores
θ(x, y).
In this complex situation, the only hope is to obtain asymptotic results, i.e.
a reasonable approximation of the sampling distribution of the estimator when
n is large enough (in the same spirit that a Central Limit Theorem gives an
approximate normal distribution of a sample mean when n is large enough).
We will see below that if today theoretical results are available, they will be of
little practical interest but will be useful to prove the consistency of the bootstrap
alternative.
Asymptotic results 49

3.3.2 Sampling distributions

The first available result is for the bivariate DEA case (p = q = 1) and is
due to Gijbels, Mammen, Park and Simar (1999). For instance, for the input
orientation, they obtain the asymptotic result, when n → ∞,
2
  asy.
n 3 θDEA (x, y) − θ(x, y) ∼ F (·, ·) (3.6)

where F (·, ·) is a regular distribution function known up to some unknown

constants. These constants depend on the DGP and are related to the curvature of
the frontier and the value of the joint density f (x, y) at the corresponding frontier
point (θ(x, y)x, y) in the input orientation). Gijbels, Mammen, Park and Simar
(1999) provide also the tabulation of the quantiles of a pivotal correspondent
of F (·, ·). This allows to build confidence intervals for the efficiency score
θ(x, y).
The multivariate DEA case (p, q ≥ 1) was much more difficult to handle,
due to the difficulty of characterizing the dominating facet of (x, y) in Ψ  DEA
and due to the radial nature of the Farrell-Debreu measures. Kneip, Simar and
Wilson (2003) obtain the following result, when n → ∞,
 θ  asy.
2
DEA (x, y)
n (p+q+1) −1 ∼ Q(.) (3.7)
θ(x, y)
where no closed analytical form for Q(.) is available, but only a mathematical
expression showing its existence and its good properties (in particular that it
is a non-degenerate distribution function). Of course a similar result could be
obtained for the difference θDEA (x, y) − θ(x, y) rather than for the ratio and
also for the output oriented case. But this result is of little practical importance
because the limiting distribution is difficult to manipulate. However this result
is of fundamental theoretical importance to prove the consistency of the appro-
priate bootstrap approach (see below). Hence, the bootstrap will appear to be
the only practical alternative to do inference in this setup.
The multivariate FDH case (p, q ≥ 1) was easier to handle and Park, Simar
and Weiner (2000) derive the following result, when n → ∞,
1   asy.
n (p+q) θF DH (x, y) − θ(x, y) ∼ W eibull(·, ·) (3.8)

here, again, the limiting Weibull depends on some unknown parameters depend-
ing on the DGP but which can be estimated. This result allows to obtain bias
corrected estimators and confidence intervals for the efficiency scores, however,
Park, Simar and Weiner illustrate how imprecise is the asymptotic distribution
when p + q is large with moderate sample sizes (for instance, they recommend
n to be larger than say 1000 if p + q = 5). Again, similar results are derived
for the output oriented case.
50 Statistical inference in nonparametric frontier estimation

Therefore, as a conclusion, even if these results have a real theoretical im-

portance and even if they are potentially useful to estimate asymptotic bias and
variance, as well as asymptotic confidence intervals, they remain asymptotic
results which may be misleading when used with small or moderate sample
sizes. Moreover, additional noise is introduced when estimates of the unknown
parameters of the limiting distributions are used in constructing estimates of
confidence intervals. We note also that in the DEA case there are no closed
analytical forms for the multivariate case. So the bootstrap seems to be an
unavoidable alternative.

3.4 Bootstrap techniques and applications

The bootstrap is indeed an attractive alternative to the theoretical limiting
distributions described above. The bootstrap is intended to provide approxi-

 y) − θ(x, y) or of θ(x, y) where
mations of the sampling distributions of θ(x,
θ(x, y)

(x, y) is the unit under interest. The nonparametric estimator θ(x, y) can be
the FDH estimator (if we do not assume convexity of Ψ) or the DEA estimator.
The presentation below is for the DEA case, we will summarize later the avail-
able results for the FDH case. We follow the presentation of Simar and Wilson
(2000a, 2006a).
The bootstrap is a data-based simulation method for statistical inference.
As reported by Efron and Tibshirani (1993, p.5), the use of the term bootstrap
derives from the phrase to pull oneself up by one’s bootstrap, widely thought to
be based on one of the eighteenth century Adventures of Baron Munchausen, by
Rudolph Erich Raspe15 . The essence of the bootstrap idea (Efron 1979, Efron
and Tibshirani, 1993) is to approximate the sampling distributions of interest
by simulating (or mimicking) the DGP.
The basic idea behind the bootstrap can be summarized as follows. Consider
the simple problem where a DGP (a statistical model) P generates a random
sample X = {X1 , . . . , Xn } of size n. Suppose we want to investigate the
sampling distribution of an estimator θ of some unknown parameter θ. In
general, θ is one particular characteristic of the DGP P and θ = θ(X  ) is a
statistics function of the random sample X . The knowledge of L(θ(X  )), the

sampling distribution of θ(X ), is all what we need to evaluate the bias, the
 ) and to derive bounds of confidence intervals of any
standard deviation of θ(X
desired level for θ.

15 TheBaron had fallen to the bottom of a deep lake. Just when it looked like all was lost, he thought to pick
himself up by his own bootstraps.
Bootstrap techniques and applications 51

Except in very few simple problems (like estimating the mean and the vari-
ance of a normal model) the sampling distribution L(θ(X  )) is unknown or
only asymptotic approximations are available. The aim of the bootstrap is to
provide an approximation of this distribution which will be easy to obtain by
using Monte-Carlo approximations. Under regularity conditions the only thing
that will be required to implement the bootstrap is a consistent estimator of the
DGP P.
Indeed, if this DGP P would be known, it would be very easy to approximate
 ) without any mathematical developments, by
the sampling distribution of θ(X
a simple Monte-Carlo experiment that the computer could perform for us. We
could indeed simulate a large number of times a random sample X  from P
and then compute the corresponding value of θ(X  ) in each Monte-Carlo trial.
By repeating this exercise a larger number of time the Monte-Carlo empirical
  ) would provide a Monte-Carlo ap-
distribution of the observed values θ(X
proximation of the true but unknown sampling distribution L(θ(X  )). This is
a direct consequence of the strong law of large number and the quality of the
approximation depends only on the number of replications in the Monte-Carlo
exercise (that the user can chose as large as she/he wants): no mathematics is
needed here, only some computing time on the computer that will perform this
simulation.
The bootstrap principle is now easy to explain: since P is unknown, we will
plug-in an appropriate consistent estimator P in the place of P in the Monte-
Carlo experiment above. Here we will call a bootstrap sample a random sample
X  generated from P. If some care is taken on how to generate these bootstrap
samples, it can be proven, when the bootstrap works, that the empirical (Monte-
  ), which is conditional on P,
Carlo) bootstrap distribution of θ(X  approximates

the unknown L(θ(X )). In fact as we will see below, to build confidence intervals
 )−
it is more appropriate to rather approximate the unknown distribution of θ(X

θ by the bootstrap distribution of θ(X   ) conditional on the estimate
) − θ(X
 
P. The error of estimation θ(X ) − θ is sometimes refereed as the estimation
  ) − θ(X
error in the real world whereas, θ(X  ) is the error of estimation in the
bootstrap world where the true unknown P and θ have been replaced by the
known observed P and θ(X  ).
In many simple applications, the easiest way to generate a random sample X 
according an estimate P of P, is to mimic what has been done in the real world.
In the real world X = {X1 , . . . , Xn } is generated from P, so a nonparametric
estimator of P could be chosen as the empirical process which gives a mass
1/n at each observed sample point Xi ∈ X . So a bootstrap sample will be
defined as X  = {X1 , . . . , Xn }, where each Xj is obtained by drawing with
replacement from the n values {X1 , . . . , Xn }. This is sometimes refereed as
the naive bootstrap and is very easy to implement.
52 Statistical inference in nonparametric frontier estimation

When we say that the bootstrap works we mean that the bootstrap approxima-
tion is consistent, or in other words that when the sample size n of X increases,
  )− θ(X
the bootstrap distribution of θ(X  ) conditional on P converge to the true
 ) − θ. This is the crucial point for the bootstrap. It is often
distribution of θ(X
true in statistics that the bootstrap (when correctly implemented) is consistent,
but it is well known also that there are cases where the bootstrap is inconsistent.
This may depend on the model, on the properties of the estimate P but also on
the way to generate random samples X  from P.  This is particularly true in the
case of estimating boundaries or support of random variables, as it is the case
in frontier models. This issue is discussed below where consistent solutions are
provided.

3.4.1 Bootstrap in frontier models

The first use of the bootstrap in frontier models is attributed to Simar (1992).
Its development for nonparametric envelopment estimators was introduced by
Simar and Wilson (1998).16
The definition of the DGP is a crucial step for the bootstrap procedure. If the
DGP is not defined, it is not clear which process the bootstrap is mimicking. We
have seen above that the DGP can be denoted by P = P(Ψ, f (·, ·)) to stress the
fact that it is fully characterized by the knowledge of Ψ and of the joint density
f (x, y) in the input-output space.
We describe the general setting adopting an input oriented framework with
VRS (Variable Returns to Scale) DEA. For practical purposes, it is advantageous
to express the input-oriented efficiency in terms of the Shephard (1970) input-
distance function introduced in Section 2.3 and defined in (2.12) as δ(x, y) =
(θ(x, y))−1 . Extensions to the output-oriented framework are straightforward.
The objective is thus to investigate the sampling distribution of (δ(x,  y) −
δ(x, y)) for a given DMU (x, y).
Due to our knowledge of the DGP P, we can produce a consistent estimator
P of P from the data X : P = P(Ψ,  f(·, ·)). So, in the true world, P and δ(x, y)
are unknown ((x, y) is a given fixed point of interest), but in the bootstrap world,
the consistent estimate P and δ(x, y) are known and can take the place of P
and of δ(x, y).
Therefore, we can generate data sets from P.  Denote by X  = {(X  , Y  ),
i i
i = 1, ...., n} a data-set generated by P.  This pseudo-sample defines the
corresponding quantities Ψ   and δ (x, y) which can be viewed as estimators of

16 Someother bootstrap procedures have been presented in literature, but their inconsistence have been
demonstrated, see below.
Bootstrap techniques and applications 53

 y). They are defined by:

 and δ(x,
the corresponding quantities Ψ
 
n
  = (x, y) ∈ Rp+q | y ≤
Ψ γi Yi ; x ≥ γi Xi ;
i=1

n 
γi = 1; γi ≥ 0; i = 1, ..., n (3.9)
i=1

 x 
δ (x, y) = sup δ | ( , y) ∈ Ψ
 .
δ
The latter can be calculated through the following linear program:
 
n 
n
(δ (x, y))−1 = min θ > 0 | y ≤ γi Yi ; θx ≥ γi Xi ;
i=1 i=1

n 
γi = 1; γi ≥ 0; i = 1, ..., n . (3.10)
i=1

Conditionally on X , the sampling distribution of δ (x, y)) is (in principle)

known since P is known, although it may be difficult to compute analytically.
Monte Carlo methods can be used to easily approximate the sampling distrib-
ution of δ (x, y).
Using P to generate B samples Xb , for b = 1, ...., B, and applying the linear
program described above, for the given unit (x, y) we obtain a set of pseudo
estimates {δb (x, y)}B 
b=1 . The empirical distribution {δb (x, y)}b=1 is the Monte
B

Carlo approximation of the distribution of δ (x, y) conditional on P. 

If the bootstrap method is consistent, the available bootstrap distribution
of δ (x, y) will “mimic” the original unknown sampling distribution of the
 y). More precisely:
estimator of interest δ(x,

(δ (x, y) − δ(x,

 y))|P  y) − δ(x, y))|P.
 ∼ (δ(x,
approx.
(3.11)

Since the left hand side of (3.11) is available (though the Monte-Carlo exercise),
it can be used to provide properties usually obtained from the right-hand side.
In particular we can use the bootstrap approximation to estimate the bias of
the DEA estimator or to estimate the quantiles of the sampling distribution of
 y) − δ(x, y)) in order to build confidence intervals.
(δ(x,
In Table 3.1 below, the analogy between the original inferential problem
and the bootstrap is described in terms of an analogy between the real world,
where we want to make inference about the parameter δ(x, y) but most of the
desired quantities are unknown, and the bootstrap world, where we mimic the
real world but where everything is known and so can be computed or simulated
54 Statistical inference in nonparametric frontier estimation

by Monte-Carlo methods. In practice, as described in the next subsections, the

available GP(t) will serve to approximate the unknown GP (t) (for correcting
bias or evaluate standard deviation of estimates) and HP(t) will be used to
estimate the quantiles of the unknown HP (t) providing bootstrap confidence
intervals.

Table 3.1. Summary of the bootstrap principle for inference on δ(x, y).

REAL WORLD BOOTSTRAP WORLD

P = P Ψ, f (·, ·) unknown =P Ψ

Given X , P  , f(·, ·) is known
δ(x, y) unknown 
δ (x, y) is known

DGP P generates X DGP
P  generates X

⇓ ⇓

δ (x, y) estimator of δ(x, y) 
δ
(x, y) estimator of 
δ (x, y)

Sampling distribution Bootstrap distribution

GP (t) = ProbP (δ (x, y) ≤ t) GP 

 = ProbP(δ (x, y) ≤ t)
(t)

Moments Moments
EP (
δ (x, y)) EP 

(δ (x, y))
V arP (
δ (x, y)) V arP 

(δ (x, y))

W = δ (x, y) − δ(x, y) W
=  δ
(x, y) − 

δ (x, y)
−1
HP (t) = ProbP (W ≤ t) ⇒ aβ = HP (β) HP
 (t) = Prob  (W

≤ t) ⇒ aβ = H −1 (β)

P P
⇓ ⇓
CI for δ(x, y) Bootstrap CI for δ(x, y)
δ (x, y) − a1−α/2 , 
 δ (x, y) − aα/2  a1−α/2 , 
δ (x, y) −  δ (x, y) − 
aα/2

3.4.2 
Correcting the bias of δ(x, y)
An estimator is a random variable since it is computed as a function of a
random sample. An unbiased estimator has the desirable property that its mean
is equal to the target value of the parameter being estimated. In our case, we
know by construction that the DEA estimator δ(x, y) is a biased estimator of
 y) < δ(x, y)
δ(x, y) (for the input Shephard distance considered here, δ(x,
Bootstrap techniques and applications 55

 y)) < δ(x, y)). Formally the bias of

with probability one, so the true EP (δ(x,

δ(x, y) is defined as:
 y)) = EP (δ(x,
bias(δ(x,  y)) − δ(x, y), (3.12)

which of course cannot be computed, because the sampling distribution of

 y), denoted GP (t) in Table 3.1, is unavailable (even the asymptotic ap-
δ(x,
proximation is too complicated to handle).
However the bootstrap approximation is available and so we will estimate
 y) by the following:
the bias of δ(x,
 y)) = E (δ (x, y)) − δ(x,
bias(δ(x,  y). (3.13)

P

In practice, the expectation is given by the mean of the Monte-Carlo realizations

of {δb (x, y)}B
b=1 . So our bootstrap estimate of the bias is obtained through:

 1 
B
bias(δ(x, y)) ≈ δb (x, y) − δ(x,
 y). (3.14)
B b=1

In the same way the standard deviation of the DEA estimator δ(x,  y) is
obtained as the square-root of the variance of the bootstrap distribution denoted
GP(t) in Table 3.1. Namely:

2 B 1 B 2
 y)) ≈ 1
std (δ(x, δb,2 (x, y) − δb (x, y) . (3.15)
B b=1 B b=1

A bias corrected estimator is then obtained by defining:

y) = δ(x,
δ(x,  y) − bias(δ(x,
 y))

 y) − 1 B
= 2 δ(x, δ (x, y). (3.16)
B b=1 b

However, it is well known that correcting for the bias introduces additional
noise (increasing the variance of the estimator). As a rule of thumb, Efron and
 y))| >
Tibshirani (1993) recommend not to correct for the bias unless |bias(δ(x,
 y))/4. In practice, due to inherent bias of the DEA estimator, the
std(δ(x,
bias-correction has almost always to be performed. Numerical examples are
provided in the second part of this book.

3.4.3 Bootstrap confidence intervals for δ(x, y)

The construction of confidence intervals is obtained by determining the quan-
tile of HP (t), the sampling distribution of W = δ(x, y) − δ(x, y). Indeed, if
56 Statistical inference in nonparametric frontier estimation

HP (·) were known, it would be easy to find, for instance, the values a0.025 and
a0.975 such that:
 y) − δ(x, y) ≤ a0.975 ) = 0.95,
ProbP (a0.025 ≤ δ(x,

leading to the confidence interval of level 0.95 for δ(x, y):

 y) − a0.975 ≤ δ(x, y) ≤ δ(x,
ProbP (δ(x,  y) − a0.025 ) = 0.95. (3.17)

Since the quantiles aβ are unknown, the quantiles of the bootstrap distribution
of W  = δ (x, y) − δ(x,
 y), denoted by H (t) in Table 3.1, will provide the
P
appropriate approximation. If â0.025 and â0.975 are such that

ProbP(â0.025 ≤ δ (x, y) − δ(x,

 y) ≤ â0.975 ) = 0.95,

the bootstrap confidence interval for δ(x, y) is obtained as:

 y) − â0.975 ≤ δ(x, y) ≤ δ(x,
ProbP (δ(x,  y) − â0.025 ) ≈ 0.95. (3.18)

The quantiles âβ are directly obtained from the quantiles of the Monte-Carlo
distribution of the values {δb (x, y)}B
b=1 themselves as follows, for all β ∈ [0, 1]:

 y),
âβ = ĉβ − δ(x, (3.19)

where ĉβ is the β-quantile of the empirical distribution of the values

{δb (x, y)}Bb=1 . Note that the method just described and known as the basic
bootstrap method for building confidence intervals, automatically adjusts for
the bias of the DEA estimates.
The amplitude (the length) of the obtained interval will reflect the uncertainty
we have about the real value of the efficiency score of the unit operating at the
 y). This uncertainty may vary from
level (x, y) and estimated as being δ(x,
place to place, it depends where (x, y) is located: if the DEA frontier above
(x, y) is determined by many sample points (Xi , Yi ), the precision can be great
but if it is only determined by very few (even only one) sample point, the length
of the confidence interval will be much greater, reflecting the uncertainty we
have above the real position of the reference frontier for this point (x, y). We
know also from the results from Gijbels, Mammen, Park and Simar (1999)
that the sampling variation of the estimates also depends on the curvature of
the real efficient frontier above (x, y). This will be illustrated in some of the
applications described in the second part of this book.

3.4.4 Is the bootstrap consistent?

The bootstrap is consistent if the relation (3.11) holds. This will depend on the
way the pseudo-samples are generated. Since P generates X = {(Xi , Yi ), i =
Bootstrap techniques and applications 57

1, ...., n}, a naive estimator of P would be P(Ψ, f(·, ·)) where f(·, ·) would be
the empirical distribution function of (Xi , Yi ), defined as the discrete distribu-
tion that put a probability n1 on each point (Xi , Yi ). Then a bootstrap sample
X  = {(Xi , Yi ), i = 1, ...., n} would simply be obtained by randomly sam-
pling with replacement from X .
Unfortunately, it is well known from the bootstrap literature (Bickel and
Freedman, 1981, Efron and Tibshirani, 1993) that in a boundary estimation
framework, this bootstrap procedure does not provide a consistent approxima-
tion of the desired sampling distribution as in (3.11). Simar and Wilson (1999a,
b) discuss this issue in the context of multivariate frontier estimation. As illus-
trated below, the problem comes from the fact that in the naive bootstrap, the
efficient facet that determines in the original sample X the value of δ appears
too often, and with a fixed probability, in the pseudo-samples Xb and this fixed
probability does not vanish even when n → ∞.
Two solutions have been proposed to overcome this problem: either sub-
sampling, meaning that we will draw pseudo-samples of size m smaller than
n, say m = [nγ ], where γ < 1 and [a] stands for integer part of a number a
or smoothing techniques, meaning the use of a smooth estimate f(·, ·), in place
of the discrete empirical one of the naive approach. Kneip, Simar and Wilson
(2003) prove the consistency of both approaches in the case of strictly convex
attainable sets Ψ.

Subsampling techniques
Subsampling is certainly the easiest procedure to apply: we follow the pro-
cedure described above, with the only difference that the pseudo-samples Xm,b 

for b = 1, . . . , B are of size m in place of size n and so the reference sets

in (3.9) and (3.10) are determined by these pseudo-samples of size m only.
The procedure is consistent for any value of γ < 1, but the drawback is that
data-driven procedures for determining an optimal value of γ for a particular
data-set have not yet been investigated. In practice, several values of γ in the
range [0.5, 1) can be tried and values of γ where the results show some stability
indicate reasonable choices for this parameter.

Smoothing techniques
The idea of the smooth bootstrap (see Siverman and Young, 1987) is to draw
the pseudo observations (Xi , Yi ) from a smooth estimate of the density f (x, y).
We know how to produce such smooth estimates (see e.g. Silverman, 1986,
Scott, 1992 or Simonoff, 1996) by using kernel estimators but the problem is
complicated here by the fact that the range of (x, y) is bounded by the boundary
of the unknown Ψ. Simar and Wilson (1998, 2000b) propose procedures which
58 Statistical inference in nonparametric frontier estimation

are easy to apply. These procedures will exploit the radial nature of the Farrell-
Debreu efficiency scores and of the Shephard distance functions.
To take this radial nature into account, it is easier to transform the Cartesian
coordinates (x, y) into polar coordinates for the input vector x when input
efficiency scores are investigated as it is the case in our presentation (the output
oriented case would use polar coordinates for the output vector y).
The polar coordinates
 for x are defined by its modulus ω = ω(x) ∈ R+
where ω(x) = (x
x), and its angle η = η(x) ∈ [0, π2 ]p−1 , where for j =
xj+1
1, ..., p − 1, ηj = arctan( 1 ) if x1 > 0 or ηj = π2 if x1 = 0.
x
The density f (x, y) can be transformed, or represented by a density f (ω, η, y)
on the new coordinates and the latter joint density can be decomposed as:

f (ω, η, y) = f (ω | η, y)f (η | y)f (y), (3.20)

where we suppose all the conditionals exist. So that for the frontier point x∂ (y)
on the ray defined by the input vector x has modulus for the output level y is
ω(x)
given by ω(x∂ (y)) = inf{ω ∈ R+ |f (ω|y, η) > 0}, and δ(x, y) = .
ω(x∂ (y))
We see by the latter expression that the transformation in polar coordinates
induces a conditional pdf for δ(x, y) given (y, η), namely f (δ|y, η), with sup-
port over [1, ∞). Hence, in a certain sense we have transformed the density
f (x, y) expressed in Cartesian coordinates into a density on “polar-type” coor-
dinates f (δ, η, y) = f (δ | η, y)f (η | y)f (y). Consequently, now, the DGP is
characterized by P = P(Ψ, f (δ, η, y)). The reader can see here the analogy
with the simple model (3.1) in Section 3.1, where u was the univariate random
inefficiency term; here this term is replaced by δ which has a conditional density
f (δ | η, y).
The idea of the smooth bootstrap is to use as DGP in the bootstrap world P =
P(Ψ,  f(δ, η, y)), where f(δ, η, y) will be a smooth continuous density estimate
of the unknown density from the sample of observed values (δi , ηi , Yi ) obtained
by the polar transformation described above of the original data (Xi , Yi ) and
where the unknown δi , have been replaced by the estimates δi = δ(X  i , Yi )
(which are the distance functions in the bootstrap world, i.e., with respect to the
attainable set Ψ). 
Simar and Wilson (2000b) propose an algorithm to simulate pseudo-data
(δi , ηi , Yi ) and to transform them back in Cartesian coordinates to obtain the


bootstrap pseudo-sample X  = {(Xi , Yi ), i = 1, . . . , n}. The procedure is

rather complicated because in the kernel estimation we have to take into account
a boundary condition for δ that have to be greater than 1. The procedure implies
a reflection method (Schuster, 1985, Silverman, 1986) in the p + q dimensional
space: we add to the original points (δi , ηi , Yi ) the points (2 − δi , ηi , Yi ). By
Bootstrap techniques and applications 59

doing so, we achieve consistency of the density estimate even near its boundary
points (see Simar and Wilson, 2000b for further details).

The homogeneous smooth bootstrap

The bootstrap procedure is simplified if we are ready to make an additional
assumption on the DGP. In particular we assume that the distribution of the
inefficiencies is homogeneous over the input-output space. This is the analog
to the assumption of homoscedasticity in regression models. Formally we
suppose that:
f (δ | y, η) = f (δ). (3.21)
This may be a reasonable assumptions in many practical situations.
With this homogeneity assumption, the problem of the bootstrap is similar to
the bootstrap in homoscedastic regression models where the bootstrap is based
on the residuals (see Freedman, 1981). In the present context, the residuals
corresponds to the estimated distance from the fitted frontier of Ψ  which are

the DEA efficiency scores δi , i = 1, . . . , n.
So we generate a pseudo observation, (δi , ηi , Yi ), conditional to the original
observed values for the input mix and for the output level (ηi , Yi ) = (ηi , Yi )
and we only have to generate inefficiencies by generating a univariate value for
the input Shephard distance δi . This is done by generating δi from a smooth
density estimate of f (δ) obtained from the n data values {δi ; i = 1, . . . , n}.

Nonparametric smooth density estimation

A standard nonparametric smooth density estimate is obtained by a kernel
density estimate (Silverman, 1986):

1 n  δ − δ 
f(δ) =
i
K , (3.22)
nh i=1 h

where K(·) is a kernel function (usually a continuous standardized density

with mean 0 and variance 1) and h is the bandwidth controlling the scale of
the kernel function. The density estimate is the mean of n densities centered
at the observed values δi with standard deviation given by the bandwidth h.
The choice of the kernel function is not crucial in defining the density estimate
(the results are very stable to this choice and often a standard gaussian density
is chosen for K(·)) however the smoothing parameter h has to be carefully
determined. Too small values of h (under smoothing) localize too much the
kernels and the average f(δ) will be quite irregular; at the limit if h → 0, the
density estimate converges to the discrete empirical density with mass 1/n at
each observed points f(δ). Too large values of h will over smooth the estimate;
at the limit if h → ∞, the density estimate converges to a flat uniform density.
60 Statistical inference in nonparametric frontier estimation

Silverman (1986) shows that if f (·) is gaussian, an optimal value for h,

minimizing the mean integrated squared error between f(δ) and f (δ), is given
by:

h = 1.06sn n−1/5 ,

where sn is the empirical standard deviation of the n values δi . This is known
as the normal reference rule. It has been shown by Silverman that the choice

h = 1.06 min(sn , rn /1.34)n−1/5 , (3.23)

where rn is the interquartile range of the n data points, is more robust to de-
partures form the gaussian assumption for f (·). This latter rule is referred as
the robust normal reference rule. It is very popular and give generally reason-
able values for the bandwidth. Other empirical rules have been proposed in
the literature, like the Sheather and Jones (1991) method which tries to be still
more robust to departures from the normal assumption, by using higher order
empirical moments of the data points.
As a matter of fact the problem is slightly more complicated here for two
reasons: (i) there is a spurious mass at one in the sample of values δi , i =
1, . . . , n, and (ii) there is a boundary effect since δ ≥ 1 by definition and the
estimate in (3.22) does not verify this constraint. As suggested in Simar and
Wilson (2006a), the first problem is solved by deleting the spurious ones, in this
step of bandwidth and density estimation, and the second problem is addressed
by using the reflection method (see Silverman, 1986 for details). Formally, we
consider only the m values of δi > 1, for i = 1, . . . , m with m < n, then we
consider the set of the 2m values {2 − δ1 , . . . , 2 − δm , δ1 , . . . , δm } which are
now symmetrically distributed around 1. Then we compute the kernel density
estimate with this series (without any boundary condition). Analog to (3.22),
we have:
1  m  δ − δ   δ − 2 + δ 
i i
gh (δ) = K +K , (3.24)
2mhm i=1 hm hm

The optimal bandwidth by using the empirical rule (3.23) here is obtained by:

hm = 1.06 min(s2m , r2m /1.34)(2m)−1/5 ,

where now s2m and r2m are computed from the 2m reflected data. As pointed
above, the distribution of these reflected values is by construction symmetric
around 1 and we observed in many applications a bell-shape for the distribution
of these reflected values. Therefore the robust normal-reference rule giving hm
offers in many applications a reasonable value for the bandwidth not far from
the optimal one.
Bootstrap techniques and applications 61

Note that automatic data-driven techniques based on cross-validation exist

for selecting the optimal bandwidth in (3.24) (see e.g. Silverman, 1986). The
most usual one is based on leave one-out least-squares cross-validation function
that find the optimal hm that minimizes the following criterion:
 +∞ 2m
1 
CV (h) = gh (δ) dδ − g2 (δ), (3.25)
−∞ 2m i=1 h,(i)

where gh,(i) (δ) is the leave one-out estimator of g(δ) based on the 2m values
except δi . Another automatic data-driven technique based on likelihood cross-
validation is described in Section 5.3, where a variable bandwidth is obtained
in a different context by using a k-Nearest Neighbor approach.
Note that in many applications the solution of (3.25) is not very far from the
simple empirical rule (3.24). Note also that, as pointed in Simar and Wilson
(1998), the bootstrap results are relatively stable to small changes in the selected
bandwidth. Therefore, as a reasonable first guess for hm , we might suggest the
use of the easy rule (3.24).
Finally, as suggested by Simar and Wilson (2006a), the value of hm (whatever
being the rule used to obtain it) has to be adjusted for scale and sample size:
 2m  s 
n
h = hm . (3.26)
n s2m
The density estimate is then obtained through:

2gh (δ) ifδ > 1,
f(δ) = (3.27)
0 otherwise.

How to generate the δi
from f(δ) and build a bootstrap sample X
?
Simar and Wilson (1998, 2006a) provide an easy to implement algorithm
where it is shown that the density estimate f(δ) is even not needed to generate
the δi . We only need the selected value of h and the original DEA scores
{δi ; i = 1, . . . , n}.
The algorithm is going as follows:
[1] we first draw a random sample of size n with replacement (as in the
naive bootstrap) from the set of the 2n reflected original DEA scores
{2 − δ1 , . . . , 2 − δn , δ1 , . . . , δn }, obtaining {δ̃i ; i = 1, . . . , n}.
[2] Then we smooth the naive bootstrap resampled values by perturbating δ̃i
with a random noise generated from the kernel density with scale given
by the bandwidth h. So we obtain:
˜
δ̃ i = δ̃i + h εi , i = 1, . . . , n,
62 Statistical inference in nonparametric frontier estimation

where εi is a random draw from a standard normal distribution (because

we have used gaussian kernel).
[3] Then we have to refine by correcting for the mean and the variance of the
smoothed values (see Efron and Tibshirani, 1993, for details):
˜ 
 δ̃ i − δ̃
δi = δ̃ +  , i = 1, . . . , n,
1 + h2 /s2

where δ̃ and s2 are the empirical mean and variance of the n values δ̃i .
[4] Finally we come back to measures greater than one by reflecting the values
smaller than one. For i = 1, . . . , n we define:

2 − δi if δi < 1,
δi =
δi otherwise.

[5] A bootstrap sample X  is now obtained by generating inefficient inputs

Xi , inside the DEA attainable set and conditional on the original input
mix ηi and the original output level Yi . This is achieved by defining:
δi
X  = {(Xi , Yi ) | Yi = Yi and Xi = Xi , i = 1, . . . , n}. (3.28)
δi
The denominator of the ratio multiplying the input vector Xi projects the
original observed data point Xi on the DEA efficient facet on the ray
defined by Xi , then the numerator projects the frontier point inside the
DEA attainable set, on the same ray, by the random bootstrap factor δi .
This is done for each data point i = 1, . . . , n.
As already explained above, by redoing the above steps 1–5 B times, we end
up with B bootstrap samples Xb . Then for any fixed point of interest (x, y),
we can build the Monte-Carlo sequence of pseudo estimates {δb (x, y)}B b=1 by
 
solving (3.10) with reference set Xb . The empirical distribution {δb (x, y)}B

b=1
is the bootstrap approximation of the sampling distribution of δ(x,  y).
Note that the computation burden can be important: to derive this series of B
values (where in practice B should be at least equal to 2000, to get reasonable
Monte-Carlo approximations even in the tails of the distribution), we have to
solve B DEA linear programs (3.10) (plus n DEA programs to get the original
DEA efficiency scores δi ). In many applications the point of interest (x, y)
will be each of the original data points (xi , yi ), i = 1, . . . , n. If confidence
intervals of the efficiency scores for each data point are desired, then we will
have to solve n(B +1) linear programs. However, there exist software packages
which have already implemented the bootstrap algorithm described above. We
Bootstrap techniques and applications 63

refer in particular to FEAR by P. Wilson (Wilson, 2005a, b, c) and to DEAsoft

developed and marketed by Performance Improvement Management Ltd.
In the Applications Part of the book we will illustrate this algorithm.

The debate on how correctly implement the bootstrap

In this simplified homogeneous version of the bootstrap it is easy to under-
stand why the naive bootstrap is not consistent (the same argument applies in
the more general bootstrap algorithms mentioned above). This issue opened
a debate on the consistency of the bootstrap in frontier models, reported in
the Journal of Productivity Analysis17 . Things are now very clear and can be
summarized as follows.
Suppose in the homogeneous case we use a naive bootstrap yielding the
values {δ̃i , i = 1 . . . , n} drawn randomly with replacement from the original
set {δi ; i = 1, . . . , n}. As observed in Simar and Wilson (1999a), it is easy to
verify that:
   n
 y)|P
Prob δ (x, y) = δ(x,  =1− 1− 1 > 0,
n
i.e. a fixed number depending only on the sample size n whatever the real DGP
P is. This problem does not vanish when n is increasing since:
 
 y)|P
lim Prob δ (x, y) = δ(x,  = 1 − e−1 ≈ 0.632. (3.29)
n→∞

The naive bootstrap is inconsistent because there is no reason why this proba-
bility should be equal to this fixed number, independently of any feature of the
real DGP P. In fact, if f (δ) is continuous on [1, ∞), the probability should be
zero since in this case:
 
 y) − δ(x, y)|P = 0.
Prob δ(x,

3.4.5 Applications of the bootstrap

The bootstrap has been introduced in several applications related to non-
parametric frontiers estimation. These useful applications include the correc-
tion for the bias and confidence intervals for efficiency scores; applications to
Malmquist indices and its various decomposition (Simar and Wilson, 1999c);
tests procedure to assess returns to scale (Simar and Wilson, 2002); statisti-
cal procedures to compare the means of several groups of producers can be
found in Simar and Zelenyuk (2003) and test for the equality of the densities

17 SeeFerrier and Hirschberg, 1997; Simar and Wilson 1999a; Ferrier and Hirschberg, 1999; Simar and
Wilson 1999b.
64 Statistical inference in nonparametric frontier estimation

of inefficiencies of two groups of firms is implemented in Simar and Zelenyuk

(2004).
In addition, there may be uncertainty about the structure of the underlying
statistical model in terms of whether certain variables are relevant or whether
subsets of variables may be aggregated. Tests of hypotheses about the model
structure have been introduced in Simar and Wilson (2001).
We apply most of these bootstrap-based procedures in the applications of
the Part II of this work. In particular we will explain how to use the bootstrap
algorithm to test returns to scale in Section 6.3 and to test the equality of the
mean of efficiency scores of two groups of units in Section 6.5.

3.4.6 Bootstrapping FDH estimators

For the FDH case, only very recently Jeong and Simar (2005) have proven
that the subsampling procedures provide consistent approximations of the sam-
 y) − δ(x, y). As described above sub-samples are
pling distribution of δ(x,
random samples of size m = [nγ ] where γ < 1, with replacement from the
original data {(Xi , Yi ), i = 1, . . . , n}. The procedure is very easy and very fast
to implement. Some Monte-Carlo experiments have shown that the procedure
is rather robust to the choice of the size of the sub-samples (choice of γ < 1) as
far as bias and variance of the estimators are concerned. However, for building
confidence intervals, the choice of this tuning parameter γ seems crucial for
obtaining sensible coverage probabilities. The problem comes from the fact
that an FDH efficiency score is characterized by only one data point, and this
data point will reappear too often in the bootstrap sample. This comes from the
discontinuous nature of the FDH boundary. Hence, Jeong and Simar (2005)
advocate the use of a smoothed FDH frontier for making inference. The dis-
continuity of the FDH frontier (in the p = q = 1 case, the “stair-case” aspect of
the frontier) may be a drawback of this estimator, because in most applications
we may expect the true frontier being continuous. They propose a linearized
version of the FDH (the Linearized FDH, called LFDH) by linear interpolating
adjacent FDH efficient vertices. When p = q = 1 the procedure is trivial and
can be done by hand; in multivariate setup they propose an algorithm to identify
the FDH vertices to be interpolated by using the Delannay triangulation (or tes-
sellation). See Simar and Wilson (2006a) for a non mathematical presentation
of the LFDH.
Chapter 4

NONPARAMETRIC ROBUST ESTIMATORS:

PARTIAL FRONTIERS

The main objective of this chapter is the presentation of robust frontiers as a

way for overcoming some of the drawbacks of traditional nonparametric tech-
niques (DEA/FDH based estimators). In particular, we deal with order-m and
order−α frontiers, and related efficiency measures. Known for being more ro-
bust to extremes or outliers in the data, these “partial” or “robust” frontiers have
several interesting properties that make them a very useful tool for empirical
applications. Among these properties we recall the same rate of convergence
of parametric estimators, that for practical works means that the “curse of di-
mensionality” (the request of thousands of observations for avoiding a great
statistical imprecision) can be avoided. In the first section, following Daraio
and Simar (2005a) which generalise previous results of Cazals, Florens and
Simar (2002), we reformulate the activity analysis framework under a proba-
bilistic perspective. Section two outlines the basic concepts and properties of
order-m frontiers. Besides, the following section presents another probabilistic
concept of frontier, the frontiers of order-α, and illustrates a new probabilistic
measure of efficiency. Then we summarize the main properties of order-m and
order-α frontiers. Afterwards, the presentation of the output oriented frame-
work is outlined. Finally, we illustrate a method recently introduced by Florens
and Simar (2005) to parametrically approximate robust and nonparametric fron-
tiers. The main advantage of this approach, in an applied perspective, is the
obtainment of robust estimators of coefficients of parametric models that can
easily be interpreted by the analyst. In the last section of the chapter we propose
a procedure to extend the parametric approximation in a full multivariate set-up
(multi-output and multi-input).
66 Nonparametric robust estimators: partial frontiers

4.1 A re-formulation based on the probability of being

dominated
Daraio and Simar (2005a), extending the ideas of Cazals, Florens and Simar
(2002), propose an alternative probabilistic formulation of the production
process.
The production process, presented in Section 2.3 within an activity analysis
framework, can be described by the joint probability measure of (X, Y ) on
Rp+ × Rq+ . This joint probability measure is completely characterized by the
knowledge of the probability function HXY (·, ·) defined as:

HXY (x, y) = Prob(X ≤ x, Y ≥ y). (4.1)

The support of the probability HXY (·, ·) is the production set Ψ and HXY (x, y)
can be interpreted as the probability for a unit operating at the level (x, y) to
be dominated. Daraio and Simar (2005a) point out that this function is a non-
standard distribution function, having a cumulative distribution form for X and
a survival form for Y . In the input oriented framework, this joint probability
can be decomposed as follows:

HXY (x, y) = Prob(X ≤ x | Y ≥ y) Prob(Y ≥ y)

= FX|Y (x|y) SY (y), (4.2)

where FX|Y (x|y) is the conditional distribution function of X and SY (y) is the
survivor function of Y ; we suppose the conditional distribution and survival
functions exist (i.e., SY (y) > 0 and FX (x) > 0). The conditional distribution
FX|Y is non-standard due to the event describing the condition (i.e.,Y ≥ y
instead of Y = y, the latter is assumed in a standard regression framework).
We can now define the efficiency scores (in a radial sense) in terms of the
support of these probabilities. The input oriented efficiency score θ(x, y) for
(x, y) ∈ Ψ is defined for all y with SY (y) > 0 as:

θ(x, y) = inf{θ | FX|Y (θx|y) > 0} = inf{θ | HXY (θx, y) > 0}. (4.3)

The idea here is that the support of the conditional distribution FX|Y (· | y)
can be viewed as the attainable set of input values X for a unit working at the
output level y. It can be shown that under the free disposability assumption, the
lower boundary of this support (in a radial sense) provides the Farrell-efficient
frontier, or the input benchmarked value.
A nonparametric estimator is then easily obtained replacing the unknown
FX|Y (x | y) by its empirical version:
n
i=1 1I(Xi ≤ x, Yi ≥ y)
F
X|Y ,n (x | y) = n , (4.4)
=1 1I(Yi ≥ y)
A re-formulation based on the probability 67

where 1I(·) is the indicator function that has to be read as follows: 1I(k) = 1 if
k is true, 1I(k) = 0 otherwise.
The resulting estimator of the input efficiency score for a given point (x, y)
coincides with the FDH estimator of θ(x, y):
θF DH (x, y) = inf{θ | (θ x, y) ∈ Ψ
 F DH } (4.5)
= inf{θ | FX|Y ,n (θx | y) > 0}. (4.6)
In the output oriented framework, the probability function HXY (·, ·) may be
decomposed as follows:
HXY (x, y) = Prob(Y ≥ y | X ≤ x) Prob(X ≤ x)
= SY |X (y|x) FX (x), (4.7)
where SY |X (y|x) = Prob(Y ≥ y | X ≤ x) denotes the conditional survivor
function of Y and FX (x) = Prob(X ≤ x) denotes the distribution function of
X that we assume exists, i.e. FX (x) > 0.
The output efficiency score may be defined accordingly:
λ(x, y) = sup{λ | SY |X (λy|x) > 0} = sup{λ | HXY (x, λy) > 0}. (4.8)
As for the input oriented case, a nonparametric estimator of λ(x, y) is ob-
tained by plugging in Equation (4.8) the empirical conditional survival function
SY |X,n (y|x) given by:
 XY,n (x, y)
H
SY |X,n (y|x) = , (4.9)
 XY,n (x, 0)
H
where,
 1
n
HXY,n (x, y) = 1I(xi ≤ x, yi ≥ y). (4.10)
n i=1
Again, this estimator coincides with the FDH estimator of λ(x, y).
The FDH estimator Ψ  F DH , as well as its convex version Ψ DEA , are very
sensitive to extremes and outliers, since, as estimators of the “full” set Ψ, they
envelop all the data points of the observed set X (this is seen by looking to
the inf and sup operator in (4.6) and (4.8)). The corresponding frontiers of
 F DH and Ψ
Ψ  DEA , can be viewed as estimators of the “full” frontier of Ψ. As
an alternative, partial frontiers can be investigated. They do not correspond
to the boundary of Ψ and are such that the full frontier can be viewed as a
limiting case of the partial frontiers. These frontiers correspond to another
benchmark frontier against which DMU will be compared. The advantage is
that their nonparametric estimators will not envelop all the data points and so
will be more robust to extreme and outlying data points. Two partial frontiers
have been investigated in the literature: the order-m frontiers and the order-α
quantile frontiers. They are introduced in the two next sections.
68 Nonparametric robust estimators: partial frontiers

4.2 Order-m frontiers and efficiency scores

The order-m frontiers, and their derived efficiency scores, have been pro-
posed by Cazals, Florens and Simar (2002). As pointed above, the support of
FX|Y (· | y) defines the attainable set of input values X for a unit working at
the output level y. Instead of looking at the lower boundary of this support, we
may define as a benchmark value, the average of the minimal value of inputs for
m units randomly drawn according FX|Y (· | y), i.e., units producing at least
the output level y. This defines the input order-m frontier.
In the simplest univariate case (all firms produce one unit of output), order-m
frontiers (input oriented framework) can be introduced as follows. Consider
a fixed integer m ≥ 1. The order-m lower boundary of X is defined as the
expected value of the minimum of m random variables X 1 , ..., X m drawn from
the distribution function of X. The formal definition is the following:
 ∞
φm = E[min(X 1 , ..., X m )] = [1 − FX (x)]m dx. (4.11)
0

The value φm is the expected minimum achievable input-level among m DMUs

drawn from the population of firms that, in this simplest case, produce one unit
of output. It can be proven that lim (m→∞) φm = φ and that, for all finite
value of m, φm ≥ φ.
The nonparametric estimator of φm is given by plugging the empirical dis-
tribution function of X in equation (4.11):
 ∞
φm,n = E[min(X
 1
, ..., X m )] = [1 − FX,n (x)]m dx. (4.12)
0

The relations between φ and φm remain valid between their empirical counter-
parts: φn and φm,n . For all finite values of m we have: φm,n ≥ φn .
We remark that in the standard case, φn ≤ Xi , i = 1, ..., n but this is no
more the case in the order-m frontier estimator, φm,n , even for large values of
m. The reasons for this different behavior of φm,n with respect to φn are mainly
due to the expected operator in the definition of φm (see equation (4.11)) and
to the finiteness of m.
The extension at the bivariate case is straightforward: we consider the process
generating the input levels X by the conditional distribution of X given that
Y ≥ y. The full frontier function φ(y) is defined as the minimal achievable
input level for producing at least the output y. It may be written as:

φ(y) = θ(x, y) x = inf{x | FX|Y (x | y) > 0}. (4.13)

Now, given a fixed integer value of m ≥ 1, we can define the (expected) order-
m lower boundary of X for DMUs producing at least y, as the expected value
of the minimum of m random variables X 1 , ..., X m drawn from the distribution
Order-m frontiers and efficiency scores 69

function of X given that Y ≥ y. The formal definition is the following:

 ∞
1
φm (y) = E[min(X , ..., X )|Y ≥ y] = m
[1 − FX|Y (x|y)]m dx (4.14)
0

Again, for all values of y and for all finite values of m, φm (y) ≥ φ(y) and for
all y, lim (m→∞) φm (y) = φ(y).
See Figure 4.1 for an illustration.

0.8
FX|Y(x | Y > = y)

m = 7 data points x with output Y > = y

0.6
φ7(y)
0.4

0.2 φ(y)

0 *5 *
0 10* * * 15 * 20 * 25 30 35
values of input x


Figure 
4.1. Input order-m frontier in the bivariate case. For any value of y, φm (y) =
E min(X 1 , . . . , X m ) | Y ≥ y . Here the stars X 1 , . . . , X 7 are m = 7 draws from
FX|Y (x|Y ≥ y) .

The expected frontier of order-m, φm (y), has an economic meaning: it is

the expected minimum value of input achievable among a fixed number of m
firms drawn from the population of firms producing at least a level of output y;
it represents another reasonable benchmark value for a firm producing a level
of output y.
The nonparametric estimation of φm (y) can be done by plugging in the
empirical distribution function in equation (4.14):
 ∞
φm,n (y) = E[min(X
 1
, ..., X m )|Y ≥ y] = [1 − FX|Y (x|y)]m dx (4.15)
0

where X 1 , ..., X m are m i.i.d. (independent and identically distributed) R.V.

(Random Variables) generated by the empirical distribution function of X given
that Y ≥ y whose distribution function is FX|Y (x|y).
The relations between the order-m frontier and the full frontier remain valid
with their estimators φm,n (y) and φn (y), the FDH estimator of the frontier.
70 Nonparametric robust estimators: partial frontiers

The extension to the multivariate inputs case, is done by defining order-m

efficiency measures.
For a given level of outputs y in the interior of the support of Y , consider
now m i.i.d. random variables Xi , i = 1, . . . , m generated by the conditional
p-variate distribution function FX|Y (x | y) and define the set:

Ψm (y) = {(x, y
) ∈ Rp+q

+ | x ≥ Xi , y ≥ y, i = 1, . . . , m}. (4.16)

Then, for any x, we may define:

θ̃m (x, y) = inf{θ | (θx, y) ∈ Ψm (y)}. (4.17)

Note that θ̃m (x, y) may be computed by the following formula:

 
 Xj 
i
θ̃m (x, y) = min max . (4.18)
i=1,...,m j=1,...,p xj

θ̃m (x, y) is a random variable since the Xi are random variables generated by
FX|Y (x | y).
The order-m input efficiency measure is defined, according to Daraio and
Simar (2005a), as follows:

θm (x, y) = E(θ̃m (x, y) | Y ≥ y), (4.19)

 ∞
= (1 − FX|Y (ux | y))m du (4.20)
0
 ∞
= θ(x, y) + (1 − FX|Y (ux | y))m du, (4.21)
θ(x,y)

A nonparametric estimator of θm (x, y) is straightforward: we replace the true

FX|Y (· | y) by its empirical version, FX|Y,n (· | y). We have:
 θ̃m (x, y) | Y ≥ y)
θ̂m,n (x, y) = E(
 ∞
= (1 − FX|Y,n (ux | y))m du, (4.22)
0
 ∞
= θ̂n (x, y) + (1 − FX|Y,n (ux | y))m du (4.23)
θ̂n (x,y)

Hence, in place of looking for the lower boundary of the support of FX|Y (x | y),
as was typically the case for the full-frontier and for the efficiency score θ(x, y),
the order-m efficiency score can be viewed as the expectation of the minimal
input efficiency score of the unit (x, y), when compared to m units randomly
drawn from the population of units producing more outputs than the level y.
This is certainly a less extreme benchmark for the unit (x, y) than the “absolute”
minimal achievable level of inputs: it is compared to a set of m peers producing
Order-m frontiers and efficiency scores 71

more than its level y and we take as benchmark, the expectation of the minimal
achievable input in place of the absolute minimal achievable input.

Note that the order-m efficiency score is not bounded by 1: a value of

θm (x, y) greater than one indicates that the unit operating at the level (x, y) is
more efficient than the average of m peers randomly drawn from the population
of units producing more output than y. Then for any x ∈ Rp+ , the expected
minimum level of inputs of order-m is defined as x∂m (y) = θm (x, y) x which
can be compared with the full-frontier x∂ (y) = θ(x, y) x.

Order-m frontiers are estimators of the frontier, that for finite m, do not
envelop all the observed data points and therefore, are less sensitive to extreme
points and/or to outliers. As m increases and for fixed n, θ̂m,n (x, y) → θ̂n (x, y).

Daraio and Simar (2005b) define convex and local convex order-m frontiers
as well as a practical method to compute them.

Economic meaning of order-m input efficiency measures

Consider the firm (x, y); it produces a level of output y using a quantity x of
inputs. We recall that φm (y) is not the efficient frontier of the production set,
but it gives the expected minimum input among a fixed number of m potential
competing firms producing more than y. The comparison of x with φm (y) is
important, from an economic point of view, as it gives a clear indication of
how efficient the firm is, compared with these m potential firms. The value m
represents the number of potential firms (drawing from the population of firms)
producing at least the output level of y, against which we want to benchmark
the analyzed firm.

Let us give some examples of the economic meaning of the order-m input
efficiency measures. If a firm (x, y) has an efficiency score θm,n (x, y) = 0.9
(1.4), means that it uses 10% more inputs -radial extension- (uses 40% less
inputs - proportionate reduction) than the expected value of the minimum input
level of m other firms drawn from the population of firms producing a level of
output ≥ y. On the contrary, if θm,n (x, y) = 1, the firm (x, y) uses the same
level of inputs than the expected value of the minimum input level of m other
firms drawn from the population of firms producing at least y of output, i.e.
the firm is on the efficient boundary of the order-m frontier in the input space
direction.
72 Nonparametric robust estimators: partial frontiers

Computational aspects
For the computation of order-m efficiency θ̂m,n (x, y) the univariate integral
(4.23) could be evaluated by numerical methods18 , even when the number of
inputs p ≥ 1.
However, numerical integration can be avoided by an easy Monte-Carlo
algorithm, proposed by Cazals, Florens and Simar (2002), that we describe
below, as fast for small values of m such as m = 10, but much slower when m
increases:
[1] For a given y, draw a sample of size m with replacement among those Xi
such that Yi ≥ y and denote this sample by (X1,b . . . , Xm,b ).
  X j 
b (x, y) = min i,b
[2] Compute θ̃m i=1,...,m maxj=1,...,p xj
.

[3] Redo [1]-[2] for b = 1, . . . , B, where B is large.

1 B
[4] Finally, θ̂m,n (x, y) ≈ B
b
b=1 θ̃m (x, y).

The quality of the approximation can be tuned by increasing B, but in most

applications, say B = 200, seems to be a reasonable choice.

4.3 Order-α quantile-type frontiers

The partial order-m frontiers have the advantage of sharing very nice prop-
erties and since m is an integer, the mathematics behind this object is not too
complicated. Even if the order m of the frontier has some economic interpreta-
tion (benchmarking against m competitors), in practice and as discussed above,
m serves as a trimming parameter which allows to tune the percentage of points
that will lie above the order-m frontier. The idea behind order-α quantile-type
frontier is to go the other way around: determine the frontier by fixing first the
probability (1 − α) of observing points above this order-α frontier.
Using the probabilistic formulation developed above in Section 4.1, it is easy
to adapt the order-m ideas to order-α quantile type-frontiers. These estimators
were introduced for the univariate case by Aragon, Daouia and Thomas-Agnan
(2003) and extended to the multivariate setting by Daouia and Simar (2004).
As for order-m frontiers, we develop the presentation for the input oriented
case. A summary for the output oriented case is provided in Section 4.5. Con-
sider first the case where we only have one input and several outputs y ∈ Rq+ .
In the preceding section, for a firm operating at the level (x, y), the benchmark
is the order-m partial frontier determined by the expected minimal input among
m peers randomly drawn in the population of firms producing at least a level y

18 For
the numerical integration we use the build-in Matlab “quad” procedure (based on adaptive Simpson
quadrature).
Order-α quantile-type frontiers 73

of outputs. Here, for the same unit, the benchmark will be the order-α quantile
frontier defined as the input level not exceeded by (1 − α) × 100-percent of
firms among the population of units producing at least a level y of outputs.
Formally:

φα (y) = inf{x | FX|Y (x | y) > 1 − α}. (4.24)

This frontier can be viewed as a nonstandard conditional quantile frontier, where

the word “nonstandard” means that we focus on the unusual conditioning Y ≥ y
in the cdf FX|Y (· | y) considered here.
In Figure 4.2 we illustrate the concept for α = 0.90. In this picture, φ(y) is the
full frontier level, it is given by the left boundary of the support of FX|Y (· | y),
φα (y) corresponds to the (1 − α) quantile of FX|Y (· | y).

0.8
FX|Y(x | Y >=y)

0.6

0.4
φ0.90(y)

0.2 φ(y)
1 − α = 0.10
0
0
*5 * * * * 15 *
10 20
* 25 30 35
values of input x

Figure 4.2. Input order-α frontier in the bivariate case, α = 0.90.

As for the order-m frontier, this concept can be easily extended to the multiple
inputs case by defining the order-α input efficiency score for a unit operating
at the level (x, y), as follows:

θα (x, y) = inf{θ | FX|Y (θx | y) > 1 − α}. (4.25)

It is a radial version of quantile type function φα (y) of the univariate input

case: in this latter case we have φα (y) = θα (x, y)x. From the expression
(4.25), it appears clearly that θα (x, y) converges to the usual Farrell-Debreu
input efficiency score θ(x, y) when α → 1. The order-α efficiency score has
a nice interpretation: for instance if θα (x, y) = 1, then the unit is said to be
efficient at the level α × 100% since it is dominated by firms producing more
output than y with a probability 1 − α. If θα (x, y) < 1, then the unit (x, y)
74 Nonparametric robust estimators: partial frontiers

has to reduce its input to the level θα (x, y)x to reach the input efficient frontier
of level α × 100%. Note that here θα (x, y) can be greater than one indicating
that a firm (x, y) can increase its input by a factor θα (x, y) to reach the same
frontier. Therefore, this latter firm is considered as super-efficient with respect
to the order-α frontier level.
Nonparametric estimators are easily obtained by plugging, as for the order-m
frontiers, the empirical cdf in the expression above, we have:
φα,n (y) = inf{x | FX|Y,n (x | y) > 1 − α}, (4.26)
and for the multivariate input case, we obtain:
θα,n (x, y) = inf{θ | FX|Y,n (θx | y) > 1 − α}. (4.27)

Here again it appears clearly that when α → 1, θα,n (x, y) converges to the
FDH input efficiency score θF DH (x, y).
The nonparametric estimators of order-α frontier or efficiency scores shares
√
the same properties than their order-m analogs. In summary, they are n-
consistent estimator of their population analogs, they are asymptotically unbi-
ased and normally distributed with a known expression for the variance (see
Aragon, Daouia and Thomas-Agnan (2005) for the properties of φα,n (y) and
Daouia and Simar, 2004 for those of θα,n (x, y)). Using tools of robustness
theory, it is shown in Daouia and Simar (2004) that the order-α frontiers are
more robust to extremes than the order-m frontiers.

Computation of the estimators of the order-α efficiency scores

n
Let My = nSY,n (y) > 0 = i=1 1I(Yi ≥ y), and define:
Xik
Xi = max , i = 1, · · · , n.
k=1,···,p xk
y
For j = 1, · · · , My , denote by X(j) the j-th order statistic of the observations
y y y
Xi such that Yi ≥ y: X(1) ≤ X(2) ≤ . . . ≤ X(M y)
. Then we have,
 y My
1I(Xi ≤ θ)
i|Yi ≥y j=1 1I(X(j) ≤ θ)
FX|Y,n (θx|y) = =
My My
 y

 0 if θ < X(1)
j y y
= if X(j) ≤ θ < X(j+1) , j = 1, · · · , My − 1


My
y
1 if θ ≥ X(My ) .
It follows,
 y
X((1−α)M if (1 − α)My ∈ N
θα,n (x, y) = y
y)
X([(1−α)M ]+1) otherwise,
Order-α quantile-type frontiers 75

where N denotes the set of all nonnegative integers. If x is univariate, φα,n (y) =
θα,n (x, y)x. Therefore, the computation of θα,n (x, y) is very fast and very easy
since it only implies sorting routines.

Monotone estimators of the partial frontiers

As discussed in Cazals, Florens and Simar (2002) and Aragon, Daouia and
Thomas-Agnan (2005), for the univariate input case, the partial frontiers φm (y)
and φα (y) are monotone function of y. However, the nonparametric estimators
φm,n (y) and φα,n (y) can be not monotone in y in finite samples. Daouia and
Simar (2005) propose an easy way to isotonize these estimators to achieve the
appropriate monotonicity. It is shown that these isotonized versions are even
more robust to extreme and outliers than the original nonparametric estimators.

A new measure of efficiency

As pointed by Daouia and Simar (2004), for every attainable point (x, y) ∈ Ψ,
there exists an α such that θα (x, y) = 1. This α could serve as an alternative
measure of input efficiency score. If FX|Y (x | y) is continuous in x, this
quantity is given (input orientation) by:
αinput (x, y) = 1 − FX|Y (x | y). (4.28)
In other words, one may set the estimated performance measure for a unit
operating at the level (x, y) to be the order α of the estimated quantile frontier
which passes through this unit. This new concept of efficiency, the α efficiency,
is illustrated in Figure 4.3. Suppose that we want to measure the efficiency score
of a unit located at the point A (this is a unit which produces a level of output y
using a level of input indicated by the point A on the x-axis). Its input efficiency
score is equal to αinput (A) = 0.30 since 70% of the units producing at least
the level y of output are using less input than unit A.
This idea has been first proposed by Aragon, Daouia and Thomas-Agnan
(2003) in the univariate case: they analyze the properties of these measures and
the properties of their nonparametric estimators. The multivariate extension
comes from Daouia and Simar (2004). These nonparametric estimators are ob-
tained by using the empirical counterparts of the distribution function. We have
to take into account the discreteness of empirical distributions. It can be shown
that the correct expression for the nonparametric estimator of αinput (xi , yi ), in
the input orientation, is given by:
1
 input (xi , yi ) = 1 − FX|Y,n (xi |yi ) +
α . (4.29)
Myi
In Figure 4.4 we illustrate all the nonparametric and robust measures introduced
in the previous sections (input oriented framework). The illustration is presented
76 Nonparametric robust estimators: partial frontiers

0.8 αinput(A) = 1− 0.70 = 0.30

FX|Y(x | Y > = y)

0.6

0.4

0.2
φ(y)

0 *5 *
0 10* * * 15 * 20 * 25 30 35
Data point A values of input x

Figure 4.3. An illustration of αinput measure of efficiency in the bivariate case.

for α = 0.90 and m = 7. In this figure, φ(y) is the full frontier level, it is given
by the left boundary of the support of FX|Y (· | y), φα (y) corresponds to the
(1 − α) quantile of FX|Y (· | y) and φm (y) is the expectation of the minimum
of m virtual data points (here 7) generated by FX|Y (· | y). We represent by the
stars on the x-axis, 7 potential values of these random data points and we show
where φ7 (y) could be around. The measure αinput is the new probabilistic
efficiency measure defined above.

0.8 αinput(A)=1− 0.70=0.30

FX|Y(x | Y > = y)

m =7 data points x with output Y >=y

0.6
φ7(y)

0.4
φ0.90(y)
φ(y)
0.2
1− α = 0.10

0 *5 *
0 10* * * 15 * 20 * 25 30 35
Data point A values of input x

Figure 4.4. Illustration of full and partial frontiers in the input orientation. Here m = 7 and
α = 0.90: the solid curve is the conditional cdf FX|Y (x | y) = Prob(X ≤ x | Y ≥ y). The
stars on the x-axis represent 7 potential observations generated by FX|Y (· | y).
Properties of partial frontier estimators 77

4.4 Properties of partial frontier estimators

In this section we summarize the most important properties of partial fron-
tiers.

4.4.1 Statistical properties

The asymptotic properties of order-m and order-α frontiers and related ef-
ficiency scores have been analyzed by Cazals, Florens and√Simar (2002) and
Daouia and Simar (2004). In particular, they show the n-consistency of
√
θ̂m,n (x, y) to θm (x, y) for m fixed, as n → ∞, and the n-consistency of
θ̂α,n (x, y) to θα (x, y) for α fixed, as n → ∞ respectively. In addition, they
prove that the related nonparametric estimators are asymptotically unbiased and
asymptotically normally distributed.

4.4.2 Robust estimators of the full frontier

Another interesting feature of the partial frontiers is that when m → ∞ (or
when α → 1), the partial frontiers φm (y) (φα (y), respectively) converge to the
full frontier φ(y). For multivariate inputs we have the same properties for the
efficiency scores: θm (x, y) (θα (x, y), respectively) converges to θ(x, y), the
Farrell-Debreu efficiency score computed with respect to the full frontier.
We have also similar properties for the nonparametric estimators φm,n (y)

(φα,n (y), respectively) and θm,n (x, y) (θα,n (x, y)) which converge to their
FDH correspondent φn (y) or θF DH (x, y) as m → ∞ (α → 1, respectively).
These properties allow the use of partial frontiers as robust estimators of the
full frontier itself. The idea is to consider the order of the frontier (either m or
α) as a function of n such that m = m(n) → ∞ and α = α(n) → 1 as n → ∞
at appropriate rates, so that the partial frontiers converge to the full frontier and
we can hope that their nonparametric estimators will also converge to the full
frontier. For finite n however, the partial order frontiers do not envelop all the
data points and hence are more robust to extreme and outlying points.
All the theory behind this idea has been provided in Cazals, Florens and
Simar (2002) and in Daouia and Simar (2004): the resulting nonparametric
estimators of the partial frontiers converge indeed to the full frontier itself as
n → ∞, sharing the same limiting Weibull distribution than the FDH estimator.
For instance for the input oriented order-α efficiency score we have:
 
n1/(p+q) θα(n),n (x, y) − θ(x, y)
approx.
∼ W eibull(µp+q
N W,0 , p + q) as n → ∞,

where µN W,0 is a constant described in Park, Simar and Weiner (2000). Order-
m estimators share the same properties and the same is true for the output
oriented measures.
See Daraio and Simar (2005a), Daouia and Simar (2004, 2005) and Daouia,
Florens and Simar (2005) for illustrations of these robustness properties in
78 Nonparametric robust estimators: partial frontiers

simulated samples. These robustness properties can also be exploited in the

detection of outliers as explained below.

4.4.3 Advantages of using partial frontiers

Partial frontiers and related measures of efficiency show some nice statistical
properties together with several “appealing” economic features that deserve
some comments.
First of all, for their ability of not enveloping all data points, these robust
measures of frontiers and the related efficiency scores are less influenced and
hence more robust to extreme values and outliers. For empirical applications,
this property means that we can avoid one of the more important limitation of
the traditional nonparametric estimators, related to their deterministic nature.
Second, as a consequence of their statistical properties, robust measures of
efficiency do not suffer of the curse of dimensionality shared by most nonpara-
metric estimators and by the DEA/FDH efficiency estimators. Again this is a
very important property for empirical works since it states that we can work
with samples of moderate size and we do not require large samples to avoid
imprecise estimation (e.g. large confidence intervals). In a lot of real data
applications, in fact, there is not the availability of such a huge amount of data.
Third, and even more important is the economic interpretation of order-m
measures of efficiency, and the appealing notion of order-α and in particular
of α measures of efficiency. In particular, the parameter m has a dual nature.
It is defined as a “trimming” parameter for the robust nonparametric estima-
tion. It defines also the level of benchmark one wants to carry out over the
population of firms. We propose to use m in its dual meaning to provide both
robust estimations and a potential competitors analysis. The first task can be
accomplished by plotting the percentage of points outside order-m frontier in
function of m. By consulting this graph we may choose the value of m which
corresponds to the target (or desired) degree of robustness, i.e. the percentage
of high performers of the population we want to exclude in our more realistic
benchmarking comparison (comparative analysis or performance evaluation).
The second application of m concerns the evaluation of a potential com-
petitors scenario, in which for each firm we choose the potential competitors
(m) against which we want to benchmark its performance and in this choice
we take into account the turbulence of the competitive environment faced by
the firm. The choice of m, hence, is related to the dynamics, the structure and
characteristics of the considered market. We might also build several simulated
competitive scenarios, by setting several values of m. In setting the values of
m, however, there are neither fixed rules nor automatic procedure, as m can be
any positive integer number. Nevertheless, for the choice of the “most reason-
able” values of m, the statistical properties of estimators have to be considered
together with the economic meaning of parameters. Even if m is independent
Properties of partial frontier estimators 79

from n (the number of analyzed firms, the sample size), the values of m might
be fixed by considering the possible number of potential competitors we want
“more realistically” benchmark our firm against. Furthermore, in most empir-
ical applications of order-m efficiency measures, we noted that for m ≥ 200
the order-m efficiency score is almost equal to the FDH efficiency score, i.e.
the asymptotic result limm→∞ θm,n (x, y) = θF DH,n (x, y) in practice, hap-
pens already for values such that m ≥ 200. Leading by these considerations,
we can define a grid of values for m to use in the sensitivity analysis, to build
simulated competitive scenarios, that can be particularly useful for the analysis
of industries and markets with an intensive and dynamic competition.
The economic meaning of order-α and α measures of efficiency is very inter-
esting and useful. If order-α measures of efficiency could be roughly considered
as a kind of “continuous version” of their order-m brothers, α measures have
also an immediate economic meaning. They are based on the idea that there ex-
ists for each firm in the comparison set a quantile frontier which passes through
it, on which the firm is efficient (either along the input dimension or along the
output dimension). If the quantile on which the firm is efficient (in the input
orientation) is 0.2, this means that there are 80% (1 − 0.2 = 0.8) of the firms in
the comparison set (firms producing at least the same level of outputs) which
outperform the considered firm by using less inputs. So that we can interpret
1 − α as a firm’s probability of being dominated on the input dimensions by
the other firms producing at least the same level of output. Accordingly, is the
interpretation for the output oriented case.

4.4.4 Detection of outliers

Partial frontiers (order-m and order-α) have been shown to be also very
useful to detect outliers or extreme data. Indeed, the problem of extremes and
outliers is a relevant question that has been addressed by several authors (among
others by Wilson, 1995).
Simar (2003a), proposes the use of order-m frontiers to accomplish this
difficult task. Of course the same procedure may apply order-α measures.
Here, in the presentation, we follow Simar (2003a) and detail the procedure for
the input orientation. The adaptation to the output orientation is straightforward.
The main idea is that if some data points remain outside the partial order frontier
even when m increases (or when α increases), this may warn these data points
as potential outliers. Being outside the partial frontier is warned by values of
θm,n (Xi , Yi ) (or θα,n (Xi , Yi )) larger than 1.
The first issue to address is the choice of the values of m (or α). In choosing
these values, it has also to be decided from what threshold level of the order-m
(order-α) efficiency score we decide to flag a potential outlier. Some reasonable
threshold values distant from one could be chosen on a grid of values 1 + τ ,
where τ = 0.20, 0.30, 0.40, 0.50 (for the output oriented case, a point outside
80 Nonparametric robust estimators: partial frontiers

the partial frontier being detected by efficiency scores less than one, the grid
would be defined by 1 − τ ).
Thus for the order-m scores, the main steps of the computations are:
[1] Compute for each data point (Xi , Yi ), for i = 1, ..., n, its leave-one-out
input efficiency score i.e. its order-m input efficiency score leaving out
(i)
the observation (Xi , Yi ) from the reference set. Denote by θm,n (Xi , Yi )
the “leave-one-out” efficiency score and the corresponding reference set
by X (i) .
(i)
[2] Compute θm,n (Xi , Yi ), for i = 1, ..., n, for several reasonable value of
m, e.g. m = 10, 25, 50, 75, 100, 150.
[3] Compute also the number of points used to estimate the conditional distri-
bution function FX|Y (x | y ≥ Yi ), i.e. the number of points in X (i) with
y ≥ Yi . Denote this number as Ninput (Xi , Yi ). It is the number of points
used to estimate the p-variate distribution function. If Ninput (Xi , Yi ) is
small or even equal to zero, the correspondent point (Xi , Yi ) lies at the
border of the sample values X . Ninput (Xi , Yi ) thus indicates how the
point (Xi , Yi ) is near to the border of the support of data points.
[4] For each values in τ , plot the percentage of points in the sample X with
θm,n
(i)
(Xi , Yi ) ≥ 1 + τ (4.30)
as a function of m. This curve represents the percentage of points outside
the order-m frontier as function of m for all the threshold values defined
by the grid τ .
The computations for the output-oriented case are mutatis mutandis the same
as above.
The interpretation of the results is based on the following statement: any
point is outlying the cloud of points of data set X in the input direction (output
direction) when its order-m input (output) efficiency score is greater (smaller)
than one. The data points with input order-m efficiency measure greater than
one, even if m increases, or with small values of Ninput should be flagged as
being extremes. In the output oriented case, the points with order-m efficiency
score smaller than one, even if m increases, or with small values of Noutput
should also be flagged as extremes. When data points are detected as extremes
in both directions (input and output), they are warned as potential outliers.
For doing this, we need to choose at which level we consider that m is
“large”. This may be achieved by looking at the plots obtained above, showing
the percentage of points outside order-m frontiers for the different threshold
values τ . By construction, these curves should decrease when m increases,
and if there are no outliers, they should converge approximately linearly to the
Summary of the results for the output oriented case 81

percentage of points having a leave-one-out FDH score greater than one (smaller
than one for the output-oriented case). As a consequence, any strong deviation
from linearity should indicate the potential existence of outliers: if the curves
show an elbow effect (sharp negative slope, followed by a smooth decreasing
slope) they indicate that the points remaining outside the order-m frontier for
this value of m have to be further analyzed by the procedure described above
and eventually warned as potential outliers.
Beside this, we have also to select m such that a reasonable percentage of
points remains outside the frontier.
√
It has been suggested (Barnett and Lewis,
1995) to use the rule of thumb nn as a reasonable upper bound for the percentage
of outliers in a sample of size n.
Of course, once the potential outliers have been identified, they have to be
carefully analyzed to understand “why” they are outliers. Very often outliers
(when not due to errors) contains useful information on the process under analy-
sis (missing variables in the model, etc. . . ).

4.5 Summary of the results for the output oriented case

We briefly summarize here the corresponding concepts for the output oriented
case.

Order-m frontiers and related efficiency scores

Note that when y is univariate, the frontiers can be defined in terms of pro-
duction functions. For instance, the full frontier (the production function) for
an input level x is defined as:
ψ(x) = λ(x, y) y = sup{y | SY |X (y | x) > 0}. (4.31)
Then, the expected order-m frontier is defined, for a fixed integer value of
m ≥ 1 as the expected value of the maximum of m random variables Y 1 , ..., Y m
drawn from the conditional distribution function of Y given that X ≤ x.
Its formal definition is:
 ∞
ψm (x) = E[max(Y 1 , ..., Y m )|X ≤ x] = (1 − [FY |X (y | x)]m ) dy.
0
(4.32)
Its nonparametric estimator can be defined and computed by:
 ∞
ψm,n (x) = E[max(Y
 1
, ..., Y m )|X ≤ x] = (1 − [FY |X,n (y|x)]m ) dy.
0
(4.33)
The asymptotic properties of this estimator are the same, mutatis mutandis, than
the estimator in the input-oriented framework.
The multivariate extension y ∈ Rq+ can be presented as follows. For a given
level of inputs x in the interior of the support of X, consider m i.i.d. random
82 Nonparametric robust estimators: partial frontiers

variables Yi , i = 1, . . . , m generated by the conditional q-variate distribution

function FY |X (y | x) = Prob(Y ≤ y | X ≤ x) and define the set:

Ψm (x) = {(x
, y) ∈ Rp+q

+ | x ≤ x, Yi ≤ y, i = 1, . . . , m}. (4.34)
Then, for any y, we may define:
λ̃m (x, y) = sup{λ | (x, λy) ∈ Ψm (x)}
 
Y j 
i
= max min . (4.35)
i=1,...,m j=1,...,q yj
The order-m output efficiency measure is defined as follows. For any y ∈
Rq+ , the (expected) order-m output efficiency measure denoted by λm (x, y) is
defined for all x in the interior of the support of X as:
λm (x, y) = E(λ̃m (x, y) | X ≤ x), (4.36)
where we assume the existence of the expectation.
As above, for any y ∈ Rq+ and for all x in the interior of the support of X,
if λm (x, y) exists, we have:
 ∞
λm (x, y) = 1 − (1 − SY |X (uy | x))m du
0
 λ(x,y)
= λ(x, y) − (1 − SY |X (uy | x))m du. (4.37)
0

From this, it is easily seen that limm→∞ λm (x, y) = λ(x, y). A nonparametric
estimator of λm (x, y) is given by:
 ∞
 m,n (x, y) =
λ 1 − (1 − SY |X,n (uy | x))m du
0
 
λn (x,y)
 n (x, y) −
= λ (1 − SY |X,n (uy | x))m du. (4.38)
0
Let us give an example of the economic meaning of the order-m output effi-
ciency measure λ  m,n (x, y) by looking at a firm operating at the level (x, y)
and such that λm,n (x, y) = 1.8. This firm produces a level of output -in radial
extension- that is equal to 0.56, i.e. 1/1.8, times the expected value of the
maximum level of output of m other firms drawn from the population of firms
using a level of inputs ≤ x. A value of λm,n (x, y) = 0.5 would indicate that
the firm produces 2 = 1/0.5 times more output than the expected value of the
maximum level of output of m other firms drawn from the same population.
As for the input oriented case, the computation of the order-m efficiency
 m,n (x, y) may be done either by numerical integration, calculating the uni-
λ
variate integral in (4.38), or adapting the Monte Carlo algorithm presented in
Section 4.2 (for the input oriented case) as follows:
Summary of the results for the output oriented case 83

[1] For a given x, draw a sample of size m with replacement among those Yi
such that Xi ≤ x and denote this sample by (Y1,b , . . . , Ym,b ).
  Y j 
i,b
[2] Compute λ̃bm (x, y) = maxi=1,...,m minj=1,...,q yj
.

[3] Redo [1]-[2] for b = 1, . . . , B, where B is large.

1 B
[4] Finally, λ̂m,n (x, y) ≈ B
b
b=1 λ̃m (x, y).

Again, the approximation can be tuned by increasing B.

Order-α frontiers and related efficiency scores

Let x ∈ Rp+ and y be univariate, then the order-α production function can
be defined as:
ψα (x) = sup{y | SY |X (y | x) > 1 − α}. (4.39)
It is the level of output not exceeded by (1 − α) × 100-percent of firms among
the population of units using less input than x. For the multivariate output case,
we define the output oriented order-α efficiency score of a unit operating at the
level (x, y) as:
λα (x, y) = sup{λ | SY |X (λy | x) > 1 − α}. (4.40)
Again λα (x, y) converges to λ(x, y), the usual Farrell-Debreu output efficiency
score, when α → 1. In the univariate output case we have ψα (x) = λα (x, y)y
and in general, λα (x, y) gives the proportionate reduction (if < 1) or increase
(if > 1) in outputs needed to move the unit (x, y) to the order-α output frontier,
so that it is dominated by firms using less input than x with a probability 1 − α.
The nonparametric estimators are obtained by plugging the empirical version
of the survival function SY |X (y | x) in the expression above. For the univariate
output case we have:
ψα,n (x) = sup{y | SY |X,n (y | x) > 1 − α}. (4.41)
In the multivariate setup we have:
 α,n (x, y) = sup{λ | S
λ Y |X,n (λy | x) > 1 − α}. (4.42)
Mutatis mutandis, √ the estimators shares the same properties as for the in-
put oriented case ( n-consistency and asymptotic normality). We have also
 α,n (x, y) converges to the FDH estimator λ
λ  F DH (x, y), if α → 1.

The computations of λα,n (x, y) is obtained through the following simple
sorting algorithm. Indeed, define:
Yik
Yi = min , i = 1, · · · , n,
k=1,···,q yk
84 Nonparametric robust estimators: partial frontiers

and let Nx = nFX,n (x) = ni=1 1I(Xi ≤ x) be non null. For j = 1, · · · , Nx ,
denote by Y(j)
x the j-th order statistic of the observations Y such that X ≤ x:
i i
Y(1) ≤ Y(2) ≤ . . . ≤ Y(N
x x x
x)
. We have:
 
i|Xi ≤x 1I(Yi ≥ λy) i|Xi ≤x 1I(λ ≤ Yi )
S
Y |X,n (λy|x) = =
Nx Nx
Nx
j=1 1I(λ ≤ Y(j) )
x
=
Nx


 1 if λ ≤ Y(1)
x
Nx −j
= if Y(j)
x < λ ≤ Yx
(j+1) , j = 1, · · · , Nx − 1


Nx
0 if λ > Y(N x
x)
.

It follows,

 α,n (x, y) =
Y(αN
x
x)
if αNx ∈ N ∗
λ (4.43)
Y([αN
x
x ]+1)
otherwise,

where N ∗ denotes the set of positive integers and [αNx ] denotes the integral
part of αNx .

A new probabilistic measure of efficiency

As we have seen for the input oriented case, for every attainable point (x, y) ∈
Ψ, there exists an α such that λα (x, y) = 1, and the value of this α could become
an alternative measure of output efficiency score. Under the continuity of the
survival function SY |X (y | x), this new probabilistic measure can be defined
as αoutput (x, y) = 1 − SY |X (y | x). Put it in another way, we may set the
estimated performance measure for a unit operating at the level (x, y) to be the
order α of the estimated quantile frontier which passes through this unit. This
new measure of efficiency is illustrated in Figure 4.5 which summarizes also all
the other robust and nonparametric efficiency measures in the output oriented
framework. Suppose that we want to measure the efficiency score of a unit
located at the point A (this unit uses a level of input x for producing a level of
output y indicated by the point A on the x-axis). Its output efficiency score is
equal to αoutput (A) = 0.75 since 25% of the units using a level of input less or
equal to x are producing a level of output higher than those of unit A.
A nonparametric estimator of this new efficiency measure (which takes into
account the discreteness of the empirical survival function) for the output ori-
ented case is given by:
1
 output (xi , yi ) = 1 − SY |X,n (yi |xi ) +
α . (4.44)
Nx
Parametric approximations of robust nonparametric frontiers 85

Figure 4.5 also illustrates the main efficiency measures introduced in this section
for the output oriented framework. Here α = 0.90, m = 7, and ψ(x) is the full
frontier level, it is given by the right boundary of the support of SY |X (· | x).
ψα (x) corresponds to the (1 − α) quantile of SY |X (· | x) and ψm (y) is the
expectation of the maximum of m virtual data points generated by SY |X (· | x).
The stars on the x-axis are 7 potential values of these random data points and
ψ7 (y) shows where ψm (y) could be around.

αoutput(A)=1− 0.25 = 0.75

0.8
SY|X(y | X <= x)

0.6 m data points y with input X <= x

ψm(x)
ψα(x) ψ(x)
0.4

0.2
1− α

0 * * * * **
0 2 4 6 8 10 12 14 16
Data point A values of output y

Figure 4.5. Illustration of full and partial frontiers in the output orientation. Here m = 7 and
α = 0.90: the solid curve is the conditional survival function SY |X (· | x) = Prob(Y ≥ y |
X ≤ x). The stars on the x-axis represent 7 potential observations generated by SY |X (· | x).

4.6 Parametric approximations of robust and

nonparametric frontiers
Nonparametric estimators are very appealing because they rely on very few
assumptions on the shape of the frontier (like free disposability and eventually
concavity) and on the stochastic process which generates the data on Ψ (except
some general regularity conditions). However, this flexibility and generality has
some drawbacks. For instance it is not easy in these nonparametric models to
make a sensitivity analysis of the production of outputs to some particular inputs
or to infer about elasticities, and other coefficients. On the contrary, parametric
models for the production frontier (or for a cost or an input function) allow
easier and richer economic interpretation. But of course, we have the risk of
misspecification resulting in a risk of inconsistent estimates.
The parametric approaches to model a production frontier are straightfor-
ward and estimates are obtained trough least-squares methods or maximum
likelihood techniques. However, as pointed by Florens and Simar (2005) these
standard parametric approaches for frontier estimation suffer from additional
86 Nonparametric robust estimators: partial frontiers

drawbacks that will be briefly described below. Florens and Simar (2005) pro-
pose a new method which overcomes most of these drawbacks by providing the
best parametric approximation of a frontier which is non parametrically esti-
mated in a first stage. Using the robust version of the nonparametric estimators
in this first step (order-m as in Florens and Simar, 2005, or order-α as in Daouia,
Florens and Simar, 2005), we obtain at the end√estimators of the parameters of
the models sharing nice statistical properties ( n-consistency and asymptotic
normality). So the inference on the parameters, or functions of the parameters
is very easy.
In this section, we will present the ideas in the output orientation where the
output y is univariate y ∈ R+ and the input x ∈ Rp+ , hence we are interested
in a production function. The same could be done for an input function where
x ∈ R+ and y ∈ Rq+ , by using input orientation. In Section 4.7, we will see
how these concepts can be adapted to a full multivariate setup with multiple
inputs and outputs.
Consider a suitable parametric family of production function defined on Rp+ :

{ϕ(· | θ) where θ ∈ Θ ⊆ Rk }. (4.45)

Cobb-Douglas or Translog models are often used for this parametric family.
The parametric model for the frontier can be written as:

y = ϕ(x; θ) − u,

where u ≥ 0. The classical estimation of this model depends on the assumptions

we want to impose on the error term u. If no particular assumption is made on u
(except independence between u and x), shifted least-squares can be used (also
called Modified OLS, MOLS). In this framework, a standard OLS is used to
estimate the shape of the frontier and then all the residuals are shifted to satisfy
u ≥ 0 and so to identify the intercept. If some parametric density is chosen for
u, we can improve the estimation of the intercept by estimating the parameters
of this density and identifying the intercept by some moment conditions on
the residuals (this is the Corrected OLS, COLS). In addition, with this full
parametric model, maximum likelihood estimators (MLE) can be computed.
For details see Greene (1980), Deprins and Simar (1985) or Kumbhakar and
Lovell (2000).
The advantages√of these methods are that they are easy to implement and
that they achieve n-consistency, even for the intercept if COLS or MLE is
used. Nevertheless, the main drawback of these regression-type estimators is
that they require strong assumptions on the stochastic part of the model. Since
u is independent of x, we have E(y|x) = ϕ(x; θ) − E(u), where E(u) = µ is
a constant. At a shift µ, ϕ(x; θ) is the regression of y on x. As a consequence,
whatever the method used, ϕ(x; θ̂) will capture the shape of the “middle” of
Parametric approximations of robust nonparametric frontiers 87

the cloud of points, whereas the frontier (and its characteristics) are properties
of the boundary of the observed cloud of points. Besides, in addition to the
independence assumption between u and x, these methods require specific
distributional parametric assumption for the efficiency term u (at least for COLS
and MLE), where in general, very few is known a priori on the shape of this
distribution. Finally, as most of the deterministic models, the procedure is very
sensitive to extremes or outliers.

4.6.1 Two stage methods

Coming back to ideas suggested in a descriptive way by Simar (1992), Flo-
rens and Simar (2005) propose a two stage method for estimating the parametric
model, which overcomes most of the drawbacks just mentioned above. It is
based on the following steps. In a first stage we project the cloud of data points
on a nonparametric frontier, to capture in a flexible way the shape of the bound-
ary of the cloud of points. After that, in a second stage, we fit the appropriate
parametric model on these projected points. Of course a nonparametric full
frontier estimator could be used in the first step but then it would suffer of the
same drawback (of traditional methods) of being sensitive to outliers. Florens
and Simar (2005) suggest as an alternative the use of the robust partial order-
m frontiers in the first stage. Daouia, Florens and Simar (2005) extend this
approach to order-α quantile type frontiers as first stage estimators. All the
statistical theory of these two stage methods has been provided in these two
papers. We briefly summarize the results after defining some useful notation.
Let us denote with ϕ(x) the true but unknown production frontier, we want to
estimate θ providing the best parametric approximation of the frontier function
ϕ(·). The pseudo-true value of θ can be defined as:
 n 

2
θ0 = arg min (ϕ(xi ) − ϕ(xi ; θ)) . (4.46)
θ
i=1

Of course if the parametric model is true, then θ0 is the true value of the para-
meter.
We can similarly define the pseudo-true values for the partial frontiers ϕm (·)
or ϕα (·), as follows:
 n 

2
θ0m = arg min (ϕm (xi ) − ϕ(xi ; θ)) , (4.47)
θ
i=1
 n 

2
θ0α = arg min (ϕα (xi ) − ϕ(xi ; θ)) . (4.48)
θ
i=1

The estimation procedure is then straightforward. We plug-in the nonparametric

estimator in place of the unknown ϕ, ϕm , ϕα in the above expressions. This
88 Nonparametric robust estimators: partial frontiers

leads to:
 n 

2
θ̂n = arg min (ϕF DH,n (xi ) − ϕ(xi ; θ)) , (4.49)
θ
i=1
 n 

2
θ̂nm = arg min (ϕm,n (xi ) − ϕ(xi ; θ)) , (4.50)
θ
i=1
 n 

2
θ̂nα = arg min (ϕα,n (xi ) − ϕ(xi ; θ)) . (4.51)
θ
i=1

The statistical properties of these estimators are derived in Florens and Simar
(2005) and in Daouia, Florens and Simar (2005). They can be summarized as
follows:
θ̂n − θ0 = op (1)
√ m approx.
n(θ̂n − θ0m ) ∼ Nk (0, Vm )
√ α approx.
n(θ̂n − θ0α ) ∼ Nk (0, Vα )
where an explicit expression is obtained for Vm and Vα . In practice however,
it is simpler to use a bootstrap algorithm, that we describe below19 , to provide
consistent estimators of these variances. Note that for the full frontier parame-
ters we only have consistency and not the asymptotic normality. This is another
argument to favor the use of partial frontiers in the first step calculations. In
addition, as explained below, by choosing m or α large enough, we estimate
also the full frontier itself.
Note also that here no particular assumption is made on the error term when
fitting the parametric model: we do not need a particular parametric distribution,
we do not require homoscedasticity, the error term can be related to the level
of the inputs x. Hence, clearly, most of the drawbacks of the regression-type
estimators are overcome with this two-stage approach.
We have seen above that ϕm(n),n and ϕα(n),n are robust estimator of ϕ, the
full frontier itself, if m(n) → ∞ and α(n) → 1 when n → ∞. Daouia, Florens
and Simar (2005) prove that if m(n) and α(n) are such that:
lim m(n) = ∞, lim m(n)(log n/n)1/2 = 0, lim n(1 − α(n)) = 0.
n→∞ n→∞ n→∞

Then, as n → ∞ the partial frontier estimates converge to the parameters of

the full frontier model:
θ̂nm(n) − θ0 = op (1),
θ̂nα(n) − θ0 = op (1).

19 See also the appendix of Florens and Simar (2005) for additional details.
Parametric approximations of robust nonparametric frontiers 89

For finite n, partial frontier estimators provide more robust estimators of θ0

because the nonparametric estimator of the frontiers will not envelop all the
data points, as is the case, instead, for the FDH estimator.
In practice, we select m, or α by tuning the number of points staying out of the
frontier, then we estimate the corresponding θ̂nm or θ̂nα and use the asymptotic
normal distribution for inference, with a bootstrap estimate of the variance.
Below we describe the bootstrap algorithm while in the next section we show
how these ideas can be extended to a full multivariate setup.

4.6.2 The bootstrap algorithms

The evaluation of the covariance matrices Vm and Vα is rather complicated.
For practical purposes of inference on the value of θ0m and θ0α , it will be easier to
approximate these matrices by a bootstrap method. As pointed in Florens and
Simar (2005), the naive bootstrap is consistent here because in the estimation
of θ0m or θ0α , we do not estimate the boundary of the support of the distribution
of (x, y).
We present the algorithm for the order-α parametric approximation, the same
could be done for the order-m approximation
√ case. The aim is to approximate
the asymptotic distribution of n(θnα − θ0α ) by its bootstrap analog, i.e., the
√
bootstrap distribution of n(θnα,∗ − θ0α,∗ ), where θnα,∗ and θ0α,∗ are the bootstrap
analogs of θnα and of θ0α , respectively.
We have to be careful when defining these bootstrap analogs. They depend
on how the pseudo-true values are defined. Florens and Simar give in their
appendix three versions of the algorithm but here due to our definition of the
pseudo-true value given in equation (4.48), the algorithm can be written as
follows:

[1] Draw a random sample of size n with replacement from X = {(xi , yi )|i =
1, . . . , n} to obtain the bootstrap sample Xb∗ = {(x∗i,b , yi,b
∗ )|i = 1, . . . , n}.

√
[2] With this sample Xb∗ , compute n(θb,n
α,∗ α,∗
− θ0,b ) where:

n  2
1
α,∗
θ0,b = arg min ϕα,n (x∗i,b ) − ϕ(x∗i,b ; θ) , (4.52)
θ n
i=1
n  2
1
θb,n
α,∗
= arg min ϕ∗α,n (x∗i,b ) − ϕ(x∗i,b ; θ) , (4.53)
θ n i=1

where for any x ∈ Rp , ϕα,n (x) is the nonparametric estimation of ϕα (x)

obtained from the original sample X and ϕ∗α,n (x) is the nonparamet-
ric estimation of ϕα (x) obtained from the bootstrap sample X ∗ . These
90 Nonparametric robust estimators: partial frontiers

estimators20 are evaluated in (4.52) and (4.53) at the bootstrap values x∗i,b ,
i = 1, . . . , n.

[3] Redo steps [1] and [2], a large number of times, b = 1, . . . , B.

√ α,∗
The empirical distribution of n(θb,n − θ0,bα,∗
), b = 1, . . . , B approximates
√ α
the distribution of n(θn − θ0 ) and can be used for doing inference (bootstrap
α

confidence intervals, percentiles methods, . . . ). In particular, its variance matrix

approximates Vα . Illustrative examples will be given in Part II of this book (see
Section 7.5).

4.7 Multivariate parametric approximations

The two stage approach presented above can indeed be extended to the mul-
tivariate setup by using parametric approximations of distance functions. A
modelization of Shephard distance function by flexible parametric models has
been proposed by Grosskopf, Hayes, Taylor and Weber (1997) and its statistical
properties have been investigated by Coelli (2000). However, what they pro-
posed is based on regression-type techniques of estimation, and shares all the
drawbacks described in the preceding section. On the contrary, we suggest, fol-
lowing Florens and Simar (2005), to use a two-stage approach: first estimate in
a nonparametric way the distance functions, and then, fit the estimated values by
an appropriate parametric model. We will see below how to apply this approach
to translog distance functions, extensively used in the literature. The translog
function can indeed be seen as a quadratic approximation of the real function
in the log scale. The Cobb-Douglas function, besides, is just a particular case
of the translog, where only the linear approximation is considered.
We make the presentation in the output oriented case, that is easy to translate
for the input orientation. Now we analyse a full multivariate setup, i.e. x ∈ Rp+
and y ∈ Rq+ . Let δ(x, y) be the Shephard output distance function (introduced
as the inverse of Farrell efficiency score in Section 2.3, see (2.13)):

δ(x, y) = inf{δ|(x, y/δ) ∈ Ψ} ≡ (λ(x, y))−1 .

From its definition and the discussion in Section 2.3, it is easily seen that
the distance function shares the following properties: (i) for all (x, y) ∈ Ψ,
δ(x, y) ≤ 1; (ii) δ(x, y) = 1, if and only if (x, y) is on the efficient boundary
of Ψ; (iii) δ(x, y) is homogeneous of degree one in y: δ(x, ηy) = ηδ(x, y) for
all η > 0. Of course, the parametric models proposed for approximating these
distances should be constrained to satisfy their properties.

20 In the common case where ϕ (x; θ) is linear in θ, the solutions of (4.52) and (4.53) are obtained by simple
α
OLS techniques. Otherwise nonlinear least squares methods have to be used.
Multivariate parametric approximations 91

Also order-m output distance function , δm (x, y), and order-α output dis-
tance function δα (x, y) can be considered by taking the inverse of the corre-
sponding λ measures defined in the preceding sections. These partial frontiers
have to be preferred if we want to be more robust to extreme data points.
Consider now a parametric family of functions defined on Rp+ ×Rq+ , denoted
by {ϕ(·, ·; θ) | θ ∈ Θ ⊂ Rk }, such that:

∀(x, y) and ∀η > 0, ϕ(x, ηy; θ) = ηϕ(x, y; θ).

Our aim is to estimate θ providing the best parametric approximation of the

distance function δ(·, ·).
As in Florens and Simar (2005) we can define the pseudo-true value of θ as:
 n 

θ0 = arg min (δ(xi , yi ) − ϕ(xi , yi ; θ))2 . (4.54)
θ
i=1

Finally the estimation is obtained by solving the following equation:

 n 

2
θ̂n = arg min (δ̂F DH (xi , yi ) − ϕ(xi , yi ; θ)) . (4.55)
θ
i=1

The same procedure can be followed with partial frontiers to get more robust
estimators of θ0 . This is achieved by using δ̂m or δ̂α in place of δ̂F DH in
Equation (4.55).
By Florens-Simar (2005) and Daouia-Florens-Simar (2005) the resulting
estimators share the same statistical properties as in the√preceding section,
namely consistency for the full frontier approximation, n-consistency and
asymptotic normality for the partial frontiers parameters.
In the following subsections we present two examples of parametric mod-
els for ln δ(x, y) based on the Generalized Cobb-Douglas and the Translog
functions.

4.7.1 Generalized Cobb-Douglas parametric model for

ln δ(x, y)
In this case, the candidate parametric model is linear in the logs of the argu-
ments:

ln δ(x, y) ≈ ln ϕ(x, y; θ) = α0 + α
ln x + β
ln y. (4.56)

The homogeneity of order one in y implies that for all η > 0,

ln ϕ(x, ηy; θ) = ln η + ln ϕ(x, y; θ) so that β
iq = 1,

where iq is a vector of ones of length q.

92 Nonparametric robust estimators: partial frontiers

The pseudo-true values θ0 = (α0 , α, β) are defined as the solution of the

following problem:
 n

θ0 = arg min ln δ(xi , yi ) − [α0 + α
ln xi + β2
ln yi2
α0 ,α,β2
i=1
2 
+(1 − β2
iq−1 ) ln yi1 ] ,

where y 1 is the first component of y and y 2 and β2 are the((q − 1) × 1) vectors

of the last q − 1 components of y and β. Then the last parameter is identified
by the homogeneity condition β1 = 1 − β2
iq−1 . Equivalently the pseudo-true
values can be defined as:
 n 
 2
θ0 = arg min ln(1/yi1 )δ(xi , yi ) − [α0 + α
ln xi + β2
ln ỹi2 ] ,
α0 ,α,β2
i=1

where ỹi2 = yi2 /yi1 and β1 = 1 − β2
iq−1 . Finally the problem for defining the
pseudo-true value can be written as:
 n 
 2
θ0 = arg min − ln yi∗,1

− [α0 + α ln xi + β2
ln ỹi2 ] , (4.57)
α0 ,α,β2
i=1

where y ∗ is the value of y projected on the output efficient frontier: y ∗ =

y/δ(x, y). Note that ỹ 2 ≡ ỹ ∗,2 = y ∗,2 /y ∗,1 . Again β1 = 1 − β2
iq−1 .
Of course, in the last expression δ(xi , yi ) are unknown, so are the yi∗,1 . Fol-
lowing the ideas of Florens and Simar (2005) we replace them by their nonpara-
metric estimates (full frontier, or preferably robust partial frontier estimates):
ŷ ∗ = y/δ̂F DH,n (x, y) or ŷ ∗ = y/δ̂m,n (x, y) or ŷ ∗ = y/δ̂α,n (x, y).
For example, for the full-frontier approach, the estimators is defined by the
following estimation equation:
 n 
 2
θ̂n = arg min − ln ŷi∗,1

− [α0 + α ln xi + β2
ln ỹi2 ] ,
α0 ,α,β2
i=1

i
and then β̂1,n = 1 − β̂2,n m α
q−1 . The same could be done for θ̂n and for θ̂n .
Let’s now compare this approach with the classical approach by Grosskopf,
Hayes, Taylor and Weber (1997) and Coelli (2000), where the following model
is estimated by COLS, MOLS or MLE:
− ln yi1 = α0 + α
ln xi + β2
ln ỹi2 + ui ,
where ui > 0 is considered as the inefficiency term. The difference comes from
the left-hand side term of the equation and from the stochastics on the efficiency
Multivariate parametric approximations 93

distribution: in our approach we use the projected values on the nonparametric

frontier and we do not make any assumption on the error term.
The drawbacks of this classical regression-based approach are well known:
restrictive and non realistic assumptions on u (e.g, homoscedasticity), and prob-
lems of consistency (u is not independent from ỹ 2 ), see Coelli (2000). With
our approach we overcome all these difficulties and get more sensible results
because the real (non-parametric estimated) values of the distance functions are
directly fitted by the appropriate model.

4.7.2 Translog parametric model for ln δ(x, y)

Here the parametric model can be written as:
 
1 ln x
ln δ(x, y) ≈ ϕ(x, y; θ) = α0 + α ln x + β ln y + [ln x
ln y
] Γ



,
2 ln y

where Γ = Γ
is symmetric. The homogeneity of degree one in y imposes

p + q + 1 constraints:

β
iq = 1, 1 constraint
Γ12 iq = 0, p constraints
Γ22 iq = 0, q constraints

 
Γ11 Γ12
where Γ = .
Γ21 Γ22

Here the pseudo true values of θ are defined through the equation:
 n

θ0 = arg min − ln yi∗,1 − [α0 + α
ln xi + β̃2
ln ỹi2
α0 ,α,β2 ,Γ11 ,

Γ12 ,

Γ22 i=1
2 
1
12 ln ỹ 2 + 1 ln ỹ
,2 Γ

22 ln ỹ
,2 ]
+ ln x
i Γ11 ln xi + ln x
i Γ i i i .
2 2
!
  c11 c
2
where: β = (β1 β̃2
)
,
12
Γ12 = a Γ and Γ22 =
c2 Γ
22

with a ∈ Rp , Γ
12 is (p × (q − 1)), c = (c11 c
)
∈ Rq and Γ
22 is ((q − 1) ×

2
(q − 1)). Here the missing parameters are identified through the homogeneity
conditions (see below).

12 , Γ
The estimation of α0 , α, β2 , Γ11 , Γ
22 is straightforward: we replace the
∗,1 1
unknown yi = yi /δ(xi , yi ) by the nonparametric estimates

ŷi∗,1 = yi1 /δ̂F DH,n (xi , yi )

94 Nonparametric robust estimators: partial frontiers

or by yi1 /δ̂m,n (xi , yi ) or by yi1 /δ̂α,n (xi , yi ) if more robust estimators are desired.
The estimator for (β1 , a, c) will be recovered from the homogeneity con-
straints:

β
iq = 1 =⇒ β̂1 = 1 − β̂2
iq−1



12 iq−1
Γ12 iq = 0 =⇒ â = −Γ



22 iq−1
Γ22 iq = 0 =⇒ ĉ2 = −Γ .
ĉ11 = −ĉ2 iq−1

Finally, at the
  end of the procedure we have now the full parameter estimates:
 . By applying the results of Florens and Simar (2005) and of
α̂0 , α̂, β̂, Γ
Daouia, Florens and Simar (2005), we have all the desired statistical inference
based on the appropriate normal asymptotic distribution. Here too, the bootstrap
is used in practice to make inference on the parameters θ for the partial frontiers
approximations (order-α or order-m). All the details of the algorithm have been
presented in Section 4.6.2 above. They are easily adaptable to this multivariate
setup. Again the bootstrap can be used either to estimate the variance Vm or
Vα of the estimators in the asymptotic normal approximation, or to provide
directly percentile confidence intervals. The method is illustrated on real data
in Section 7.5.
Chapter 5

CONDITIONAL MEASURES OF EFFICIENCY

This chapter deals with the important topic of the introduction of external-
environmental variables in frontier models. It is useful to explain why these
factors are important for comparative efficiency analysis and it shows the po-
tential of some recently introduced diagnostic tools for capturing their impact
on the performance of the analysed firms.
The evaluation of the influence of external-environmental factors on the effi-
ciency of producers is a relevant issue related to the explanations of efficiency,
the identification of economic conditions that create inefficiency, and finally to
the improvement of managerial performance.
The meaning and the economic role played by external-environmental vari-
ables are strictly linked to the economic field firms are operating in. The choice
of the environmental variables has to be done on a case-by-case basis, having
a good knowledge of the production process characteristics and by taking into
account the economic field of application.
From an economic point of view, we are interested in the evaluation of the
influence of Z variables on the performance of the firms. To be able to evaluate
this influence we have, firstly, to introduce the variables in the frontier estimation
problem and then we have to address some questions, like the following: “Is
the production process (and then the efficiency scores of firms) affected by the
Z variables?”; if the answer to this question is yes, “How we can evaluate their
influence?”.
The aim of this chapter is therefore to present “how” environmental variables
can be introduced in the probabilistic formulation of efficient frontier estimation
(described in the previous chapter) and to propose “a way” to operationalize
their introduction. The first purpose (introduction of Z variables) is addressed
in a following section which presents a full set of conditional (to Z) measures
of efficiency in a full frontiers (DEA/FDH) setting and in a robust (or partial)
96 Conditional measures of efficiency

frontiers (order-m and order-α) setting. The second purpose (measuring the
influence of Z) is handled through the introduction of an econometric method-
ology to follow in practical applications, based on the evaluation of the global
influence of Z on the production process, and on a decomposition of the condi-
tional (full or partial) efficiency of firm into some indicators with an interesting
economic meaning. In particular, we propose to decompose the conditional
efficiency of a firm into the unconditional (full or robust) efficiency score, an
externality index - that measures the environmental conditions in which the
firm operates in, i.e. favorable vs. unfavorable- and a producer intensity in-
dex, that measures the individual level of exploitation of the “environmental
conditions”, i.e. opportunities vs. threats of the “environment”. This decom-
position is particularly useful to facilitate comparative economic analysis of
DMU’ performance as it is shown in the applications reported in Part II of this
book.
Here we complete the approach proposed by Daraio and Simar (2005a,
2005b) for full and order-m frontiers (extending previous results by Cazals,
Florens and Simar, 2002), applied to order-α frontiers by Daouia and Simar
(2004). In particular we discuss at length and explain in details how to evaluate
the impact of multivariate Z on full and robust productive efficiency of DMU.
Furthermore, a new conditional probabilistic efficiency measure is introduced
in Section 5.2.3.
After a brief overview on the relevant literature on the introduction of external-
environmental variables in nonparametric frontier models reported in the next
section, Section 5.2.3 introduces a complete set of “conditional” measures of
efficiency, i.e. efficiency measures which take into account these external-
environmental variables, and presents their computational aspects. This presen-
tation will let the applied economist to catch the basic functioning mechanism
of the techniques. After that, Section 5.3 describes how to select the bandwidth
to compute the conditional estimators for both univariate and multivariate Z.
Afterwards, Section 5.4 explains in details and with simple illustrations the
econometric methodology useful to interpret the plots and decompose the con-
ditional efficiency measures. Finally, a series of simulation exercises illustrates
the usefulness of the methodology and how to practically implement the pro-
posed approach.

5.1 Explaining efficiency in the literature

The measurement of productive efficiency is only a first step of an efficiency
study. A natural complement is the investigation on explanatory variables of
the distribution of efficiency scores. Put it in another way, it is important to
know and measure in what extent external- environmental variables affect the
performance of the DMUs we are comparatively gauging.
Explaining efficiency in the literature 97

Hence, the basic idea underlying the introduction of external- environmental

variables is that of enriching the analysis of efficiency comparison by taking
into account extra information, related to the production process, but that are
neither inputs nor outputs under the control of the producer.
The nature of this extra information is such that it cannot be considered
neither as an input nor as an output of the production process but it is able
to affect the performance of the firms, in this sense we said above that these
variables are “related” to the production process.
For instance, as external/environmental variables we can consider some ad-
ditional information provided by the variables Z ∈ Rr which are “exogenous”
to the production process itself, but may explain part of it; some environmental
variables can not be under the control of the manager of the firm, at least in the
short run and so on.
Lovell (1993, p. 53) distinguishes the inputs/outputs of the production
process as “variables under the control of the decision maker during the time
period under consideration”, from explanatory variables that are “variables over
which the decision maker has no control during the time period under consid-
eration”.
Bartelsman and Doms (2000), in their empirical review, find that some of the
external/environmental factors behind the patterns (source of asymmetries) that
are thought to be important include form of ownership, quality of the workforce,
labor relations, technology international exposure, and the regulatory environ-
ment.
Other exogenous variables that can influence the production process are: in-
put and output quality indicators, regulatory constraints, competitive environ-
ment (competitive vs monopolistic), service network characteristics; structure
of property rights; type of ownership (private-public or domestic-foreign); en-
vironmental factors (conditions of the environment), location characteristics,
age and other characteristics of plants, and so on.
In a broader perspective, Morroni (2006) indicates as environmental con-
ditions the characteristics of information and knowledge, the techniques and
equipment available, the individual aims, the individual capabilities, the degree
of uncertainty, the structural change and the institutional and market condi-
tions21 .
Traditionally, the main focus of the literature on efficiency estimation has
been the evaluation of the performance of producers with a similar input/output
structure. In the nonparametric approach, with the exception of some pioneer-
ing works by Seitz (1966, 1971) and Timmer (1971), the explanation of the
distribution of efficiency scores through some explaining factors or “factors

21 For
a detailed discussion on the interplay between environmental conditions and internal conditions see
Morroni (2006), p. 31 ff.
98 Conditional measures of efficiency

behind patterns” (as Bartelsman and Doms, 2000 call them) has started to be
studied in literature only during the last decades22 .
An explicit reference to the importance of environmental variables can be
found in Lewin and Minton (1986) which define a research agenda for deter-
mining organizational effectiveness, whose main points are: 1) to be capable of
analytically identifying relatively most effective organizations in comparison
to relatively least effective organizations; 2) to be capable of deriving a single
summary measure of relative effectiveness of organizations in terms of their
utilization of resources and their environmental factors to produce desired out-
comes; 3) to be able to handle noncommensurate, conflicting multiple outcome
measures, multiple resource factors and multiple environmental factors outside
the control of the organization being evaluated; and not be dependent on a set
of a priori weights or prices for the resources utilized, environmental factors or
outcome measures; 4) to be able to handle qualitative factors such as participant
satisfaction, extent of information processing available, degree of competition,
etc.; 5) to be able to provide insight as to factors which contribute to relative
effectiveness ratings; 6) to be able to maintain equity in the evaluation.
As we have seen above, the exploration of the reasons for productivity/effi-
ciency differentials across production units is a relevant issue. When these
external factors Z ∈ Rr are continuous mainly two approaches have been
proposed in literature but both are flawed by restrictive prior assumptions on
the DGP and/or on the role of these external factors on the production process.
The first family of models is based on a one-stage approach (see e.g. Banker
and Morey, 1986a; Banker and Morey, 1986b for categorical external factors;
Färe, Grosskopf, Lovell and Pasurka, 1989; Färe, Grosskopf and Lovell, 1994,
p. 223-226), where these factors Z are considered as free disposal inputs and/or
outputs which contribute to define the attainable set Ψ ⊂ Rp+ × Rq+ × Rr , but
which are not active in the optimization process defining the efficiency scores.
In this case, the efficiency scores conditional to the Z are:
θ(x, y|z) = inf{θ | (θx, y, z) ∈ Ψ}, (5.1)
and the estimator of Ψ is defined as above by adding the variables Z in defining
the FDH and /or the DEA enveloping set. Here the variable Z is considered
as an input if it is conducive (favorable, advantageous, beneficial) to efficiency
and as an output if it is detrimental (damaging, unfavorable) to efficiency. The
drawback of this approach is twofold: first we have to know a priori what is
the role of Z on the production process, and second, we have to assume the free
disposability (and eventually convexity, if DEA is used) of the corresponding
extended attainable set Ψ.

22 For
the introduction of external-environmental variables in parametric frontier models, see Kumbhakar
and Lovell (2000).
Explaining efficiency in the literature 99

The second family of models is based on a two-stage approach. Here the

estimated efficiency scores are regressed, in an appropriated limited dependent
variable parametric regression model (like truncated normal regression models)
on the environmental factors Z. Recently, some models in this family propose
also three-stage and four-stage analysis as extension of the two-stage approach.
In particular, Fried, Schmidt, and Yaisawarng (1999) propose a four-stage
procedure that is based on a traditional DEA without external factors from which
radial Farrell technical efficiency scores are computed as well as inputs slacks
or outputs surpluses. In the second stage, a system of equations is specified in
which the dependent variable for each equation is the sum of slacks (or surpluses
in the output oriented case). To estimate the set of equations OLS (Ordinary
Least Squares), SUR (Seemingly Unrelated Regression) or tobit regression
are suggested to be used, according to the type of data. The third stage uses
the parameters estimated in the second stage to predict the total input slacks
(or output surpluses). These predicted values represent the “allowable” slack
(surplus) due to the operating environment, and are used to calculate adjusted
values for the primary inputs or outputs. Finally, the fourth stage consists in
re-run the DEA model under the initial input-output specification, using the
adjusted data set.
Fried, Lovell, Schmidt, and Yaisawarng (2002) propose as alternative a three-
stage approach. In the first stage a DEA is applied to obtain an initial evaluation
of the performance of producers. In the second stage a stochastic (parametric)
frontier is used to account for environmental effects and statistical noise. In the
third stage, DEA is applied on the adjusted data, taking into account environ-
mental effects and statistical noise.
As pointed out by Simar and Wilson (2003), also these models - as the
others in the two-stage family - are flawed by the fact that the usual inference
on the obtained estimates of the regression coefficients is not available. They
state: “none of the studies that employ two-stage approaches have described
the underlying data-generating process. Since the DGP has not been described,
there is some doubt about what is being estimated in the two-stage approaches”.
In addition, DEA estimates are by construction biased estimators of the unbiased
efficiency scores. Finally, a more serious problem arises from the fact that DEA
efficiency estimates are serially correlated and that the error term in the second
stage is correlated with the regressors. As stated by Simar and Wilson (2003)
“consequently, standard approaches to inference [...] are invalid”.
Simar and Wilson (2003) give a list of references where the two-stage ap-
proach has been used and propose two bootstrap-based algorithms to obtain
more accurate inference. However, even when corrected by the bootstrap, this
two-stage approach has two inconveniences.
First, it relies on a separability condition between the input × output space
and the space of Z values: the extended attainable set is the cartesian product
100 Conditional measures of efficiency

Ψ × Rr and hence, the value of Z does not influence neither the attainable
set Ψ nor the position of the frontier of the attainable set: Z acts only on the
stochastic process pushing firms far from the frontier.
Second, the regression in the second stage relies on some parametric as-
sumptions (like linear model and truncated normal error term in most studies.
Recently, Park, Simar and Zelenyuk, 2006, have proposed a nonparametric ap-
proach using maximum likelihood techniques). In the next section we see how
to avoid these limitations.

5.2 Introducing external-environmental variables

The probabilistic formulation of the production process described in Section
4.1 allows to overcome the limitations of previous approaches since external-
environmental factors can easily be introduced in the DGP.
The basic idea developed in Daraio and Simar (2005a) is that the joint distri-
bution of (X, Y ) conditional on Z = z defines the production process if Z = z.
By analogy with (4.1), the support of HX,Y |Z (x, y|z) = Prob(X ≤ x, Y ≥
y | Z = z) defines Ψz , the attainable production set when Z = z. For an input
conditional measure of efficiency, this joint distribution can be decomposed as
follows:

HX,Y |Z (x, y|z) = FX|Y,Z (x | y, z)SY |Z (y|z), (5.2)

for all y such that SY |Z (y|z) = Prob(Y ≥ y | Z = z) > 0 and where

FX|Y,Z (x | y, z) = Prob(X ≤ x | Y ≥ y, Z = z). Hence, for all y such
that SY |Z (y|z) > 0, Ψz can also be defined by the support of FX|Y,Z (· | y, z).
Then, as above in (4.3), the lower boundary of the latter will define the lower
boundary achievable for a unit producing an output level y with an environment
described by the value z. Formally we have:

θ(x, y | z) = inf{θ | FX|Y,Z (θx | y, z) > 0}. (5.3)

Note again that the conditioning on Y is the event Y ≥ y (because Y represents

the outputs) and the conditioning on Z is defined, as in a regression framework,
by Z = z. Hence Ψz can be described as:

Ψz = {(x
, y) ∈ Rp+q

+ |x ≥x
∂,z
(y) for (x, y) ∈ Ψ}, (5.4)

where x∂,z (y) is the efficient level of input, conditional on Z = z, for an output
level y: x∂,z (y) = θ(x, y | z) x, where (x, y) ∈ Ψ. Clearly, Ψz ⊆ Ψ.

5.2.1 Conditional full frontier measures

Daraio and Simar (2005a) propose a plug-in nonparametric estimator of
FX|Y,Z (· | y, z) in the expression (5.3) defining θ(x, y | z). Due to the equality
Introducing external-environmental variables 101

in the conditioning on Z this requires some smoothing techniques. At this

purpose they suggest a kernel estimator defined as follows:
n
i=1 1I(xi ≤ x, yi ≥ y)K((z − zi )/h)
F X|Y,Z,n (x | y, z) = n , (5.5)
i=1 1I(yi ≥ y)K((z − zi )/h)

where K(·) is the kernel and h is the bandwidth of appropriate size23 . Hence,
we obtain the “conditional FDH efficiency measure” as follows:
θF DH (x, y | z) = inf{θ | FX|Y,Z,n (θx | y, z) > 0}. (5.6)
Daraio and Simar (2005a) pointed out that for kernels with unbounded support,
like the gaussian kernel, it is easy to show that θF DH (x, y|z) ≡ θF DH (x, y):
the estimate of the full-frontier efficiency is unable to detect any influence of the
environmental factors. Therefore, in this framework of conditional boundary
estimation, kernels with compact support have to be used.
For any (symmetric) kernel with compact support (i.e., K(u) = 0 if |u| > 1,
as for the uniform, triangle, Epanechnikov or quartic kernels, see e.g. Silver-
man, 1986), the conditional FDH efficiency estimator is given by:
θF DH (x, y|z) = inf{θ | FX|Y,Z,n (θx | y, z) > 0} (5.7)
 
 Xj 
i
= min max . (5.8)
{i|Yi ≥y,|Zi −z|≤h} j=1,...,p xj

In this framework, the estimation of conditional full frontiers does not depend
on the chosen kernel but only on the selected bandwidth. This will be different
for the conditional order-m and order-α measures defined below.
The conditional attainable set Ψz is estimated by:
z
p+q
 F DH }, (5.9)
F DH = {(x , y) ∈ R+ | x ≥ x̂ (y) for (x, y) ∈ Ψ
∂,F DH,z
Ψ

where x̂∂,F DH,z (y) is the estimated conditional efficient level of inputs:

x̂∂,F DH,z (y) = θF DH (x, y|z) x for (x, y) ∈ Ψ

 F DH .

Daraio and Simar (2005b) define also a conditional DEA estimator and ad-
dress its computational issue. Consistency and asymptotic properties of these
conditional estimators are investigated in Jeong, Park and Simar (2006).

5.2.2 Conditional order-m measures

The conditional order-m input efficiency measure is defined in Daraio and
Simar (2005a), where only free disposability is assumed. For a given level of

23 Issues about the practical choice of the bandwidth are discussed in Section 5.3.
102 Conditional measures of efficiency

outputs y in the interior of the support of Y , we consider the m i.i.d. random

variables Xi , i = 1, . . . , m generated by the conditional p-variate distribution
function FX|Y,Z (x | y, z) and we define the conditional random set:

z (y) = {(x, y
) ∈ Rp+q | x ≥ Xi , y
≥ y, i = 1, . . . , m}.
Ψ (5.10)
m +

Note that this set depends on the value of z since the Xi are generated through
the conditional distribution function. For any x ∈ Rp+ , the conditional order-
m input efficiency measure given that Z = z, denoted by θm (x, y|z) is then
defined as:
z
θm (x, y|z) = EX|Y,Z (θ̃m (x, y) | Y ≥ y, Z = z), (5.11)

where θ̃mz (x, y) = inf{θ | (θx, y) ∈ Ψ
z (y)} and the expectation is relative
m
to the distribution FX|Y,Z (· | y, z). It is shown by Daraio and Simar (2005a)
(Theorem 3.1) that θm (x, y|z) converges to θ(x, y|z) when m → ∞.
A nonparametric estimator of θm (x, y|z) is provided by plugging the non-
parametric estimator of FX|Y,Z (x|y, z) proposed in (5.5), which depends on the
kernel and on the chosen bandwidth. Formally, the estimator can be obtained
by:
θm (x, y|z) = E

X|Y,Z (θ̃m (x, y) | y, z)
z
(5.12)
 ∞
= (1 − FX|Y,Z,n (ux | y, z))m du. (5.13)
0

This involve the computation of a one-dimensional numerical integral.

Since θm (x, y|z) → θF DH (x, y|z) when m → ∞, the order-m conditional
efficiency score can again be viewed as a robust estimator of the conditional
efficiency score θ(x, y|z) when choosing m = m(n) → ∞ with n → ∞. For
finite m, the corresponding attainable set will not envelop all the data points
and so is more robust to extremes or outlying data points.
Cazals, Florens and Simar (2002) prove the consistency of this estimator and
derive its asymptotic law. Daraio and Simar (2005b) propose also a conditional
convex order−m estimator.

Computational aspects
As we described in Section 4.2 the computation of order-m efficiency θ̂m,n
(x, y) could be done either by evaluating the univariate integral (4.23) via nu-
merical methods, or by an easy Monte-Carlo algorithm.
Similarly, the conditional order-m efficiency θ̂m,n (x, y | z) can be computed
either evaluating the integral (5.13) by numerical methods or using an adapted
version of the Monte-Carlo algorithm recalled above. This Monte-Carlo algo-
rithm for the conditional input order-m efficiency works as follows. Suppose
that h is the chosen bandwidth for a particular kernel K(·):
Introducing external-environmental variables 103

[1] For a given y, draw a sample

of size m with replacement, and with a
probability K((z − zi )/h)/ nj=1 K((z − zj )/h), among those Xi such
that Yi ≥ y. Denote this sample by (X1,b , . . . , Xm,b );
  X j 
b,z (x, y) i,b
[2] compute θ̃m = mini=1,...,m maxj=1,...,p xj
.

[3] Redo [1]-[2] for b = 1, . . . , B, where B is large.

1 B
[4] Finally, θ̂m,n (x, y | z) ≈ B
b,z
b=1 θ̃m (x, y).

However, for large m, the numerical univariate integral in (5.13) is much

faster to compute.

5.2.3 Conditional order-α measures

An alternative to the order-m partial frontier to obtain robust conditional
measure, robust to extreme data points and outliers, is the order-α quantile-type
frontier described in Section 4.3. The extension of these conditional measures
to order-α has been proposed by Daouia and Simar (2004). We continue the
presentation with the input orientation.
In place of looking for the conditional input efficiency score with respect
to the full conditional frontier, as above, θ(x, y | z) = inf{θ | FX|Y,Z (θx |
y, z) > 0}, we rather consider the (1 − α) quantile of the same conditional
cumulative distribution function (cdf). Formally, the conditional order-α input
efficiency measure given that Z = z, denoted by θα (x, y|z) is defined for all y
in the interior of the support of Y as:
θα (x, y|z) = inf{θ | FX|Y,Z (θx|y, z) > 1 − α}. (5.14)
We have the same interpretation we had for the order-α efficiency scores except
that here we condition additionally on the environmental variables Z = z. In
other words, θα (x, y|z) is the proportionate reduction (if < 1) or increase (if >
1) of inputs, the unit operating at the level (x, y) and confronted to environmental
conditions z, should achieve to reach the conditional to z input efficient frontier
of level α × 100%. Being on this frontier (θα (x, y|z) = 1) means the unit is
dominated by firms producing more output than y and confronted with the same
environmental conditions z, with a probability 1 − α. Daouia and Simar (2004)
analyse the properties of this measure and as expected, it is easy to prove that
θα (x, y|z) converges to θ(x, y|z) when α → 1.
A nonparametric estimator is given by plugging, in (5.14), the nonparametric
estimator of FX|Y,Z (·) introduced above. It can be shown, with the notations
introduced in Section 4.3, that:

1 
My  
FX|Y,Z,n (θx|y, z) = y
1I(X(j) y
≤ θ)K (z − Z[j] )/hn
Qy,z j=1
104 Conditional measures of efficiency
 y
 if θ < X(1)
 0
y y
= k if X(k) ≤ θ < X(k+1) , k = 1, · · · , My − 1

 y
1 if θ ≥ X(M y)
,
y
where, for j = 1, · · · , My , Z[j] denotes the observation Zi corresponding to
y n
the order statistic X(j) , Qy,z = 1I(Yi ≥ y)K( z−Z i
hn ), and finally k =
   i=1
y
(1/Qy,z ) kj=1 K (z − Z[j] )/hn .
Thus the nonparametric estimator of the conditional order-α input efficiency
measure given that Z = z is given by:
 y
 X(1)
 if 0 ≤ 1 − α < 1
θα,n (x, y|z) = y
X(k+1) if k ≤ 1 − α < k+1 , (5.15)


k = 1, · · · , My − 1.
Again this is easy and very fast to compute since it is only based on enumerative
algorithms. Daouia and Simar (2004) have proven the consistency of these
conditional estimators.

A new measure of conditional input efficiency

Extending the ideas presented in Section 4.3, for every attainable point
(x, y) ∈ Ψ, there exists an αz such that θαz (x, y|z) = 1. This αz could serve as
an alternative measure of conditional input efficiency score. If FX|Y,Z (x | y, z)
is continuous in x, this quantity is given, for the input orientation, by:

αzinput (x, y) = 1 − FX|Y,Z (x | y, z). (5.16)

In other words, one may set the estimated conditional performance measure
for a unit operating at the level (x, y) facing the environmental conditions z,
to be the order αz of the estimated conditional quantile frontier which passes
through this unit. This new measure of conditional efficiency, the conditional
α efficiency, may be estimated by the following quantity:
input
 z,n
α = 1 − k−1 , (5.17)
y
where k is the index such that X(k) = 1, k was defined above and we set
0 = 0.
This new measure has an appealing economic interpretation. The quantity
(1−αzinput (x, y)) is the firm (x, y)’s probability of being dominated in the inputs
space, given its level of outputs, taking into account its external-environmental
conditions Z = z.
Hence, a simple indicator of the impact of external factors on firms perfor-
αinput
z (x,y)
mance may be αQz = αinput (x,y)
, where αinput (x, y), defined in (4.28), is the
Introducing external-environmental variables 105

unconditional corresponding measure. Let us give some examples on how to

interpret this indicator.

– If αQz > 1, then (1 − αzinput (x, y)) < (1 − αinput (x, y)) In this case, for
the firm operating at the level (x, y) the probability of being dominated
given the condition Z = z is lower than that of being dominated without
taking into account the external conditions.

– If αQz = 1, then (1 − αzinput (x, y)) = (1 − αinput (x, y)) Here, the
probabilities are equal, hence it seems that the external conditions do not
play any role.

– If αQz < 1, then (1 − αzinput (x, y)) > (1 − αinput (x, y)) In this situation,
the probability of being dominated given the condition Z = z is higher
than that of being dominated without taking into account the external
conditions.

An application of this new measure and related indicators on real data is reported
in Chapter 8, where their usefulness is also shown.
In Table 5.1 we report all the nonparametric and robust measures of efficiency
introduced in this book as well as the main related references. The presentation
in the table is done for the input oriented framework. However, most of the cited
references report or just outline the output oriented correspondent measures that
will be described in the following section.

5.2.4 Summary for the output oriented case

In this section we briefly show how to adapt the conditional measures devel-
oped above for the input orientation to the output orientation. We have seen in
Section 4.1 that the Farrell-Debreu output efficiency score can be characterized
as λ(x, y) = sup{λ | SY |X (λy | x) > 0}, where SY |X (y | x) = Prob(Y ≥
y | X ≤ x). The FDH nonparametric estimator of λ(x, y) is then provided by
the empirical version of SY |X (y | x):

n
i=1 1I(Xi ≤ x, Yi ≥ y)
S
Y |X,n (y | x) = n .
i=1 1I(Xi ≤ x)

Then, the FDH estimator of the output efficiency score for a given point (x, y)
can be written as λ̂n (x, y) = sup{λ | SY |X,n (λy | x) > 0}. Mutatis mutandis,
all the output oriented measures share the same properties as their input oriented
correspondent.
106 Conditional measures of efficiency

Table 5.1. A summary of nonparametric and robust efficiency measures presented in this book
with the most important references. Input orientation.

Unconditional Conditional
Measures Measures

Full θDEA (x, y) θDEA (x, y | z)

Frontiers Farrell (1957) Daraio and Simar (2005b)
Charnes, Cooper and Rodhes (1978)

θF DH (x, y) θF DH (x, y | z)

Deprins, Simar and Tulkens (1984) Daraio and Simar (2005a)

Robust θm (x, y) θm (x, y | z)

Frontiers Cazals, Florens and Simar (2002) Cazals, Florens and Simar (2002)
and Daraio and Simar (2005a,b) and Daraio and Simar (2005a,b)

θα (x, y) θα (x, y | z)

Aragon et. al. (2002) Daouia and Simar (2004)
and Daouia and Simar (2004)

ninput
α input
z,n
α
Aragon et. al. (2003) this book
Daouia and Simar (2004)
and this book

Conditional full-frontier

When conditioning on environmental factors we have to replace SY |X (y | x) by

SY |X,Z (y | x, z) = Prob(Y ≥ y | X ≤ x, Z = z) so we obtain the definition
of the output conditional full-frontier efficiency measure:
λ(x, y|z) = sup{λ | SY |X,Z (λy|x, z) > 0}. (5.18)
A nonparametric estimator of the conditional full-frontier efficiency λ(x, y|z)
is given by plugging in its formula a nonparametric estimator of SY |X,Z (y|x, z).
We can use the following smoothed estimator:
n
i=1 1I(Xi ≤ x, Yi ≥ y)K ((z − Zi )/hn )
S
Y |X,Z,n (y|x, z) = n , (5.19)
i=1 1I(Xi ≤ x)K ((z − Zi )/hn )

where K is the kernel and hn is the bandwidth of appropriate size. Again,

kernels with compact support have to be used. The computation of this estimator
Introducing external-environmental variables 107

is very easy. We obtain:

 n (x, y|z) = sup{λ | S
λ Y |X,Z,n (λy|x, z) > 0}
 
Y j 
i
= max min .
{i|Xi ≥x,|Zi −z|≤h} j=1,...,q yj

Conditional order-m frontier

The conditional order-m output efficiency measure is introduced as follows. For

a given level of inputs x, consider the m i.i.d. random variables Yi , i = 1, . . . , m
generated by the conditional q-variate survival function SY |X,Z (y | x, z) and
define the set:

Ψzm (x) = {(u, v) ∈ Rp+q

+ | u ≤ x, Yi ≤ v, i = 1, . . . , n}. (5.20)

Note that this random set depends on the value of z since the Yi are generated
through SY |X,Z (y | x, z). Then, for any y, we may define λ̃zm (x, y) = sup{λ |
(x, λy) ∈ Ψzm (x)}. The conditional order-m output efficiency measure is
defined as:

λm (x, y|z) = EY |X,Z (λ̃zm (x, y) | X ≤ x, Z = z).

It can be computed as:

 ∞
λm (x, y|z) = [1 − (1 − SY |X,Z (uy | x, z))m ]du, (5.21)
0

where again it can be shown that limm→∞ λm (x, y|z) = λ(x, y|z). A non-
parametric estimator of λm (x, y|z) is given by:
 ∞
 m (x, y|z) =
λ 1 − (1 − SY |X,Z,n (uy | x, z))m du
0
 
λn (x,y|z)
 n (x, y|z) −
=λ (1 − SY |X,Z,n (uy | x, z))m du. (5.22)
0

The Monte Carlo algorithm presented in Section 5.2.2 can be easily adapted
to the output orientation.

Conditional order-α frontier

Similarly, we can define the conditional order-α output efficiency measure given
that Z = z as:

λα (x, y|z) = sup{λ | SY |X,Z (λy|x, z) > 1 − α}. (5.23)

108 Conditional measures of efficiency

A nonparametric estimator of λα (x, y|z) is provided by plugging in its formula

the nonparametric estimator of SY |X,Z (y|x, z). Formally, it is defined as:
 α,n (x, y|z) = sup{λ | S
λ Y |X,Z,n (λy|x, z) > 1 − α}.

Here also we have limα→1 λ  α,n (x, y|z) = λ  n (x, y|z). The computation of
the estimator can be described as follows, by using the notations introduced in
Chapter 4. For j = 1, · · · , Nx , denote by Z[j]
x the observation Z corresponding
i
n
to the order statistic Y(j)
x , and let R
x,z = i=1 1I(Xi ≤ x)K( z−Z i
hn ) > 0. Then
it can be shown that:
1  Nx  
SY |X,Z,n (λy|x, z) = 1I(λ ≤ Y(j)
x
)K (z − Z[j]
x
)/hn
Rx,z j=1


 1 λ ≤ Y(1)
if x

= Lk+1 if Y(k)
x < λ ≤ Yx
(k+1) , k = 1, · · · , Nx − 1


0 if λ > Y(N
x
x)
,
Nx  
j=k+1 K (z − Z[j] )/hn . The estimator is then
where Lk+1 = (1/Rx,z ) x

computed as follows:

Y(k)
x if Lk+1 ≤ 1 − α < Lk , k = 1, · · · , Nx − 1
 α,n (x, y|z) =
λ (5.24)
Y(Nx ) if 0 ≤ 1 − α < LNx .
x

A new measure of conditional output efficiency

Here, for every attainable point (x, y) ∈ Ψ, there exists an αz such that
λαz (x, y|z) = 1, this αz could serve as an alternative measure of conditional
output efficiency score. If SY |X,Z (y | x, z) is continuous in y, this quantity is
given, as for the input orientation, by:
αzoutput (x, y) = 1 − SY |X,Z (y | x, z). (5.25)
It may be estimated by the following quantity:
output
 z,n
α = 1 − Lk+1 , (5.26)
where k is the index such that Y(k)
x = 1, L
k+1 was defined above and we set
LNx +1 = 0.

5.3 Bandwidth selection

It is well known in nonparametric smoothing that the choice of the kernel is
not so important (in the sense that the results are very robust to this choice) but
Bandwidth selection 109

that the choice of the bandwidth may be crucial. We have already discussed
in Section 3.4 the bandwidth selection issue for bootstrapping DEA efficiency
scores, where we reported also some simple rules of thumb which can be easily
put in place and seem work pretty well. Here, for the computation of conditional
measures of efficiency we propose a very simple and easy to compute rule
based on a k-Nearest Neighbor (k-NN) method to select the bandwidth. We
will present the ideas in the simplest case where Z is univariate and where a
family of continuous kernels with compact support is available (like, triangular,
quartic or Epanechnikov kernels) and then we will adapt the presentation for
multivariate Z with an easy to implement kernel with compact support based
on truncated multivariate normal kernels.

5.3.1 Univariate case

The idea is that the smoothing in computing our Z-conditional efficiency
estimators, e.g., (5.6) for the input oriented case, comes from the smoothing
in the estimation of the conditional distribution function FX|Y,Z,n (x | y, z),
see equation (5.5). This is due to the continuity of the variable Z. Hence,
we suggest in a first step to select a bandwidth h which optimizes in a certain
sense the estimation of the density of Z. We propose to use the likelihood cross
validation criterion for the choice of h.
The method of likelihood cross-validation is a natural development of the
idea of using likelihood to judge the adequacy of fit of a statistical model. The
rationale behind the method, as applied to density estimation, is as follows.
Suppose that, in addition to the original data set, an independent observation Y
from f were available. Then the log likelihood of f as the density underlying
the observation Y would be log f (Y ); regarding fˆ as a parametric family of
densities depending on the window width h, but with the data X1 , ..., Xn fixed,
would give log fˆ(Y ), regarded as a function of h, as the log likelihood of
the smoothing parameter h. Now, since an independent observation Y is not
available, we could omit one of the original observations Xi from the sample
used to construct the density estimate, and then use Xi as the observation Y .
(−i) (−i)
This would give log likelihood log fˆh (Xi ), where fˆh is defined as:
n  
(−i) 1 x − Xj
fˆh (x) = K ,
(n − 1)h j=1,j =i h

Since there are no reasons to motivate the choice of which observation to leave
out, the log likelihood is averaged over each choice of omitted Xi , to give the
following score function:

n  
(−i)
CV (h) = n−1 log fˆh (Xi ) .
i=1
110 Conditional measures of efficiency

The likelihood cross-validation choice of h is then the value of h which maxi-

mizes the function CV (h) for the given data24 .
The likelihood cross validation criterion we apply in our methodology is
based on a k-NN method: this allows to obtain bandwidths which are localized,
insuring we have always the same number k of observations Zi in the local
neighbor of the point of interest z when estimating the density of Z.
Therefore, we define a grid of values of k to evaluate the leave-one-out
kernel density estimate of Z. We set the range of possible values of k in order
to have as lower bound a k equal to 5% of observations, and as upper bound a
reasonable value of k which corresponds to the 25% of the n observations. For
this grid of values of k, we evaluate the leave-one-out kernel density estimate
(−i)
of Z, fˆk (Zi ) for i = 1, . . . , n and find the value of k which maximizes the
score function:

n  
(−i)
CV (k) = n−1 log fˆk (Zi ) ,
i=1
where

n  
(−i) 1 Zj − Zi
fˆk (Zi ) = K ,
(n − 1)hZi j=1,j =i
hZi

and hZi is the local bandwidth chosen such that there exist k points Zj verifying
|Zj − Zi | ≤ hZi .
Afterwards, in a second step, in order to compute FX|Y,Z,n (x | y, z) (and
SY |X,Z,n (y | x, z) for the output-oriented case), we have to take into account for
the dimensionality of x and y, and the sparsity of points in larger dimensional
spaces. Consequently, we expand the local bandwidth hZi by a factor 1 +
n−1/(p+q) , increasing with (p + q) but decreasing with n.
The issue of choosing an optimal bandwidth in this setup is still an open
research question but the empirical method that we propose here turns out to
provide very sensible results, as shown by our simulated examples (see below)
and in most of our applications with real data (see Part II of this book). See also
the comments at the end of the next section devoted to the multivariate case.

5.3.2 Multivariate case

Silverman (1986), Scott (1992) and Simonoff (1996) have investigated the
problem of multivariate density estimation with multivariate gaussian kernels
and they have proposed reasonable empirical rules for the choice of the optimal
bandwidths which can be shown to approximate, in some favorable cases, the
optimal choice of the bandwidth.

24 For more details see e.g. Silverman (1986).

Bandwidth selection 111

The problem is that each component of Z has its own dispersion and so the
bandwidths should be scaled accordingly for each component. More gener-
ally, if one wants to estimate a r-dimensional density by kernel smoothing we
must
" choose a kernel function K(u) where u ∈ Rr , such that K(u) ≥ 0 and
Rr K(u) du = 1. Then we have to select a bandwidth matrix H which has to
be a (r × r) positive definite matrix. The scaled kernel function can then be
written as KH (u) = |H|−1 K(H −1 u) where |H| stands for the determinant of
the matrix H. Then a density estimate for Z could be written as:

 1
n
1  n
f (z) = KH (Zi − z) = K(H −1 (Zi − z)).
n i=1
n|H| i=1

A popular and simple choice of the kernel

#r
is the product kernel with a diagonal
matrix for the bandwidths: K(u) = j=1 Kj (uj ) and H = diag(h1 , . . . , hr ),
where the Kj (uj ) are standard univariate kernels and where generally the uni-
variate bandwidths are scaled to the standard deviation of each components
of Z, hj = sj h with sj being the standard deviation of Z j . In this case an
estimator of the multivariate density can be written as:

1 
n $
r
Zij − z j
f(z) = r
#r K( ),
nh j=1 sj i=1 j=1 hsj

where only a univariate h has to be selected. More generally, if we want to

take into account for some dependence among the components of Z, we could
use a multivariate kernel as follows. The basic kernel function in this case is
the standard r-variate normal density K(u) = 1/(2π)r/2 exp{−u
u/2}, and
we may choose as bandwidth matrix the following scaled matrix H = hS 1/2 ,
where S is the empirical covariance matrix of the r components of Z. The
density estimate appears then to be:

1 n
1
f(z) = 1/2
exp{− 2 (Zi − z)
S −1 (Zi − z)}.
n(2π) h |S|
r/2 r
i=1
2h

Empirical rules of the literature suggest to choose either

h = {4/(d + 2)}1/(d+4) n−1/(d+4)
which is optimal for independent multivariate normal if independent normal
kernels are used, or even simpler h = n−1/(d+4) . Of course h could also be
chosen by some k-nearest neighbor techniques.
The above approach is simple but is not very convenient in our setup be-
cause we need kernels with compact support and we prefer to take benefit for
the possible dependence between the components of Z and so avoid product
kernels with compact support. This is the reason why we develop a truncated
112 Conditional measures of efficiency

multivariate normal kernel approach. The idea is very simple, we truncate the
basic gaussian kernel K(u) on a sphere of radius one and in order to obtain a
continuous kernel at the boundary (which is preferable when estimating con-
tinuous densities), we rescale it so that the truncated kernel is equal to zero on
the boundary sphere defined by u
u = 1. After some analytical manipulations,
this leads to a new basic kernel bounded on the sphere of radius one defined as:
exp{−u
u/2} − exp{−1/2}
K ∗ (u) = r/2
1I(u
u ≤ 1),
C − exp{−1/2} Γ(1+r/2)
π

where C = (2π)r/2 Prob(χ2r ≤ 1) is a constant and χ2r is a chi-square random

variable with r degrees of freedom. K ∗ (u) is a regular continuous density
centered in zero and bounded on the unit sphere. Then, as before, we choose
the bandwidth matrix scaled by the covariance structure H = hS 1/2 . By doing
so we end up with the following kernel function:

∗ 1
KH (u) = r/2
hr |S|1/2 (C − exp{−1/2} Γ(1+r/2)
π
)
1
−1
×( exp{− u S u} − exp{−1/2})1I(u
S −1 u ≤ h2 ). (5.27)
2h2
This is a truncated normal distribution, truncated at the ellipsoid u
S −1 u ≤ h2
of “radius” h, the density being scaled so that it is continuous (equal to zero)
on its boundary. Finally, the expression for the density estimate of Z may be
written as above as:
1 n
f(z) = K ∗ (Zi − z). (5.28)
n i=1 H

Here again only one bandwidth has to be selected, we will use the k-nearest
neighbor principle: we select a local bandwidth hz such that the ellipsoid
centered in z with shape matrix S −1 and “radius” hz contains exactly k data
points Zi . An optimal value for k is then obtained as explained above by
likelihood cross-validation. Of course, as for the univariate case, once hi is
determined, we correct it in a second step to expand the local bandwidth hi by
a factor 1 + n−1/(p+q) .
In all our applications below, it is the method we have used even for the
particular case of r = 1. The method provided very sensible results and nice
estimators of the density of Z (uni- and multivariate) and the conditional effi-
ciency scores that derive from this kernel and bandwidth choice showed a great
stability to small changes in the change of k. We compared also with some of
the empirical rules proposed above and the results were often similar although
in some cases (dependence among the Z’s) the latter could be a wrong choice.
An econometric methodology 113

The same remark about the stability of the results with respect to the band-
width choice in this setup was already made by Simar and Wilson (1999c) for
Malmquist indices, and by Daraio and Simar (2005a,b) for the conditional and
convex efficiency measures, compared-for the univariate bandwidth case-with
the Sheather and Jones (1991) method.

5.4 An econometric methodology

Once we have handled the introduction of Z variables in the frontier esti-
mation framework, both in the full frontier and in the robust partial frontiers
(order-m or order-α), we propose a method to evaluate their influence on the
production process.
After that, we propose a decomposition of the conditional efficiency (again,
full and robust) very useful to evaluate the effect of Z on the efficiency of the
firm, disentangling the following aspects of the firm’s performance: Internal
or managerial efficiency (unconditional efficiency); Externality effects of the
environment (level of Z): positive vs. negative externality; Individual Intensity
on environmental conditions (level of exploitation of the environment).

5.4.1 Global effect of Z on the production process

For analyzing the global influence of Z on the production process, the com-
parison of θ̂n (x, y | z) with θ̂n (x, y) (and of λ̂n (x, y | z) with λ̂n (x, y) for the
output oriented case) is certainly of interest. When Z is univariate, a scatter plot
of the ratios25 Qz = θ̂n (x, y | z)/θ̂n (x, y) (or of Qz = λ̂n (x, y | z)/λ̂n (x, y))
against Z and its smoothed nonparametric regression line would be helpful to
describe the influence of Z on efficiency. As will be shown below, what we
propose here is an exploratory tool for highlighting the influence of Z on the
efficiency structures. At this stage of our research, we do not provide inferen-
tial tools for this analysis but show the usefulness of the approach in practical
problems.
The framework of our nonparametric regression smoothing is the following.
For the input oriented case, we have:
Qzi = g(Zi ) + i , i = 1, ..., n, (5.29)

θ̂n (Xi , Yi | Zi )
where Qzi = , i is the usual error term with E(i |Zi ) = 0, and
θ̂n (Xi , Yi )
g is the mean regression function, since E(Qzi | Zi ) = g(Zi ).
In this exploratory phase we choose the simple smoothed nonparametric
regression estimator introduced by Nadaraya (1964) and Watson (1964). The

25 We can do the same with the differences θ̂ (x, y | z)−θ̂ (x, y), but since efficiency scores are proportions,
n n
ratios seem very natural.
114 Conditional measures of efficiency

Nadaraya-Watson estimator of the regression function g is the following:

n z−Zi z
i=1 K( h )Qi
g(z) = n z−Zi
. (5.30)
i=1 K( h )

Note that this estimator can be written in the following form:


n
wi
g(z) = ( n )Qzi , (5.31)
i=1 j=1 wj

where wi = K( z−Z i z
h ), and hence, it is a linear combination of the Qi . There-
fore, the mean smoothed regression function is a weighted average of the Qzi ,
where the weights are represented by the kernels.
Accordingly, for the output oriented case, we have the same Nadaraya-
Watson estimator of the regression function g, where:

 n (Xi , Yi | Zi )
λ
Qzi = .
 n (Xi , Yi )
λ

The choice of the bandwidth h, for the smoothed nonparametric regression,

has been done applying a least-squares cross-validation automatic procedure26 .
In the following applications we use a Gaussian kernel K.

Interpreting the effect of Z

In an input-oriented framework, if the smoothed nonparametric regression
is increasing, it indicates that Z is detrimental (unfavorable) to efficiency and
when this regression is decreasing, it specifies a Z factor conducive (favorable)
to efficiency.
In the first case (unfavorable Z) the environmental variable acts like an
“extra” undesired output to be produced asking for the use of more inputs in
production activity, hence Z has a “negative” effect on the production process.
In this case θ̂n (x, y | z), the efficiency computed taking Z into account, will
be much larger than the unconditional efficiency θ̂n (x, y) for large values of Z
then for small value of Z. This is due to the fact that for firms with an high level
of Z, the efficiency score without taking into account Z is much smaller than

26 For more details, see e.g. Härdle (1990).

An econometric methodology 115

that–one computed taking into account Z; in this last case, the effect of Z let the
efficiency score going up. Consequently, the ratios Qz = θ̂n (x, y | z)/θ̂n (x, y)
will increase, on average, with Z.
In the second case (favorable Z), the environmental variable plays a role of a
“substitutive” input in the production process, giving the opportunity to “save”
inputs in the activity of production; in this case, Z has a “positive” effect on
the production process. It follows that the conditional efficiency θ̂n (x, y | z)
will be much larger than θ̂n (x, y) for small values of Z (less substitutive inputs)
than for large values of Z. Here again, this is due to the fact that firms with
a small value of Z do not exploit the positive effect of Z, and then, when
we take into account Z, their efficiency score goes up. Therefore, the ratios
Qz = θ̂n (x, y | z)/θ̂n (x, y) will, on average, decrease when Z increases.
Since we know that full-frontier estimates, and the derived estimated effi-
ciency scores, are very sensitive to outliers and extreme values, we do also
the same analysis for the more robust order-m and order-α efficiency scores.
Thus, in the empirical illustrations reported in the following section, we present
also the nonparametric smoothed regression of the ratios Qzm = θ̂m,n (x, y |
z)/θ̂m,n (x, y) on Z and of Qzα = θ̂α,n (x, y | z)/θ̂α,n (x, y) respectively.
Mutatis mutandis, the same could be done in the output oriented case, with
similar conclusions to detect the influence of Z on efficiency. In this case,
the influence of Z goes in the opposite direction: an increasing regression
corresponds to favorable environmental factor and a decreasing regression in-
dicates an unfavorable factor. In an output oriented framework, a favorable
Z means that the environmental variable operates as a sort of “extra” input
freely available: for this reason the environment is “favorable” to the produc-
tion process. Consequently, the value of λ̂n (x, y | z) will be much smaller
(greater efficiency) than λ̂n (x, y) for small values of Z than for large values of
Z. Here again, as for the input oriented case, this is due to the fact that firms
with small values of Z do not take advantage from the favorable environment,
and then, when Z is taking into account their efficiency scores improves, i.e.
the value of λ̂n (x, y | z) is smaller, indicating a greater efficiency. The ratios
Qzi = λ̂n (Xi , Yi | Zi )/λ̂n (Xi , Yi ) will increase with Z, on average.
In the case of unfavorable Z, the environmental variable works as a “compul-
sory” or unavoidable output to be produced to face the negative environmental
condition. Z in a certain sense penalizes the production of the outputs of inter-
est. In this situation, λ̂n (x, y|z) will be much smaller than λ̂n (x, y) for large
values of Z. Here, firms with an high level of Z are “more” negatively in-
fluenced by the environment with respect to firms with a low level of Z: for
that reason their efficiency score taking Z into account is much higher with
respect to their unconditional efficiency. As a result, the regression line of
Qzi = λ̂n (Xi , Yi | Zi )/λ̂n (Xi , Yi ) over Z will be decreasing.
116 Conditional measures of efficiency

Two simple illustrations to fix the ideas

In this paragraph we describe two simple production processes in which
the environmental variable Z is favorable (boiler case) and unfavorable (fridge
case). We show also the usefulness of robust measures to reveal conditional
local effects when there are extreme observations in the comparison set.

The Boiler case

The production process consists in an electric kettle which has to heat some
waters at 100 grades centigrade. We observe the performance of n firms, using
electric energy in kW h as input, to produce a certain amount of heated water.
Suppose that the output (heated water) is equal for all firms and is equal to one
liter, i.e. Y = 1. The input X is the energy in kW h, and the environmental
variable Z is the “external temperature”.
This is a simple production process in which the variable Z has a positive
influence on the production process, as if the external temperature (Z) is high,
the firm will need less energy (X) to heat the water (Y ), and hence Z acts like
a “substitutive input”.

Z
z0 * 8A *
* * *
φ (y) * *
z1 8B
* * *
* * * *
*
z2
* 8C
*
X
O P Q R S T U

Qz C
* *8
***
* B
* ** * *
>>1 * *
* 8 *
* * * A
* * * * *** *
1 * * * * * **8* * *
Z
z2 z1 z0

Figure 5.1. The Boiler case. Z favorable.

In Figure 5.1 we show a simple example to explain why the smoothed re-
gression line of the ratios Qzi = θ̂n (Xi , Yi | Zi )/θ̂n (Xi , Yi ) on Z is decreasing
when Z is favorable. Hence, Figure 5.1 (bottom panel) explains the decreasing
trend of the smoothed regression of the ratios Qz on Z whereas the top panel
An econometric methodology 117

shows the “isoquant” of the production process which shows a “substitutive

effect” existing between Z and X. To illustrate, Figure 5.1 represents also the
following firms:

A = (x0 , Y, z0 ) that uses a level of input x0 = OQ,

B = (x1 , Y, z1 ) that uses a level of input x1 = OS,
C = (x2 , Y, z2 ) that uses a level of input x2 = OU .

The efficient frontier FDH (φ(y)) is given by the minimum value of input X
used among the analyzed firms (here y is equal for all firms); it corresponds to
the value OP .
For the firms A, B and C, the conditional and unconditional FDH efficiency
scores, together with their ratios Qz , are the following:
OP OP
θA (x0 , Y |z0 ) = ; θA (x0 , Y ) = ⇒ QzA = 1;
OQ OQ
OR  OP
θB (x1 , Y |z1 ) = ; θB (x1 , Y ) = ⇒ QzB > 1;
OS OS
OT OP
θC (x2 , Y |z2 ) = ; θC (x2 , Y ) = ⇒ QzC >> 1.
OU OU

Note that firm A has the highest value of Z (compared with firm B and firm
C). Due to the “substitution effect” between Z and X, in correspondence to
this value of Z = z0 , we have the lowest value of the minimum of X (in this
case OP ). Firm B has a level of Z = z1 lower than z0 but higher than z2 . As
a result, for firm B the minimum value of X taking Z into account is OR, that
is higher than OP but lower than OT (the minimum value of X for the firm C
taking Z into account). As a consequence, we have the corresponding order of
the ratios QzC > QzB > QzA .

The Fridge case

In this case the production process consists in a refrigerator that has to pro-
duce several liters of glaced water (the output Y ); the input X is the energy
measured in kW h and the environmental factor (Z) is the external temperature.
This is a typical case of unfavorable Z. The bottom panel of Figure 5.2
illustrates the explanation of the increasing trend of the smoothed regression of
Qzi = θ̂n (Xi , Yi | Zi )/θ̂n (Xi , Yi ) on Z (bottom panel) through an “isoquant”
of a production process (in the top panel) which shows a Z as an undesired
output to be produced requiring the use of more input X.
Following the same arguments as in the Boiler case, the reader will indeed
understand that in this case of unfavorable Z, the nonparametric regression line
of Qzi over Z will be increasing as shown in Figure 5.2.
118 Conditional measures of efficiency

Z φ(y)
A A’
z0 8 8
*
z1 * *
8B *
*
* *
z2 8 C * * * * * *
* *
X
0 P Q R S T T’

Qz
* * 8
* * * A = A’
**
* *
* * *
*
* ** >>1
* * * * * *8*
* * * * * * B
1 * ** * * * 8 * *
C
z2 z1 z0 Z
0

Figure 5.2. The Fridge case. Z unfavorable.

All what has been said for the input-oriented case applies mutatis mutan-
dis also for the interpretation of the output-oriented case, where an increasing
nonparametric regression line points to a positive effect of Z on the production
efficiency whilst a decreasing nonparametric regression line points to a neg-
ative impact of Z. By construction, λ(x,   y)) scores are ≥ 1 (≤ 1),
y) (θ(x,
   
λ(x, y|z) ≤ λ(x, y) (θ(x, y|z) ≥ θ(x, y)), and therefore Qzi ≤ 1 (Qzi ≥ 1).
The same kind of reasoning applies again for the order-m and order-α con-
ditional and unconditional efficiency scores. Note that here the ratios Qzm,i
and Qzα,i are not bounded by 1 and λ  m (x, y|z) (θm (x, y|z)) is not necessarily
≤λ  m (x, y) (≥ θm (x, y)), as well as for Qz ratios.
α,i
Moreover, robust ratios have the advantage of being able to show the impact
of external factors even if some extreme observations may mask it when using
full frontier ratios. This case is illustrated in Figures 5.3 and 5.4 which show in
the top panel the case of a production process consisting in a fridge as we have
seen above, in which there are 3 units (marked in the picture as bigger stars
with dotted square around) consisting in fridges of a new generation, that are
perfectly isolated and then are not influenced by the external temperature. In
this case, we see from Figure 5.3 (bottom panel) that full frontier ratios are not
able to capture the effect of the external factor: the smoothing nonparametric
regression line is straight due to the influence of these extremes. On the contrary,
An econometric methodology 119

Figure 5.4 (bottom panel) shows what the impact of external factor is, because
partial frontiers are not influenced by these fridges of new generation. We will
see in Section 6.4 an illustration of this case on real data.

Figure 5.3. The Fridge case with outliers. Full frontier case

Figure 5.4. The Fridge case with outliers. Partial frontiers case

5.4.2 A decomposition of conditional efficiency

Using the methodology described above on the introduction of Z and the
evaluation of its influence, we can also provide a decomposition of the perfor-
mance of the firm, as represented by its conditional efficiency scores.
We propose to disentangle the performance of the firm (x, y), as measured by
the Conditional Efficiency index, named CE z (x, y), in three main indicators:
120 Conditional measures of efficiency

an indicator of the internal or managerial efficiency, i.e. the Unconditional

Efficiency score named U E(x, y); an indicator of the level of Z owned by the
firm, the Externality Index, named EI z (x, y); and finally, an Individual Index
which measures the firm intensity in catching the opportunities or threats by the
environment, named II z (x, y).
Expressing the proposed decomposition in formulae we have:

CE z (x, y) = U E(x, y) ∗ EI z (x, y) ∗ II z (x, y). (5.32)

We define the ratio of conditional on unconditional efficiency score, Qz , as

follows:
CE z (x, y)
Qz = . (5.33)
U E(x, y)
The Externality Index is the expected value of the ratio Qz given that Z = z,
and is given by:
 z |Z = z);
EI z (x, y) = E(Q (5.34)
whilst the Individual Index is given as follows:
Qz
II z (x, y) = . (5.35)
EI(x, y)
In the input orientation, the conditional and unconditional measures of effi-
ciency are given by:
 y|z); U E(x, y) ≡ θ(x,
CE z (x, y) ≡ θ(x,  y). (5.36)

Note that the externality index defined in equation (5.34), may be estimated us-
ing the Nadaraya-Watson nonparametric estimator defined above (see equation
(5.30)) or another nonparametric estimator.
For the output oriented case, in equation (5.32) we have only to substitute:

CE z (x, y) ≡ λ(x, 
y|z); U E(x, y) ≡ λ(x, y). (5.37)

The same is for robust (order-m and order-α) measures, where we have
to substitute in equation (5.32) the relative conditional and unconditional effi-
ciency scores in the selected orientation.
For the interpretation of the proposed indicators, and the evaluation of the
influence of Z at the firm level, a major role is played by the ratio Qz , the ratio
of conditional on unconditional measure of efficiency. If we are in an input or
output orientation, and Qz = 1, this means that conditional and unconditional
efficiency scores are equal (this applies both to full and robust efficiency scores).
This value of the ratio points to a situation in which the external factors do not
affect the performance of the analysed firm. On the contrary, if we are in an
Simulated illustrations 121

input oriented framework and Qz > 1 this means that taking Z into account
lead to an higher efficiency score of the firm.
The externality index EI z (x, y) = E(Q z |Z) represents the expected influ-
ence of Z on the performance of the firm: it depends on the own level of Z.
It should be interpreted taking into account the global effect of Z on the pro-
duction process and what said above for the ratio Qz . If EI z = 1 this means
that the firm considered operates in a situation in which we expect that given
the level of the environment, Qz should be equal to one. When EI z > 1 this
means that the firm works at a level of environment with an expected Qz > 1.
Finally, the Individual Index tells us how the firm performed with respect to
the expected value of its performance; i.e. an II z = 1 means that the firm’s
Qz is exactly equal to E(Q  z |Z). If the II z > 1 this means that the effect
of the environment on the efficiency score of the firm under consideration is
higher with respect to its expected value. On the contrary, if the II z < 1 we are
considering a firm for which the environmental externality is lower then what
expected for its level of Z.
Summing up, considering the above indications, consulting the smoothed
nonparametric regression plot of Qz over Z, and taking into account the mini-
mum and maximum level of Z, we are able to interpret the effect of Z at firm
level, on the efficiency score of firm, by decomposing the conditional efficiency
score in its main components: unconditional efficiency, externality index and
individual index.
The same interpretation, given for the input-oriented case, mutatis mutandis,
can be done in the output-oriented framework, recalling that λ(x, y) ≥ 1, and
hence Q ≤ 1.
z

Accordingly is also the interpretation for the decomposition of the Robust

Conditional Efficiency index (of order-m and order-α). In the robust case, note
that the ratio of conditional and unconditional efficiency can be higher, equal or
lower than one (both in the input and in the output orientation). In Chapter 8 we
illustrate how these indicators may be useful in an application on Aggressive
Growth mutual funds data.

5.5 Simulated illustrations

In this section we illustrate the econometric methodology presented above
using simulated examples which describe multiple input-output production
processes conditioned by univariate and multivariate external-environmental
factors. These illustrations show to the reader how to implement in practice our
method. See also Chapters 6 to 8 for applications done on real data sets.
122 Conditional measures of efficiency

5.5.1 Univariate Z
In this example, we simulate a multi-input (p = 2) and multi-output (q = 2)
production process in which the function describing the efficient frontier is (as
in Park , Simar and Weiner, 2000) the following:
y (2) = 1.0845(x(1) )0.3 (x(2) )0.4 − y (1) (5.38)
where y (j) (x(j) ) denotes the jth component of y (of x) for j = 1, 2. We
(j) (j)
draw Xi independent uniforms on (1, 2) and Ỹi independent uniforms on
(0.2, 5). Then the generated random rays in the output space are characterized
(2) (1)
by the slopes Si = Ỹi /Ỹi . Finally, the generated random points on the
frontier are defined by:
(1) (2)
(1) 1.0845(Xi )0.3 (Xi )0.4
Yi,ef f = (5.39)
Si + 1
(2) (1) (2) (1)
Yi,ef f = 1.0845(Xi )0.3 (Xi )0.4 − Yi,ef f . (5.40)
The efficiencies are generated by exp(−Ui ) where Ui are drawn from an expo-
nential with mean µ = 1/3. Finally, in a standard setup (without environmental
factors), we define Yi = Yi,ef f ∗ exp(−Ui ).
On this data set, we introduce the dependency on an environmental factor
Z, adapting Case 1 of Daraio and Simar (2005a). Z is uniform on (1, 4) and
such that it has a quadratic negative impact on the production process till a Z
value of 2.5 and then a quadratic positive impact (here we consider an output
oriented framework):
(1) (1)
Yi = [1 + (Z − 2.5)2 ] ∗ Yi,ef f ∗ exp(−Ui ) (5.41)
(2 ) (2)
Yi = (1 + |Z − 2.5|) ∗ Yi,ef f ∗ exp(−Ui ). (5.42)
We simulate n = 100 observations according to this scenario.
In the nonparametric estimation, we have chosen a truncated gaussian kernel
for the smoothing; we remark that the results are very stable if other kernels with
compact support are used. Figure 5.5 illustrates the likelihood cross validation
plot for the choice of the number of the Nearest Neighbourhood (NN), that in
this case is 18.
For the choice of the values of m and α, the inspection of Figure 5.6 is
particularly useful as it shows a sensitivity analysis on the percentage of points
outside the partial frontiers in a sensitive way, i.e. after a threshold value of 0.15
(we applied the procedure described in Section 4.4.4, in particular in Equation
(4.30) we use a τ = 0.15). We have chosen, then, m = 50 and α = 0.985
such that the percentages of points outside the partial frontiers be close to
zero. These values are such that both order-α and order-m efficiency scores
Simulated illustrations 123

−1.12

−1.13

−1.14
LIKELIHOOD CV criterion

−1.15

−1.16

−1.17

−1.18

−1.19

−1.20

−1.21

−1.22
10 15 20 25 30 35 40 45 50
VALUE of N (Nearest Neighbour)

Figure 5.5. Simulated example with univariate Z. Likelihood cross validation plot for the
choice of the number of the Nearest Neighbourhood (NN). Here the number of k − NN which
maximizes the likelihood cross validation criterion is 18.

Percentage of points out of α−frontiers

0.3
percentages

0.2

0.1

0
0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 1.02
values of α
Percentage of points out of m−frontiers
percentages

0.3

0.2

0.1

0
0 20 40 60 80 100 120 140 160
values of m

Figure 5.6. Simulated example with univariate Z. Plots of the percentage of points outside
order-m and order-α frontiers.
124 Conditional measures of efficiency

are very close to FDH efficiency scores because in this scenario we do not have
outliers. In practice, the choice of these two “tuning” parameters (m and α)
may also be governed by their economic interpretation.
The results are displayed in Figure 5.7. In all panels (top for the FDH case,
middle for the α frontier case and bottom for the m frontier) as expected, we
see that the ratios (of conditional and unconditional FDH, α and m efficiency
scores) allow to detect the “U − shaped” effect of Z on the production process.

5.5.2 Multivariate Z
Two independent components
In this exercise the multi-input (p = 2) and multi-output (q = 2) data set
is simulated according to the same scenario described in the previous Section
5.5.1, but the Z variable now is drawn from a bivariate normal distribution

0.25 0
with mean µ = [2.5 2.5] and covariance matrix Σ = . The
0 0.25
dependence of the production process from Z is introduced as follows:
(1) (1)
Yi = (1 + |Z1 − 2.5|3 ) ∗ Yi,ef f ∗ (1 + Z2 ) ∗ exp(−Ui )
(2) (2)
Yi = (1 + |Z1 − 2.5|3 ) ∗ Yi,ef f ∗ (1 + Z2 ) exp(−Ui ),

(.)
where Yi,ef f are generated as in Section 5.5.1, and the Ui are drawn from an
exponential with mean µ = 1/2. Again, we simulate n = 100 observations
according to this scenario which defines a U-shaped pattern around 2.5 for Z1
and a linear pattern for Z2 .
In the nonparametric estimation, we have chosen a truncated gaussian kernel
for the smoothing as before, which can be easily generalised for the multivariate
case. Figure 5.8 illustrates the likelihood cross validation plot for the choice
of the number of the Nearest Neighbourhood (N N = 30) for the estimation
of the density of Z, which we use for computing the conditional measures of
efficiency. Figure 5.9 shows the estimation of Z done using our k-NN approach
as well as its contour plot. As it appears from the contour plot, Z1 and Z2 are
independent.
To set the value of m and α to compute the robust partial efficiency scores
for the simulated dataset whose DGP has been described above, we plot in
Figure 5.10 the percentage of points which are outside the partial frontiers after
a threshold value of 0.15. By inspecting Figure 5.10 we choose the value of
m = 35 and the value of α = 0.965 so that we leave outside a percentage of
points close to zero.
Some results are displayed in Figures 5.11 and 5.12. In particular, Figure 5.11
shows the global impact of the external variables Z on the simulated production
process providing, the surface of Qz on Z1 and Z2 .
Simulated illustrations 125

Effect of Z on Full frontier

1.1

0.9

0.8
Qz

0.7

0.6

0.5

0.4

1 1.5 2 2.5 3 3.5

values of Z

Effect of Z on Order−α frontier

1.1

0.9

0.8
Qzα

0.7

0.6

0.5

0.4

1 1.5 2 2.5 3 3.5

values of Z

Effect of Z on Order−m frontier

1.1

0.9

0.8
Qzm

0.7

0.6

0.5

0.4

1 1.5 2 2.5 3 3.5

values of Z

Figure 5.7. Simulated example with univariate Z. Smoothed nonparametric regression of Qz

on Z (top panel), of Qzα on Z (middle panel) and of Qzm on Z (bottom panel).
126 Conditional measures of efficiency

− 0.80

− 0.81

− 0.82
LIKELIHOOD CV criterion

− 0.83

− 0.84

− 0.85

− 0.86

− 0.87

− 0.88

− 0.89
10 15 20 25 30 35 40 45 50
VALUE of N (Nearest Neighbour)

Figure 5.8. Simulated example with multivariate Z. Likelihood cross validation plot for the
choice of the number of the Nearest Neighbourhood (NN). Here the number of k − NN which
maximizes the likelihood cross validation criterion is 30.

Figure 5.12 provides a more detailed information on the impact of Z on the

simulated production process. It plots the ratios Qz against Z1 (top panel) and
Z2 (bottom panel) at the three quartiles of the other component of Z. As we
expected, we recover a positive U-shaped effect of Qz on Z1 and a positive
linear effect on Z2 .
The same effects were obtained also for Qzα and Qzm (here we do not have
outliers).

Two correlated components

Figure 5.13 shows another simulated bivariate Z variable in which Z1 and Z2
are correlated (0.75). We introduced this new Z in the same scenario described
above (U-shaped effect of Z1 , linear effect of Z2 ) and compute the conditional
and unconditional nonparametric and robust efficiency indices as well as their
ratios. Figure 5.14 shows the results. Since Z1 and Z2 are highly correlated
(0.75), in a normal bivariate setup, the two effects are confounded: the linear
effect is added to the quadratic one giving the result we can see; a sort of
“mixture” of the two effects with a slightly more pronounced curvature for Z1 .
This example shows that if we have independent Z variables, we can finely
recover their marginal impact on the production process, but if the Z variables
are correlated (as they are in Figure 5.14) then their marginal impact has to be
more carefully examined.
Simulated illustrations 127

Density of Z

0.8

0.6

0.4

0.2

0
6
5
4 4
3
2 2
1
0 0
values of Z2 values of Z1

Contour plot for the density of Z

4

3.5

3
values of Z2

2.5

1.5

1 1.5 2 2.5 3 3.5 4

values of Z1

Figure 5.9. Density of Z and Contour plot of the density of Z.

128 Conditional measures of efficiency

Percentage of points out of α−frontiers

0.4
percentages

0.3
0.2
0.1
0
0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 1.02
values of α
Percentage of points out of m−frontiers
0.4
percentages

0.3
0.2
0.1
0
0 20 40 60 80 100 120 140 160
values of m

Figure 5.10. Simulated example with multivariate Z. Plots of the percentage of points outside
order-m and order-α frontiers.

0.92

0.90

0.88
Qz

0.86

0.84

0.82
6
5
4 4
3
2 2
1
Z2 0 0
Z1

Figure 5.11. Simulated example with multivariate Z. Surface of Qzα on Z1 and Z2 .

Simulated illustrations 129

0.95
Qz

0.9

0.85

0.8
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Z1

0.9

0.88
Qz

0.86

0.84

0.82
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Z2

Figure 5.12. Simulated example with multivariate Z. Top panel smoothed nonparametric
regression of Qz on Z1 for Z2 ’s quartiles. Bottom panel smoothed nonparametric regression of
Qz on Z2 for Z1 ’s quartiles. The dashed line corresponds to the first quartile, the solid line to
the median and the dashdot line to the third quartile.

Density of Z Contour plot for the density of Z

3.5
1.5

3
1
values of Z2

2.5
0.5

0 2
4
3 4
3 1.5
2 2
values of Z2 1 1 1.5 2 2.5 3 3.5
values of Z1 1
values of Z1

Figure 5.13. Density of Z and Contour plot of the density of Z. Z1 and Z2 are correlated.
130 Conditional measures of efficiency

1.4

1.2
Qzα

0.8
1 1.5 2 2.5 3 3.5 4
Z1

1.4

1.2
Qzα

0.8
1 1.5 2 2.5 3 3.5 4 4.5
Z2

Figure 5.14. Simulated example with multivariate Z. Z1 and Z2 are correlated. Top panel
smoothed nonparametric regression of Qzα on Z1 for Z2 ’s quartiles. Bottom panel smoothed
nonparametric regression of Qzα on Z2 for Z1 ’s quartiles. Dashed line = first quartile, solid line
= median, and dashdot line = third quartile.

Sensitivity to outliers
To complete the simulated illustration of our econometric methodology, we
introduce 5 outliers in the multivariate Z (independent components) simulation
setting described at the beginning of this section.
These extremes points are introduced at the following values of X: (1.25,1.5),
(1.25, 1.75), (1.5,1.5), (1.75, 1.25) and (1.5, 1.25), the corresponding values for
the slopes in the Y space are (0.25, 0.75, 1, 3, 5). The corresponding values
of Z have been drawn from a bivariate normal with mean µ and covariance Σ
(as above). Finally the outliers in the output direction were projected outside
the true frontier multiplying by a factor of 2.5.
The results are displayed in Figures 5.15 and 5.16. As it clearly appears, the
FDH estimator, in presence of the 5 outliers, fails to detect the correct quadratic
effect of Z1 on the production process (the curvature is missed, see Figure 5.15),
but the order−α estimator is able to reproduce the simulated effect of Z1 (see
Figure 5.16). We obtain a similar result for the order−m case.
Simulated illustrations 131

0.96

0.94

0.92
Qz

0.9

0.88

0.86
6
5
4 4
3
2 2
Z2 1 Z1
0 0

0.95
Qz

0.9

0.85

0.8
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Z1
0.95
Qz

0.9

0.85
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Z2

Figure 5.15. Simulated example with multivariate Z and 5 outliers. Surface of Qz on Z1

and Z2 (top graph). Bottom graph: smoothed nonparametric regression of Qz on Z1 for Z2 ’s
quartiles (top panel) and on Z2 for Z1 ’s quartiles (bottom panel); dashed line = first quartile,
solid line = median and dashdot line = third quartile.
132 Conditional measures of efficiency

1.02

0.98
Qzα

0.96

0.94

0.92
6

4 5
4
3
2 2
1
Z2 0 0 Z1

0.98
Qzα

0.96

0.94

0.92
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Z1
1

0.98
Qzα

0.96

0.94

0.92
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Z2

Figure 5.16. Simulated example with multivariate Z and 5 outliers. Surface of Qzα on Z1
and Z2 (top graph). Bottom graph: smoothed nonparametric regression of Qzα on Z1 for Z2 ’s
quartiles (top panel) and on Z2 for Z1 ’s quartiles (bottom panel); dashed line = first quartile,
solid line = median and dashdot line = third quartile.

This illustration confirms that it is always useful to compare the results ob-
tained using full frontier efficiency estimators with those obtained applying
robust partial frontiers measures.
Chapter 6

INSURANCE INDUSTRY:
IN SEARCH FOR ECONOMIES OF SCALE,
SCOPE AND EXPERIENCE IN THE
ITALIAN MOTOR-VEHICLE SECTOR

6.1 Introduction
The analysis of the Italian insurance market is interesting because of the
dramatic changes that have occurred in the business during the past two decades.
With more than 43 millions of circulating vehicles, 754 vehicles per thousand
of inhabitants and 103 vehicles per km of road, 13,842 million of Euro of
direct premiums and 416 million of Euro of losses in 2001 (ANIA -Italian
Association of Insurance companies- data), the automobile liability business
is the most important insurance line in Italy, accounting for about 60.7% of
Non-life insurance business and for almost 24% of total insurance premiums
(2001). Its nature of compulsory insurance, the increasing of its tariffs and
their influence on the inflation rate, together with the growing role played by
biological damage reimbursements and by frauds, have opened a deep and
sometimes harsh debate both in the political and technical environments, on
the measures to be adopted in order to take under control premium levels, to
increase efficiency and to promote competition.
The insurance industry in Italy, as well as in other European countries, has
been traditionally subject to stringent regulation affecting pricing, contractual
provisions, establishment of branches, solvency standards, and numerous ad-
ditional operational details. Competitive intensity was very low, with minimal
price and product competition and stable profit margins (Swiss Re 1996, 2000b).
The implementation of the EU’s Third Generation Directives, beginning on
July 1st, 1994, represented a major step in creating conditions in the EU resem-
bling those in a single deregulated national market. In the last years, numerous
events have taken place that have reduced existing barriers among financial in-
stitutions and countries and increased the competitive pressures in the market.
The following are some of the most relevant factors: a) changes in the na-
136 Economies of scale, scope and experience in the Italian motor-vehicle sector

tional and the EU law systems reducing the barriers among financial institutions
and increasing information transparency between the insurer/distributor and the
customer; b) development of new communication and information technologies
and of new management methods (with the consequent increase of information
transmission between the insurer and the distributors); c) increased financial so-
phistication of customers, who are more interested in financial problems, better
educated, and more demanding; d) internationalization of markets; e) increased
importance of insurance products in families saving portfolios; f) growing in-
adequacy of social security pension systems and the consequent increase in the
demand for life insurance products; and g) birth of new financial intermediaries.
Despite of these changes, an analysis by Swiss Re (2000b) finds that per-
sonal lines insurance markets have remained localised. One reason explaining
the slowness of the emergence of cross-border competition is that the European
Directives did not completely eliminate the ability of host countries to influence
insurance markets. For example, EU member countries can still utilize taxa-
tion to discriminate between domestic companies and those based in other EU
countries (Hess and Trauth 1998). In addition, there are significant differences
in contract law across European nations (Swiss Re 1996), impeding contract
standardisation. Domestic insurers also are likely to have an advantage in their
home markets because of cultural affinities, established brand names and dis-
tribution networks, and buyer perceptions that such firms have higher quality
or financial stability than foreign firms. Finally, foreign insurers may be at a
disadvantage in comparison with domestic insurers in terms of their knowledge
of the underwriting characteristics of buyers, exposing foreign firms to higher
informational asymmetry and adverse selection problems in comparison to do-
mestic firms. It is interesting to see in more details how these recent changes
have affected the Italian insurance market.
The Italian discipline of the insurance sector has seen its fundamental year
in 1912 (Law 04/04/1912, no. 305) when the Istituto Nazionale delle Assi-
curazioni (INA) has been created and the affirmation of the principles of “au-
thorisation of admission” and of “control on tariffs” has been ratified (this law
regulated only the life business). With the transfer, in 1923, of the control of
the insurance sector to the Ministry of Industry, begins a long period in which
insurance companies experiment a kind of subjection in respect to the public
administration. Only from the Seventies we observe a deep process of legisla-
tion, mainly due to the European Community regulations. We had three basic
directives.
First directives regulate and harmonise the discipline of the “freedom of
establishment”: a company with its head office in a country member of the EU
can open branches in other EU countries, to which is given the control on the
activity of the branch, according to the principle of the “host country control”.
Introduction 137

Second directives, as well as partially modify the first ones, deal with the
“freedom of services”, in particular with reference to industrial and commercial
risks, and to automobile insurance. The disposition of these directives is the
possibility for insurers to operate in other EU countries without trade barriers
and without the obligation to open a head office in loco.
Third directives, as well as partially modify the first and the second ones,
ratify: a) the application of the principle of the “home country control”; b) the
“single EU licence” allowing to operate in whole EU; c) the deregulation in the
control of tariffs.
The Legislator’s intervention incentives the passage from a strong protec-
tionist context to a wider and free market for insurance. In particular, the July,
1st 1994 has been a milestone for the insurance sector. With the coming into
force of the third directives on life and non-life business, in fact, a common
European market for insurance services has been created. The aim of this disci-
pline is to promote competition among insurers and benefit customers in terms
of a widening supply, reduction of tariffs and increase of quality of services.
The motor-vehicle insurance business, in particular, has known the strongest
deregulation process. Starting from July, 1st 1994 the public Authorities could
not control tariffs and insurance policy conditions anymore. The companies
started to be free to fix prices according to customers’ risk attitudes, and intro-
duced the new tariffs system based on the bonus/malus mechanism. In Italy,
seven direct selling companies for telephone and on-line selling were set up
and services started to be improved with the opening of call centres working 24
hours a day.
Today, in Italy the insurance sector is under the control of three Authorities:
a) the Consob (National Commission for the stock exchange market), on the
subject of transparency on the company’s information (Law 58/98); b) the
ISVAP (Italian Control and Vigilance Authority of the insurance sector), on the
subject of stability and of transparency on premiums and tariffs (Law 576/82);
c) the Antitrust Authority, on the subject of competition (Law 287/90).
In the last years the debate between insurers and Customers Associations has
been very harsh.
On the one side, companies maintain that the Automobile market is very
competitive, as the high differentiation in the tariffs demonstrates. Moreover,
the newly-born companies dealing with telephone and on-line selling - having
lower operating costs - have played a vital role in improving market efficiency
by limiting tariffs increase.
On the other side, Customers Associations have strongly argued that after
1994 tariffs have grown, according to the estimation measures, between the
50% and 100%.
138 Economies of scale, scope and experience in the Italian motor-vehicle sector

Some main events occurred:

1) the Government, due to the impact of insurance Automobile prices on in-

flation, in 2000 froze tariffs (Law 26/05/2000, n. 137), decision censured
because of its incompatibility with European laws;
2) the Italian Antitrust Authority sanctioned (measure no. 8546 - I377 Bul-
letin no. 30, 14 August 2000) a quite large number of companies - 41
insurers - for violation of the competition discipline (fined, as reported,
361,5 million Euro);
3) the exhaustive enquiry no. 11891 of the Italian Antitrust Authority on
the motor vehicle insurance, dated 17 April 2003 (Bulletin no. 16-17
2003), showed that besides the information exchange as collusive behav-
iour (already assessed by the measure no. 8546 of 2000), the Italian motor
vehicle insurance market presents anomalies which certainly do not in-
centivate companies to compete.27 In Italy, the liberalisation process did
not push insurance companies to increase the competition among them,
as happened in other European countries. The conclusions reached by
the Antitrust Authority asked for a radical innovation of the distribution
channels, a renewal of the remuneration-incentive system of agents and
of the refunding system.
4) Recently (measure no. 14926 of the 30 November 2005), the Antitrust
Authority sanctioned the Italian National Association of Insurers (ANIA,
Associazione Nazionale delle Imprese Assicurative, fined as reported 2
million Euro) for violation of the competition discipline and collusive
behaviour of its members with the Italian Associations of professionals in
charge of insurance damage liquidation (Associazioni dei periti).

From a theoretical perspective, a principal objective of financial services

deregulation28 is to improve market efficiency and enhance consumer choice
through increased competition. Efficiency gains can occur as the result of
the market consolidation. Consolidation, in fact, has the potential to improve
efficiency in an industry if it forces poorly performing firms exiting the market
or if it allows firms to take advantage of scale economies reducing unit costs of

27 For an analysis of the information exchange and of the Italian Antitrust intervention in the insurance market

under a law and economic perspective see Porrini (2004).

28 Useful references on the deregulation of the U.S. banking industry are Berger, Kashayap and Scalise
(1995), and Barth, Dan Brumbaugh and Wilcox (2000). The deregulation of European banking is discussed
in Barth, Nolle and Rice (1997); and the deregulation in the Japanese financial system is analysed by Deckle
(1988) and Goto (1999). The deregulation of the European insurance industry is illustrated in Hogan (1995),
Swiss Re (1996), and Hess and Trauth (1998); and the deregulation of the Japanese insurance industry is
presented in Swiss Re (2000a).
Data description 139

production. Moreover, firms offering different product lines also may realize
economies of scope. Nevertheless, the empirical evidence on economies of
scale and scope is contradictory: while some studies found efficiency gains
others show no efficiency gains or efficiency losses.29
It is also interesting to analyse the impact of the age of companies on their
performance, as proxy for their ability in survivor in a growingly competitive
market, and as a proxy for the experience acquired along time.
The main aim of this chapter is to provide new empirical evidence on classic
industrial organisation topics such as economies of scale, scope and of ex-
perience, analysing the Italian motor-vehicle insurance business. Besides we
provide also a bootstrap-based test on returns to scale and a bias-corrected esti-
mation of the efficiency scores of Italian insurers (along with 95% confidence
intervals). Analysing data of 2000, we provide also a test on the comparison of
structural efficiency of ‘fined’ vs. ‘non fined’ companies. Fined companies are
insurers hit in 2000 by the Antitrust measure no. 8546 (recalled above) for anti-
competitive behavior, while the non fined ones are companies not sanctioned
by the Antitrust measure.
The chapter is organised as follows. The next section presents the data
analysed, the inputs and outputs chosen as well as a normalized principal com-
ponents analysis to explore the dataset. After that, a procedure allowing the
aggregation of inputs and outputs is illustrated. The section that follows shows
the results of the bootstrapping exercise for a sensitivity analysis of the effi-
ciency scores and for a test on returns to scale. Then, economies of scale, scope
and experience are analysed and, finally, the main results are summarised in the
concluding section.

6.2 Data description

To give an outline on the dynamics of the Italian insurance market from 1982
to 2001 we show in Figure 6.1 and in Figure 6.2 the trend of nominal and real
rate of growth of gross premiums (direct business).
Table 6.1 defines the Nominal and Real rate of growth of premiums illustrated
in Figures 6.1 to 6.2.
A definition of the lines of business plotted can be found in Table 6.2.
Regarding the nominal and real rates of growth of gross premiums (direct
business) in the period 1992-2001, it is interesting to note that the automo-
bile business (Motor vehicles line - MV) has experimented a series of ups
and downs until the first half of Nineties, with a downward convergence in the

29 Useful surveys are Berger and Humphrey (1997); Cummins and Weiss (2001), and Amel, Barnes, Panetta
and Salleo (2002). See also Harker and Zenios (2000) for an overview on the performance of financial
institutions and its linkages with efficiency, innovation and regulation.
140 Economies of scale, scope and experience in the Italian motor-vehicle sector

40

30

20

10
MV.NRG
ONL.NRG
TNL.NRG
0 TL.NRG
TOT.NRG

1980 1985 1990 1995 2000

Year

Figure 6.1. Nominal rate of growth - gross premiums (direct business) by line of business,
years 1982-2001.

MV.RRG
40 ONL.RRG
TNL.RRG
TL.RRG
TOT.RRG

30

20

10

1980 1985 1990 1995 2000

Year

Figure 6.2. Real rate of growth - gross premiums (direct business) by line of business, years
1982-2001.
Data description 141

Table 6.1. Definition of Nominal and Real rate of growth.

Variable Definition
NRG Nominal Rate of growth year t
[(premium year t/premium year t-1) *100-100].
It is the increase of nominal premium in the business considered.
RRG Real Rate of growth=NRG-CPI
CPI Consumer Price Index
(percentage variation on previous year:
up to 1995 of Italy, from 1996 of Europe,
the harmonised Consumer Price. Source: Bank of Italy)

Table 6.2. Definition of rate of growth by line of business.

Variable Definition
M V N RG Motor vehicles premiums Nominal Rate of growth
M V RRG Motor vehicles premiums Real Rate of growth
ON L N RG Other Non life premiums Nominal Rate of Growth
ON L RRG Other Non life premiums Real Rate of Growth
T N L N RG Total Non Life premiums Nominal Rate of Growth
T N L RRG Total Non Life premiums Real Rate of Growth
T L N RG Total Life premiums Nominal Rate of Growth
T L RRG Total Life premiums Real Rate of Growth
T OT N RG Total premiums Nominal Rate of Growth
T OT RRG Total premiums Real Rate of Growth

1992-1994 period, and that after the deregulation period, both the nominal and
the real rate of growth of premiums have started to increase again (with a stop
in 2000 and 2001, when the Government imposed a price freeze on motor-third
party liability). This is particularly significant because the rate of growth of
premiums of other non-life lines continued to decline also after 1994.
It is interesting to note the difference in the evolution of the real rate of growth
of automobile business in comparison with the real rate of growth of the other
non-life lines (see Figure 6.2). In fact, as most of them are mature lines, we
would expected that they experienced a similar dynamics.
In the analysis which follows we use ANIA (Italian Association of Insurance
Companies) official data. We had access to the balance sheet and income
statements data as well as several features of the companies: date of foundation,
information on the activity of the firm, i.e. if it is a generalist insurer operating
142 Economies of scale, scope and experience in the Italian motor-vehicle sector

both in life and non-life business, if it is specialist of a business (life or non-life),

if it works in the auto liability business and so on.
In our empirical analysis we selected insurers operating in Italy, active and
working in the motor-vehicle business in the year 2000, and operating in the
market also in 1999 in order to compute the changes in reserves necessary in
the construction of the outputs (see the discussion in the next paragraph). We
selected a recent dataset available, in order to show the power of advanced robust
and nonparametric methods for the evaluation of the performance of insurers.
The number of retained companies is of 78. The 71.79% of our sample (i.e. 56
firms) are active only in the non-life business, i.e. are specialist insurers (SPEC);
the remaining 28.21% (i.e. 22 insurers) is made by generalist insurers (GEN),
active both in the life and nonlife business. Furthermore, 37 firms of our sample
(the 47.44%) were hit by the Italian Antitrust intervention (measure no. 8546
of 2000) for collusive behavior based on exchange of information. Hereafter
we call these fined companies as FIN, in opposition to non fined companies,
called NFIN. We notice also that most generalist insurers have been fined by
the Antitrust intervention.30

6.2.1 Definition of outputs and inputs

The definition of the inputs and outputs of insurers is a critical step in an effi-
ciency analysis. This is due to the fact that insurers offer services to customers
and most of their outputs are intangible (as it is the case for financial services).
The main approaches used to measure the outputs and inputs of insurers have
been reviewed in Cummins and Weiss (2001), where also a review of studies
applying frontier methodologies to the insurance industry is provided.
The main services provided by insurers are (the presentation below follows
Cummins and Weiss, 2001, p.790):

1. Risk-pooling and risk-bearing. Insurance provides a mechanism through

which consumers and businesses exposed to losses can engage in risk
reduction through pooling. The actuarial, underwriting, and related ex-
penses incurred in risk pooling are important components of value added
in the industry. Insurers also add value by holding equity capital to bear
the residual risk of the pool.

2. “Real” financial services relating to insured losses. Insurers provide a va-

riety of real services for policyholders including financial planning, risk
management, and the supply of legal defence in liability disputes. By
contracting with insurers to provide these services, policyholders take ad-

30 Thetotal number of companies hit by the Italian Antitrust Authority is 41, hence in our sample we have
more than 90% of the overall fined companies.
Data description 143

vantage of insurers’ specialised expertise to reduce the costs associated

with managing risks.

3. Intermediation. For life insurers, financial intermediation is a principal

function, accomplished through the sale of asset accumulation products
such as annuities. For non-life insurers, intermediation is an important but
incidental function, resulting from the collection of premiums in advance
of claim payments.

Transactions flow data, such as the number of applications processed, the

number of policies issued, the number of claims settled, etc., are not easily
available. However, a satisfactory proxy for the amount of risk-pooling and
real insurance services provided is the value of real losses incurred (Cummins,
Turchetti, Weiss, 1996; Berger, Cummins, and Weiss, 1997; Cummins, Weiss,
and Zi, 1999). Losses incurred are defined as the losses that are expected to
be paid as the result of providing insurance coverage during a particular period
of time. Because the objective of risk-pooling is to collect funds from the
policyholder pool and redistribute them to those who incur losses, proxying
output by the amount of losses incurred seems quite appropriate.
Losses are also a good proxy for the amount of real services provided, since
the amount of claims settlement and risk management services also are highly
correlated with loss aggregates.
Insurers perform also services in connection with claims occurring in prior
years that have not yet been settled or claims resulting from contingent events.
As a proxy for these services, in the definition of the output we consider, for
each year, the sum of the paid claims and the claims reserves of the year, minus
the claims reserves of previous year. By doing so, in the definition of the output,
we consider as proxy of financial services the sum of paid claims and changes
in reserves.
Finally, we identify 3 inputs and 2 outputs as well as some external factors
described in Table 6.3.
A descriptive statistics on inputs, outputs and external factors is offered in
Table 6.4. Inputs and outputs are expressed in thousands of Euros, Market shares
are percentages. The management index is given by the ratio of management
expenses on the total volume of premiums. It is an indicator of the intensity of
the company’s managerial activity.
In the efficiency analysis which follows we adopt an input oriented frame-
work. This choice is mainly due to our interest in understanding how efficiently
insurers were in using their resources (inputs) given that, at least in the short
run, the quantity and quality of outputs cannot be affected in a substantial way.
Moreover, this choice is also consistent with all the international literature of
the field (see e.g. Cummins and Weiss, 2001) as well as with previous studies on
144 Economies of scale, scope and experience in the Italian motor-vehicle sector

Table 6.3. Definition of Inputs, Outputs and External Factors.

Variable code Type of variable Description

in1 Input 1 Labour
in2 Input 2 Physical Capital
in3 Input 3 Financial capital
ou1 Output 1 Incurred losses and
changes in reserves
ou2 Output 2 Invested assets
zx1 Ext. factor 1 Age at 2000
(number of years active
on the market till 2000)
zx2 Ext. factor 2 Market share
zx3 Ext. factor 3 Management index
(manag. exp/tot premium)

Table 6.4. Descriptive statistics on inputs, outputs and external factor considered in the analysis.
Italian motor vehicle insurers (78 obs).

Variable mean stand. dev. min value max value

in1 128341.846 293996.564 301.000 2266839.000
in2 101892.795 211575.880 4.000 1078654.000
in3 1656299.051 4631613.110 6056.000 36607503.000
ou1 343555.308 581724.310 52.000 3229235.000
ou2 1351788.590 3621866.838 6046.000 27495409.000
zx1 56.128 49.605 1.000 179.000
zx2 0.663 1.167 0.001 6.760
zx3 0.348 0.580 0.083 4.698

the Italian insurance market (Turchetti and Daraio, 2004; Cummins, Turchetti
and Weiss, 1996).
In order to avoid the curse of dimensionality of nonparametric estimation
(here we have only 78 observations in a space at 3+2 dimensions) we tried to
reduce the dimensional space of the analysis. After an exploratory analysis,
in Section 6.2.3, we provide an illustration of a statistical methodology which
might be useful for the dimension reduction in productivity analysis.
Data description 145

6.2.2 An exploratory investigation

Before starting the efficiency analysis, it is always useful to have an idea
about the data we are going to deal with. At this purpose there are several
multivariate statistical tools which may be of interest to see multivariate dataset
(for a clear presentation of these tools in an applied perspective, see Härdle
and Simar, 2003). One of the most known tool is the normalized principal
component analysis (PCA). This kind of analysis, which aims at reducing the
information contained in a multivariate space providing illustrations in two
dimensions, was used as a preliminary investigation to an efficiency analysis
exercise in Deprins and Simar (1988). A first picture shows the correlation
structure existing among the variables; in a second picture, all the individuals are
projected on a reduced two dimensional space. The inspection of the correlation
matrix reported in Table 6.5 tells us, as it is often the case in productivity
analysis, that a strong correlation exists (higher than 0.80 in all the cases) among
all inputs and outputs. See also Table 6.7 where the correlations of the first two
principal components with the original variables are reported. It appears that
the two first principal components summarize 87% of the information. This is
the information provided by the cumulated percentage of variance explained by
the first two eigenvalues reported in Table 6.6. Hence, we can provide sensible
two dimensional pictures, without loosing too much information. Figure 6.3
displays the projections of the original variables: the coordinates are the factor
loadings, which are also the correlations of the principal components with the
original variables. We see that the first axis characterizes a “size” effect (level)
of the activity done by the insurers (it has, in fact, negative correlation with
all the activity variables, inputs and ouputs), so it could be interpreted as a
“dimensional factor”; the second axis characterizes the “management index”
of units. The two axis are orthogonal (no correlation) which indicates very

Table 6.5. Correlations matrix. Italian motor vehicle insurers (78 obs).

in1 in2 in3 ou1 ou2 zx1 zx2 zx3

in1 1.00 0.85 0.98 0.92 0.98 0.50 0.93 -0.10
in2 0.85 1.00 0.83 0.91 0.86 0.55 0.91 -0.12
in3 0.98 0.83 1.00 0.88 1.00 0.49 0.91 -0.10
ou1 0.92 0.91 0.88 1.00 0.90 0.55 0.98 -0.15
ou2 0.98 0.86 1.00 0.90 1.00 0.50 0.92 -0.10
zx1 0.50 0.55 0.49 0.55 0.50 1.00 0.56 -0.19
zx2 0.93 0.91 0.91 0.98 0.92 0.56 1.00 -0.15
zx3 -0.10 -0.12 -0.10 -0.15 -0.10 -0.19 -0.15 1.00
146 Economies of scale, scope and experience in the Italian motor-vehicle sector

Table 6.6. Eigenvalues and percentages of variances explained. Italian motor vehicle insurers
(78 obs).

eigenvalues % of variance cumulated %

5.9479 0.7435 0.7435
1.0201 0.1275 0.8710
0.6388 0.0799 0.9508
0.2461 0.0308 0.9816
0.1089 0.0136 0.9952
0.0275 0.0034 0.9987
0.0102 0.0013 0.9999
0.0004 0.0001 1.0000

Table 6.7. Correlations of the first two pc’s with the original variables (Factors loadings).
Italian motor vehicle insurers (78 obs).

Original variable % First pc Second pc

in1 -0.9699 0.0907
in2 -0.9256 0.0204
in3 -0.9573 0.1045
ou1 -0.9651 0.0164
ou2 -0.9688 0.0975
zx1 -0.6198 -0.2918
zx2 -0.9763 0.0191
zx3 0.1629 0.9514

few linear relationship between the dimension and the management index of
the units. The age of the companies (number of years they are active on the
market) is mainly related to the level of activity and interestingly seems not
related to the management index.
Figure 6.4, left panel, provides a two dimensional picture of the insurance
companies mapped on the two principal components. The interpretation of
this picture is facilitated (highlighted) by looking simultaneously at Figure 6.3
which give the weights of the original variables in the principal components.
Hence, on the left of Figure 6.4, we have the big insurance companies (in
terms of their level of activity), as an example, company n. 3 is Assicurazioni
Generali, company number 1 is RAS, n. 10 is SAI, n.8 is Fondiaria and so on
(these companies are among the biggest and most known insurance companies
Data description 147

Correlations circle
1 zx3

0.8

0.6

0.4
in3
in2
0.2 in1
Factor 2

0
zx2 ou2
−0.2
ou1

−0.4 zx1

−0.6

−0.8

−1
−1 − 0.5 0 0.5 1
Factor 1

Figure 6.3. Projection of the variables for the Italian Motor-vehicle insurers dataset, year 2000.

Projection of Units Projection of Units

0.2 47
1.5
0.15
1 37
0.1
58
0.5 77
78 0.05
Factor 2

Factor 2

36
3 14 47 41 71 6876
1 10 2 11 625 58 35 32 65
0 36
76
68
41
71
33
65
72 3829
7767 0 6933
67 72
5366
70
935
17
24
12 32
19
16
7
4666
53
70
40
34
18
73
75
27
30
63
23
56
51
64
57
61
31
50
44
60
59
74
62
21
120
355
43
48
42
52
49
45 26
28 29
38 1950
40
34
18
30
6364
27
515673
75
22
4 8 395 15 69
54 17
12 24 28 7 31235761
59
44
60
74
62
16 48
−0.05 21 43 55
−0.5 9 13 46
54 42 52
26
20 4549
−0.1
−1 22

−0.15
−1.5
−0.2

−1.5 −1 −0.5 0 0.5 1 1.5 0 0.1 0.2 0.3

Factor 1 Factor 1

Figure 6.4. Projection of the individual insurers for the Italian Motor-vehicle dataset, year
2000. Right panel a zoom.

in Italy). On the right, small units are present (the origin is the average point).
On the north of the picture we find companies with an high management index,
on the south, on the contrary, low management index companies. It appears that
the company n. 37 is particularly active along the management dimension as well
as company n. 78 and company n. 47. These companies are, respectively, Direct
Line, DB Assicura and Dialogo Assicurazioni, three relatively new companies
working through call center, which do not have a big activity dimension but put
148 Economies of scale, scope and experience in the Italian motor-vehicle sector

in place innovative management policies. According to human capital theories,

in fact, companies with an higher management index should outperform those
with a lower associated management index.
The right panel of Figure 6.4 is a zoom on the preceding picture which
displays the companies near the origin (which is the center of gravity of the
cloud of points). Generally speaking, observations which are far from the
origin deserve special attention, as they show an atypical behavior. Sometimes
extreme points flag potential outliers.

6.2.3 Aggregation of inputs and outputs

The high correlation levels we found among the outputs and among the
inputs suggest that a suitable aggregation of inputs and outputs for the efficiency
analysis is possible and in this case also recommended, because the curse of
dimensionality is always present in nonparametric estimation. Indeed the curse
of dimensionality implies that working in smaller dimensions tends to provide
better estimates of the frontier. Moreover, if we are able to reduce the frontier
estimation at the simplest situation (one input one output) it is also possible
provide graphical bi-dimensional illustrations of the estimated efficient frontier.
The basic principle (suggested by Mouchart and Simar, 2002) is to find an
“input factor”, a linear combination of the inputs, which best summarizes the
information provided by all the inputs (in this case p = 3).
Since this factor has to be interpreted as a proxy for all inputs, it should be
positively correlated with all inputs.
In addition, since inputs might be expressed in different units of measure, we
must correct for the scale of the inputs (by dividing each input by its standard
deviation or its mean). We recall that Farrell-Debreu efficiency scores are radial
measures and their DEA estimates have the property of being scale-invariant,
hence this means that the normalization of the data we perform here does not
affect the results of the following efficiency analysis.
Then, we are looking for a vector a ∈ R3+ such that the projection of the
(scaled) input data matrix X on the vector a, “best” represents the data matrix X
(in terms of minimizing the sum of squares of the residuals). Mathematically,
the vector of the n projections is determined by the resulting input factor Fin :
Fin = X a = a1 X1 + a2 X2 + a3 X3 (6.1)
where X : (n × 3) is the (scaled) inputs data matrix. It can be shown that
the optimal direction vector a is the first eigenvector of the matrix X
X cor-
responding to its largest eigenvalue λ1 (see Table 6.9, first column). Note that
here, we are in a different situation than in the PCA’s case: in this case, the data
have not been centered, so that the eigenvalues do not represent the factors’
variances. Eigenvalues are rather the “inertia” (or moment of the second order)
of the factor. So that the ratio λ1 /(λ1 + λ2 + λ3 ) indicates the percentage
Data description 149

Table 6.8. Input Factor inertia. Italian motor vehicle insurers (78 obs).

%inertia %cumul
0.9429 0.9429
0.0509 0.9937
0.0063 1.0000

Table 6.9. Eigenvectors of the matrix X
X. Italian motor vehicle insurers (78 obs).

0.5643 0.2178 0.7963

0.4857 -0.8676 -0.1068
0.6676 0.4471 -0.5954

Table 6.10. Correlations between the input factor (Fin ) and inputs. Italian motor vehicle
insurers (78 obs).

Fin in1 in2 in3

Fin 1.0000 0.9876 0.9089 0.9854
in1 0.9876 1.0000 0.8528 0.9803
in2 0.9089 0.8528 1.0000 0.8326
in3 0.9854 0.9803 0.8326 1.0000

of inertia which is explained by this first factor (see Table 6.8, first column).
When this ratio is high (close to 1), it indicates that most of the information
contained in the original 3-dimensional data matrix X, is well summarized by
the first factor Fin . Correlations between Fin and X1 , ..., X3 indicate also how
well this new one-dimensional variable represents the original ones (see Table
6.10).
In the output space, the same can be done with the 2 (scaled) output variables,
providing one output factor:

Fout = Y b = b1 Y1 + b2 Y2 (6.2)

where Y : (n × 2) is the data matrix of the (scaled) outputs. Here the vector
b ∈ R2+ is the first eigenvector of the matrix Y
Y , corresponding to its largest
eigenvalue (see Table 6.12, first column).
150 Economies of scale, scope and experience in the Italian motor-vehicle sector

Table 6.11. Output Factor inertia. Italian motor vehicle insurers (78 obs).

%inertia %cumul
0.9591 0.9591
0.0409 1.0000

Table 6.12. Eigenvectors of the matrix Y
Y . Italian motor vehicle insurers (78 obs).

0.5527 -0.8334
0.8334 0.5527

Table 6.13. Correlations between the output factor (Fout ) and outputs. Italian motor vehicle
insurers (78 obs).

Fout ou1 ou2

Fout 1.0000 0.9505 0.9915
ou1 0.9505 1.0000 0.9020
ou2 0.9915 0.9020 1.0000

The results of this factor analysis are shown below. We obtain:

a = (0.5643 0.4857 0.6676)
(6.3)

b = (0.5527 0.8334)
(6.4)

Therefore, in both cases, the factors are a sort of ‘average’ of the (scaled)
original variables. The percentage of inertia explained by the first factor is very
high in both cases (0.9429 for the input factor; 0.9591 for the output case, see
Table 6.11): it is certainly appropriate to summarize the information of the full
data matrix by these two one-dimensional factors, without loosing too much
information. The correlation between the factors and the original variables is
also high (above 0.83 for the input case; above 0.90 in the output case, see Table
6.13). This analysis shows that we may describe the production activity of all
these units by only one input factor and one output factor. Nevertheless, to
reach this conclusion in a rigorous way, it may be useful to apply the Simar and
Wilson (2001)’s bootstrap based procedures to test for aggregation possibilities
(restrictions) on inputs and outputs in efficient frontier models. In our case
here we did not perform this test due to the high correlations found between the
original variables and their aggregate factors.
Testing returns to scale and bootstrapping efficiency scores 151

6.3 Testing returns to scale and bootstrapping efficiency

scores
We applied the bootstrap algorithm described above in Section 3.4 to test
the model of returns to scale (RTS) in this application, following the method
proposed by Simar and Wilson (2002).
We have seen in Section 2.3 that returns to scale are properties of the boundary
Ψ∂ of the production set. The less restrictive model for RTS is the varying return
to scale (VRS) situation where the returns are allowed to be eventually locally
increasing, then constant and finally non-increasing.
The first procedure proposed by Simar and Wilson (2002) is to test the more
restrictive model of constant returns to scale (CRS) against the VRS: H0 : Ψ∂
is globally CRS against H1 : Ψ∂ is V RS.
Under VRS, the attainable set is estimated by the free-disposal convex
hull of the cloud of points. So we can compute the DEA efficiency scores
θV RS,n (Xi , Yi ) under this model by solving Equation (2.16). Under the more
restrictive CRS model, the DEA 
estimator is calculated according to Equation
(2.16), where the constraint ni=1 γi = 1 on the multipliers is dropped, and the
resulting DEA scores are denoted by θCRS,n (Xi , Yi ).
The VRS estimators are consistent whatever being the hypothesis on RTS,
the CRS are only consistent if the CRS hypothesis is true. So, if the CRS
hypothesis is true, the two sets of estimators would be very similar. We will
use below as test statistics the mean of the ratios of the efficiency scores:
n 
1 θCRS,n (Xi , Yi )
T (Xn ) = . (6.5)
n i=1 θV RS,n (Xi , Yi )

By construction θCRS,n (Xi , Yi ) ≤ θV RS,n (Xi , Yi ) and we will reject the null
hypothesis if the test statistics T is too small. The p-value of the null-hypothesis
is then obtained by computing:
p − value = Prob(T (Xn ) ≤ Tobs |H0 is true), (6.6)
where Tobs is the value of T computed on the original observed sample Xn .
Of course, we cannot compute this probability analytically but we can ap-
proximate this value by using the bootstrap algorithm described in Section 3.4.
We simulate B pseudo-samples Xn∗,b of size n under the null (i.e. using the
CRS estimate of the frontier for generating the pseudo-samples), and for each
bootstrap sample we compute the value of T ∗,b = T (Xn∗,b ). The p-value is then
approximated by the proportion of bootstrap samples with values of T ∗,b less
than the original observed value Tobs :

B
1I(T ∗,b ≤ Tobs )
p − value ≈ . (6.7)
b=1
B
152 Economies of scale, scope and experience in the Italian motor-vehicle sector

In the application here, with one input-factor and one output-factor as defined
above, we obtain for this test (with B = 2000) a p-value of 0.0055 < 0.05,
hence we reject the null hypothesis of CRS.
Before accepting the VRS hypothesis, Simar and Wilson (2002) suggest to
perform the following test where the null-hypothesis is less restrictive than the
CRS: we test the non-increasing returns to scale (NIRS) model against the VRS:
H0
: Ψ∂ is globally N IRS against H1 : Ψ∂ is V RS.
The procedure is similar to the preceding one where CRS has to be replaced by
NIRS and θN IRS,n (Xi , Yi ) is computed as in Equation (2.16)
where the equality
constraint on the multipliers is replaced by the inequality ni=1 γi ≤ 1. The
computations for this second test lead to a p-value of 0.0405 < 0.05. Hence,
we reject H0
(even if we are close to a border line case) and choose to accept
H1 , i.e. the hypothesis of VRS.
We can visualize the efficient DEA-VRS frontier in Figure 6.5 with a zoom
on the core of the cloud of points in Figure 6.6. Several interesting information
can be obtained by inspecting these figures. For instance, we see that companies
n. 3 and n. 1 are estimated as efficient; they are also isolated with no other
companies to be benchmarked against.

DEA–VRS frontier
25

20
Output factor

15

10

10
14 4
5 2 8
5 6
15
12 11
917 25 39
26
35
28
38
137224
9
32
46
50
33
69
63
2016
19
31
67
22
21
27
30
23
18
34
48
45
42
72
55
57
74
53
51
64
49
54
60
40
65
59
75
73
43
61
66
52
58
56
76
71
70
44
68
41
62
36
37
47
0 78
77
0 5 10 15 20 25 30 35
Input factor

Figure 6.5. Output factor versus Input factor and DEA frontier for the Italian Motor-vehicle
dataset, year 2000.

In Table 6.14 (second column) we show the FDH (input oriented) efficiency
score computed using the input factor and the output factor (we notice that
the computation of the FDH efficiency scores with 3 inputs and 2 outputs led
Testing returns to scale and bootstrapping efficiency scores 153

DEA−VRS frontier
0.16
57
0.14 74 53
64

0.12
49
60
0.1 54
Output factor

65
40
59
0.08
75 61
73 43
0.06 58 66
52
0.04 56
76
44 70
0.02 62 37
71
47 36 41
6878
0 77
0 0.05 0.1 0.15 0.2
Input factor

Figure 6.6. Output factor versus Input factor and DEA frontier for the Italian Motor-vehicle
dataset, year 2000: a zoom.

to an average efficiency of 0.9956 with 73 efficient observations on a total of

78 observations); the DEA VRS (input oriented) computed with 3 inputs and
2 outputs (third column); the DEA VRS (input oriented) efficiency estimates
computed on the input and output factor (fourth column). In the fifth column
we report the Bias-corrected DEA VRS efficiency scores computed on the input
and output factors; column sixth reports the bias of the DEA efficiency scores,
while the seventh column reports the standard deviation of the estimates. When
the Bias is larger than the standard deviation (std), the Bias-corrected estimates
have to be preferred to the original estimates. Finally, columns eight and nine of
Table 6.14 display the Basic Bootstrap (Simar and Wilson, 2000b) 95-percent
Confidence Intervals: Lower and Upper Bound. We use the algorithm described
in Section 3.4, with B = 2000 bootstrap replications.
As noticed above for the FDH case, the DEA estimation in the third column
using 3 inputs and 2 outputs is not very reliable. Indeed the rate of convergence,
being n2/(p+q+1) (here 782/6 = 4.26) in this high dimensional space, corre-
sponds roughly as if we had 18 observations in a full parametric model, where
the rate of convergence is n1/2 . This remind us that it is not reasonable to use
FDH and DEA estimates with 78 observations in a five dimensional space.
154 Economies of scale, scope and experience in the Italian motor-vehicle sector

Table 6.14. Bootstrap results: input efficiency scores with VRS model (B=2000) for Italian
motor- vehicle insurers.

DEA DEA DEA

No. FDH Eff. Eff. Eff. Bias Std Lower Upper
Obs. Eff Score Score Bias Bound Bound
(VRS) (VRS) Corr.
IF OF p=3 IF OF IF OF
q=2
1 1.0000 1.0000 1.0000 0.6149 0.3851 0.2927 0.6675 0.9838
2 0.8048 1.0000 0.6725 0.6022 0.0703 0.0516 0.5473 0.6672
3 1.0000 1.0000 1.0000 0.3445 0.6555 0.9137 0.5629 0.9817
4 0.9780 0.9711 0.6709 0.5733 0.0976 0.0675 0.5242 0.6630
5 1.0000 1.0000 1.0000 0.9028 0.0972 0.0672 0.8229 0.9882
6 1.0000 1.0000 0.8127 0.7336 0.0791 0.0549 0.6685 0.8017
7 0.9992 0.9455 0.7009 0.6657 0.0352 0.0227 0.6302 0.6967
8 0.9959 0.9639 0.5964 0.5262 0.0702 0.0498 0.4774 0.5887
9 1.0000 0.8669 0.7897 0.7424 0.0473 0.0299 0.6985 0.7850
10 1.0000 1.0000 0.8605 0.6833 0.1772 0.1204 0.6411 0.8485
11 0.7003 0.8172 0.5989 0.5469 0.0520 0.0360 0.5001 0.5930
12 1.0000 1.0000 1.0000 0.9198 0.0802 0.0469 0.8578 0.9833
13 1.0000 0.8956 0.8334 0.7915 0.0419 0.0269 0.7493 0.8287
14 1.0000 1.0000 0.8503 0.7599 0.0904 0.0662 0.6906 0.8421
15 0.9507 1.0000 0.9244 0.8371 0.0873 0.0610 0.7637 0.9164
16 0.7835 0.9408 0.5246 0.4954 0.0292 0.0166 0.4708 0.5206
17 1.0000 1.0000 0.9362 0.8678 0.0684 0.0410 0.8119 0.9291
18 0.9958 0.8831 0.7829 0.7483 0.0346 0.0232 0.7084 0.7795
19 1.0000 0.9331 0.6818 0.6432 0.0386 0.0216 0.6114 0.6768
20 0.6772 0.9357 0.6227 0.5912 0.0315 0.0188 0.5627 0.6194
21 0.8122 0.8727 0.5568 0.5314 0.0254 0.0159 0.5049 0.5536
22 0.6753 0.8246 0.5219 0.4971 0.0248 0.0151 0.4729 0.5187
23 0.8136 0.7629 0.6594 0.6304 0.0290 0.0194 0.5969 0.6564
24 0.9619 1.0000 0.6887 0.6470 0.0417 0.0262 0.6088 0.6846
25 1.0000 1.0000 0.7987 0.7356 0.0631 0.0415 0.6790 0.7894
26 1.0000 0.9763 0.7376 0.6959 0.0417 0.0266 0.6572 0.7333
27 1.0000 1.0000 0.8595 0.8214 0.0381 0.0248 0.7783 0.8550
28 1.0000 1.0000 0.7721 0.7314 0.0407 0.0266 0.6905 0.7676
29 0.7258 1.0000 0.5381 0.5101 0.0280 0.0182 0.4819 0.5347
30 0.8664 0.9381 0.6980 0.6673 0.0307 0.0206 0.6317 0.6949
31 1.0000 1.0000 1.0000 0.9409 0.0591 0.0314 0.8953 0.9910
32 1.0000 1.0000 0.8145 0.7722 0.0423 0.0257 0.7316 0.8077
33 1.0000 1.0000 1.0000 0.9467 0.0533 0.0303 0.9011 0.9921
34 0.7335 0.9400 0.5576 0.5327 0.0249 0.0166 0.5043 0.5551
35 0.9650 0.9557 0.6347 0.6009 0.0338 0.0219 0.5673 0.6307
36 0.8304 0.8654 0.3978 0.3562 0.0416 0.0340 0.3195 0.3951
37 0.3164 1.0000 0.1333 0.1248 0.0085 0.0060 0.1157 0.1325
38 1.0000 0.9093 0.7699 0.7308 0.0391 0.0253 0.6913 0.7645
39 0.9270 1.0000 0.6543 0.6017 0.0526 0.0354 0.5538 0.6462
Testing returns to scale and bootstrapping efficiency scores 155
Table 6.14 (continued)
Bootstrap results: input efficiency scores with VRS model (B=2000) for Italian motor-vehicle
insurers.

DEA DEA DEA

No. FDH Eff. Eff. Eff. Bias Std Lower Upper
Obs. Eff Score Score Bias Bound Bound
(VRS) (VRS) Corr.
IF OF p=3 IF OF IF OF
q=2
41 0.4386 0.8005 0.2214 0.2001 0.0213 0.0174 0.1795 0.2200
42 1.0000 0.6656 0.6113 0.5801 0.0312 0.0203 0.5490 0.6082
43 0.4324 0.8238 0.4298 0.4000 0.0298 0.0173 0.3778 0.4275
44 1.0000 0.9418 0.5347 0.5044 0.0303 0.0200 0.4730 0.5318
45 0.9107 0.8545 0.5460 0.5185 0.0275 0.0180 0.4906 0.5434
46 1.0000 1.0000 0.7709 0.7294 0.0415 0.0231 0.6945 0.7636
47 1.0000 1.0000 1.0000 0.5434 0.4566 0.2952 0.6623 0.9829
48 0.9960 1.0000 0.7091 0.6766 0.0325 0.0214 0.6404 0.7053
49 0.9945 1.0000 0.8066 0.7544 0.0522 0.0297 0.7141 0.7991
50 1.0000 1.0000 0.7544 0.7141 0.0403 0.0228 0.6796 0.7481
51 0.4368 0.8265 0.4334 0.4024 0.0310 0.0175 0.3817 0.4312
52 0.8549 1.0000 0.6321 0.5945 0.0376 0.0241 0.5603 0.6297
53 0.8908 1.0000 0.6860 0.6382 0.0478 0.0275 0.6050 0.6826
54 0.6055 0.8666 0.4999 0.4693 0.0306 0.0182 0.4432 0.4958
55 1.0000 1.0000 0.8891 0.8378 0.0513 0.0321 0.7928 0.8835
56 1.0000 0.6736 0.5558 0.5246 0.0312 0.0197 0.4954 0.5527
57 0.7657 0.9732 0.6324 0.5924 0.0400 0.0241 0.5610 0.6285
58 1.0000 1.0000 0.7580 0.7124 0.0456 0.0290 0.6715 0.7548
59 0.8796 0.8330 0.6009 0.5636 0.0373 0.0223 0.5324 0.5954
60 1.0000 1.0000 0.8322 0.7812 0.0510 0.0304 0.7378 0.8254
61 1.0000 0.9327 0.7927 0.7378 0.0549 0.0316 0.6970 0.7873
62 1.0000 0.9068 0.6665 0.6187 0.0478 0.0354 0.5680 0.6617
63 0.9749 1.0000 0.8978 0.8523 0.0455 0.0271 0.8112 0.8929
64 0.8991 0.9479 0.6794 0.6307 0.0487 0.0275 0.5983 0.6757
65 0.8763 0.7432 0.6278 0.5894 0.0384 0.0231 0.5567 0.6227
66 0.5308 0.9085 0.4184 0.3927 0.0257 0.0161 0.3701 0.4159
67 0.7492 0.9009 0.5731 0.5415 0.0316 0.0172 0.5157 0.5671
68 1.0000 1.0000 0.7950 0.7083 0.0867 0.0704 0.6359 0.7891
69 0.5065 0.6566 0.4914 0.4657 0.0257 0.0149 0.4433 0.4882
70 1.0000 0.6527 0.5300 0.5000 0.0300 0.0197 0.4690 0.5270
71 1.0000 0.9485 0.4667 0.4386 0.0281 0.0193 0.4081 0.4636
72 0.9062 1.0000 0.5130 0.4850 0.0280 0.0179 0.4588 0.5097
73 0.6892 0.9238 0.6833 0.6359 0.0474 0.0275 0.6005 0.6795
74 1.0000 1.0000 1.0000 0.9280 0.0720 0.0405 0.8804 0.9941
75 1.0000 1.0000 1.0000 0.9303 0.0697 0.0401 0.8787 0.9937
76 0.7796 0.9257 0.3733 0.3525 0.0208 0.0131 0.3329 0.3712
77 0.9491 1.0000 0.7029 0.6113 0.0916 0.0705 0.5496 0.6960
78 1.0000 1.0000 0.7336 0.5876 0.1460 0.1044 0.5457 0.7238
156 Economies of scale, scope and experience in the Italian motor-vehicle sector

It is interesting to note that the widest intervals correspond to observations

no. 3 and no. 1 which are the most remote observations in Figure 6.5. As we have
seen before, Figure 6.5 tells us that there are simply no other observations near
no. 3 and no. 1 that would help estimate the frontier in their region. Interestingly,
the bootstrap is able to automatically take into account the information available
for the estimation of each efficiency score: if there are a lot of firms in the same
region of a certain unit, then the confidence interval of the efficiency score of
this unit will be very thin; on the contrary, if the unit is isolated then its efficiency
score will have a wide confidence interval.

6
Boxplots

1
0 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48 51 54 57 60 63 66 69 72 75 78
No. Obs.

Figure 6.7. Boxplots of the input Shephard efficiency scores (VRS). B=2000. Italian motor-
vehicle insurance business (78 obs).

Figure 6.7 illustrates the boxplots of the input Shephard efficiency scores
(the inverse of the Farrell efficiency measures) coming out from our bootstrap
exercise: it clearly appears that firms no. 37, 41 and 40 are the most inefficient.
Figure 6.6 shows, in fact, that unit no. 37 uses a productive mix (as represented
by the ray coming from the origin which passes through the unit) that is dom-
inated by the productive mix used by the other firms in the sample. Close to
the mix of this unit is firm no. 41, which is on a ray between firm no. 37 and
firm no. 40. Slightly better is the situation of firms no. 54 (efficiency score
0.4999), 43 (eff. score 0.4298), 66 (0.4184), 76 (0.3733), 71 (0.4667), which
use a technology that combines inputs and outputs more efficiently than the
previous firms. Then, we observe firms no. 57, 59, 65, 52 and 70 whose mix
dominates the previous firms and which in turns is dominated by the technology
Economies of scale 157

of firms no. 49, 60, 61, 58, 62 and finally by the most efficient technology of
units 47, 74 and 75.
The inspection of the last two columns of Table 6.14 reveals that the original
efficiency estimates lie outside the estimated confidence intervals. This is due
to the fact that the original estimates are biased and that the confidence intervals
estimates correct for the bias. Finally we note also that for unit no. 3 and no.1 the
relative estimated confidence intervals do not contain the corresponding bias-
corrected efficiency scores. This is related to the fact that the bias corrections
are made in terms of the original efficiency estimates, while reciprocals are
used to construct the confidence interval estimates, and because the sample
information in this region is very poor.

6.4 Economies of scale

Government policies encouraging consolidation make sense economically if
larger firms tend to be more efficient, if there are unrealised scale economies,
and/or if consolidation leads to more vigorous competition that increases market
efficiency.
Scale economies might be present in the insurance industry not only because
fixed costs are spread over a wider base as firm size increases but also because
insurance involves the diversification of risk, which is more effective in larger
risk pools. On the other hand, if insurance is primarily a variable cost industry
and insurers can use reinsurance to reduce income volatility, significant scale
economies may not be present.
For insurance markets where there is the presence of large numbers of very
small firms, there is a common belief that those small firms are scale inefficient
and not sufficiently robust to compete effectively as the EU moved towards
deregulation. It is often argued that larger insurers would provide better value
and service to insurance customers and would be more competitive with other
EU insurers. Thus, changes in the government policy are often based on the
implicit (and untested) assumption that there are significant unrealised scale
economies in the industry and that favouring the creation of larger firms would
lead to more viable insurers and a more competitive market.
Economies of scale are present if average costs per unit of output decline
as the volume of output increases. The usual source of scale economies is
the spreading of the firm’s fixed costs over a larger volume of output. Fixed
costs are present for insurers due to the need of relatively fixed factors of pro-
duction such as computer systems, managerial expertise, and financial capital.
Economies of scale also can arise if operating at larger scale allows managers
to become more specialised and therefore more proficient in carrying out spe-
cific tasks. Operating at larger scale can reduce the firm’s cost of capital if
income volatility is inversely related to size. This source of scale economies
may be particularly applicable to insurers, because the essence of insurance is
158 Economies of scale, scope and experience in the Italian motor-vehicle sector

risk diversification through pooling. These arguments lead to the prediction that
insurance operations are likely to encounter ranges of production characterised
by increasing returns to scale, permitting some insurers to reduce unit costs by
increasing production, at least within certain limits.
In this section we provide information on whether these critical assumptions
are correct and whether consolidation is likely to be beneficial in Italy, by
applying the econometric methodology described in Chapter 5. In particular
we use the conditional robust efficiency scores, where the external factor is
represented by a proxy of the size of the insurer, i.e. Z is the market share,
to shed lights on the impact of size on the performance of the Italian insurers.
This choice seems to be reasonable as market shares gives an approximation
of the volume of the activity carried out by the insurers and hence of their
dimension. We do not use total costs or incurred losses to proxy size as we use
these variables in the construction of inputs and outputs.
Note that the analysis we carry out in this section is quite different than the
analysis of Returns to Scale (RTS) done in the previous section. In general,
RTS are properties of the frontier of the production set, and are calculated as-
suming the convexity of the production set and using a deterministic estimator
(DEA) which suffers from the curse of dimensionality and of the influence of
extremes/outliers. Here we propose to use robust estimators (such as order-m)
which do not assume any convexity for the production set, and are less influ-
enced by extreme points. Moreover, the econometric methodology developed
in Chapter 5 gives us the possibility of measuring the impact of size at the level
of the individual firm as well as globally, offering the opportunity of capturing
also local effects if they are at place.
Interestingly, Figure 6.8 shows the scatterplot and a straight smoothed non-
parametric regression line of the ratios Qz = θ̂n (x, y | z)/θ̂n (x, y) on Z. On
the contrary, Figure 6.9 - top panel - illustrates an increasing nonparametric
regression line of the Qzm = θ̂n,m (x, y | z)/θ̂n,m (x, y) on Z, till around a mar-
ket share of 1. This trend is confirmed by the ratios Qzα reported at the bottom
panel of Figure 6.9. Here we chose a level of m = 35 and α = 0.97 robust
at around 10%, and our data driven nearest neighborhood approach selected a
k − N N = 29.
As extensively explained in Chapter 5, an increasing (decreasing) nonpara-
metric regression line denotes a negative (positive) effect of the external factor
(Z) on the performance of benchmarked firms, in an input oriented framework.
This case corresponds exactly to the situation described in Section 5.4 where
Figure 5.3 showed no effect, while Figure 5.4 showed a negative impact (here
the impact is till Z = 1). In this situation, even if in a inputs-outputs space
there are not heterogeneous (extreme) units, these extremes may appear in a
more complete space in which external factors have a role. This confirms the
Economies of scale 159

Effect of Z on Full frontier

1.05
1
0.95
0.9
0.85
Qz

0.8
0.75
0.7
0.65
0.6

0 1 2 3 4 5 6
values of Z

Figure 6.8. Economies of scale in the Italian motor-vehicle insurance business. Full frontier
case (78 obs). Z = market share.

Effect of Z on Order−m frontier

1
0.9
Qzm

0.8
0.7
0.6
0 1 2 3 4 5 6
values of Z
Effect of Z on Order − α frontier
1
0.9
Qzα

0.8
0.7
0.6
0 1 2 3 4 5 6

Figure 6.9. Economies of scale in the Italian motor-vehicle insurance business. Partial Frontiers
case (78 obs) Z = market share.

usefulness of partial frontiers in revealing the impact of external factors masked

in the full frontier case by “different” or extreme points.
By inspecting Figure 6.9 it appears that there are two groups of insurers:
a smaller group with big insurance companies (market share greater than 1.5)
that is not affected by its dimension, and a larger group with the majority
of Italian motor-vehicle companies in it, which is on the increasing part of
the nonparametric regression line companies. For this latter group there is a
negative influence of size on their performances. As most of Italian insurers are
agglomerated on this side of the scatter plot, at this exploratory level, it seems
160 Economies of scale, scope and experience in the Italian motor-vehicle sector

that there is a prevalence of diseconomies of scale in the Italian motor-vehicle

insurance industry.

6.5 Economies of scope

Economies of scope may arise from a reduction in the search, information,
monitoring, and/or transaction costs borne by the company. Costs may be
significantly reduced when several financial products are sold by the same firm.
In order to investigate on the presence of economies of scope, we compared
the performance of generalist versus specialist insurers. Generalist insurers are
companies active in both life and non life business; while specialist insurers are
active only in the non life business. We tested if the difference on the average
efficiency scores of the two groups of insurers was statistically significant by
applying the bootstrap-based procedure described in Section 3.4.
Here we use the bootstrap to estimate the p−value of a null hypothesis H0 ,
i.e. that generalist (group 1) and specialist (group 2) insurers have the same
average efficiency score. The null hypothesis is rejected at the desired level
(here α = 0.05) when the p−value is too small, i.e. lower than the desired
level.
For group 1 and 2 we can postulate that:

H0 : E(θ1 ) = E(θ2 )
against
H1 : E(θ1 ) > E(θ2 ),
where E(θ1 ) is the mean efficiency of generalist insurers and E(θ2 ) is the mean
efficiency of specialist insurers. The mean of DEA input efficiency estimates
for the n1 = 22 generalist insurers is of 0.7497 (with a std. dev. of 0.2138),
while the mean of specialist insurers n2 = 56 is of 0.6658 (with a std. dev.
of 0.1786). The full sample size is n = n1 + n2 = 78 and the full sample is
denoted by Xn = Xn1 ∪ Xn2 . The overall mean efficiency is 0.6894 (with a
std. dev of 0.1915).
The test statistic we use is the following:
  i , yi )
n−1
1 i|(xi ,yi )∈Xn1 θ(x
T (Xn ) =   i , yi )
,
n−1
2 i|(xi ,yi )∈Xn2 θ(x

where θ(x  i , yi ) is the input oriented DEA VRS efficiency estimator of the unit
(xi , yi ) computed using the full sample as the reference set. When the null is
true, then by construction T (Xn ) will be “close” to 1. On the contrary, when
the alternative is true, T (Xn ) will be “far” from 1 (i.e. larger than one). The
p-value for this test has the following form:
Economies of scope 161

p − value = Prob(T (Xn ) ≥ Tobs | H0 ).

The value of the test statistic obtained from our sample is the following: Tobs =
1.1261. The p-value cannot be computed analytically but again the bootstrap
algorithm of Chapter 3 allows to approximate this value. We generate B pseudo-
values Xn,b , b = 1, ..., B under the null hypothesis (i.e., considering that the
2 subsamples come from the same DGP, so we resample from the full sample
Xn ). We compute the test statistics T ,b = T (Xn,b ) for each bootstrap sample,
then the p-value is computed as:


B
1I(T ∗,b ≥ Tobs )
p − value ≈ . (6.8)
b=1
B

In our application here, with B = 2000 replications, we obtained a p-value

of 0.8780. As a result, the null hypothesis of equality of structural (mean)
efficiency between generalist and specialist insurers cannot be rejected.
We found that the average of generalist insurers (working both in the life
and in the nonlife business) has an efficiency score (DEA, VRS input oriented)
of 0.7497 which seems higher with respect to that one of specialist companies
working in the nonlife business (0.6658), nevertheless, this observed difference
is not statistically meaningful.
Table 6.15 shows some descriptive statistics on the insurers operating in the
Italian motor-vehicle business. It appears that Generalist insurers are bigger
and older than the Specialist ones, with an average market share of 1.84 % and
an average Age of 94 years against an average market share of 0.20 % and an
average Age of 41 for the Specialist ones.
We know also that almost all the Generalist insurers have been fined by
the Italian Antitrust Authority. To better understand if there is a substantial
difference in the efficiency of fined (FIN) versus not fined (NFIN) insurers, we
compare the profile of fined specialist (FIN SPEC) against not fined specialist
(NFIN SPEC) insurers. Table 6.15 also displays some descriptive statistics on
these two groups: it shows that fined specialist insurers (19 obs.) are bigger
and older than not fined companies (37 obs.); they present an average market
share of 0.43 % and an average age of 52 years against, respectively 0.08 %
and 36 years, for not fined specialist insurers.
By applying the same bootstrap-based procedure described above we tested if
there is a significantly difference between the average performance of specialist
insurers hit by the Antitrust measure (FIN SPEC) and not fined insurers (NFIN
SPEC). The whole mean (VRS input oriented) efficiency estimate is of 0.6820
(with a standard deviation of 0.1890), 56 specialist companies. The average
efficiency of fined specialist - 19 units- is 0.7146 (with a st. dev. of 0.2260),
162 Economies of scale, scope and experience in the Italian motor-vehicle sector

Table 6.15. Some descriptive statistics on Age and size (Mktshare) of Italian motor-vehicle
insurers.

Sample Descr. Stat. Age Mktshare

All (78) Mean 56.1 0.6633
St Dev. 49.6 1.1666
Min 1.0 0.0001
Max 179.0 6.7600
Median 38.0 0.1400
GEN (22) Mean 94.1 1.8373
St Dev. 47.4 1.6540
Min 19.0 0.0100
Max 175.0 6.7600
Median 78.0 1.8050
SPEC (56) Mean 41.2 0.2022
St Dev. 42.3 0.3051
Min 1.0 0.0001
Max 179.0 1.7100
Median 25.5 0.0950
FIN SPEC (19) Mean 51.7 0.4321
St Dev. 51.3 0.4304
Min 1.0 0.0001
Max 179.0 1.71
Median 27 0.2600
NFIN SPEC (37) Mean 35.8 0.0841
St. Dev 36.4 0.0855
Min 1.0 0.0001
Max 121.0 0.3100
Median 24.0 0.0600

while not fined specialist insurers -37 units- have an average efficiency of 0.6652
0.7146
(with a st. dev. of 0.1679). Tobs = = 1.0743.
0.6652
The estimated p-value of the null hypothesis (equality of means), with B =
2000 bootstrap replications, is 0.7755.
Hence, the H0 of equality of the mean efficiency of fined and not fined
specialist insurers cannot be rejected; even if fined insurers seem to be more ef-
ficient than the not fined ones, this difference is not significant at any statistically
meaningful level.
Economics of experience 163

Therefore, we didn’t find any statistically significant difference in the perfor-

mance of generalist insurers and insurers specialized in the non-life business,
and any meaningful difference between “fined” and “not fined” specialist insur-
ers. From our analysis, it seems that in 2000, motor-vehicle insurers working
also in the life business do not have exploited economies of scope arising by
the selling of several different financial products. It emerged also that in 2000
fined specialist companies -working only in the nonlife business- did not show
a different structural (average) efficiency with respect to not fined specialist
insurers.
It seems that in 2000, fined companies (who did not respect the competition
discipline, as assessed by the Italian Antitrust Authority31 ) did not increase
their technical efficiency notwithstanding their anti-competitive behavior: to
improve the performance more structural innovations are required.

6.6 Economies of experience

In an industry where the principal good sold is the promise of the compensa-
tion of a future possible loss, the trust in the seller is one of the most important
factors guiding the customers in their choice. The insured, in fact, has to rely
on the ability of the insurer to be alive and to be able to refund him in case of
a loss in the future. The reputation and the age of the company or, better, its
reputation during the time, are good indicators to look at before granting trust
to an insurer. Moreover, the age could be considered as an indicator of consol-
idated knowledge of the market, as older companies, being in the business for
a longer period, should benefit from experience economies or, in other words,
should know better than younger companies how the business works and how
to manage it.
We explore this issue by using the conditional robust nonparametric effi-
ciency measures (presented in Chapter 5) where the conditioning factor (Z) is
represented by the age of insurers.
Figure 6.10 is highly emblematic: there are no maturity in the business ef-
fects for the Italian insurance industry. We have in fact a straight nonparametric
regression line indicating that the external factor Z -the companies’ age- does
not affect the efficiency of Italian companies. Hence, the performance of in-
surers is not influenced by their age, as computed with respect to their date of
foundation.
In this application the tuning parameters have been set to m = 35 and
α = 0.97 to reach a level of robustness around 10%. The number of k − N N
provided by our data driven procedure is 15.

31 Thewhole documentation is available on line, in Italian, on the website of the Italian Antitrust Authority
at: http://www.agcm.it/.
164 Economies of scale, scope and experience in the Italian motor-vehicle sector

Effect of Z on Full frontier

2
Qz

20 40 60 80 100 120 140 160

Effect of Z on Order – α frontier

2
Qzα

20 40 60 80 100 120 140 160

Effect of Z on Order–m frontier

2
Qzm

20 40 60 80 100 120 140 160

values of Z

Figure 6.10. Economies of experience in the Italian motor-vehicle insurance business, year
2000 (78 obs.). Z = Age in years from foundation.

6.7 Conclusions
In this chapter we explored some classical issues in the industrial organization
literature, adding empirical evidence on the presence of economies of scale,
economies of scope and experience on a sample of Italian insurers, active in the
motor-vehicle business, for the year 2000.
The methodology presented in part I of this work offers a rigorous and easy-
to-interpret tool for evaluating the performance of insurers. It may be applied
to monitor the dynamics of the performance of insurers at a regional, national,
European and international level.
We analysed the sensitivity of efficiency scores through bootstrapping and
find out also that the Returns to Scale of the frontier of the production set of
Italian insurers are variable.
We showed that the application of the probabilistic approach to introduce
external-environmental variables is useful to monitor the influence of these
factors on the performance of insurers.
Using our approach we tested, with a certain rigor, if some commonly as-
sumed economic principles on the insurance industry are empirically well-
grounded.
Conclusions 165

We found that there is a prevalence of diseconomies of scale in the Italian

motor-vehicle insurance industry. This finding if far from sustain government
policies encouraging consolidation. In fact, policies encouraging growth in
firm size make sense on scale efficiency ground only if there are many insurers
operating with increasing returns to scale. The evidence showed in this chapter
(for the year 2000) goes in the opposite direction.
We did not find a statistically significant difference in the performance of
generalist insurers and insurers specialized in the non-life business, neither a
meaningful difference between “fined” and “not fined” specialist insurers. From
our analysis, it seems that in 2000, nonlife insurers working also in the life
business have not exploited economies of scope arising by the selling of several
different financial products. It emerged also that in 2000 fined companies do not
show a structural higher efficiency than not fined specialist insurers. It appears
that the anti-competitive behaviour of fined companies did not led to an increase
of their technical efficiency: to improve the efficiency and the performance more
structural innovations are required, in line with an increase of competition and
transparency to offer customers cheaper and better services.
With the evidence presented in this chapter, we showed the importance of
having better measures of the extent and dispersion of efficiency and produc-
tivity of insurers. Such information may be helpful to regulators in evaluating
the impact of changing market structure, to insurers and other firms seeking to
acquire subsidiaries or establish alliances, and to securities markets and buyers
evaluating the prospects of particular insurers and national markets.
Chapter 7

THE ECONOMICS OF SCIENCE.

AGE, SCALE AND CONCENTRATION EFFECTS
IN A PUBLIC RESEARCH SYSTEM

7.1 Introduction
The notion of efficiency is highly problematic in the economics of science.
While policy makers and scientists are ready to accept that research activity
should be organised in such a way to avoid inefficiencies and waste of resources,
the exact definition of what accounts for efficiency is far from being accepted.
Several theoretical and methodological problems are still unsolved.
Any notion of efficiency relates a vector of inputs to a vector of outputs. Un-
fortunately, in scientific research all the elements of efficiency- inputs, outputs
and the functional relation between the two- are affected by different kind of
issues:
definitional problems concern the definition of inputs and outputs and the
identification of the unit of analysis;
measurement problems pertain to the methodologies for collecting inputs
and outputs data as well as comparative issues;
specification problems involve endogeneity, assumptions made and dy-
namic relations.
In the evaluation of productivity, the definition of what accounts for inputs or
outputs of scientific research is one of the most crucial point. From a substantive
perspective all factors can be considered both as input and as output. There are
no definitive answers to this problem. They have to be defined case by case, so
that any factor can be considered as input or as output, taking into account the
purpose of the analysis. The methodology we apply in this chapter takes the
definition of inputs and outputs as given.
A related problem is the identification of the unit of analysis of the scientific
research. While it is true that all researchers are members of an institute or
168 Age, scale and concentration effects in a public research system

department defined by discipline or thematic field, it has been convincingly

argued that the appropriate unit of analysis is the laboratory or team (Laredo and
Mustar, 2001). Researchers may be members of several projects, that cut across
administrative boundaries of institutes. In empirical works, due to the difficulty
in analysing laboratories, most studies focus on the institute or department or
the university (Ramsden, 1994; Johnston, 1994; Adams and Griliches, 2000).
In fact, data at the level of individual researchers are not available, unless case
studies on limited scale were put in place. Therefore, the research center or
institute as unit of analysis is a good compromise.
In addition to definitional problems, there are also measurement problems.
Although standardised international procedures exist for the definition and mea-
surement of research inputs (e.g. full time equivalent), very often scientific
institutions do not strictly follow these procedures and differ in the meaning
they attach to the various input activities. Consequently, much care should be
placed in making comparisons. Furthermore, the methodologies for collecting
inputs and outputs data have been developed largely independently from each
other.32
As far as the functional relation between inputs and outputs is concerned,
specification problems arise. One of the most fundamental problem is endo-
geneity: the level of inputs in terms of funds and number of researchers is a
function of past level of output, so that any specification that does not take
these effects into consideration is likely to produce misleading results. Another
important problem is that scientific production does not follow the assumptions
needed to adopt the production function approach. As a matter of fact, a lot of
published studies adopt a production function approach even though its con-
ceptual foundation in science is extremely weak (see Bonaccorsi and Daraio,
2004 for a discussion). In order to overcome these limits, a large literature has
applied the nonparametric efficiency analysis approach (mainly using DEA) to
assess the technical efficiency of academic research.33
Another problem connected to the specification is the dynamic relation be-
tween inputs and outputs. While in most production processes the time se-
quence that relates the use of productive resources to the outcome is fixed and
predictable, in science the outcome of research follows from inputs with a time
lag structure that is both unknown and variable over time.

32 See Luwel (2004) for a discussion and suggestions for a more integrated approach to construct input and
output data.
33 See among others Coelli (1996), Korhonen, Tainio and Wallenius (2001), Thursby and Kemp (2002).

Studies applying DEA to education include Bessent and Bessent (1980); Bessent, Bessent, Kennington and
Reagan (1982); Charnes, Cooper and Rhodes (1978); Färe, Grosskopf and Weber (1989), Grosskopf, Hayes,
Taylor and Weber (1999), Grosskopf and Moutray (2001). Rousseau and Rousseau (1997, 1998) apply DEA
to construct scientometrics indicators and assess research productivity across countries. See also Bonaccorsi
and Daraio (2004) for a selective review.
Introduction 169

Scientific production is a multi-input, multi-output relation, in which, differ-

ently from standard production activity, both inputs and outputs are qualitatively
heterogeneous and sometimes incommensurable, the relation is dynamic and
not deterministic and the output is lagged but with a non fixed structure.34
Ideally, at the level of inputs, one should include:
human resources (e.g. the number of researchers differentiated by age,
level of qualification or seniority; technical staff, administrative staff and
so on);
financial resources (e.g. governmental research funds, funds raised from
the market, and so on);
physical resources (e.g. physical capital, equipment, laboratories, libraries
and so on);
cumulated stock of knowledge (e.g. the number and quality of publications
in the past).
In practice, it is extremely difficult to collect data on all types of inputs. In most
cases very crude data on the number of researchers and on research funds are
the only available information.
At the level of outputs, most analyses work with count data (e.g. number
of publications), although it is clear that the quantity of papers does not have
a necessary relation with their quality (for instance measured by normalised
received citations) or importance.35 Moreover, it should be recognised that
the outputs of a scientific institute are not limited to publications but include
also teaching, training, patents, applied research for industry and other parties,
services for the public administration, consulting and the like (see Schmoch,
2004). For this reason, efficiency analyses limited to publication data are still
considered with scepticism. Even though bibliometrics methods are widely ac-
cepted in the evaluation of research productivity (Daniel and Fisch, 1990; Narin
and Hamilton, 1996; van Raan, 1993, 1997), they are viewed with suspicion
by some of those being evaluated (Collins, 1991). It is desirable not only that
these methods cover many different aspects of research outputs (Martin 1996)
but also that the evaluees have a place in helping to create appropriate methodol-
ogy by identifying the relevant categories of output (Lewison, 1998). Again, in

34 For a review of the econometric approaches to Science and Technology (S&T) systems see Bonaccorsi
and Daraio (2004).
35 On general bibliometric theory and methodology see Narin (1987), Narin, Olivastro and Stevens (1994),

Okubo (1997), Mullins, Snizek and Oehler (1988) and Moed, Glanzel and Schmoch (2004). The bibliometric
literature has discussed at length the characteristics of count data; see e.g. Garfield and Dorof (1992),
Holbrook (1992a, b), Kostoff (1994). Citation data are examined among others in Schubert, Glanzel and
Braun (1988), and in Schubert and Braun (1993, 1996). The contributions by Rosenberg (1991), May (1993),
Taubes (1993) and King (2004), among others, examine the quality of national scientific production.
170 Age, scale and concentration effects in a public research system

practice the collection of data on all these outputs is extremely difficult, unless
with field surveys on a limited scale. As a result, for large scale investigations,
data on the number of international publications are considered as acceptable.
Any meaningful measure of productivity therefore should be generated by a
model of multi-input multi-output production without a fixed functional specifi-
cation. Despite the problems recalled above, the idea that scientific production
must exhibit some relation between the resources employed and the output
produced is generally accepted. For practical and policy objectives simple
measures of the ratio of output to input are considered an indicator of scien-
tific productivity. As an example, the crude number of paper per researcher,
within relatively homogeneous fields, is considered as an acceptable indicator
of productivity.
Some related and relevant issues in the economics of science, from a policy
making perspective, empirically controversial are:

(a) the existence of economies of scale in scientific production, i.e. the pos-
itive effect of the concentration of resources over large (institutions or)
institutes on scientific productivity;

(b) the effects of the territorial agglomeration (concentration) of scientists on

scientific productivity. In some countries a policy of locating laborato-
ries and research institutes in the same territorial area has been actively
pursued, in search for economies of agglomeration;

(c) the exploration of the relation existing between age structure of researchers
and scientific productivity. If the effect of age on individual productivity
has been largely treated in literature (see for all Levin and Stephan, 1991),
there is few evidence on the effect of age structure at the level of research
institute.

Hence, this chapter aims at discussing theoretically size, age and concentra-
tion effects in science. It also provides empirical evidence analysing the case
of the Italian National Research Council (CNR). Founded in 1923, the CNR
(Consiglio Nazionale delle Ricerche) is the most important national research
institution in Italy, spanning many scientific and technological areas.36
The chapter is structured as follows. In the following of this section we dis-
cuss size, agglomeration and age effects in science. In Section 7.2 we illustrate
the data used in the application. The following sections report the results of
the applications of the econometric methodology described in Chapter 5 to the

36 Studies on the Italian CNR include Bonaccorsi and Daraio (2003a, b, 2005). On the efficiency of the
Italian university system see Bonaccorsi, Daraio and Simar (2006).
Introduction 171

institutes of the CNR, and provide empirical evidence on these controversial is-
sues. The chapter ends by deriving some policy implications from the described
empirical evidence.

Economies of scale, critical mass and concentration

In Chapter 6 we have already discussed on the existence of economies of
scale in the Italian insurance market. Here we discuss this topic as applied to
the scientific productivity of public research institutes. We recall that in man-
ufacturing production, economies of scale refer to the fact that an increase of
 percent in all factors of production determines an increase in output of more
than  percent. This notion, applied to science means that, as far as increasing
returns to scale are in place, research units should be of large size, in order to
optimise the use of their resources and increase productivity. The higher the
size of units, the higher scientific productivity. This notion is often invoked
to support policies of concentration of resources in larger institutes, forcing
small institutes to merge or disappear, or policies of merger and consolidation
of scientific institutions. The keyword for these policies is critical mass. Public
policies based on critical mass and large institutes induce levels of concentra-
tion of resources that go beyond the usual level. Concentration is a structural
property of institutional systems that allocate research funds in proportion to
publishing output. Since publication activity follows a strongly asymmetric
distribution, it is not surprising that research funds are not allocated on a uni-
form basis. As a consequence, a small number of institutions that follow a
consistent policy of hiring scientists with a strong publication record absorb a
large share of funds. Policies aimed at concentration do not simply follow the
structural asymmetry of distribution of publication activity, but aim to actively
improve productivity. In Italy, for example, the legislative reform of the Na-
tional Research Council (Reorganization Decree no. 19 of 1999) has induced a
profound change in the administrative structure. The number of institutes has
been reduced from 314 in 1999 to 108 in 2001. Many of the smallest institutes
were, in effect, the result of fragmentation processes, created around a few re-
searchers and crystallised over time. Given that the administrative burden is,
at least to a certain extent, a fixed cost associated to service indivisibility, the
existence of a minimum efficient scale for administrative costs is plausible. It
should be noted, however, that policy decision makers are often driven by a
more general notion that research activity itself, and not merely its adminis-
trative side, is subject to increasing returns to scale. This belief is based on
the idea that research, like manufacturing, is subject to (a) division of labour;
(b) indivisibility in the use of a minimum number of diverse competencies; (c)
utilisation of large physical infrastructure. These reasons are sufficient condi-
tions for the emergence of increasing returns to scale in several industries in
172 Age, scale and concentration effects in a public research system

the manufacturing sector (Scherer 1980; Milgrom and Roberts 1992; Martin
2002). However, the feasibility of these conditions for scientific research is not
guaranteed for several reasons.
In science, the knowledge stored in publications allows division of cognitive
labour to take place in different places and periods of time. Publication is one
of the most important mechanism for promoting division of cognitive labour.
This means that placing scientists within the same organisational boundaries
is neither a necessary nor a sufficient condition for benefiting from improved
division of labour. There may be a form of division of labour that requires
the establishment of formal collaboration and coordination of tasks between
scientists. Moreover, it is useful to distinguish among division of labour among
peers, and of scientists at various stages of careers, and of scientists and tech-
nicians or assistants. The former type takes the form of personal links, based
on mutual recognition and professional esteem. Only occasionally one can find
the entire web of personal peer relationships included within the boundaries of
a single organisation. A different type of division of labour takes place when
the pattern of personal relations is based on apprenticeship and scientific lead-
ership and requires long periods of joint work and supervision, normally (but
not necessarily) within the same institution. Because both types of division
of labour require personal in-depth supervision, the size of resulting units is
limited by the ability of research directors to monitor closely the work of their
research students and collaborators and to contribute to their training. In most
scientific fields this amounts to say that the maximum size is quite small, in the
order of units or one or two dozens. Summing up, it is unlikely that division of
labour per se is a source of increasing returns to scale at the level of institutes.
Indivisibility is another condition invoked for sustaining critical mass policy.
In many fields the scientific production requires the combination and coordi-
nation of many scientists from different areas, bringing competencies from
complementary fields. However indivisibility is more important at the level of
team or laboratory than at the level of institute or department. This is because
the minimum size of a team or laboratory may be extremely variable across
specific areas within the same fields. In general, this means that economies
of scale may be important up to a threshold level, then become irrelevant. If
the threshold level is quite small, the practical implication is that even small
institutes may be highly efficient, provided that their teams or labs meet the
minimum requirement.
Access to physical infrastructure is another argument commonly associated
to the call for critical mass and concentration of resources in large institutions.
However, it cannot be invoked as a general argument in favour of large institutes
as the research instrumentation required varies according to the field of research.
On this important issue the evidence is ambiguous and contradictory.
Introduction 173

Agglomeration economies
The notion of scientific districts, clusters, poles of excellence or science areas
has been prominent in national and regional science policy in the last twenty
years. The examples of Silicon Valley and Route 128 (Saxenian, 1996) and the
emergence of technopoles and regional clusters (Castells and Hall, 1994; Cooke
and Morgan, 1998) have catalysed the attention of analysts and policy makers
in all advanced countries. At a regional level the notion of cluster identifies
the co-presence and interaction of diverse subjects such as research and educa-
tional institutions, firms, innovative public administrations, financial services,
technology transfer and other intermediary organisations (Acs, 2000; Scott,
2001). At this level the emphasis is not on clustering of research activities per
se, but on clustering of complementary innovative activities in the same area.
This general notion, however, has also inspired policies of location of research
activities by some large public research institutions. In several countries large
public research institutions have pursued a policy of creating geographical con-
centrations of institutes in the same area. For example in Italy CNR promoted
the creation of Research Areas, large agglomerations of institutes in different
fields within the same physical infrastructure. In France most research institutes
at CNRS and INSERM are located in close areas. Behind these policies there
is the idea that proximity favours scientific productivity, insofar as it maximises
personal interaction, face-to-face communication, on-site demonstrations and
transmission of tacit knowledge, as well as it facilitates identification of com-
plementary competencies, unintentional exchange of ideas, café phenomena,
and other serendipitous effects. The focus of our discussion is therefore the no-
tion that concentrating research activities in the same area may bring benefits to
scientific productivity. Here we do not enter into a discussion on more general
policies for clustering and agglomeration of innovative activities.37
Underlying these policies there are some well grounded economic ideas. As
it sometimes happens, the original idea is an old one, but it was rediscovered and
enlarged more recently. The implicit economic analogy is with the concept of
external economies, or Marshallian agglomeration economies (Pyke, Becattini
and Sengenberger, 1986). Alfred Marshall observed that the concentration of
a large number of manufacturing firms in the same area (industrial district) is
not due to chance, but reflects the presence of local externalities in the form
of availability of specialised suppliers, highly trained workforce, sources of
innovative ideas. Costs of production are therefore lower in an agglomerated
area than outside it. More importantly, firms in a district enjoy a particular
industrial atmosphere and benefit from processes of collective invention. The
large literature on geography and trade (Krugman, 1991) and the geographical

37 For
more on this point and on the recent developments of the Economic Geography, see Clark, Feldman
and Gertler (2000).
174 Age, scale and concentration effects in a public research system

dimension of knowledge spillover effects (Jaffe, Trajtenberg and Henderson,

1993; Audretsch and Feldman, 1996; Zucker, Darby and Armstrong, 1998)
gave new emphasis to this idea and spurred this line of policy. Here the main
emphasis is on the fact that the diffusion of knowledge may take place via
codification and distance transmission, but in most cases requires also personal
acquaintance and face to face interaction. This is made easier and cheaper by
physical proximity.
However, in concluding her review on innovation, spillovers and agglomer-
ation, Feldman (2000, p.389) states: “Most importantly, we still have a limited
understanding of the way in which knowledge spillovers occur and benefit in-
novative activity. Marshall (1949: 152-153) tells us that knowledge ‘is in the
air’, and although we may cite Marshall, this answer is simply not very sat-
isfying. To date, the mechanisms of externalities and knowledge spillovers
have not yet been made explicit. Many researchers have tried to estimate the
geographic boundaries of knowledge spillovers. The consensus is that knowl-
edge spillovers are geographically bounded within a limited space over which
interaction and communication is facilitated, search intensity is increased, and
task coordination is enhanced. [...] In addition, there is also a literature that
documents the importance of social interaction, local networks and personal
communication in Knowledge transmission, but we do not know how social
interaction is initiated, how it evolves into a working relationships and how
economically useful knowledge is created”.
Hence, there are many reasons for a policy of agglomeration of research
activities in the same geographic area. At the same time the importance of
agglomeration is an inherently empirical matter and should be evaluated case
by case. This chapter gives a contribution to this debate by exploring the
existence of economies of agglomeration in conjunction with economies of
scale, analysing the research institutes of a public research system: the Italian
National Research Council (CNR).

Age effects
The existence of age effects in scientific production is one of the few con-
solidated stylised facts in the economics and sociology of science. The decline
of scientific productivity with age may depend on a variety of factors. On the
one hand, as time goes by the initial differences among scientists in individ-
ual productivity get larger. Most theories of scientific productivity postulate a
stochastic and cumulative mechanism (Simon, 1957) or a Matthew effect (Mer-
ton, 1968), whereby those that gain recognition initially in their careers receive
reward and resources, which will be used to carry out further research. If this
is true, initial differences in individual productivity will tend to be larger over
time. Allison and Stewart (1974) found that the Gini index for publications and
citations of scientists monotonically increases over time in a series of cohorts
Introduction 175

from the date of the PhD, with the exception of biologists. This evidence is
interpreted as strongly supporting the notion of reinforcement or positive feed-
backs38 . Another way of looking at the problem of age is to model productivity
as the outcome of a number of features that interact multiplicatively, rather
than additively. For example a model may assume that several elements or
mental factors play a role (e.g. technical ability, finding important problems,
and persistence). As it happens in any multiplicative model, the distribution
of productivity is more skewed than the distribution of any of its determinants.
As a result, a cohort of scientists starting with a given distribution will end up
with a more dispersed distribution and the variance will increase over time. On
the other hand, it is plausible that scientists work on research not only for the
sake of intrinsic pleasure of scientific puzzle solving, but also in the expecta-
tion of receiving future income. If this investment motivation is correct, it will
inevitably happen, as in any theory of human capital accumulation with finite
horizon, that the level of investment will decrease when scientist approach the
date of retirement. Models of human capital are central in the theory of life cy-
cle of scientists. This life cycle effect was found by Levin and Stephan (1991)
for most scientific areas with the exception of particle physics. The impact
of age at the level of research organisations is less clear, however. Within an
institute, for example, experienced scientists might compensate their individual
decline with a well organised activity of training of junior researchers, so that
productivity at the level of institute is not depressed. Being less creative at
the individual level, they might be still prolific in supporting young researchers
and identifying promising research avenues that they do not pursue personally.
Furthermore, aged scientists may have acquired capabilities in managing and
coordinating research teams and laboratories. More generally, little is known
on the pattern with which people of different age are mixed within research
institutes and the resulting impact on scientific productivity.
These problems are becoming critical in science policy given the alarming
evidence on the increasing average age of researchers in most European coun-
tries. For example, in Italy the proportion of professors and researchers in the
age class 24-44 was 60% in 1984 and only 29% in 2001. Those that entered the
academic system in the age class 24-34 were 19% of the total in 1984 and only
5% in 2001 (Avveduto, 2002). To face the problem of ageing of researchers,
there are suggestions that a massive effort should be made by hiring waves of
new researchers in a concentrated period of time, in order to reduce drastically
the average age. While by definition the problem of ageing worsens over time
in the absence of recruitment of many young researchers, it is not at all clear
what should be the time path of recruitment. This chapter analyses thoroughly
the effects of the age structure of researchers on scientific productivity.

38 For more details on positive feedbacks and research productivity in science, see David (1995).
176 Age, scale and concentration effects in a public research system

7.2 Data description

We use data reported in the official Report of 1998, which includes both input
data and output data and unofficially data integrated and used in previous studies
(Bonaccorsi and Daraio, 2003a,b; 2005) to which we refer for a discussion of
the relevance and limitations of these data. Input data include, for example,
research funds, funds from external sources or total costs while output data
include total number of publications and number of international publications.
The research areas considered in the analysis are listed in Table 7.1.

Table 7.1: Research areas.

CODE RESEARCH AREA N. OBS.

A10 Mathematics 8
A12 Law and Politics 5
A13 History, philosophy and philology 5
A5 Economics, sociology and statistics 7
MA1 Agriculture 24
MA2 Environment and Geology and mining 26
MA3 Biotechnologies and Medicine and biology 27
MA4 Chemistry 26
MA5 Physics 28
MA6 Engineering and Innovation and technology 31

Table 7.2 shows the variables in the dataset (all variables refer to CNR in-
stitutes). We follow the definition of variables described in the CNR Report.
Monetary variables are left in million of Italian lira (1 euro= 1936,27 lira).
In order to assess the existence of economies of scale related to agglomeration
effects in the scientific production of CNR institutes, we apply the econometric
methodology described in Chapter 5 of this work, considering the impact of a
bivariate external factor (Z) composed by a proxy of size of the institute and a
proxy of the concentration of the institutes in the same geographic area.
To account for the influence of proximity between research institutes we
constructed the Geographical Agglomeration Index (GAI) as follows. To each
institute we assigned one point for each other CNR institute located in the
same city that is not of the same research aggregation; and two points for each
other CNR institute located in the same city that is also of the same research
aggregation of the institute considered. Then we obtained a GAI that goes from
39 to 1, varying between 39 and 33 for the institutes located in Rome, from 23
to 20 for the institute located in Naples, from 16 to 14 for the institutes located
in Pisa and so on. An institute has a GAI of 1 if it is the only CNR institute in
its own town.
Data description 177

Table 7.2: Definition of variables.

VARIABLE DEFINITION
T P ERS Total number of personnel
RESF U N Total research funds
N RESF U N Research funds obtained from the state
M RESF U N Research funds obtained from the market
T COS Total costs
LABCOS Labour costs
T RES Total number of Researchers
T ECH Number of Technicians
ADM Number of Administrative Staff
ADT ECH Number of Adm. Staff and Technicians
T PUB Total number of publications
P IN T P U B Percent international publications
IN T P U B Number of International Publications
P U B P ERS Publications per capita
IP U P ERS International Publications per capita
P U B RES Publications per researcher
IP U RES International Publications per researcher
P M ARF U N Percent of funds raised from the market
P IN V Percent of Total costs allocated to investment
COP U B Cost per publication
COP U BIN T Cost per international publication
AV IM Average Impact factor
IN ST AG Institute age
(based on an estimate of the date of foundation)
GAI Geographical Agglomeration Index
T RES AG Average age of Researchers (all types)

In order to use this measure in our analysis which requires the use of con-
tinuous variables, we let the GAI become continuous by simply adding a small
number randomly chosen from the continuous uniform distribution on the in-
terval [-0.499,+0.499]. This transformation does not affect the GAI indicator
and is suitable for our procedure.
Ideally, the dimension of a research institute is measured by the space it
has, by the physical infrastructure (e.g. no. of computers), and mostly by
the people that work in it. However, we do not have data on the area and
physical infrastructure of institutes, and we use the number of researchers and
the number of technicians and administrative staff as inputs in our analysis. We
have obtained very similar results using as proxy of size the total number of
178 Age, scale and concentration effects in a public research system

people working in the institutes (T P ERS), using also other proxies such as
total costs (T COS) and labour costs (LABCOS) of institutes.

7.3 Scale and concentration effects

In our analysis we use an input oriented framework as our interest lies on
the allocation of resources given the level of output obtained by the research
institutes. The inputs and outputs used are presented in Table 7.3.

Table 7.3: Definition of Inputs and Outputs.

Variable Description
Input 1 No. of Researchers (T RES)
Input 2 No. of technicians and administrative staff (ADTECH)
Input 3 Normalised Research Funds (N RESF U N )
Output Normalised no. of international publications (N IN T P U B)
Ex. factor1 Geographical Agglomeration Index (GAI)
Ex. factor2 Total no. of Personnel (T P ERS)

As output we use the number of international publications of each institutes

divided by the mean number of publications of the research area the institute is
in. As pointed out in Schubert and Braun (1996), the comparative assessment
of scientometric indicators has to be based on normalized scientometric indi-
cators that first gauge them against a properly chosen reference standard, then
compare their relative standing. There are basically two types of approaches in
setting reference standards for cross-field normalization of scientometric indi-
cators. The first type is based on a prior classification of units into science field
categories of the required depth. In the second type for each unit to be assessed,
a specific standard is set on the basis of automatic algorithms or human exper-
tise. In this work we choose the first approach that is easier to comprehend and
accept, even if we loose in flexibility.39
Table 7.4 shows the differences of the average values of labour costs, total
costs, research funds and number of international publications by research area.
Hence, we analyse the stability of results using the variables RESF U N ,
T COS and LABCOS normalized, i.e., dividing by the mean of the research
area to which the institutes belong. We notice that we obtained very similar
results. In the following we report the results obtained using as inputs T RES,
ADT ECH, N RESF U N , and as output N IN T P U B. Z1 is GAI and Z2
is T P ERS.

39 For more details on normalization methods see Schubert and Braun (1996).
Scale and concentration effects 179

Table 7.4: Average values of Inputs and Outputs by research area.

RES. LABCOS T COS RESF U N IN T P U B

AREA
A10 763.25 1268.50 505.25 31.71
A12 1391.60 1954.20 562.60 8.83
A13 1161.00 1540.11 379.11 13.21
A5 900.71 1380.57 479.86 11.54
MA1 1318.25 1885.00 566.75 21.15
MA2 2036.08 3176.35 1140.27 25.27
MA3 2323.78 3926.26 1602.48 53.00
MA4 2044.88 2855.08 810.19 44.29
MA5 3015.18 4274.64 1259.46 73.77
MA6 2232.00 7053.75 1302.16 33.04

A descriptive statistics on inputs, outputs and external factors is offered in

Table 7.5 where mean, standard deviation, minimum, maximum and interquar-
tile range are reported. Very often the impact factor has been used as an output.
However, we prefer do not use the average impact factor as further output for
a twofold reason. Firstly, we do not have the possibility to check if the self
reported average are correct and we do not have its value for all the institutes
in the sample. Secondly, journal impact factors as a measure of quality have
many general limitations (see Moed and van Leeuwen 1996, Seglen 1997).

Table 7.5: CNR institutes - Descriptive statistics on inputs, outputs

and external factors considered in the analysis (169 obs).

VARIABLE Mean Std Min Max Iqr

T RES 13.077 9.057 1.000 45.000 11.300
ADT ECH 13.751 12.726 1.000 69.000 11.000
IN T P U B N 1.000 0.601 0.031 3.099 0.800
P M ARF U N 13.862 10.315 1.000 53.000 12.300
T RES AG 44.363 5.149 33.333 53.455 8.600
IN ST AG 28.107 9.123 3.000 43.000 9.000
GAI 14.880 11.090 1.000 39.000 14.000
LABCOS N 1.053 0.752 0.059 4.090 0.900
T COS N 0.939 0.632 0.050 3.555 1.000
RESF U N N 0.983 0.710 0.028 5.628 0.900
RESF U N 984.083 864.974 45.000 7329.000 718.000
LABCOS 2127.367 1740.406 96.000 9128.000 1849.800
T COS 3111.450 2483.750 197.000 16457.000 2529.700
180 Age, scale and concentration effects in a public research system

Density of Z

3 10−3
1.5

0.5

0
60
40 60
20 40
20
0 0
−20 −20
values of Z2 values of Z1

Contour plot for the density of Z

40

35

30
values of Z2

25

20

15

10

−5
−5 0 5 10 15 20 25 30 35 40
values of Z1

Figure 7.1. Estimation of the density of Z (Z1 = GAI,Z2 = T P ERS) (top panel) and its
contour-plot (bottom panel) CNR institutes (169 obs).
In order to have a robust measure of the global impact of dimension and
geographical concentration on the performance of CNR institutes we choose a
level of robustness at 10% and we obtain a level of α = 0.985 and a level of
m = 50. The number of nearest neighbors chosen by our datadriven approach
is k − N N = 33. Figure 7.1 illustrates the quality of the estimation of the
density of Z carried out by our bandwidth selection procedure, as well as Z’s
contour plot. The contour plot shows that the two external factors, size (Z2 =
T P ERS) and geographical concentration (Z1 = GAI) are not correlated.
The main results of our investigation are reported in Figures 7.2 to 7.5.
As we can see (Figure 7.2), the ratio of full frontier efficiency estimates Qz , is
Age effects on CNR scientific productivity 181

influenced by some outliers rather than the robust partial efficiency ratios Qzα
and Qzm . In the full frontier case, in fact, it appears an inverse U-shape pattern
determined by the influence of extreme values, that is not confirmed by the
robust partial efficiency estimates. This illustration shows the usefulness of
robust measures of efficiency.
As we have at length explained in Chapter 5 by means of the simulated
examples, in order to detect the global effect of the external factors on the
performance of the analysed units, it is of interest the analysis of the behavior
of the surface of the ratios of conditional and unconditional efficiency measures:
Qz , Qzα and Qzm , on Z, as well as the nonparametric smoothed regression of
the ratios Qz (Qzα and Qzm ) on Z1 at the Z2 ’s quartiles, and viceversa, on Z2 at
the Z1 ’s quartiles. In particular, these plots are also able to shed lights on the
interactions between external factors.
We recall that in an input oriented framework, as is the case here, an increas-
ing nonparametric regression line indicates an unfavorable external factor, a
decreasing nonparametric regression line points to a favorable external factor,
while a straight nonparametric regression line denotes no effect of the external
factor.
Figure 7.2 has shown the usefulness of applying robust (partial) estimators
of efficiency to check if the impact of external factors on full frontier estimators
is influenced by the presence of outliers, as it is the case here. Therefore, the
results reported in Figure 7.3 are not very reliable. Another striking result is
evident from Figures 7.4 and 7.5, top panels. We see that there is no influence
of the geographical concentration on the performance of CNR institutes till a
GAI (Z1 ) of 25. In this region of the plots there is the 82.84% of CNR institutes.
Only 29 institutes out of 169 have a GAI greater than 26, and are in the region
of GAI in which there is a slightly decreasing nonparametric regression line,
meaning that there could be a positive influence of geographical concentration
on their performance. Even more interesting is the inspection of the bottom
panels of Figures 7.4 and 7.5. It appears that there is an increasing trend
of the smoothed nonparametric regression line for each number of personnel
(T P ERS). Interestingly enough, size negatively affects the performance of
all CNR institutes.
In this chapter we confirm the results found in a previous work on Italian and
French research institutes (see Bonaccorsi and Daraio, 2005). As it comes out
from our analysis, scientific productivity seems not favoured by the concentra-
tion of resources into larger institutes geographically agglomerated.

7.4 Age effects on CNR scientific productivity

With respect to age, our interest is in evaluating the impact of age distribu-
tion on productivity. This effect is different from the one assumed in the life
182 Age, scale and concentration effects in a public research system

1.7

1.6

1.5
Qz

1.4

1.3

150
100 50
40
30
50 20
Z2 10
0 0
Z1

1.8

1.6
Qzα

1.4

1.2

1
150
100 50
40
30
50 20
Z2 10
0 0
Z1

1.8

1.6
Qzm

1.4

1.2

1
150
100 50
40
30
50 20
Z2 10
0 0
Z1

Figure 7.2. Scale and concentration effects on CNR institutes (169 obs).Surface of Qz on Z1
and Z2 (top panel), surface of Qzα on Z1 and Z2 (middle panel), and surface of Qzm on Z1 and
Z2 (bottom panel). Z1 = GAI, Z2 = T P ERS. α = 0.985 and m = 50.
Age effects on CNR scientific productivity 183

1.8
Qz

1.6

1.4

0 5 10 15 20 25 30 35 40 45
Z1
1.6

1.5
Qz

1.4

1.3

0 20 40 60 80 100 120 140

Z2

Figure 7.3. Scale and concentration effects on CNR institutes (169 obs). Smoothed nonpara-
metric regression of Qz on Z1 = GAI for Z2 ’s quartiles(top panel) and on Z2 = T P ERS
for Z1 ’s quartiles (bottom panel); dashed line = first quartile, solid line = median and dashdot
line = third quartile.

1.8

1.6
Qzα

1.4

1.2

1
0 5 10 15 20 25 30 35 40 45
Z1
1.8

1.6
Qzα

1.4

1.2

1
0 20 40 60 80 100 120 140
Z2

Figure 7.4. Scale and concentration effects on CNR institutes (169 obs). Smoothed nonpara-
metric regression of Qzα on Z1 = GAI for Z2 ’s quartiles(top panel) and on Z2 = T P ERS
for Z1 ’s quartiles (bottom panel); dashed line = first quartile, solid line = median and dashdot
line = third quartile.
184 Age, scale and concentration effects in a public research system

1.5

1.4

1.3
Qzm

1.2

1.1

1
0 5 10 15 20 25 30 35 40 45
Z1
1.8

1.6
Qzm

1.4

1.2

1
0 20 40 60 80 100 120 140
Z2

Figure 7.5. Scale and concentration effects on CNR institutes (169 obs). Smoothed nonpara-
metric regression of Qzm on Z1 = GAI for Z2 ’s quartiles(top panel) and on Z2 = T P ERS
for Z1 ’s quartiles (bottom panel); dashed line = first quartile, solid line = median and dashdot
line = third quartile.

cycle hypothesis. We do not have longitudinal data on scientists’ production

and do not assume that younger scientists are monotonically more productive
than older ones. Rather, we assume that institutes with higher average age have
a lower turnover. They are presumably less able to attract young scientists and
are more likely to be isolated from the latest developments in science. As a
first approximation we consider the average age of scientists within institutes
(T RES AG). Institutes with higher average age of scientists are old institutes
in which the entry of young scientists has not compensated the effect of ageing
of incumbents.
Here again, we apply the conditional measures of efficiency to shed empirical
light on this interesting aspect. Figure 7.6 shows that there is a negative effect of
the average age of researchers on productivity indicators (increasing regression
line), although this effect seems to be negligible after the age of 46 years.40
The general negative effect of average age on productivity does not neces-
sarily mean that older scientists are less productive in absolute terms. Rather,

40 Here we set a level of robustness at 10% obtaining a value of α = 0.98 and m = 50. The number of

K − N N provided by our data driven procedure is 48.

Age effects on CNR scientific productivity 185
Effect of Z on Full frontier
7

4
Qz

34 36 38 40 42 44 46 48 50 52
values of Z
Effect of Z on Order−α frontier
5

4.5

3.5

3
Qzα

2.5

1.5

0.5

34 36 38 40 42 44 46 48 50 52
values of Z
Effect of Z on Order−m frontier
5

4.5

3.5

3
Qzm

2.5

1.5

0.5

34 36 38 40 42 44 46 48 50 52
values of Z

Figure 7.6. Age effects on CNR institutes. FDH (top panel), order-α (middle panel) and
order-m (bottom panel) estimates (169 obs). Z = T RES AG.
186 Age, scale and concentration effects in a public research system

institute with higher average age of researchers might have a lower proportion
of younger scientists, that they wouldn’t or couldn’t attract. The average age
of an institute reflects its attractiveness and scientific vitality. In fact, the aver-
age age of existing personnel is lowered each time a young researcher enters
the institute. The higher the scientific prestige of the institute, the resources
available for job positions and the prospects for career, the higher the number
of young candidates wishing to enter. The average age may be considered a
summary statistics for turnover and attractiveness. The policy implication of
this finding is then straightforward: to increase their scientific productivity,
research institutes have to attract young and talented researchers.

7.5 Robust parametric approximation of multioutput

distance function
In this section we apply the recently introduced two stage semi-parametric
approach by Florens and Simar (2005) and Daouia, Florens and Simar (2005)
and its multivariate extension (described in Chapter 4) to estimate parametrically
a multioutput distance function on the Italian CNR research institutes.
It allows to overcome the main drawbacks of the traditional fitting of para-
metric models of Shephard (output/input) distance function (e.g., Grosskopf,
Hayes, Taylor and Weber, 1997), solving additional consistency issues (Coelli,
2000).
The basic idea of this new approach is as follows. In a first stage we estimate
the efficient frontier using nonparametric and robust nonparametric methods
(like FDH, order-m or order-α frontier estimators). In a second stage, we fit,
by standard OLS, the appropriate parametric model on the obtained nonpara-
metric/or robust nonparametric frontier.
This approach provide much more sensible estimators of the parametric fron-
tier model and allows for some noise by tuning the robustness parameters (α
and m). Florens-Simar (2005) and Daouia-Florens-Simar (2005) analyse also
the statistical properties of the two-stage semiparametric
√ estimators, and find
consistency for full frontier (FDH) estimators and n and asymptotic normality
for partial frontiers parameters (order-m and order-α).
In the application which follows we have chosen a flexible functional form
for the output distance function of the CNR institutes, imposing relatively fewer
a priori restrictions on the structure of production. One such flexible form is
the translog form (Christensen, Jorgenson and Lau, 1973), which is a second
order-Taylor expansion, usually around the mean, of a generic function, where
all variables appear in logarithms.
Robust parametric approximation of multioutput\\ distance function 187

As we presented in Section 4.7.2, the translog model for ln δo (x, y) is the

following:
 
1 ln x
ln δo (x, y) ≈ ϕ(x, y; θ) = α0 + α
ln x + β
ln y + [ln x
ln y
] Γ .
2 ln y

where Γ = Γ
is symmetric. Homogeneity of degree one in y imposes p + q + 1

constraints:

β
iq = 1, 1 constraint
Γ12 iq = 0, p constraints
Γ22 iq = 0, q constraints

 
Γ11 Γ12
where Γ = .
Γ21 Γ22

In the previous section our aim was to analyse the impact of scale and ag-
glomeration, and then of age, on how resources are allocated to CNR institutes
(on how their inputs are managed). Therefore, we adopted an input oriented
framework. Now we consider the resources allocated to the research institutes
as given (as they are not very high compared to other European countries) and we
focus on how much the CNR, as a whole, could improve its efficiency in the pro-
duction of international publications per 100 researchers (IN T P U B100res)
and increasing the services done (including external contracts and other ser-
vices approximated by the share of funds coming from external sources -
P M ARF U N ) given the level of cost per 100 employees (COST 100emp)
and the research funds per 100 researchers (RESF U N 100res) owned by its
research institutes. Hence, we adopt an output oriented framework.
The inputs and outputs used in this section are described in Table 7.6. Due
to the heterogeneity in the structure of costs and in the publication practices
across scientific disciplines, we have normalized all variables dividing by the
mean of the scientific areas.
Table 7.6: Definition of inputs and outputs used for the translog
estimation of CNR output distance function.

VARIABLE DEFINITION
Inputs
COST 100emp Labour cost per 100 employees
(thousands of Euros)
RESF U N 100res Research funds per 100 researchers
(thousands of Euros)
188 Age, scale and concentration effects in a public research system

Table 7.7. Output distance function parameter estimates for CNR research institutes.

Param Variable θF DH θα θm θCOLS

α0 Intercept -0.6170 -0.4717 -0.4971 -1.4013
α1 ln(x1 ) -0.3186 -0.1176 -0.1803 0.4965
α2 ln(x2 ) -0.4687 -0.3798 -0.3975 -0.6711
β1 ln(y1 ) 0.4238 0.4870 0.4622 0.4722
β2 ln(y2 ) 0.5762 0.5130 0.5378 0.5278
2
Γ11.11 ln(x1 ) 0.0957 0.1284 0.1129 0.2122
Γ11.12 ln(x1 )ln(x2 ) 0.5622 0.4260 0.5172 0.1088
Γ11.22 ln(x2 )2 -0.0467 -0.0321 -0.0413 -0.1216
Γ12.11 ln(x1 )ln(y1 ) -0.0553 0.0264 -0.0279 -0.2229
Γ12.12 ln(x1 )ln(y2 ) 0.0553 -0.0264 0.0279 0.2229
Γ12.21 ln(x2 )ln(y1 ) 0.0655 0.0302 0.0551 0.0449
Γ12.22 ln(x2 )ln(y2 ) -0.0655 -0.0302 -0.0551 -0.0449
Γ22.11 ln(y1 )2 0.1053 0.1041 0.1079 0.1077
Γ22.12 ln(y1 )ln(y2 ) -0.1053 -0.1041 -0.1079 -0.1077
Γ22.22 ln(y2 )2 0.1053 0.1041 0.1079 0.1077

Outputs
IN T P U B100res Number of international publications
per 100 researchers
P M ARF U N Percentage of funds
raised from the market

As usually done in empirical works (see e.g. Coelli and Perelman 1996,
1999, 2000; Perelman and Santin, 2005; see also Färe and Primont, 1996), we
have mean-corrected all variables prior to estimation, i.e., each output and input
variable has been divided by its geometric mean. By doing that, the first order
coefficients may be interpreted as distance elasticities evaluated at the sample
mean.
The empirical results for the estimated model are presented in Table 7.7 in
which the parameters are estimated fitting a translog function on the FDH fron-
tier (third column), α = 0.975 and m = 35 frontiers (respectively fourth and
fifth columns), the latter robust at around 10%, and finally the COLS standard
approach.
Robust parametric approximation of multioutput\\ distance function 189

Table 7.8. Bootstrapped confidence intervals for order-α frontier translog approximation.
α = 0.975.

Param Variable θα Norm Norm BB BB Mean Std

α= low up low up boot boot
0.975 distrib distrib
α0 Intercept -0.4717 -0.6207 -0.3226 -0.6706 -0.3785 0.0348 0.0760
α1 ln(x1 ) -0.1176 -0.2868 0.0517 -0.2776 0.0732 -0.0224 0.0864
α2 ln(x2 ) -0.3798 -0.5405 -0.2190 -0.4581 -0.1397 -0.0551 0.0820
β1 ln(y1 ) 0.4870 0.3816 0.5924 0.4039 0.5978 -0.0024 0.0538
β2 ln(y2 ) 0.5130 0.4076 0.6184 0.4022 0.5961 0.0024 0.0538
Γ11.11 ln(x1 )2 0.1284 -0.0569 0.3137 -0.1320 0.2923 0.0096 0.0945
Γ11.12 ln(x1 )ln(x2 ) 0.4260 0.1246 0.7275 0.0089 0.6093 0.0945 0.1538
Γ11.22 ln(x2 )2 -0.0321 -0.1284 0.0643 -0.1078 0.0886 -0.0258 0.0492
Γ12.11 ln(x1 )ln(y1 ) 0.0264 -0.1421 0.1948 -0.0545 0.2864 -0.0634 0.0859
Γ12.12 ln(x1 )ln(y2 ) -0.0264 -0.1948 0.1421 -0.2864 0.0545 0.0634 0.0859
Γ12.21 ln(x2 )ln(y1 ) 0.0302 -0.0544 0.1149 -0.0856 0.0843 0.0206 0.0432
Γ12.22 ln(x2 )ln(y2 ) -0.0302 -0.1149 0.0544 -0.0843 0.0856 -0.0206 0.0432
2
Γ22.11 ln(y1 ) 0.1041 0.0673 0.1410 0.0599 0.1364 0.0061 0.0188
Γ22.12 ln(y1 )ln(y2 ) -0.1041 -0.1410 -0.0673 -0.1364 -0.0599 -0.0061 0.0188
2
Γ22.22 ln(y2 ) 0.1041 0.0673 0.1410 0.0599 0.1364 0.0061 0.0188

Tables 7.8 and 7.9 display the results of our bootstrap exercise, done follow-
ing the procedure described in Florens and Simar (2005), to build confidence
intervals on the estimated parameters (see Chapter 4). Here we set B = 1000
bootstrap loops. The covariance matrices used to estimate the normal approxi-
mation confidence interval are not reproduced here to save space.
The Scale Elasticity (SE), evaluated at the sample mean, is given by:


p
SE = − ∂ ln δo (x, y) /∂ ln xk .
k=1
It is the negative of the sum of the input elasticities. Therefore, increasing
(decreasing) scale economies are indicated by a value of SE greater (less) than
one. The scale elasticity, at the approximation point, is equal to the following
values, according to the estimator used in the first stage.

SEF DH = 0.3186 + 0.4687 = 0.7873,

190 Age, scale and concentration effects in a public research system

Table 7.9. Bootstrapped confidence intervals for order-m frontier translog approximation.
m = 35.

Param Variable θm Norm Norm BB BB Mean Std

m= low up low up boot boot
0.35 distrib distrib
α0 Intercept -0.4971 -0.6370 -0.3573 -0.7159 -0.4440 0.0526 0.0714
α1 ln(x1 ) -0.1803 -0.3019 -0.0587 -0.3277 -0.0847 0.0175 0.0620
α2 ln(x2 ) -0.3975 -0.5347 -0.2603 -0.4458 -0.1823 -0.0538 0.0700
β1 ln(y1 ) 0.4622 0.3663 0.5580 0.3600 0.5254 0.0133 0.0489
β2 ln(y2 ) 0.5378 0.4420 0.6337 0.4746 0.6400 -0.0133 0.0489
Γ11.11 ln(x1 )2 0.1129 -0.0159 0.2417 0.0293 0.3190 -0.0151 0.0657
Γ11.12 ln(x1 )ln(x2 ) 0.5172 0.2959 0.7386 0.2474 0.7073 0.0231 0.1129
Γ11.22 ln(x2 )2 -0.0413 -0.1176 0.0349 -0.0983 0.0586 -0.0190 0.0389
Γ12.11 ln(x1 )ln(y1 ) -0.0279 -0.1370 0.0812 -0.1399 0.0898 -0.0131 0.0557
Γ12.12 ln(x1 )ln(y2 ) 0.0279 -0.0812 0.1370 -0.0898 0.1399 0.0131 0.0557
Γ12.21 ln(x2 )ln(y1 ) 0.0551 -0.0115 0.1216 -0.0171 0.1242 -0.0025 0.0340
Γ12.22 ln(x2 )ln(y2 ) -0.0551 -0.1216 0.0115 -0.1242 0.0171 0.0025 0.0340
2
Γ22.11 ln(y1 ) 0.1079 0.0825 0.1333 0.0756 0.1298 0.0020 0.0130
Γ22.12 ln(y1 )ln(y2 ) -0.1079 -0.1333 -0.0825 -0.1298 -0.0756 -0.0020 0.0130
2
Γ22.22 ln(y2 ) 0.1079 0.0825 0.1333 0.0756 0.1298 0.0020 0.0130

SEF DH ∈ CI normapprox = [0.6112, 0.9633],

SEF DH ∈ CI basicboot = [0.6090, 0.9739];
SEα=0.975 = 0.4974;
SEα=0.975 ∈ CI normapprox
= [0.2920, 0.7027],
SEα=0.975 ∈ CI basicboot = [0.1966, 0.5990];
SEm=35 = 0.5778,
SEm=35 ∈ CI normapprox
= [0.4046, 0.7510],
SEm=35 ∈ CI basicboot = [0.3561, 0.6948].

As we explained in Sections 4.6 and 4.7 the COLS estimator (here we obtain
a SE of 0.1746) is really a bad choice; the FDH estimator for the first stage
may be influenced by outliers, hence it is always useful to compare the previous
results with those of partial robust estimators (order-m or order-α).
Conclusions 191

We observe that all results are of the same order. These results show that in
all cases the SE is less than one, indicating the presence of decreasing returns
to scale at the mean of the Italian CNR institutes. Hence, it appears that the
result of decreasing returns to scale for the CNR as a whole seems quite stable
for robust and full frontiers estimates.
The inspection of Tables 7.8 and 7.9 show other interesting evidence:

there is a positive and significant interaction between the inputs (COST

100emp and RESF U N 100res), see Γ11.12 and its related confidence
interval.

there is a significant trade off between the research activity (measured

by the IN T P U B100RES) and the external services offered to the mar-
ket (as measured by P M ARF U N ), see the estimations as well as the
confidence interval of Γ22.12 .

Nevertheless, these are just a rough approximation of the considered trade-

offs as we have not analytically computed elasticities of complementarity and
substitution.41 This computation can be easily handled by applying the stan-
dard microeconomic toolbox of partial derivatives and the like (see e.g. Mas-
Colell, Whinston and Green, 1995; or Varian, 1992) to this distance function
framework. It is quite interesting, from an applied perspective, to measure the
elasticity of the distance function with respect to inputs and outputs; the rate
of substitution between outputs42 , the elasticity of each output with respect to
each input, and finally the elasticity between the outputs themselves.
Our main aim in this chapter was to illustrate the powerful application of the
new two-stage approach proposed by Florens and Simar (2005), generalised
to the multiple output case in Section 4.7 as a better alternative (in terms of
statistical inference) than the traditional parametric approaches followed in the
literature.

7.6 Conclusions
In the following we report the main findings of the analysis carried out on
the Italian CNR research institutes.
A striking result of our analysis is that size negatively affects the performance
of all CNR institutes. We graphically illustrated that the majority of Italian
research institutes operates with decreasing returns to scale.

41 For a comparative detailed review of the elasticities of complementarity and substitution see Cantarelli
(2005). On elasticities of substitution and complementarity see also Bertoletti (2005).
42 See Grosskopf, Margaritis and Valdmanis (1995) for the computation of marginal rates of transformation

and Morishima elasticities of substitution among the units of a public system, within a distance function
framework.
192 Age, scale and concentration effects in a public research system

We did not find a strong support for agglomeration effects. The argument
that scientific productivity is favoured by the geographical agglomeration of
institutes in the same area did not receive empirical support from our data.
Based on detailed evidence at the micro level on research institutes we
showed that scientific productivity declines with the average age of researchers
of the institute. Nevertheless, the key problem is not the declining individual
productivity, but rather the fact that as time goes on, it becomes increasingly
difficult to create the research climate within scientific institutions that attracts
young and talented scientists. The turnover of scientific personnel must be kept
high on a permanent basis.
The result on decreasing returns to scale was also confirmed by the robust
estimation of the scale elasticity for the average of CNR institutes (around
0.50). By applying the new two step procedure introduced by Florens and
Simar (2005), Daouia, Florens and Simar (2005) and extended in Section 4.7,
we were able to estimate the confidence intervals for the average scale elastic-
ity of CNR institutes both for order-α and order-m robust estimators. Finally,
by doing so, we showed that the parametric approximation of robust and non-
parametric frontiers is also feasible in a multi output framework, by estimating
robust parameters of a multi output Translog distance function which have better
properties than the traditional parametric estimates.
The calculation of elasticities of substitution and of marginal rate of trans-
formation and the like is straightforward and is left to other applications.
Chapter 8

MUTUAL FUNDS PERFORMANCE:

EXPLORING THE EFFECTS OF MANAGER
TENURE, FUND AGE, AND THEIR INTERACTION

8.1 Introduction
The development of personal finance and the recent movements of retirement
planning have renewed the interest on wealth allocation across asset categories
and detailed investments. Consequently, mutual fund investment companies
have become an increasingly popular way (channel) for capital appreciation
and income generation.
However, the identification of superior performing funds remains a contro-
versial topic due to the volatile nature of individual fund performance and the
methodological problems that compromise the findings of empirical studies.
There is a growing literature on the evaluation of the performance of mutual
funds which deals both on its methodological aspects and on its empirical facets
(features).43
Recently, several works have applied efficiency and productivity techniques
for evaluating the performance of mutual funds. Studies which apply the para-
metric approach include, for instance: Annaert, van den Broeck and Vennet
(1999), Briec and Lesourd (2000). Among the applications of the nonpara-
metric efficiency analysis approach there are: Murthi, Choi and Desai (1997);
Morey and Morey (1999); Sengupta (2000); Basso and Funari (2001); Wilkens
and Zhu (2001); Daraio and Simar (2004). Indeed the estimation of efficient
boundaries arises in portfolio management, as well as in the production frame-
work. In fact, in Capital Assets Pricing Models (CAPM, Markowitz, 1952,
1959) the goal is to study the performance of investment portfolios. Risk
(volatility or variance) and average return on a portfolio act like inputs and

43 Several
surveys are available. See e.g. Shukla and Trzcinca (1992), Ippolito (1993) Grinblatt and Titman
(1995), Cesari and Panetta (2002).
194 Exploring the effects of manager tenure, fund age and their interaction

outputs in production models. In these Models, the boundary of the attainable

set of portfolios gives a benchmark relative to which the efficiency of a portfo-
lio can be measured. Sengupta (1991) and Sengupta and Park (1993) analyse,
from a methodological perspective, the connections between the nonparametric
estimation of frontiers and CAPM. For a theoretical analysis of nonparametric
efficiency measurement approaches for portfolio selection see Briec, Kerstens
and Lesourd, 2004.
Within the nonparametric approach in efficiency analysis, Daraio and Simar
(2004) propose a robust nonparametric approach to evaluate and explain the per-
formance of mutual funds. They show how robust nonparametric (order−m)
frontiers could be useful to address some issues in portfolio performance eval-
uation. This robust approach adds some new advantages to the traditional
advantages of the nonparametric approach (namely: the absence of specifica-
tion of the functional form for the input-output relationship; the evaluation of
performance without the specification of a theoretical model as benchmark; the
computation of an efficiency index for each mutual fund; no need to assume
the normality of return distribution).
As shown above in Chapter 5, the first advantage is, of course, the robustness:
as the robust indicators are based on estimators that do not envelop all funds,
they are more robust to outliers and extreme values which can strongly influence
the nonparametric estimation of portfolio efficiency. The level of robustness
can be set by means of the tuning parameters m or α (if order-α frontiers are
used). The second advantage is the absence of the curse of dimensionality, that
is typical of nonparametric estimators; and the third one is the possibility of
comparing samples with different size, in an indirect way, avoiding the sample
size bias, of which nonparametric indicators (DEA/FDH) suffer.
Another benefit consists in the possibility of extending the analysis, consid-
ering the conditional influence of environmental variables Z in a robust way
as well as decomposing the performance of a firm in an indicator of the in-
ternal or managerial efficiency, an externality index and an individual index
which measures the firm intensity in catching the opportunities or threats by the
environment.
In this chapter we develop further the empirical analysis done in Daraio and
Simar (2004) on US mutual funds. We apply the comprehensive probabilistic
approach described in Chapter 5, and the newly introduced probabilistic effi-

ciency measures, α(x, y) and α z (x, y), providing new empirical evidence on
several interesting issues in the literature of mutual funds performance evalua-
tion.
We base our analysis on a cross-section dataset of US Aggressive Growth
funds (annual data for the period June 2001-May 2002). In Daraio (2003) an
application of the bootstrap on these data is proposed as well as an analysis of
the distribution of efficiency scores. Daraio and Simar (2004) have explored
Introduction 195

economies of scale and the influence of market risks on these funds and com-
pared with other US mutual funds categories by objective.
In this chapter we focus our empirical investigation on the impact of some new
management variables on the performance of mutual funds, namely manager
tenure and fund age (age from inception date) and their interactions on fund
performance.
In particular manager tenure is a measure of the manager’s survivorship at
the job (Golec, 1996). Long tenure implies that the management company
finds the manager’s ability and performance satisfactory but may also indicate
that the manager has few better opportunities because of specialized skills or
a modest performance record. Age of fund provides a measure of the fund’s
longevity or ability to survive in a highly competitive environment. It is simply
the number of months that a funds has been in operation.
According to human capital theory, managers with greater human capital
(intelligence and so on) should lead to better performance and hence should be
paid with an higher compensation. Moreover, performance, risks and fees of
mutual funds are all interrelated; consistent with several agency models (see
Golec, 1996 and the references cited there) a manager portfolio risk choices
will partly depend upon his/her risk taking preferences because the volatility of
a manager’s pay is influenced by the portfolio’s performance.
There is a rich literature on the relation between fund manager tenure and
mutual fund performance.
For years, economists have debated whether it is possible for mutual fund
managers to “beat the market”, either through superior stock selection abilities,
or by correctly predicting the timing of overall market advances and declines.
Chevalier and Ellison (1999a) examine whether mutual fund performance is
related to characteristics of fund managers that may indicate ability, knowl-
edge, or effort. During the time period they study, there is a strong correlation
between fund returns and a manager’s age, the average SAT score of his or
her undergraduate school, and whether he or she holds an MBA. For an inter-
esting analysis of the labour market for mutual fund managers and managers’
responses to the implicit incentives created by their career concerns see Cheva-
lier and Ellison(1999b) which find also that managerial turnover is sensitive to
a fund’s recent performance.
Nevertheless, the evidence is inconclusive regarding manager tenure and
performance.
On the one hand, in their study, Lemak and Satish (1996) found that longer-
term managers have a tendency to outperform shorter-term fund managers and
that longer-term fund managers assemble less risky portfolios. Golec (1996)
illustrated that manager tenure is the most significant predictor of performance
and found that longer-term managers, with more than seven years tenure, have
better risk-adjusted performance. He also showed that performance are posi-
196 Exploring the effects of manager tenure, fund age and their interaction

tively related to higher management fees and turnover ratios. Khorana (1996)
added empirical evidence on the underperformance of fund managers who are
replaced or terminated. Furthermore, he showed also that departing managers
exhibit higher portfolio turnover rates and higher expenses relative to non-
replaced managers.
On the other hand, Porter and Trifts (1998) found that managers with ten-year
track records do not perform better than those with shorter track records. On
the same line, Detzel and Weigand (1998) and Fortin, Michelson and Wagner
(1999) showed that manager tenure is not related to performance.
Summing up, whether mutual fund managers produce greater returns is con-
troversial because most studies’ funds, sample periods, and the methodological
assumptions of the adopted performance measures are not comparable.
We add new empirical evidence on this very interesting and contentious issue
using our nonparametric and robust approach. In this chapter, we reduce the
extent of fund changes and the survivorship bias which affect empirical studies
focusing on long time periods, providing a cross-sectional analysis. The chapter
unfolds as follows. In the next section we make a description of the data. In
the following sections we report the results of our empirical investigation and
then we conclude.

8.2 Data description

Our original data consist of the US Mutual Funds universe, collected by
the reputed Morningstar and updated at 05-31-2002. Among this universe we
selected the Aggressive-Growth (AG) category of Mutual Funds.
According to the description given by Morningstar, Aggressive Growth(AG)
are funds that seek rapid growth of capital and that may invest in emerging
market growth companies without specifying a market capitalization range.
They often invest in small or emerging growth companies.
We analyse 117 out of the 129 AG funds analysed in Daraio and Simar (2004),
for which two new management variables were available: manager tenure and
inception date.
The definition of the variables used in the analysis follows.

Total Return is the annual return at the 05-31-2002, expressed in percentage

terms. Morningstar calculation of total return is determined each month
by taking the change in monthly Net Asset Value (NAV), reinvesting all
income and capital-gains distributions during that month, and dividing
by the starting NAV. Since most of returns were negative for the analised
period, we add 100 to their amounts. We notice also that this transforma-
tion does not affect the efficiency analysis that we carry out in an input
oriented framework using total return as output.
Data description 197

Risk is the standard deviation of Return, it depicts how widely the returns
varied over a certain period of time. It is computed using the trailing
monthly total returns for 3 years. All of the monthly standard deviations
are then annualized. Standard deviation of return is an absolute measure
of volatility. It offers a probable range within which a fund’s realized
return is likely to deviate from its expected return.
Transaction costs are made by the sum of Expense Ratio, Loads and Turnover
Ratio.
Expense Ratio is the percentage of fund assets paid for operating expenses and
management fees, including 12b-1 fees (the annual charge deducted from
fund assets to pay for distribution and marketing costs), administrative
fees, and all other asset-based costs incurred by the fund except brokerage
costs. Sales charges are not included in the expense ratio.
Loads have been obtained by summing the Front-End Load and the Deferred
Load of each fund. The Front-End Load is the initial sales charge which
consists in a one-time deduction from an investment made into the fund.
The amount is generally relative to the amount of the investment, so that
larger investments incur smaller rates of charge. The sales charge serves
as a commission for the broker who sold the fund. The Deferred Loads
are also known as back-end sales charges and are imposed when investors
redeem shares. The percentage charged, generally declines the longer
shares are held.
Turnover ratio is a measure of the fund’s trading activity which is computed by
taking the lesser of purchases or sales and dividing by average monthly
net assets. It gives an indication of trading activity: funds with higher
turnover (implying more trading activity) incur greater brokerage fees for
affecting the trades. It is also an indication of management strategy: a low
turnover figure would indicate a “buy-and-hold strategy”; high turnover
would indicate an investment strategy involving “considerable buying and
selling” of securities.
Market risks reflects the percentage of a fund’s movements that can be ex-
plained by movements in its benchmark index. Morningstar compares
all equity funds to the S&P 500 index and all fixed-income funds to the
Lehman Brothers Aggregate Bond Index. It is calculated on a monthly ba-
sis, based on a least-squares regression of the funds returns on the returns
of the fund’s benchmark index.
Manager tenure is the number of years that the current manager has been the
portfolio manager of the fund. For funds with more managers the average
tenure is reported.
198 Exploring the effects of manager tenure, fund age and their interaction

Fund inception date is the date on which the fund began his operations. We
use this information to compute the age in months of mutual funds at April
2002.

As done in previous studies, we adopt an input oriented framework. We use

as inputs: Risks (standard deviation of return), Turnover and Expense ratio; as
output the Total return, and as External variables: Manager tenure, funds age
and their interrelation. We use also as descriptors market risks and fund size
(net assets in million of US dollars).
Some descriptive statistics on the inputs, output, external factors and de-
scriptors are offered in Table 8.1, were the average, the standard deviation, the
minimum, the maximum and the interquartile range (Iqr) are reported. The lat-
ter (Iqr) gives a robust measure of the spread of the data, since it is not affected
by changes in the upper and lower 25%; it is given, in fact, by the difference
between the 75% and the 25% percentiles of the analysed variable.
Inspection of Table 8.1 shows that the average fund has a risk (standard
deviation of return) of 35, has a turnover ratio of 154 with an expense ratio of
1.73; it has a total return of 82 (which is negative, as we added 100 to total
return) and an average market risk and fund size of respectively 47 and 469.

Table 8.1. Descriptive statistics on inputs, output, external factors for Aggressive Growth mu-
tual funds (AG117).

Variable Mean St. Dev. Min Max Iqr

Risk 34.656 8.827 14.730 81.050 9.765
( Input 1)
Turnover 153.479 101.439 15.000 642.000 130.750
(Input 2)
Expense Ratio 1.726 1.334 0.480 14.700 0.823
(Input 3)
Total Return 81.766 10.099 40.120 103.760 14.270
( Output)
Manager tenure 4.994 3.735 0.513 30.140 3.274
( Ex. factor 1)
Fund Age 113.084 103.155 29.103 543.517 48.445
( Ex. factor 2)
Market risks 47.402 16.159 6.000 100.000 14.250
( Ex. factor 3)
Fund size 468.833 1215.700 0.200 8828.100 215.750
( Ex. factor 4)
Impact of mutual fund manager tenure on performance 199

The average fund manager tenure is of around 5 years, while its age is about
113 months from inception date. The range of variation are particularly broad:
manager tenure goes from half a year to 30 years, age of fund from 29 to 544
months.44
The variables presented in Table 8.1 are highly skewed; this is also confirmed
by the scatterplot matrices reported in Figures 8.1 and 8.2 where, along the
diagonals of the matrices, are also reported the histograms of all variables.
Figure 8.3 shows the scatterplots of Z1 (manager tenure) and Z2 (fund age)
against the various inputs.
From these plots, it emerges that there are no particular structures or rela-
tionships among the variables, and some extreme points are clearly highlighted.
This evidence calls for the use of robust methods that we will apply in the fol-
lowing of the chapter to avoid the influence of these outlying points on the
efficiency comparisons.

8.3 Impact of mutual fund manager tenure on performance

In the introductory section we have seen that the role of manager tenure on
the performance of mutual funds is not a stylised fact but is debated in the
literature with contradictory evidences. In this section we apply the methodol-
ogy described in Chapter 5 to shed light on the role of US Aggressive Growth
Mutual funds first globally (i.e., considering the mutual funds as a whole) and
then individually, investigating the impact of fund manager on each AG mutual
fund, trying to find patterns in the analysed funds.
The following Figure 8.4 illustrates the smoothed nonparametric regression
of Qzm on manager tenure (Z) (top panel) and of αQz on Z (bottom panel).
We set the level of m = 25 and α = 0.95 so that we are robust at 10%. Qzm
(and αQz ) are the ratios of conditional on unconditional efficiency measures of
order−m (respectively order−α) computed in an input oriented framework. We
remember that in this framework an increasing nonparametric regression line
states a globally negative impact of the external factor, a decreasing nonpara-
metric regression line shows a positive impact on mutual funds performance
while a straight line points to an absence of global effect of the factor on the
analysed funds. As confirmed also by Figure 8.5 - which shows a zoom on the
impact on AG mutual funds which have a manager tenure lower than 13 years,
globally speaking, we see that there is no impact of manager tenure on mutual
funds performance.

44 We notice that in order to use the methodology described in Chapter 5 we have made more continuous the

last two variables (manager tenure and fund age) by simply adding a small number randomly chosen from
the continuous uniform distribution on the interval [-0.499,+0.499]. This transformation does not affect the
values and is suitable for our procedure which requires the use of continuous variables.
200 Exploring the effects of manager tenure, fund age and their interaction

AG 117
100

X1 50

0
1000
X2

500

0
20
X3

10

0
200
Y

100

0
0 50 100 0 500 1000 0 10 20 0 100 200
X1 X2 X3 Y

Figure 8.1. Scatterplot matrix of inputs (X) and output (Y) for Aggressive Growth Mutual
Funds (AG117).
AG117
150
100
Y

50
0
40
30
Z1

20
10
0
600
400
Z2

200
0
0 50 100 150 0 20 40 0 200 400 600
Y Z1 Z2

Figure 8.2. Scatterplot matrix of output (Y) and external factors (Z) for Aggressive Growth
Mutual Funds (AG117). Z1 is manager tenure and Z2 is Fund age in months.

This result supports the findings of Porter and Trifts (1998), Detzel and
Weigand (1998) and of Fortin, Michelson and Wagner (1999) who find that
longer term manager do not perform better than those with shorter track records.
Nevertheless, as we stated in the introduction, it is difficult to compare results
of evidence obtained using different dataset, coverage and in primis different
methods. Most used techniques in empirical finance are ordinary least squares
(OLS) with some multi-stage OLS (see e.g., Golec (1996) for a discussion).
On the contrary, our flexible approach, robust and nonparametric, offers the
possibility of investigate not only aggregate trends but also single efficiency
Impact of mutual fund manager tenure on performance 201

40

30

Z1
20

10

600
500
400
Z2

300
200
100
0

0 50 100 0 500 1000 0 5 10 15

X1 X2 X3

Figure 8.3. Scatterplot matrix of external factors (Z) and Inputs (X) for Aggressive Growth
Mutual Funds (AG117).
Effect of Z on Order−m frontier
1.5
Qzm

0.5

5 10 15 20 25 30

Effect of Z on Order−α frontier

1.5
Qzα

0.5

5 10 15 20 25 30
values of Z

Figure 8.4. Influence of manager tenure on the performance of Aggressive Growth (AG117)
Mutual Funds.
Effect of Z on Order−m frontier
1.5

1.25
Qzm

0.75
2 4 6 8 10 12

Effect of Z on Order−α frontier

1.5
1.25
Qzα

1
0.75
0.5

2 4 6 8 10 12

Figure 8.5. Influence of manager tenure on the performance of Aggressive Growth (AG117)
Mutual Funds. A zoom on Mutual funds with a manager tenure lower than 13 years.
202 Exploring the effects of manager tenure, fund age and their interaction

patterns. These particulars distinct funds are more interesting in empirical

finance, where financial analysts look for best outperformers and stars.
In the following we report some descriptive statistics on the various measures
of efficiency we computed. To save space we do not report all the results
available for order−α measures that are very similar to the order−m ones. We
notice that also the ranks offered in the various tables that follow (for order-m)
are the same than those obtained for order-α measures.
Table 8.2. Unconditional measures of efficiency, and various indicators, Aggressive Growth
Mutual Funds (AG117).

Qzm < 1
(15 obs) Npoint θ(x, y) (x, y)
α Nemp θm,n (x, y)
Mean 1 0.82 0.98 56 1.13
St.dev. 2 0.22 0.03 31 0.42
Min 0 0.51 0.91 14 0.65
Max 6 1.00 1.00 115 2.46

Qzm > 1
(26 obs) Npoint θ(x, y) (x, y)
α Nemp θm,n (x, y)
Mean 15 0.67 0.80 69 0.76
St.dev. 15 0.13 0.16 32 0.13
Min 1 0.51 0.41 13 0.59
Max 56 0.97 0.99 117 1.04

Qzm = 1
(76 obs) Npoint θ(x, y) (x, y)
α Nemp θm,n (x, y)
Mean 7 0.76 0.89 56 0.85
St.dev. 11 0.17 0.13 34 0.15
Min 0 0.47 0.43 1 0.53
Max 60 1.00 1.00 114 1.16

ALL
(117) Npoint θ(x, y) (x, y)
α Nemp θm,n (x, y)
Mean 8.31 0.75 0.88 59.05 0.87
St.dev. 12.46 0.17 0.14 33.76 0.23
Min 0.00 0.47 0.41 1.00 0.53
Max 60.00 1.00 1.00 117.00 2.46

Table 8.2 presents some unconditional measures of efficiency and indica-

tors by groups of AG funds which have a different impact or role of manager
tenure, namely the 15 mutual funds which have a Qzm < 1; the 26 funds
with Qzm > 1; the 76 funds with Qzm = 1 and finally all the 117 funds together.
Impact of mutual fund manager tenure on performance 203

Npoint is the number of points which dominates the analysed funds, θ(x,  y) is
the FDH input efficiency measure. α(x, y) is the input oriented probabilistic
measure introduced in Section 4.3. It is the order−α of the estimated quantile
frontier which passes through the unit (x, y). For instance α(x,  y) = 0.98
means that only 2% of funds with a return at least equal to y are using less inputs
than the unit (x, y). Hence, 1 − α(x,
 y) gives an estimates of the probability of
unit (x, y) to be dominated, given its level of y. Nemp is the number of units
used to estimate the distribution function and finally θm,n (x, y) is the input
order−m efficiency measure where m = 25 and is robust at 10%.
Table 8.3 presents some conditional measures of efficiency and indicators
for the same groups of AG funds with a different role of manager tenure (given
by the values of Qzm >, <, = 1) as in Table 8.2. ZNpoint is the number of
points which dominates the analysed funds given that Z = z, θ(x,  y|z) is the
conditional FDH input efficiency measure, α  z (x, y) is the conditional input
oriented probabilistic measure introduced in Section 5.2.3. It is the order−α
of the estimated conditional quantile frontier which passes through the unit
(x, y) given that Z = z. Hence, 1 − α  z (x, y) gives an estimates of the prob-
ability of unit (x, y) to be dominated, given its level of y and the condition
Z = z. θm,n (x, y|z) is the conditional input order−25 efficiency measure and
Qzm = θm,n (x, y|z)/θm,n (x, y). EImz = E(Q z |Z = z) is the externality index
z
defined in Chapter 5, IIm = Q /E(Q z  z |Z = z) is the individual index and
finally αQz = α  z (x, y)/α(x,
 y).
To characterize the different groups of funds according to the different impact
of manager tenure, we provide in Table 8.4 some descriptive statistics which
have to be related to the several efficiency measures and indicators presented
above.

What do we learn by looking at Tables 8.2 to 8.4

The main aim of the descriptive analysis we reported in Tables 8.2 to 8.4,
as we recalled above, is search for individual or group patterns of extremely
outperforming funds and characterize their profile. This is particular useful for
empirical finance.
It emerges that the profile of AG funds with Qzm < 1, i.e. funds whose
efficiency conditioned by the manager tenure is smaller than the unconditional
robust efficiency (see the first five lines of Tables 8.2 and 8.3) is characterized
by an higher average of manager tenure, around 7 years against a global average
of 5, a lower average age of fund, 104 months against the global average of 113
months, a lower average risk (32 against 35), a lower turnover ratio (95 against
153), lower expense ratio (1.48 against 1.73), but an higher level of market
risks (56 against 47) and a size of more than twice the average age of the whole
sample (net asset values in million of US $ of 1014 against 469).
204 Exploring the effects of manager tenure, fund age and their interaction

Table 8.3. Conditional measures of efficiency and various indicators (Z is manager tenure), by
groups of Aggressive Growth Mutual Funds (AG117).

Qzm < 1
(15 obs) ZNpoint θ(x, y|z) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 0.93 0.82 0.97 1.03 0.91 1.00 0.92 0.99
St.dev. 1.48 0.22 0.05 0.37 0.03 0.03 0.04 0.02
Min 0.00 0.51 0.85 0.61 0.84 0.92 0.85 0.94
Max 5.00 1.00 1.00 2.25 0.95 1.04 1.00 1.00

Qzm > 1
(26 obs) ZNpoint θ(x, y|z) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 8.12 0.74 0.85 0.85 1.11 1.03 1.08 1.07
St.dev. 9.49 0.17 0.16 0.13 0.07 0.02 0.07 0.05
Min 0.00 0.51 0.46 0.63 1.05 1.00 1.00 0.98
Max 32.00 1.00 1.00 1.11 1.33 1.11 1.30 1.17

Qzm = 1
(76 obs) ZNpoint θ(x, y|z) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 4.91 0.78 0.88 0.85 1.00 1.01 0.99 0.99
St.dev. 6.89 0.17 0.13 0.15 0.02 0.02 0.03 0.04
Min 0.00 0.47 0.42 0.50 0.95 0.95 0.92 0.86
Max 37.00 1.00 1.00 1.11 1.05 1.05 1.05 1.14

ALL
(117) ZNpoint θ(x, y|z) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 5.11 0.77 0.88 0.87 1.0 1.01 1.00 1.01
St.dev. 7.44 0.18 0.13 0.20 0.07 0.03 0.07 0.05
Min 0.00 0.47 0.42 0.50 0.84 0.92 0.85 0.86
Max 37.00 1.00 1.00 2.25 1.33 1.11 1.30 1.17

Quite an opposite profile is those of the AG funds with Qzm > 1, i.e. funds
which in turn have an efficiency conditioned by manager tenure higher than
the unconditional one (see the efficiency measures reported above, from the 6th
line from the top to the 10th of Tables 8.2 and 8.3). These funds have a manager
tenure of 6 years, market risk, fund age and size are quite the same than the
average of the whole sample, but they show a lower average return and higher
expense ratio and turnover ratio (182 against 153).
Impact of mutual fund manager tenure on performance 205

Table 8.4. Some descriptive statistics by group of funds with different impact of manager
tenure. Z is manager tenure.

Qzm < 1
(15 obs) Risk Turnover Expense T Return Mkt Manager Fund Size
risks Tenure age
Mean 31.87 95.07 1.48 82.26 56.00 6.53 103.53 1014.48
St.dev. 9.82 71.03 0.75 8.28 18.25 6.71 65.21 2258.49
Min 14.73 15.00 0.48 62.24 20.00 2.00 46.00 2.30
Max 49.71 290.00 2.62 93.08 91.00 30.00 318.00 8828.10

Qzm > 1
(26 obs) Risk Turnover Expense T Return Mkt Manager Fund Size
risks Tenure age
Mean 36.98 182.31 2.18 77.51 47.58 5.88 102.19 571.74
St.dev. 10.94 136.48 2.53 11.35 13.67 2.85 95.42 1292.05
Min 21.07 50.00 0.92 40.12 6.00 1.00 41.00 0.20
Max 81.05 642.00 14.70 93.42 70.00 16.00 544.00 5324.00

Qzm = 1
(76 obs) Risk Turnover Expense T Return Mkt Manager Fund Size
risks Tenure age
Mean 34.41 155.14 1.62 83.13 45.64 4.34 118.75 325.93
St.dev. 7.45 85.89 0.56 9.48 15.85 2.94 110.42 771.33
Min 17.69 44.00 0.90 64.30 8.00 1.00 29.00 0.20
Max 50.22 482.00 3.80 103.76 100.00 21.00 541.00 4914.30

ALL
(117 obs) Risk Turnover Expense T Return Mkt Manager Fund Size
risks Tenure age
Mean 34.66 153.48 1.73 81.77 47.40 4.97 113.12 468.83
St.dev. 8.79 101.00 1.33 10.06 16.09 3.73 102.70 1210.45
Min 14.73 15.00 0.48 40.12 6.00 1.00 29.00 0.20
Max 81.05 642.00 14.70 103.76 100.00 30.00 544.00 8828.10
206 Exploring the effects of manager tenure, fund age and their interaction

Table 8.5. Rank of AG mutual funds with Qzm > 1 ordered by decreasing INCR =
θm,n (x, y|z) − θm,n (x, y). Z is manager tenure.

Fund Name (x, y)

α θm,n z (x, y)
α θm,n
z
Qzm z
EIm z
IIm αQz IN CR
1 0.66 0.60 0.77 0.80 1.33 1.03 1.30 1.17 0.20
2 0.76 0.68 0.85 0.85 1.26 1.03 1.22 1.12 0.17
3 0.85 0.70 0.96 0.87 1.24 1.05 1.18 1.13 0.17
4 0.81 0.75 0.90 0.91 1.22 1.03 1.18 1.11 0.16
5 0.90 0.84 0.99 0.99 1.18 1.05 1.12 1.09 0.15
6 0.87 0.72 0.93 0.82 1.13 1.04 1.09 1.07 0.10
7 0.78 0.76 0.82 0.85 1.13 1.00 1.12 1.05 0.10
8 0.97 0.92 1.00 1.01 1.10 1.02 1.08 1.03 0.09
9 0.81 0.75 0.87 0.83 1.11 1.11 1.00 1.07 0.08
10 0.65 0.59 0.74 0.68 1.14 1.01 1.13 1.14 0.08
11 0.88 0.79 0.93 0.86 1.10 1.05 1.04 1.06 0.08
12 0.99 1.04 1.00 1.11 1.07 1.00 1.07 1.01 0.08
13 0.86 0.72 0.96 0.79 1.10 1.05 1.05 1.11 0.07
14 0.85 0.88 0.85 0.96 1.08 1.00 1.08 1.01 0.07
15 0.92 0.93 1.00 1.00 1.08 1.01 1.07 1.08 0.07
16 0.94 0.90 0.96 0.97 1.08 1.05 1.03 1.02 0.07
17 0.53 0.61 0.53 0.67 1.09 1.00 1.09 1.02 0.06
18 0.73 0.66 0.80 0.72 1.08 1.03 1.05 1.09 0.05
19 0.94 0.92 0.95 0.97 1.05 1.03 1.03 1.01 0.05
20 0.93 0.91 0.95 0.95 1.05 1.04 1.02 1.02 0.05
21 0.91 0.88 0.93 0.93 1.05 1.00 1.05 1.02 0.05
22 0.95 0.76 0.97 0.80 1.06 1.03 1.03 1.02 0.05
23 0.85 0.75 0.95 0.79 1.05 1.04 1.02 1.12 0.04
24 0.41 0.60 0.47 0.64 1.06 1.04 1.02 1.14 0.04
25 0.41 0.60 0.46 0.64 1.06 1.03 1.02 1.14 0.03
26 0.63 0.60 0.62 0.63 1.05 1.05 1.01 0.98 0.03
Funds Name: 1 SunAmerica Foc MultiGr B, 2 SunAmerica Foc MultiGr A, 3 American Heritage Growth, 4
MFS Emerging Growth C, 5 Liberty Contrarian SmCpA, 6 Putnam Voyager II B, 7 Phoenix-Engemann Agg
Gr A, 8 Mason Street Ag Gr Stk A, 9 MFS Emerging Growth B, 10 SunAmerica Foc MultiGr II, 11 Putnam
Voyager II M, 12 Fidelity Aggressive Grth, 13 Navellier Aggr Growth, 14 Alliance Quasar Instl I, 15 SSgA
Aggressive Equity, 16 Vintage Aggr Growth, 17 Atlas Emerging Growth A, 18 Evergreen Sm Co Grth B, 19
Credit Suisse Tr Gl PstVt, 20 Putnam New Opports B, 21 Commerce MidCap Gr Svc, 22 Delaware Trend C,
23 Quaker Aggressive Gr A, 24 Prudential US Emerg Gr C, 25 Prudential US Emerg Gr B, 26 Van Wagoner
Emerging Gr.
Impact of mutual fund manager tenure on performance 207

Table 8.6. Rank of AG mutual funds with Qzm < 1 ordered by decreasing DECR =
θm,n (x, y|z) − θm,n (x, y). Z is manager tenure.

Fund Name (x, y)

α θm,n z (x, y)
α θm,n
z
Qzm z
EIm z
IIm αQz DECR
1 1.00 1.48 1.00 1.25 0.84 0.99 0.85 1.00 -0.23
2 1.00 2.46 1.00 2.25 0.91 1.04 0.88 1.00 -0.21
3 1.00 1.39 1.00 1.19 0.86 0.99 0.86 1.00 -0.20
4 0.99 0.95 0.98 0.83 0.87 0.99 0.88 0.99 -0.12
5 1.00 1.20 1.00 1.09 0.91 0.95 0.96 1.00 -0.11
6 0.99 0.96 0.98 0.86 0.90 0.99 0.91 0.99 -0.10
7 0.98 0.88 0.96 0.78 0.89 0.99 0.90 0.98 -0.10
8 1.00 1.13 1.00 1.03 0.92 0.92 1.00 1.00 -0.10
9 1.00 1.10 1.00 1.03 0.93 1.03 0.91 1.00 -0.07
10 1.00 1.10 1.00 1.03 0.94 1.02 0.92 1.00 -0.07
11 0.98 0.90 0.98 0.83 0.93 1.03 0.90 1.00 -0.07
12 0.98 1.04 0.97 0.98 0.94 0.99 0.95 0.98 -0.06
13 1.00 1.06 1.00 1.00 0.95 1.02 0.93 1.00 -0.06
14 0.91 0.65 0.85 0.61 0.94 0.99 0.94 0.94 -0.04
15 0.91 0.65 0.86 0.61 0.94 1.00 0.95 0.94 -0.04
Funds Name: 1 Goldman Sachs Agg GrStrB, 2 Touchstone Aggr Grth A, 3 GMO Tax-Mgd U.S. Eq III, 4
Scudder Aggressive Gr C, 5 Putnam New Opports A, 6 Scudder Aggressive Gr B, 7 Security Ultra B, 8 Value
Line Leveraged Gr, 9 Fidelity Capital Apprec, 10 Janus Olympus, 11 AmCent Vista Inst, 12 Evergreen Sm
Co Grth I, 13 Delaware Trend Instl, 14 Alliance Quasar C, 15 Alliance Quasar B.

In order to find out the funds most positively influenced by manager tenure,
we report in Table 8.5 the rank of US AG funds, with Qzm > 1, ordered by
decreasing INCR -difference between conditional and unconditional order−m
efficiency measure (= θm,n (x, y|z) − θm,n (x, y)) as well as their individual
efficiency measures. On the contrary, Table 8.6 reveals the name of AG funds
most negatively influenced by manager tenure (with Qzm < 1), ordered by
decreasing DECR (= θm,n (x, y|z) − θm,n (x, y)) as well as their individual
efficiency measures.
Hence, we were able to find out the US AG funds that have most been
influenced by manager tenure, and analyse their different profiles. Of course,
this analysis is not conclusive, because other information would have been
useful to complete our understanding of the manager tenure effect, such as
age of managers, if they have an MBA, and so on, all information that are not
available to us. Nevertheless, this analysis is quite interesting and informative:
even if globally speaking, manager tenure does not affect the performance of
the analysed funds (see Figures 8.4 and 8.5) our procedure is able to identify the
208 Exploring the effects of manager tenure, fund age and their interaction

funds which had the major impact (positive or negative) and let us characterize
their profile. This is particularly useful in empirical finance to try to understand
the management behaviour of stars and best performers.
Another interesting phenomena is the impact of fund age on the performance
and its interaction with manager tenure, with which we deal in the next section.

8.4 Interaction between manager tenure and fund age

By using the same technique, we analysed also the impact of fund age on
AG mutual funds, and find out that there is no global effect on the performance
of the analysed funds. See Figures 8.6 and 8.7.

Effect of Z on Order−m frontier

1.5
Qzm

0.5
50 100 150 200 250 300 350 400 450 500
values of Z
Effect of Z on Order−α frontier
1.5
Qzα

0.5
50 100 150 200 250 300 350 400 450 500

Figure 8.6. Influence of fund age on the performance of Aggressive Growth (AG117) Mutual
Funds.

It seems that the ability to survive in a highly competitive environment,

as measured by the experience accumulated in a longer number of years in
operation, does not affect the performance of the AG mutual funds as a whole.
We have also to consider that we analyse a very peculiar time frame: our data
spans the terroristic attack of the 11th September 2001 which contributed to the
collapse of financial markets in most advanced countries.
Another more interesting question is if in the analysed period there has
been interaction between managerial experience (manager tenure) and funds’
longevity (measured by fund age), and if this interaction has affected the perfor-
mance of AG mutual funds. As done below, we analyse the interaction globally,
considering a bivariate case for Z, and after that individually (looking at the
efficiency of each fund) to point on very peculiar mutual funds which have
been particularly influenced (positively or negatively) simultaneously by these
factors.
Interaction between manager tenure and fund age 209

Effect of Z on Order−m frontier

1.5
Qzm
1

0.5
50 100 150 200 250
values of Z
Effect of Z on Order−α frontier
1.5
Qzα

0.5
50 100 150 200 250

Figure 8.7. Influence of fund age on the performance of Aggressive Growth (AG117) Mutual
Funds. A zoom on Mutual funds with a fund age lower than 250 months.

The global impact of manager tenure (Z1 ) and fund age (Z2 ) on AG funds
is shown in Figures 8.8 and 8.9.
We notice that the correlations between manager tenure and fund age is quite
low (it is 0.26 although significant at 95%- p. val. 0.004 < 0.05); the number
of k − N N provided by our data driven method (see Chapter 5) is of 55. See
Figure 8.10 for the estimation of the density of Z as well as its contour plot.
In particular, Figure 8.8 shows the surface of Qzm on Z1 and Z2 . Figure
8.9 illustrates the smoothed nonparametric regression of Qzm on Z1 (top panel)
for Z2 ’s quartiles; and the smoothed nonparametric regression of Qzm on Z2
(bottom panel) for Z1 ’s quartiles. The dashed line indicates the first quartile,
the solid line the median and the dashdot line the third quartile.
It appears, as we have seen in Figure 8.4, that globally there is no impact of
manager tenure till a tenure of around 20 years (note that we only have 3 funds
with manager tenure higher than 15 years) and this appears for all quartiles of
fund age (top panel of Figure 8.9). By inspecting Figure 8.9 (bottom panel)
it seems that fund age does not affect the performance of AG funds if it is
conjoint with the first quartile of manager tenure (dashed line), while is seems
that for the median of manager tenure (solid line) there is a positive combined
effect with an age higher than 400 months, and for longer manager tenure (third
quartile, dashdot line) there is a positive effect starting even for funds with age
higher than 100 months. Hence, even though there is almost no global effect of
manager tenure, our procedure allows to shed lights on the interaction between
Z1 and Z2 . An interpretation of this plot could be that longer and then more
experienced manager tenure are better able to exploit the ability/experience of
funds in facing highly competitive markets.
210 Exploring the effects of manager tenure, fund age and their interaction

1.1

1.05
Qzm

0.95

0.9
600
400 40
30
200 20
Z2 10
0 0
Z1

Figure 8.8. Influence of manager tenure and fund age on the performance of Aggressive Growth
Mutual Funds (AG117). Surface.

1.1

1.05
Qzm

0.95

0.9
0 5 10 15 20 25 30 35
Z1
1.05
Qzm

0.95
0 100 200 300 400 500 600
Z2

Figure 8.9. Influence of manager tenure (Z1 ) and fund age (Z2 ) on the performance of Ag-
gressive Growth Mutual Funds (AG117). Plots.

But let examine now the peculiar behaviour of single funds. As in the previ-
ous section, Table 8.7 shows some descriptive statistics on efficiency measures
and indicators for group of funds that have a different impact of manager tenure
jointly considered with fund age. Table 8.8 is useful to characterize the profile
of groups of US funds.
Table 8.10 illustrates the rank of US AG funds, with Qzm > 1, ordered by
decreasing INCR -difference between conditional and unconditional order−m
efficiency measure (= θm,n (x, y|z) − θm,n (x, y)) as well as their individual
efficiency measures. Table 8.9 reveals the name of AG funds with Qzm < 1,
ordered by decreasing DECR (= θm,n (x, y|z) − θm,n (x, y)) as well as their
Interaction between manager tenure and fund age 211

Density of Z

3 10−3
2

1.5

0.5

0
40
20 40
30
20
0 10
0
values of Z2 −20 −10 values of Z1

Contour plot for the density of Z

30

25

20
values of Z2

15

10

0
0 5 10 15 20 25 30
values of Z1

Figure 8.10. Influence of manager tenure (Z1 ) and fund age (Z2 ) on the performance of Ag-
gressive Growth Mutual Funds (AG117). Density of Z (top panel) and contour plot of the density
of Z (bottom panel).
individual efficiency measures. Here we remember that Z is bivariate and is
done by manager tenure and fund age. It is also interesting to compare these
results with those of the previous section. In particular, there are some funds
such as no.1,2,7,12,19,20 and 21 in Table 8.10 that were not listed in Table
8.5; for these funds, manager tenure has an effect only if taking into account
conjointly with funds’ age. The same holds true for some funds of Table 8.9,
such as no. 9, 13, 15, 17 and 18 not listed in Table 8.6.
These considerations are just examples on how to use our approach to shed
light on individual pattern of efficiency measures and their explanation. As a
matter of fact, we showed that our methodology is very appealing in providing
empirical evidence not only on global financial performance of mutual funds,
but also on single peculiar profiles.
212 Exploring the effects of manager tenure, fund age and their interaction

Table 8.7. Some descriptive statistics on efficiency measures and indicators by group of funds
with different impact of manager tenure. Multivariate Z.

Qzm < 1
(19 obs) (x, y)
α θm,n (x, y) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 0.96 1.06 0.95 0.95 0.91 1.01 0.90 0.99
St.dev. 0.06 0.40 0.06 0.34 0.04 0.02 0.05 0.02
Min 0.81 0.65 0.82 0.60 0.80 0.93 0.78 0.95
Max 1.00 2.46 1.00 2.16 0.95 1.02 0.98 1.04

Qzm > 1
(28 obs) (x, y)
α θm,n (x, y) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 0.75 0.73 0.81 0.82 1.12 1.02 1.10 1.07
St.dev. 0.17 0.12 0.19 0.13 0.10 0.00 0.10 0.06
Min 0.41 0.59 0.42 0.63 1.05 1.01 1.03 0.99
Max 0.97 0.94 1.00 1.01 1.45 1.03 1.43 1.23

Qzm = 1
(70 obs) (x, y)
α θm,n (x, y) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 0.91 0.87 0.91 0.87 1.00 1.01 0.99 1.00
St.dev. 0.10 0.14 0.10 0.14 0.02 0.00 0.02 0.04
Min 0.49 0.53 0.53 0.54 0.96 1.00 0.94 0.91
Max 1.00 1.10 1.00 1.08 1.05 1.02 1.04 1.13

ALL
(117 obs) (x, y)
α θm,n (x, y) z (x, y)
α θm,n (x, y|z) Qzm z
EIm z
IIm αQz
Mean 0.88 0.87 0.89 0.87 1.01 1.01 1.00 1.01
St.dev. 0.14 0.23 0.13 0.19 0.09 0.01 0.09 0.05
Min 0.41 0.53 0.42 0.54 0.80 0.93 0.78 0.91
Max 1.00 2.46 1.00 2.16 1.45 1.03 1.43 1.23
Interaction between manager tenure and fund age 213

Table 8.8. Some descriptive statistics by group of funds with different impact of manager
tenure. Bivariate Z.

Qzm < 1
(19 obs) Risk Turn. Exp. T Ret. Mkt Manager Fund Size
risks Tenure age
Mean 33.76 119.84 1.47 81.70 51.74 6.47 111.32 960.77
St.dev. 9.57 97.18 0.69 8.51 18.98 6.04 90.49 2179.79
Min 14.73 15.00 0.48 62.24 20.00 2.00 29.00 2.30
Max 49.71 305.00 2.62 93.08 91.00 30.00 395.00 8828.10

Qzm > 1
(28 obs) Risk Turn. Exp. T Ret. Mkt Manager Fund Size
risks Tenure age
Mean 37.95 165.32 2.19 78.69 46.50 5.71 70.96 382.11
St.dev. 9.94 74.90 2.43 10.80 15.95 2.76 27.02 845.08
Min 26.11 64.00 1.10 40.12 17.00 1.00 35.00 0.20
Max 81.05 305.00 14.70 95.34 72.00 16.00 149.00 3593.00

Qzm = 1
(70 obs) Risk Turn. Exp. T Ret. Mkt Manager Fund Size
risks Tenure age
Mean 33.58 157.87 1.61 83.02 46.59 4.26 130.47 370.00
St.dev. 7.68 108.91 0.59 9.87 15.07 3.00 118.75 891.70
Min 17.69 44.00 0.89 51.61 6.00 1.00 41.00 0.20
Max 50.22 642.00 3.80 103.76 100.00 21.00 544.00 5324.00

ALL
(117 obs) Risk Turn. Exp. T Ret. Mkt Manager Fund Size
risks Tenure age
Mean 34.66 153.48 1.73 81.77 47.40 4.97 113.12 468.83
St.dev. 8.79 101.00 1.33 10.06 16.09 3.73 102.70 1210.45
Min 14.73 15.00 0.48 40.12 6.00 1.00 29.00 0.20
Max 81.05 642.00 14.70 103.76 100.00 30.00 544.00 8828.10
214 Exploring the effects of manager tenure, fund age and their interaction

Table 8.9. Rank of AG mutual funds with Qzm < 1 ordered by decreasing DECR =
θm,n (x, y|z) − θm,n (x, y).Bivariate Z: Z1 manager tenure, Z2 fund age.

Fund Name (x, y)

α θm,n z (x, y)
α θm,n
z
Qzm z
EIm z
IIm αQz DECR
1 1.00 2.46 1.00 2.16 0.88 1.02 0.86 1.00 -0.30
2 1.00 1.48 1.00 1.18 0.80 1.02 0.78 1.00 -0.30
3 1.00 1.39 1.00 1.15 0.83 1.02 0.82 1.00 -0.24
4 0.98 0.90 0.96 0.77 0.85 1.02 0.84 0.98 -0.13
5 0.99 0.95 0.98 0.82 0.87 1.01 0.86 0.99 -0.12
6 0.99 0.96 0.98 0.84 0.88 1.01 0.87 0.99 -0.12
7 1.00 1.13 1.00 1.03 0.91 0.93 0.98 1.00 -0.10
8 1.00 1.10 1.00 1.02 0.93 1.02 0.91 1.00 -0.08
9 0.94 0.86 0.92 0.79 0.91 1.02 0.90 0.97 -0.08
10 0.98 1.04 0.96 0.97 0.93 1.02 0.91 0.98 -0.07
11 0.98 0.88 0.96 0.80 0.92 1.01 0.90 0.99 -0.07
12 1.00 1.20 1.00 1.13 0.94 1.02 0.92 1.00 -0.07
13 1.00 1.16 1.00 1.10 0.94 1.01 0.94 1.00 -0.06
14 1.00 1.06 1.00 1.00 0.95 1.02 0.93 1.00 -0.06
15 0.82 0.69 0.85 0.64 0.93 1.02 0.92 1.04 -0.05
16 0.91 0.65 0.86 0.60 0.93 1.01 0.92 0.95 -0.04
17 0.95 0.76 0.93 0.71 0.94 1.02 0.93 0.98 -0.04
18 0.81 0.77 0.82 0.73 0.95 1.02 0.93 1.00 -0.04
19 0.91 0.65 0.87 0.62 0.95 1.01 0.94 0.96 -0.03
Funds Name: 1 Touchstone Aggr Grth A, 2 Goldman Sachs Agg GrStrB, 3 GMO Tax-Mgd U.S. Eq III, 4
AmCent Vista Inst, 5 Scudder Aggressive Gr C, 6 Scudder Aggressive Gr B ,7 Value Line Leveraged Gr, 8
Fidelity Capital Apprec, 9 Prudential US Emerg Gr Z, 10 Evergreen Sm Co Grth I, 11 Security Ultra B, 12
Putnam New Opports A, 13 INVESCO Dynamics Inv, 14 Delaware Trend Instl, 15 AIM Mid Cap Opp A,
16 Alliance Quasar C, 17 Delaware Trend C, 18 Van Kampen Aggr Growth A, 19 Alliance Quasar B.
Interaction between manager tenure and fund age 215

Table 8.10. Rank of AG mutual funds with Qzm > 1 ordered by decreasing INCR =
θm,n (x, y|z) − θm,n (x, y).Bivariate Z: Z1 manager tenure, Z2 fund age.

Fund Name (x, y)

α θm,n z (x, y)
α θm,n
z
Qzm z
EIm z
IIm αQz IN CR
1 0.67 0.63 0.80 0.91 1.45 1.02 1.43 1.20 0.28
2 0.70 0.63 0.81 0.89 1.41 1.02 1.38 1.16 0.26
3 0.85 0.70 0.97 0.88 1.27 1.02 1.24 1.15 0.19
4 0.76 0.68 0.88 0.82 1.20 1.02 1.18 1.16 0.14
5 0.81 0.75 0.89 0.87 1.17 1.02 1.15 1.09 0.13
6 0.66 0.60 0.81 0.69 1.16 1.02 1.14 1.23 0.10
7 0.82 0.74 0.90 0.84 1.14 1.01 1.12 1.09 0.10
8 0.63 0.60 0.64 0.68 1.13 1.02 1.11 1.01 0.08
9 0.87 0.72 0.92 0.80 1.11 1.02 1.09 1.06 0.08
10 0.81 0.75 0.89 0.82 1.11 1.03 1.08 1.10 0.08
11 0.85 0.88 0.92 0.97 1.10 1.01 1.09 1.09 0.09
12 0.87 0.72 0.91 0.79 1.10 1.01 1.08 1.04 0.07
13 0.97 0.92 1.00 1.01 1.10 1.02 1.08 1.03 0.09
14 0.88 0.79 0.92 0.86 1.09 1.02 1.07 1.05 0.07
15 0.53 0.61 0.52 0.66 1.08 1.01 1.07 0.99 0.05
16 0.90 0.84 0.94 0.91 1.08 1.02 1.06 1.05 0.07
17 0.92 0.93 1.00 1.00 1.08 1.01 1.06 1.08 0.07
18 0.91 0.88 0.95 0.94 1.07 1.02 1.06 1.03 0.06
19 0.77 0.78 0.76 0.84 1.07 1.02 1.06 0.99 0.06
20 0.43 0.60 0.44 0.65 1.07 1.02 1.05 1.02 0.04
21 0.43 0.60 0.42 0.64 1.07 1.02 1.05 0.99 0.04
22 0.94 0.90 0.97 0.96 1.07 1.02 1.05 1.03 0.06
23 0.65 0.59 0.74 0.63 1.06 1.02 1.04 1.14 0.04
24 0.74 0.66 0.80 0.70 1.06 1.02 1.04 1.08 0.04
25 0.93 0.91 0.95 0.96 1.06 1.02 1.04 1.02 0.05
26 0.41 0.60 0.43 0.63 1.05 1.02 1.04 1.06 0.03
27 0.41 0.60 0.43 0.63 1.05 1.02 1.03 1.06 0.03
28 0.95 0.94 0.97 0.98 1.05 1.02 1.03 1.02 0.05
Fund Names: 1 AIM Small Cap Opp C, 2 AIM Small Cap Opp B, 3 American Heritage Growth, 4 SunAmerica
Foc MultiGr A, 5 MFS Emerging Growth C, 6 SunAmerica Foc MultiGr B, 7 Rochdale Alpha, 8 Van Wagoner
Emerging Gr, 9 Putnam Voyager II B, 10 MFS Emerging Growth B, 11 Alliance Quasar Instl I, 12 Putnam
Voyager II C, 13 Mason Street Ag Gr Stk A, 14 Putnam Voyager II M, 15 Atlas Emerging Growth A, 16 Liberty
Contrarian SmCpA, 17 SSgA Aggressive Equity, 18 Commerce MidCap Gr Svc, 19 ABN AMRO/Veredus
AggGr N, 20 Van Kampen Aggr Growth B, 21 Van Kampen Aggr Growth C, 22 Vintage Aggr Growth, 23
SunAmerica Foc MultiGr II, 24 Evergreen Sm Co Grth C, 25 Putnam New Opports B, 26 Prudential US
Emerg Gr C, 27 Prudential US Emerg Gr B, 28 Putnam New Opports M.
216 Exploring the effects of manager tenure, fund age and their interaction

8.5 Conclusions
In this chapter we analysed US Aggressive growth mutual funds and the
impact of management variables on their performance, namely manager tenure
and fund age.
Considered as a whole, we find that there is almost no impact of manager
tenure on mutual funds performance. This result supports the findings of Porter
and Trifts (1998), Detzel and Weigand (1998) and of Fortin, Michelson and
Wagner (1999) who find that longer term manager do not perform better than
those with shorter track records. Nevertheless, as we stated in the introduction,
it is difficult to compare results of evidence obtained using different dataset,
coverage and in primis different methods.
We analysed also the impact of fund age on AG mutual funds, and find that
there is no global effect on the performance of the analysed funds. It seems
that the ability to survive in a highly competitive environment, as measured
by the experience accumulated in a longer number of years in operation, does
not affect the performance of the AG mutual funds as a whole. We analysed
here a very peculiar time frame: our data span the terroristic attack of the 11th
September 2001 which contributed to the collapse of financial markets in most
advanced countries.
Finally, we investigated the interaction between managerial experience (man-
ager tenure) and funds’ longevity (measured by fund age), and how this inter-
action affected the performance of AG mutual funds. We found that longer
and then more experienced manager tenure are better able to exploit the abil-
ity/experience of funds in facing highly competitive markets.
Our flexible approach, robust and nonparametric, offers the possibility of
investigate not only aggregate trends but also single efficiency patterns.
By applying the methodology developed in Chapter 5, we were able to find
out the US AG funds that have most been influenced by manager tenure and their
interaction with fund age, and analyse their different profiles. Our approach is,
in fact, able to capture the interaction between the components of the external
factors (in this case manager tenure and fund age) at the level of individual
funds, even in absence of a global impact of these external factors.
Of course, this analysis is not conclusive, because other information would
have been useful to complete our understanding of the manager tenure effect,
such has age of managers, if they have an MBA, and so on, all information
that were not available to us. Nevertheless, this analysis is quite interesting
and informative: even if globally speaking, manager tenure does not affect the
performance of the analysed funds our procedure is able to identify the funds
which had the major impact (positive or negative) and let us characterize their
profile. This is particularly useful in empirical finance to try to understand the
management behaviour of stars and best performers.
Chapter 9

CONCLUSIONS

In this book we have provided a state-of-the-art presentation of robust and

nonparametric methods in efficiency analysis. We focused our analysis on
recent developments which allow the overcoming of some of the most critical
drawbacks of traditional nonparametric efficient frontier estimators (DEA/FDH
based) and their related efficiency scores.
In addition, this book adds new contributions to the efficiency literature.
It presents in a unified way the robust nonparametric measures of efficiency
based on partial frontiers (of order-m and of order-α) both unconditional and
conditional to external-environmental factors. It develops in a full multivariate
case the econometric methodology for the evaluation of the impact of external
factors on the performance. A new probabilistic robust measure of efficiency is
extended to the conditional case in the Methodology part and operationalized
in the Applications part. Another important contribution of the book, from a
methodological point of view, is the presentation of the parametric approxi-
mations of nonparametric robust frontiers and its extension to the multivariate
outputs case.
The simulated examples (with both univariate and multivariate external fac-
tors) offer useful guidelines on how to implement the proposed methods in
practice. Besides, the applications with real data (presented in Part II) show in
details how the proposed techniques could be put in place and give an idea on
the rich empirical results that are available from their applications. We provide
evidence on some classical topics in Industrial Organization such as economies
of scale and economies of scope in various fields (insurance sector, economics
of science and financial performance evaluation).
Specifically, from the insurance application we learnt how to reduce the
dimensional space to face the curse of dimensionality and the usefulness of
218 Conclusions

partial frontiers in revealing the impact of external factors masked, in the full
frontier case, by extreme observations.
In the analysis on the Italian National Research Council institutes we show
how to estimate parametric approximation of robust and nonparametric fron-
tiers in a full multivariate framework, by estimating robust parameters of a
multi output Translog distance function which have better properties than the
traditional parametric estimates.
Finally, the mutual funds application illustrates how the profile of group of
funds could be characterized starting from a global analysis of the comparison
set towards an even more detailed illustration of single extremely good or bad
performers.
These applications clearly demonstrate that the methodological toolbox pre-
sented in Part I of this book is built up by methods that are often complementary
and help all together to shed light on important aspects of the production process.
By using the real data applications, we have shown that it is always useful to
start the analysis by some correlation matrix-plots and some exploratory mul-
tivariate techniques (see Härdle and Simar, 2003). After that, according to the
kind of data and problem to be analysed an appropriate set of measures has to be
selected from the taxonomy we presented in Chapter 2. In a lot of studies, the
choice of robust methods presented in Chapters 4 and 5 may be the better so-
lution. The knowledge of the statistical properties of the estimators, presented
in Chapter 3 is a basic and fundamental step, in order to be aware of the main
problems and limitations of the traditional DEA/FDH approach, and to learn
how to bootstrap in this context to allow for a better inference in this setting.
A step further is the Chapter 4 where an alternative probabilistic formulation
of the activity analysis framework allows us to open the field to a new set of
probabilistic efficiency measures which keep a link with the traditional FDH
estimator (only asymptotically) while offering a wide range of properties useful
under an applied perspective as well as give us the opportunity of parametri-
cally approximate robust multi-output nonparametric frontiers providing better
inference also in this setting. In Chapter 5 we use this probabilistic approach to
introduce external-environmental factors in this general setting. The economet-
ric methodology we detail and extend to the full multivariate case is particularly
useful to shed light on factors behind the patterns and for the characterisation
of the profile of single DMU and groups of DMUs and not only in providing
aggregate or average tendencies.
We hope that the reading of the book has been useful for applied economists
who wanted to make use of these recently introduced techniques to evaluate and
explain the performance of DMUs in their field or research, without the burden
of limitations of traditional methods. Nowadays the implementation of these
recent techniques is facilitated by the availability of free software like FEAR
(see Wilson, 2005a,b,c).
Conclusions 219

At this point, the readers should be aware that the performance evaluation is
a complex task. A better understanding of the methods described in this book
is a necessary step in the performance evaluation. However, a full exploitation
of this toolbox is not possible without having a good knowledge of the field of
application.
References

[1] Acs, Z. (ed.) (2000), Regional innovation, knowledge and global change. London:
Pinter.

[2] Adams, J.D. and Griliches, Z. (2000), “Research Productivity in a System of Univer-
sities”, in Encaoua, D. et al.(eds.), The Economics and Econometrics of Innovation,
105-140, Kluwer Academic Publishers, Netherlands.

[3] Afriat, S.N. (1967), “The Construction of Utility Functions from Expenditure Data”,
International Economic Review, 8, 67-77.

[4] Afriat, S.N. (1972), “Efficiency Estimation of Production Functions”, International

Economic Review, 13, 568-598.

[5] Aigner, D.J. and Chu S.F. (1968), “On Estimating the Industry Production Function”,
American Economic Review, 58, 826-839.

[6] Aigner, D.J., Lovell, C.A.K., and Schmidt P. (1977), “Formulation and Estimation of
Stochastic frontier Production Function Models”, Journal of Econometrics, 6, 21-37.

[7] Alchian, A. (1965) “Some economics of Property Rights”, Il Politico 30:4 (December),
816-829.

[8] Alchian, A., and R. Kessel (1962) “Competition, Monopoly, and the Pursuit of Money”,
in Aspects of Labor Economics. Princeton, NJ: Princeton University Press for National
Bureau of Economic Research.

[9] Allen, R., Athanassopoulos, A., Dyson, R.G. and Thanassoulis, E. (1997), “Weights
restrictions and value judgements in data envelopment analysis: evolution, develop-
ment and future directions”, Annals of Operations Research, 73, 13-34.

[10] Allison, P.D. and Stewart, J.A. (1974), “Productivity differences among scientists:
evidence for accumulative advantage”, American Sociological Review, 39 (4), 596-
606.

[11] Amel, D., Barnes C., Panetta F. and Salleo C. (2002), “Consolidation and efficiency
in the financial sector:a review of the international evidence”, Tema di discussione,
Banca d’Italia n. 464.
222 REFERENCES

[12] Andersen, P. and Petersen, N.C. (1993), “A procedure for ranking efficient units in
data envelopment analysis”, Management Science, 39, 1261-1264.

[13] Angulo-Meza, L. and Pereira Estellita Lins M. (2002), “Review of Methods for In-
creasing Discrimination in Data Envelopment Analysis”, Annals of Operational Re-
search, 116, 225-242.

[14] Annaert, J., van den Broeck, J., and Vennet R.V., (1999), “Determinants of Mutual
Fund Underperformance: A Bayesian Stochastic FrontierApproach”, paper presented
at the 6EWEPA, Copenhagen, Denmark.

[15] Aragon, Y., A. Daouia and C. Thomas-Agnan (2003), A conditional quantile-based

efficiency measure, Discussion paper, GREMAQ et LSP, Université de Toulouse,
(http://www.univ-tlse1.fr/GREMAQ/Statistique/mesure.pdf).

[16] Aragon, Y., A. Daouia and C. Thomas-Agnan (2005), Nonparametric frontier estima-
tion: A conditional quantile-based approach, Econometric Theory, 21, 358–389.

[17] Audretsch, D.B. and Feldman, M. (1996) “R&D spillovers and the geography of
innovation and production”, American Economic Review, 86(3), 630-640.

[18] Avveduto, S. (2002), “Human resources in science and technology”, paper presented
to the CISS Moncalieri Workshop, December 11.

[19] Banker, R.D. (1993) “Maximum likelihood, Consistency and Data Envelopment
Analysis: A Statistical Foundation”, Management Science, 39, 10, 1265-1273

[20] Banker, R.D., Chang H., and Cooper W.W. (1996), “Simulation studies of efficiency,
returns to scale, and misspecification with nonlinear functions in DEA”, Annals of
Operations Research, 66, 233-253.

[21] Banker, R.D., Charnes, A., and Cooper W.W. (1984), “Some Models for Estimating
Technical and scale inefficiencies in DEA”, Management Science, 32, 1613-1627.

[22] Banker, R.D. and Maindiratta, A. (1988), “Nonparametric Analysis of Technical and
Allocative Efficiens in Production”, Econometrica, 56, 1315-1332.

[23] Banker, R.D. and R.C. Morey (1986a), “Efficiency analysis for exogeneously fixed
inputs and outputs”, Operations Research, 34(4), 513–521.

[24] Banker, R.D. and R.C. Morey (1986b), “The use of categorical variables in Data
Envelopment Analysis”, Management Science, 32 (12), 1613-1627.

[25] Barnett, V. and Lewis T. (1995). Outliers in Statistical Data, Chichester, Wiley.

[26] Bartelsman, E.J. and Doms M. (2000), “Understanding productivity: lessons from
longitudinal microdata”, Journal of Economic literature, Vol. XXXVIII, pp.569-594.

[27] Barth, J.R., Nolle D.E. , and Rice T.N. (1997), “Commercial Banking Structure, Reg-
ulation, and Performance: An International Comparison”, in Papadimmitriou D.B.
(ed.), Modernizing Financial Systems, New York: Oxford University Press.

[28] Barth, J.R., Dan Brumbaugh, R.Jr., and Wilcox J.A. (2000), “Policy Watch: The Re-
peal of Glass- Steagall and the Advent of Broad Banking”, Journal of Economic
Perspectives 14, 191-204.
REFERENCES 223

[29] Basso, A., S. Funari, (2001), “A Data Envelopment Analysis Approach to Measure
the Mutual Fund Performance”, European Journal of Operational Research, 135 (3),
477-492.

[30] Berger, A. N., Kashyap A.K., and Scalise J. M. (1995), “The Transformation of the
U.S. Banking Industry: What A Long, Strange Trip It’s Been,” Brookings Papers on
Economic Activity, 2, 55-218.

[31] Berger, A.N. and Humphrey, D.B. (1997), “Efficiency of Financial Institutions: Inter-
national Survey and Directions for Future Research”, European Journal of Operational
Research 98, 175-213.

[32] Berger, A.N., Cummins, J.D. and Weiss, M.A. (1997), “The Coexistence of Multiple
Distribution Systems for Financial Services: The Case of Property-Liability Insur-
ance,” Journal of Business 70, 515-546.

[33] Bergson, A. (1961), National Income of the Soviet Union since 1928. Cambridge, MA:
Harvard University Press.

[34] Bertoletti P. (2005), “Elasticities of Substitution and Complementarity: A Synthesis”,

Journal of Productivity Analysis, 24(2), 183 - 196.

[35] Bessent, A. and Bessent, W.E. (1980). “Determining the comparative efficiency of
schools through DEA”, Educational Administration Quarterly, 16, 57-75.

[36] Bessent A., Bessent W., Kennington J., and Reagan B. (1982), “An Application of
mathematical programming to assess productivity in the Houston independent school
district”, Management Science, 28, 1355-1367.

[37] Bickel, P.J. and Freedman, D.A. (1981), “Some Asymptotic Theory for the Bootstrap”,
Annals of Statistics, 9, 1196-1217.

[38] Bjurek, H. Hjalmarsson, L. and Forsund, F.R. (1990), “Deterministic Parametric and
Nonparametric Estimation of Efficiency in Service Production: A Comparison”, Jour-
nal of Econometrics, 46, 213-228.

[39] Bogetoft, P. (2000), “DEA and Activity Planning under Asymmetric Information”,
Journal of Productivity Analysis, 13 (1), 7-48.

[40] Boles, J.N. (1966), “Efficiency Squared-Efficient Computation of Efficiency Indexes”,

Proceedings of the Thirty Ninth Annual Meeting of the Western Farm Economics
Associations, 137-142.

[41] Bonaccorsi A. and Daraio, C. (2003a), “A robust nonparametric approach to the analy-
sis of scientific productivity”, Research Evaluation, 12 (1), 47-69.

[42] Bonaccorsi A. and Daraio, C. (2003b), “Age effects in scientific productivity. The case
of the Italian National Research Council (CNR)”, Scientometrics, 58 (1), 47-88.

[43] Bonaccorsi A. and Daraio, C. (2004), “Econometric approaches to the analysis of

productivity of R&D systems. Production functions and production frontiers”, in H.F.
Moed, W. Glanzel and U. Schmoch (edited by), Handbook of Quantitative Science
and Technology Research, Kluwer Academic Publishers, 51-74.
224 REFERENCES

[44] Bonaccorsi A. and Daraio, C. (2005), “Exploring size and agglomeration effects on
public research productivity”, Scientometrics, Vol. 63, No. 1, 87-120.

[45] Bonaccorsi, A., Daraio C. and Simar, L. (2006), “Advanced Indicators of Productivity
of Universities. An application of Robust Nonparametric Methods to Italian data”,
Scientometrics, Vol. 66, No. 2, 389-410.

[46] Bressler, R.G. (1966), “The Measurement of Productive Efficiency”, Proceedings of

the Thirty Ninth Annual Meeting of the Western Farm Economics Associations, 129-
136.

[47] Briec, W., Kerstens, K., and Lesourd J.B., (2004), “Single Period Markowitz Port-
folio Selection, Performance Gauging and Duality: A variation on the Luenberger’s
Shortage Function”, Journal of Optimization Theory and Applications, 120(1), 1-27.

[48] Briec,W., Kerstens K. and Vanden Eeckaut P.,(2004a), “Non-convex Technologies and
Cost Functions: Definitions, Duality and Nonparametric Tests of Convexity”, Journal
of Economics, 81 (2), 155-192.

[49] Briec,W., Kerstens K. and Vanden Eeckaut P.,(2004b), “Nonparametric Disposability

Assumptions and A New Approach to Measure Congestion”, paper presented at the
21 JMA 2004, Lille.

[50] Briec, W., and Lesourd J.B., (2000), “The Efficiency of Investment Fund Management:
An Applied Stochastic Frontier Model, in: C.L. Dunis (ed.), Advances in Quantitative
Asset Management, Kluwer, Boston, 41-59.

[51] Cantarelli D. (2005), “Elasticities of Complementarity and Substitution in Some Func-

tional Forms. A comparative Review”, Working Paper no. 3.157, University of Salerno.

[52] Castells, M. and Hall, P. (1994), Technopoles of the world. The making of the 21st
century industrial complexes. London: Routledge.

[53] Caves, D.W., Christensen, L.R. and Diewert, W.E. (1982), “The Economic Theory of
Index Numbers of the Measurement of Input, Output and Productivity”, Economerica,
50, 1393-1414.

[54] Cazals, C., Florens, J.P. and Simar, L. (2002), “Nonparametric frontier estimation: a
robust approach”, Journal of Econometrics, 106, 1-25.

[55] Cesari, R. Panetta, F. (2002), “The performance of the Italian equity funds”, Journal
of Banking and Finance, 26, 99-126.

[56] Chambers, R.G., Chung Y. and Färe, R. (1996), “Benefit and distance functions”,
Journal of Economic Theory, 70, 407-419.

[57] Chambers, R. G., and Quiggin J.,(2000), Uncertainty, Production, Choice and Agency:
The State-Contingent Approach, Cambridge University Press, New York.

[58] Chambers, R. G. (2004), “Productivity Measurement Under Uncertainty”, invited pa-

per at the 21 JMA 2004, Lille.

[59] Charnes, A. and Cooper, W.W. (1985), “Preface to Topics in Data Envelopment Analy-
sis”, Annals of Operations Research, 2, 59-94.
REFERENCES 225

[60] Charnes, A. and Cooper, W.W. (1961), Management Models and Industrial Applica-
tions of Linear Programming, Wiley, New York.

[61] Charnes, A., Cooper, W.W., Lewin A.Y and Seiford L.M (1994), (edited by), Data En-
velopment Analysis. Theory, Methodology and Applications, Kluwer Academic Pub-
lishers, Norwell USA.

[62] Charnes, A., Cooper, W.W., and Rhodes, E. (1978), “Measuring the Efficiency of
Decision Making Units”, European Journal of Operational Research, 2, 429-444.

[63] Chavas, J.-P. and Cox, T.L. (1988), “A Nonparametric Analysis of Agricultural Tech-
nology”, American Journal of Agricultural Economics, 70, 303-310.

[64] Chavas, J.-P. and Cox, T.L. (1990), “A Non-Parametric Analysis of Productivity: The
Case of U.S. and Japanese Manufacturing”, American Economic Review, 80, 450-464.

[65] Cherchye, L., Kuosmanen T. P., and Post G.T., (2000) “What is the economic meaning
of FDH? A replay to Thrall”, Journal of Productivity Analysis, 13, 263-267.

[66] Cherchye, L., Kuosmanen, T. and Post, G.T. (2001), “Nonparametric Production
Analysis under Alternative Price Conditions”, CES Discussion Paper 01.05.

[67] Chevalier, J., and Ellison G. (1999a), “Are some Mutual Fund Managers Better Than
Others? Cross-sectional Patterns in Behavior and Performance”, Journal of Finance,
54 (3), 875-898.

[68] Chevalier, J., and Ellison G. (1999b), “Career concerns of Mutual Fund Managers”,
The Quarterly Journal of Economics, 114 (2), 389-432(44).

[69] Christensen L.R., Jorgenson D. W., and Lau L.J., (1973), “Transcendental Logarithmic
Production Frontiers”, Review of Economics and Statistics, 55, 28-45.

[70] Clark, G.L., Feldman M.P. and Gertler M.S., edited by, (2000), The Oxford Handbook
of Economic Geography, Oxford University Press, NY.

[71] Consiglio Nazionale delle Ricerche (1998), CNR Report 1998. CNR Roma.

[72] Coelli, T. (1996), Assessing the Performance of Australian Universities using Data
Envelopment Analysis, mimeo, Centre for Efficiency and Productivity Analysis, Uni-
versity of New England.

[73] Coelli, T. (2000), “On the econometric estimation of the distance function repre-
sentation of a production technology”, CORE Discussion Paper 2000-42, Université
Catholiqué de Louvain, and CEPA, School of Economic Studies, University of New
England.

[74] Coelli, T. and Perelman S., (1996), “Efficiency Measurement, Multiple-output Tech-
nologies and Distance Functions: with application to European Railways”, CREPP
Working Paper no. 96/05, Université de Liege, Belgium.

[75] Coelli, T. and Perelman S., (1999), “A comparison of parametric and nonparamet-
ric distance functions: with application to European railways”, European Journal of
Operational Research, 117, 326-339.
226 REFERENCES

[76] Coelli, T. and Perelman S., (2000), “Technical efficiency of European railways: a
distance function approach”, Applied Economics, 32, 1967-1976.

[77] Coelli, T., Rao, D.S.P. and Battese, G.E. (1998), An Introduction to Efficiency Analysis,
Kluwer Academic Publishers.

[78] Collins, P.M.D. (1991), Quantitative assessment of Departmental Research, SEPSU

Policy study no. 5, London: The Royal Society.

[79] Cooke, P. and Morgan K. (1998) The associational economy. Firms, regions and in-
novation. Oxford: Oxford University Press.

[80] Cooper, W.W., Li S., Seiford L.M., Tone K., Thrall R.M., and Zhu J. (2001), “Sensi-
tivity and Stability Analysis in DEA: Some Recent Developments”, Journal of Pro-
ductivity Analysis, 15, 217-246.

[81] Cooper, W.W., Seiford L.M., and Tone K. (2000), Data Envelopment Analysis: A
Comprehensive Text with Models, Applications, References and DEA-Solver Soft-
ware,Kluwer Academic Publishers, Boston.

[82] Cornwell C., Schmidt P., and Sickles R. C. (1990), “Productivity Frontiers with cross
sectional and time series variation in efficiency levels”, Journal of Econometrics, 46,
185-200.

[83] Cummins, J.D., Turchetti, G. and Weiss, M. (1996), “Productivity and Technical Ef-
ficiency in the Italian Insurance Industry,” Working paper, Wharton Financial Institu-
tions Center, University of Pennsylvania, Philadelphia.

[84] Cummins, J.D. and Weiss, M.A. (2001), “Analyzing Firm Performance in the Insurance
Industry Using Frontier Efficiency Methods,” in Georges Dionne (ed.), Handbook of
Insurance Economics, Boston: Kluwer Academic Publishers.

[85] Cummins, J.D., Weiss, M.A., and Zi, H. (1999), “Organizational Form and Efficiency:
An Analysis of Stock and Mutual Property-Liability Insurers”, Management Science
45, 1254-1269.

[86] Daniel, H.D. and Fisch, R. (1990), “Research performance evaluation in the German
university sector”, Scientometrics, Vol. 19 (5-6), 349-361.

[87] Dantzig, G.B. (1963), Linear Programming and its Extensions, Princeton University
Press, Princeton.

[88] Daouia, A. and L. Simar (2004), Nonparametric Efficiency Analysis: A Multivariate

Conditional Quantile Approach, Discussion paper #0419, Institut de Statistique, Uni-
versité Catholique de Louvain, Louvain-la-Neuve, Belgium, forthcoming in Journal
of Econometrics.

[89] Daouia, A. and L. Simar (2005), Robust Nonparametric Estimators of Monotone

Boundaries, Journal of Multivariate Analysis, 96, 311–331.

[90] Daouia, A., J.P. Florens and L. Simar (2005), Functional Convergence of Quantile-
type Frontiers with Application to Parametric Approximations, Discussion paper 0538,
Institut de Statistique, UCL.
REFERENCES 227

[91] Daraio C. (2003), Comparative Efficiency and Productivity Analysis based on non-
parametric and robust nonparametric methods. Methodology and Applications, Doc-
toral dissertation, Scuola Superiore Sant’Anna, Pisa (Italy).

[92] Daraio C., Simar, L. (2004), “A Robust Nonparametric Approach to Evaluate and Ex-
plain the Performance of Mutual Funds”, Discussion Paper no. 0412, Institut de Sta-
tistique, UCL, Belgium, forthcoming in European Journal of Operational Research.

[93] Daraio C. and Simar, L. (2005a), “Introducing Environmental Variables in Nonpara-

metric Frontier Models: a Probabilistic Approach”, Journal of Productivity Analysis,
24 (1), pp. 93-121.

[94] Daraio C. and Simar, L. (2005b), “Conditional Nonparametric Frontier Models for
Convex and Non Convex Technologies: A unifying Approach”, Discussion Paper no.
0502, Institut de Statistique, UCL, Belgium, forthcoming in Journal of Productivity
Analysis.

[95] David, P.A. (1995), “Positive Feedbacks and Research Productivity in Science: Re-
opening another Black Box”, in Granstrand O. (eds), Economics of Technology, North-
Holland, Amsterdam.

[96] de Alessi, L. (1974) “An economic analysis of government ownership and regulation:
theory and the evidence from the electric power industry”, Public Choice, 19:1, 1-42.

[97] de Alessi, L. (1983) “Property rights, transaction costs, and X-efficiency: an essay in
economic theory”, American Economic Review, 73:1 (March), 64-81.

[98] DEAsoft, software developed and marketed by Performance Improvement Managment

Ltd, website: www.deasoftware.co.uk .

[99] Debreu, G. (1951), “The Coefficient of Resource Utilization”, Econometrica, 19, 273-
292.

[100] Deckle, R. (1988), “The Japanese ’Big Bang’ financial reforms and marketing impli-
cations”, Journal of Asian Economics, 9, 237-249.

[101] Deprins, D. and Simar, L. (1983), “On Farrel Measures of Technical Efficiency”,
Recherches économiques de Louvain, 49, 123-137.

[102] Deprins, D. and L. Simar (1985), A Note on the Asymptotic Relative Efficiency of the
M.L.E. in a Linear Model with Gamma Disturbances, Journal of Econometrics, 27,
383–386.

[103] Deprins, D. and Simar, L. (1988), “Mesures d’efficacité de réseaux de chemin de

fer”, in Gestion de l’ économie et de l’entreprise. L’approche Quantitative, Edited by
CORE, De Boeck-Wesmael, Bruxelles, pp. 321-344.

[104] Deprins, D., Simar L. and Tulkens H. (1984), “Measuring labor-efficiency in post
offices”, in Marchand, M., Pestieau, P. and Tulkens, H. (eds.) The Performance of
public enterprises - Concepts and Measurement, Amsterdam, North-Holland, 243-
267.

[105] Detzel L.F., R. Weigand (1998), “Explaining Persistence in Mutual Fund Perfor-
mance”, Financial Services Review, 7 (1), 45-55.
228 REFERENCES

[106] Diewert, W.E. and Parkan, C. (1983), “Linear Programming Tests of Regularity Condi-
tions for Production Frontiers”, in Eichhorn, W., Henn, R., Neumann, K. and Shephard,
R.W. (eds.) Quantitative Studies on Production and Prices, Wuerzburg and Vienna,
Physica-Verlag.

[107] Diewert, W E, (1992), “The Measurement of Productivity”, Bulletin of Economic

Research, vol. 44(3), 163-198.

[108] Dorfman R., Samuelson P. and Solow R. (1958), Linear Programming and Economic
Analysis, McGraw Hill Text.

[109] Efron, B. (1979), “Bootstrap Methods: another look at the Jackknife”, The Annals of
Statistics, Vol. 7, No.1, 1-26.

[110] Efron, B., and Tibshirani, R.J. (1993), An introduction to the Bootstrap, Chapman and
Hall, NY.

[111] Emrouznejad A.(2001), “An extensive Bibliogaphy of Data Envelopment Analysis

(DEA) Volume I-V”, http://www.warwick.ac.uk/ ∼ bsrlu.

[112] Färe, R. (1975), “Efficiency and the Production Function”, Zeitschrift fuer Nation-
aloekonomie, 35, 317-324.

[113] Färe, R., Grosskopf, S., Lindgren,B. and Roos, P. (1989), “Productivity Developments
in Swedish Hospitals: A Malmquist Output Index Approach”, in Charnes, A, Cooper,
W.W., Lewin, A. and Seiford, L. (eds.), Data Envelopment Analysys: Theory, Method-
ology and Applications, Boston: Kluwer Academic Publishers.

[114] Färe, R., Grosskopf, S. and Lovell, C.A.K. (1985), The Measurement of Efficiency of
Production, Boston: Kluwer-Nijhoff Publishing.

[115] Färe, R., Grosskopf, S. and Lovell, C.A.K. (1992), “Indirect Productivity Measure-
ment”, Journal of Productivity Analysis, 2, 283-298.

[116] Färe, R., Grosskopf, S. and Lovell, C.A.K. (1994), Production Frontiers, Cambridge
University Press, Cambridge.

[117] Färe, R., Grosskopf, S. and Russell, R.R. (1998), Index Numbers: Essays in Honour
of Sten Malmquist, Boston: Kluwer Academic Publishers.

[118] Färe, R., S. Grosskopf, C.A. K. Lovell and C. Pasurka (1989), “Multilateral Productiv-
ity Comparisons when some Outputs are Undesirable: a Nonparametric Approach”,
Review of Economics and Statistics 71 (1), 90-98.

[119] Färe, R., Grosskopf, S. and Weber, W. (1989), “Measuring school district perfor-
mance”, Public Finance Quarterly, 17, 409-428.

[120] Färe, R., Grosskopf, S., (2004), New Directions: Efficiency and Productivity, Kluwer
Academic Publishers.

[121] Färe, R. and Lovell, C.A.K. (1978), “Measuring the Technical Efficiency of Produc-
tion”, Journal of Economic Theory, 19, 150-162.

[122] Färe, R. and Primont, D. (1996), “The opportunity cost of duality”, Journal of Pro-
ductivity Analysis, 7, 213-224.
REFERENCES 229

[123] Färe, R. and V. Zelenyuk (2003), “On Aggregate Farrell Efficiency Scores”, European
Journal of Operations Research 146:3, 615-620.

[124] Farrell, M.J. (1957), “The measurement of the Productive Efficiency”, Journal of the
Royal Statistical Society, Series A, CXX, Part 3, 253-290.

[125] Farrell, M.J. (1959), “Convexity assumption in theory of competitive markets”, Journal
of Political Economy, 67, 377-391.

[126] Farrell, M.J. and Fieldhouse, M. (1962), “Estimating Efficient Production Functions
under Increasing Return to Scale”, Journal of the Royal Statistical Society, Series A,
CXXV, Part 2, 252-267.

[127] Feldman, M.P. (2000), “ Location and Innovation: the New Economic Geography of
Innovation, Spillovers, and Agglomeration”, in Clark, G.L., Feldman M.P. and Gertler
M.S., edited by, , The Oxford Handbook of Economic Geography, Oxford University
Press, NY, 373-394.

[128] Ferrier G.D., and Hirschberg J. G. (1997), “Bootstrapping Confidence Intervals for
Linear Programming Efficiency Scores: with an illustration using Italian Bank Data”,
The Journal of Productivity Analysis, 8, 19-33.

[129] Ferrier G.D., and Hirschberg J. G. (1999), “Can we bootstrap DEA scores?”, The
Journal of Productivity Analysis, 11.

[130] Fischer, I. (1922), The Making of Index Numbers, Boston: Houghton Mifflin.

[131] Florens, J.P. and Simar, L. (2005), “Parametric Approximations of Nonparametric

Frontier”, Journal of Econometrics, 124, 91-116.

[132] Forsund, F.R., and N. Sarafoglou (2002), “On the origins of Data Envelopment Analy-
sis”, Journal of Productivity Analysis, 17 (1/2), 23–40.

[133] Fortin R., S. Michelson, and J. Wagner (1999), “Does Mutual Fund Manager Tenure
Matter?” Journal of Financial Planning, August, 12.

[134] Freedman, D.A. (1981), “Bootstrapping regression models”, Annals of Statistics, Vol
9, 6, 1218–1228.

[135] Fried, H.O, Lovell, C.A.K. and Schmidt S.S. (1993), edited by, The measurement
of Productive Efficiency. Techniques and Applications, New York Oxford, Oxford
University Press.

[136] Fried, H.O, Lovell, C.A.K. and Schmidt S.S. (2006), edited by, The Measurement of
Productive Efficiency, 2nd Edition, New York Oxford, Oxford University Press.

[137] Fried, H.O., C.A.K. Lovell, S.S. Schmidt and S. Yaisawarng (2002), “Accounting for
environmental effects and statistical noise in Data Envelopment Analysis”, Journal of
Productivity Analysis, 17 (1/2), 157–174.

[138] Fried, H.O., S.S. Schmidt and S. Yaisawarng (1999), “Incorporating the operating
environment into a nonparametric measure of technical efficiency”, Journal of Pro-
ductivity Analysis, 12, 249–267.
230 REFERENCES

[139] Garfield, E. and Dorof, W.A. (1992), “Citation data: their use as quantitative indicators
for science and technology evaluation and policy making”, Science and Public Policy,
19 (5), 321-327.

[140] Gattoufi, S., M. Oral, and A. Reisman (2004), “Data envelopment analysis literature:
A bibliography update (1951-2001)”, Journal of Socio-Economic Planning Sciences,
Vol.38, Issues 2-3, 159-229.

[141] Gijbels, I., Mammen, E., Park, B.U., and Simar L. (1999), “On Estimation of Monotone
and Concave Frontier Functions”, Journal of the American Statistical Association, 94,
220-228.

[142] Girod, O. and Triantis, K. (1999), “The Evaluation of Productive Efficiency Useing
a Fuzzy Mathematical Programming Approach: The Case of the Newspaper Preprint
Insertion Process”, IEEE Transactions on Engineering Management, 46, 1-15.

[143] Golec, J. (1996), “The Effects of Mutual Fund Managers’ Characteristics on Their
Portfolio Performance, Risk and Fees”, Financial Services Review, 5 (2), 133-148.

[144] Goto, I. (1999), “Japan: The Finalization of the Big Bang,” International Financial
Law Review (July).

[145] Greene, W. H. (1980), “Maximum likelihood estimation of econometric frontier func-

tions”, Journal of Econometrics, 13 (1), 27–56.

[146] Greene, W.H. (1990), “A Gamma-Distributed Stochastic Frontier Model”, Journal of

Econometrics, 46:1/2, 141-164.

[147] Greene, W.H. (1993),The econometric approach to efficiency analysis, in H. Fried, K.

Lovell and S. Schmidt (eds) The Measurement of Productive Efficiency: Techniques
and Applications, Oxford University Press.

[148] Grinblatt, M. and Titman S. (1995), “Performance Evaluation”, in R.Jarrow, V. Mak-

simovic, W.T. Ziemba (eds.), Handbooks in OR & MS: Finance, vol 9, Amsterdam,
Elsevier, 581-609.

[149] Grosskopf S., Margaritis D. and Valdmanis V. (1995), “Estimating output substitutabil-
ity of hospital services: A distance function approach”, European Journal of Opera-
tional Research, 80, 575-587.

[150] Grosskopf, S.(1996), “Statistical Inference and Nonparametric Efficiency: A selective

Survey”, The Journal of Productivity Analysis, 7, 161-176.

[151] Grosskopf, S., Hayes, K., Taylor, L. and Weber, W. (1997), “Budget constrained fron-
tier measures of fiscal equity and efficiency in schooling, Review of Economics and
Statistics, 79, 116-124.

[152] Grosskopf, S. (2003), “Some Remarks on Productivity and its Decompositions”, Jour-
nal of Productivity Analysis, 20, 459-474.

[153] Grosskopf, S. and Moutray, C. (2001) “Evaluating performance in Chicago public

high schools in the wake of decentralization”, Economics of Education Review, 20,
1-14.
REFERENCES 231

[154] Grosskopf, S., Hayes, K., Taylor L. and Weber W. (1999) “Anticipating the conse-
quences of school reform: a new use of DEA”, Management Science, 45, 608-620.

[155] Hall, P. and Simar, L. (2002), “Estimating a changepoint, boundary or frontier in the
presence of observation error”, Journal of the American Statistical Association, 97,
523-534.

[156] Halme, M., Joro, T., Korhonen, P., Salo, S. and Wallenius, J. (2000), “Value efficiency
analysis for incorporating preference information in DEA”, Management Science, 45,
103-115.

[157] Hanoch, G. and Rotschild, M. (1972), “Testing the Assumptions of Production Theory:
A Nonparametric Approach”, Journal of Political Economy, 80, 256-275.

[158] Hansmann, H. (1988) “Ownership of the firm”, Journal of Law, Economics and Or-
ganization, 4:2 (Fall), 267-304.

[159] Härdle, W. (1990), Applied Nonparametric Regression, Cambridge University Press,

Cambridge.

[160] Härdle, W. and L. Simar (2003), Applied Multivariate Statistical Analysis, Springer-
Verlag, Berlin.

[161] Harker P.T. and S.A. Zenios (2000), edited by, Performance of Financial institutions.
Efficiency, Innovation, Regulation, Cambridge University Press, UK.

[162] Henderson, D.J. and L. Simar (2005), “A Fully Nonparametric Stochastic Frontier
Model for Panel Data”, Discussion paper 0525, Institut de Statistique, UCL.

[163] Hess, T. and Trauth, T. (1998), “Towards A Single European Insurance Market”,
International Journal of Business 3, 89-100.

[164] Hicks, J.R. (1935), “The theory of Monopoly: A Survey”, Econometrica, 3:1 (January),
1-20.

[165] Hogan, A.M.B. (1995), “Regulation of the Single European Insurance Market”, Jour-
nal of Insurance Regulation, 13, 329-358.

[166] Holbrook J.A.D. (1992a), “Basic indicators of scientific and technological perfor-
mance”, Science and Public Policy, 19 (5), 267-273.

[167] Holbrook J.A.D. (1992b), “Why measure science? ”, Science and Public Policy, 19
(5), 262-266.

[168] Holmstrom, B. R., and J. Tirole (1989), “The theory of the firm”, in R. Schmalensee
and R. D. Willig, eds., Handbook of Industrial Organization, Volume I. Amsterdam:
Elsevier Science Publishers.

[169] Ippolito, R.A. (1993), “On Studies of Mutual Fund Performance, 1962-1991”, Finan-
cial Analysts Journal, 49 (1), 42-50.

[170] Jaffe, A.B., Trajtenberg M. and Henderson R. (1993) “Geographic localization of

knowledge spillovers as evidenced by patent citations”, Quarterly Journal of Eco-
nomics, 577-598.
232 REFERENCES

[171] Jeong, S.O. , B. U. Park and L. Simar (2006), “Nonparametric conditional efficiency
measures: asymptotic properties”. Discussion paper 0604, Institut de Statistique, UCL.

[172] Jeong, S.O. and L. Simar (2005), “Linearly interpolated FDH efficiency score for
nonconvex frontiers”, Discussion 0501, Institut de Statistique, UCL.

[173] Johnston R. (1994), “Effects of resource concentration on research performance”,

Higher Education, 28, 25-37.

[174] Kao, C. and Liu, S.T. (1999), “Fuzzy Efficiency Measures in Data Envelopment Analy-
sis”, Fuzzy Sets and Systems, forthcoming.

[175] Khorana, A.(1996) “Top Management Turnover: An Empirical Investigation of Mutual

Fund Managers”, Journal of Financial Economics, 40, 403-427.

[176] King, D. (2004), “The Scientific impact of Nations”, Nature, 430, 311-316.

[177] Kneip, A. and Simar, L. (1996), “A general framework for frontier estimation with
panel data”, The Journal of Productivity Analysis, 7, 187-212.

[178] Kneip, A., Park B.U. and Simar L. (1998), “A Note on the Convergence of Nonpara-
metric DEA Estimators for Production Efficiency Scores”, Econometric Theory, 14,
783-793.

[179] Kneip, A., Simar, L. and Wilson, P.W. (2003), “Asymptotics for DEA Estimators in
Nonparametric Frontier Models”, Discussion Paper no. 0317, Institut de Statistique,
UCL.

[180] Koopmans, T.C. (1951), “An Analysis of Production as an Efficient Combination of

Activities”, in Koopmans T.C. (Ed.), Activity Analysis of Production and Allocation,
Cowles Commission for Research in Economics, Monograph No. 13, Wiley, New
York.

[181] Koopmans, T.C. (1957), Three Essays on the State of Economic Science. New York:
McGraw Hill.

[182] Korhonen, P., Tainio, R. and Wallenius J. (2001), “Value efficiency analysis of acad-
emic research”, European Journal of Operational Research, 130, 121-132.

[183] Kostoff R.N. (1994), “Performance measures for government-sponsored research:

overview and background”, Scientometrics, Vol. 36, n. 3.

[184] Korostelev, A., Simar L. and Tsybakov A.B. (1995), “Efficient estimation of monotone
boundaries”, The Annals of Statistics, 23, 476-489.

[185] Krugman P. (1991), “Increasing returns and economic geography”, Journal of Political
Economy, 99(3), 483-499.

[186] Kumbhakar S. C., Lovell C.A.K. (2000), Stochastic Frontier Analysis, Cambridge
University Press, UK.

[187] Kumbhakar, S.C. , Park, B.U., Simar, L. and E.G. Tsionas (2004), “Nonparametric
stochastic frontiers: a local likelihood approach”, Discussion paper 0417, Institut de
Statistique, UCL, forthcoming in Journal of Econometrics.
REFERENCES 233

[188] Kuosmanen, T. and Post, G.T. (2001), “Measuring Economic Efficiency with Incom-
plete Price Information: With an Application to European Commercial Banks”, Euro-
pean Journal of Operational Research, 134 (1), 43-58.

[189] Land, K.C., Lovell C.A.K., and Thore S. (1993), “Chance- Constrained Data Envel-
opment Analysis”, Managerial and Decision Economics, 14 (6), 541-554.

[190] Laredo, P., Mustar P. (eds), (2001), Research and Innovation policies in the new global
economy. An international comparative analysis, Edward Elgar.

[191] Leibenstein, H. (1966), “Allocative efficiency vs. ’X-efficiency”’, American Economic

Review 56 (3), 392-415.

[192] Leibenstein, H. (1975), “Aspects of the X-efficiency theory of the firm”, Bell Journal
of Economics, 6, 580-606.

[193] Leibenstein, H. (1976), Beyond economic man. Cambridge, MA: Harvard University
Press.

[194] Leibenstein, H. (1978), “X-Inefficiency exists - Reply to an Xorcist”, American Eco-

nomic Review 68 (1), 203-211.

[195] Leibenstein, H. (1987), Inside the firm. Cambridge, MA: Harvard University Press.

[196] Lemak D., P. Satish (1996), “Mutual Fund Performance and Managers’ Terms of
Service: Are There Performance Differences?”, The Journal of Investing, Winter ,pp.
59-63.

[197] Leontief, W.W. (1941), The Structure of the American Economy 1919-1939. New
York: Oxford University Press.

[198] Leontief, W.W. (1953), Studies in the Structure of the American Economy. New York:
Oxford University Press.

[199] Levin, S.G. and P. E. Stephan (1991), “Research productivity over the life cycle:
evidence for academic scientists”, American Economic Review, 81 (1), March, 114-
32.

[200] Lewin, A. Y. and Minton J.W. (1986), “Determining Organizational Effectiveness:

Another Look and an Agenda for Research”, Management Science, 32 (5), 514-538.

[201] Lewison, G. (1998), “New bibliometric techniques for the evaluation of medical
schools”, Scientometrics, Vol. 41, 5-16.

[202] Li, X.B. and Reeves, G.R. (1999), “A multiple criteria approach to data envelopment
analysis”, European Journal of Operational Research, 115, 507-517.

[203] Lindsay, C.M. (1976), “A theory of government enterprise”, Journal of Political Econ-
omy, 84, 1061-1077.

[204] Lovell, C.A.K. (1993), “Production Frontiers and Productive Efficiency,” in H.O.
Fried, C.A.K. Lovell, and S.S. Schmidt, eds., The Measurement of Productive Ef-
ficiency, New York, Oxford University Press.
234 REFERENCES

[205] Lovell, C.A.K. (2001), “Future Research Opportunities in Efficiency and Productivity
Analysis”, in A. Alvarez (ed.), La Medición de la Eficiencia Productiva, Pirámide.

[206] Lovell, C.A.K. (2003), “The Decomposition of Malmquist Productivity Indexes”,

Journal of Productivity Analysis, 20, 437-458.

[207] Luenberger, D.G. (1992), “Benefit functions and duality”, Journal of Mathematical
Economics, 21, 461-481.

[208] Luwel, M. (2004),“The Use of Input Data in the Performance Analysis of R&D Sys-
tems. Potentialities and Pitfalls”, in H.F. Moed, W. Glanzel and U. Schmoch (edited
by), Handbook of Quantitative Science and Technology Research, Kluwer Academic
Publishers, 315-338.

[209] Malmquist, S. (1953), “Index Numbers and Indifference Surfaces”, Trabajos de Esta-
tistica, 4, 209-242.

[210] Markowitz, H. (1952), “Portfolio Selection”, Journal of Finance, 77-91.

[211] Markowitz, H. (1959), Portfolio Selection. Efficiency diversification of investments,

Cowles Foundation for research in Economics at Yale University, John Wiley and
sons, Inc.

[212] Martin, B.R. (1996), “The use of multiple indicators in the assessment of basic re-
search”, Scientometrics, Vol. 36, 343.

[213] Martin, S. (2002), Advanced industrial economics, Blackwell Publishers, Malden.

[214] Mas-Colell A., Whinston M.D., Green J.R. (1995) Microeconomic Theory, Oxford
University Press, USA.

[215] May R. (1993), “The scientific wealth of nations”, Science, 275, 793-796.

[216] Meeusen, W., Van den Broeck J. (1977), “Efficiency Estimation from Cobb-Douglas
Production Functions With Composed Error”, International Economic Review, 18,
435-444.

[217] Merton, R.K. (1968), “The Matthew effect in science”, Science, 159 Jan-Mar, 56-63.

[218] Milgrom P., Roberts J. (1992), Economics, organization and management, Prentice
Hall, Englewood Cliffs.

[219] Moed H.F., W. Glanzel and U. Schmoch (2004)(edited by), Handbook of Quantitative
Science and Technology Research, Kluwer Academic Publishers.

[220] Moed, H.F., van Leeuwen T.N. (1996), “Impact factors can mislead”, Nature, 381:186.

[221] Moorsteen, R.H. (1961), “On measuring Productive Potential and Relative Efficiency”,
Quarterly Journal of Economics, 75 (3), 451-467.

[222] Morey, M.R., Morey, R.C. (1999), “Mutual fund performance appraisals: a multi-
horizon perspective with endogenous benchmarking”, Omega, Int. J. Mgmt Sci., 27,
241-258.
REFERENCES 235

[223] Morroni, M. (2006), Knowledge, Scale and Transactions in the Theory of the Firm,
forthcoming Cambridge University Press, Cambridge.

[224] Mouchart M., and L. Simar (2002), “Efficiency Analysis of Air Controllers: First
Insights”, Consulting Report No. 0202, Institut de Statistique, UCL, Belgium.

[225] Mullins N., Snizek W., Oehler K. (1988), The structural analysis of a scientific paper,
in Van Raan A.F.J., Handbook of Quantitative Studies of Science and Technology, pp.
81-105, North Holland, Amsterdam.

[226] Murthi, B., Choi, Y. and Desai, P. (1997), “Efficiency of Mutual Funds and Portfolio
Performance Measurement: a Nonparametric Measurement”, European Journal of
Operational Research, 98, 408-418.

[227] Nadaraya, E.A. (1964), “On estimating regression”, Theory of Probability Applica-
tions, 9, 141-142.

[228] Narin F. (1987), “Bibliometric techniques in the evaluation of research programs”,

Science and Public Policy, Vol. 14, n. 1, pp. 99-106.

[229] Narin, F., Hamilton, K.S. (1996), “Bibliometric performance measures”, Scientomet-
rics, Vol. 36, 293-310.

[230] Narin F., Olivastro D., Stevens K. A. (1994), “Bibliometrics - Theory, Practice and
Problems”, Evaluation Review, Vol. 18, n. 1.

[231] Niskanen, W.A. Jr. (1971), Bureaucracy and representative government. Chicago,
Aldine Press.

[232] Okubo Y. (1997), “Bibliometric indicators and analysis of research systems: methods
and examples”, STI Working Papers 1997/1, OECD, Paris.

[233] Olesen, O.B., and Petersen, N.C. (1995), “Change constrained Efficiency Evaluation”,
Management Science, 41(3), 442-457.

[234] Park, B.U. and Simar, L. (1994), “Efficient Semiparametric Estimation in a Stochastic
Frontier Model”, Journal of the American Statistical Association, 89,no. 427, 929-935.

[235] Park, B.U. Sickles, R.C., and Simar, L. (1998), “Stochastic panel frontiers: A semi-
parametric approach”, Journal of Econometrics, 84, 273-301.

[236] Park, B.U., R. Sickles and L. Simar (2003a), “Semiparametric Efficient Estimation of
AR(1) Panel Data Models”, Journal of Econometrics, vol 117, 2, 279-309. Corrigen-
dum to “Semiparametric-efficient estimation of AR(1) panel data models”, Journal of
Econometrics, vol 117, 2, 311.

[237] Park, B.U., R. Sickles and L. Simar (2003b), “Semiparametric Efficient Estimation in
Dynamic Panel Data Models”, Discussion Paper 0315, Institut de Statistique, UCL,
forthcoming in Journal of Econometrics.

[238] Park, B.U., Simar, L. and Weiner C. (2000), “The FDH Estimator for Productivity
Efficiency Scores: Asymptotic Properties”, Econometric Theory, 16, 855-877.
236 REFERENCES

[239] Park, B.U., L. Simar and V. Zelenyuk (2006), “Local Likelihood Estimation of Trun-
cated Regression and Its Partial Derivatives: Theory and Application”. Discussion
paper 0606, Institut de Statistique, UCL.

[240] Pedraja-Chaparro, R., Salinas-Jimenes, J., Smith, J. and Smith, P. (1997), “On the role
of weight restrictions in DEA”, Journal of Productivity Analysis, 8, 215-230.

[241] Perelman S., Santin D. (2005), “Measuring educational efficiency at student level with
parametric stochastic distance functions: An application to Spanish PISA results”,
paper presented at the X EWEPA, Brussels, June-July 2005.

[242] Porrini D. (2004), “Information exchange as Collusive Behaviour: Evidence from an

Antitrust Intervention in the Italian Insurance Market”, Geneva Papers on Risk and
Insurance, 29 (2), 219-233.

[243] Porter G., J. Trifts, (1998) “Performance Persistence of Experienced Mutual Fund
Managers”, Financial Services Review, 7 (1), 57-68.

[244] Pyke, F. Becattini G., Sengenberger W. (1986), “Industrial districts and inter-firm
co-operation in Italy”, International Labour Office, Geneve.

[245] Ramsden P. (1994), “Describing and explaining research productivity”, Higher Edu-
cation, Vol. 28.

[246] Ray, S.C. and Bhadra D. (1993), “Nonparametric Tests of Cost Minimizing Behavior:
A Study of Indian Farms”, American Journal of Agricultural Economics, 73 (Nov),
990-999.

[247] Ray, S.C. (2004), Data Envelopment Analysis, Theory and Techniques for Economics
and Operations Research, Cambridge University Press, US.

[248] Richmond, J. (1974), “Estimating the Efficiency of Production”, International Eco-

nomic Review, 515-521.

[249] Ritter, C. and Simar, L. (1997), “Pitfalls of Normal-Gamma Stochastic Frontier Mod-
els”, Journal of Productivity Analysis, 8(2), 167-182.

[250] Rosenberg N. (1991), “Critical issues in science policy research”, Science and Public
Policy, Vol. 18, n. 6, pp. 335-346.

[251] Rousseau, S. and Rousseau, R. (1997), “Data Envelopment Analysis as a tool for
constructing scientometric indicators”, Scientometrics, Vol. 40, 45-56.

[252] Rousseau, S. and Rousseau, R. (1998), “The scientific wealth of European nations:
taking effectiveness into account”, Scientometrics, Vol. 42, 75-87.

[253] Russell, RR. (1985), “Measures of Technical Efficiency”, Journal of Economic Theory,
35 (1), 109-126.

[254] Russell, RR. (1988), “On the Axiomatic Approach to the Measurement of Technical
Efficiency”, in W. Eichhorn, ed. (1988), Measurement in Economics: Theory and
Applications of Economic Indices, Heidelberg: Physica-Verlag.

[255] Russell, RR. (1990), “Continuity of Measures of Technical Efficiency”, Journal of

Economic Theory, 51 (2), 255-267.
REFERENCES 237

[256] Samuelson, P.A. (1948), “Consumption Theory in Terms of Revealed Preference”,

Economica, 15, 243-253.

[257] Saxenian A. (1996), Regional advantage. Culture and competition in Silicon Valley
and Route 128, Boston, Harvard University Press.

[258] Scherer, F.M. (1980), Industrial market structure and economic performance,
Houghton Mifflin, Boston.

[259] Schmidt, P. (1976), “On the Statistical Estimation of Parametric Frontier Production
Functions”, Review of Economics and Statistics, 58, 238-239.

[260] Schmidt P. and Sickles R. C. (1984), “Productivity Frontiers and Panel Data”, Journal
of Business and Economic Statistics, 2, 367-374.

[261] Schmoch U. (2004), “The Technological Output of Scientific Institutions”; in H.F.

Moed, W. Glanzel and U. Schmoch (edited by), Handbook of Quantitative Science
and Technology Research, Kluwer Academic Publishers.

[262] Schubert A., Braun T. (1993), “Reference standards for citation based assessments”,
Scientometrics, Vol. 26, n. 1, pp. 21-35.

[263] Schubert, A. and Braun, T. (1996), “Cross-field normalization of scientometric indi-

cators”, Scientometrics, Vol. 36, 311-324.

[264] Schubert A., Glanzel W., Braun T. (1988), “Against absolute methods: relative sci-
entometric indicators and relational charts as evaluation tools”, in Van Raan A.F.J.,
Handbook of Quantitative Studies of Science and Technology, pp. 137-176.

[265] Schuster, E.F. (1985), “Incorporating Support Constraints into Nonparametric Estima-
tors of Densities”, Communication in Statistics - Theory and Methods, 14, 1123-1136.

[266] Scott D. W. (1992), Multivariate Density Estimation, Theory, Practice and Visualiza-
tion, John Wiley & Sons, NY.

[267] Scott, A.J. (ed.) (2001) Global city-Regions. Oxford, Oxford University Press.

[268] Seaver, B. and Triantis, K. (1992), “A Fuzzy Clustering Approach Used in Evaluating
Technical Efficiency Measures in Manufacturing”, Journal of Productivity Analysis,
3, 337-363.

[269] Seglen, P.O. (1997), “Why the impact factor of journals should not be used for evalu-
ating research”, BMJ, 314: 498-502.

[270] Seiford, L.M. (1994), “A DEA bibliography 1978-1992”, in Charnes A., Cooper W.W.,
Lewin A., and Seiford L. (eds.), Data Envelopment Analysis: Theory, Methodology,
Applications, Kluwer Academic Publishers, 437-470.

[271] Seiford, L.M. (1996), “Data Envelopment Analysis: The Evolution of the State-of-
the-Art (1978-1995)”, Journal of Productivity Analysis, 7, 99-138.

[272] Seitz J. K.(1966), “Efficiency Measures for Steam-Electric Generating Plants”, Pro-
ceedings of the Thirty Ninth Annual Meeting of the Western Farm Economics Associ-
ations, 1966, 143-151.
238 REFERENCES

[273] Seitz J. K.(1971), “Productive Efficiency in the Steam-Electric Generating Industry”,

Journal of Political Economy 79, 878-886.

[274] Sengupta J.K.(1991), “Maximum Probability Dominance and Portfolio Theory”, Jour-
nal of Optimization Theory and Applications, 71, 341-357.

[275] Sengupta, J.K. (1992), “A Fuzzy Systems Approach in Data Envelopment Analysis”,
Computers and Mathematical Applications, 24, 259-266.

[276] Sengupta, J.K., and Park, H.S. (1993), “Portfolio Efficiency tests Based on Stochastic
Dominance and Cointegration”, International Journal of Systems Science, 24, 2135-
2158.

[277] Sengupta, J.K. (1994), “Measuring Dynamic Efficiency Under Risk Aversion”, Euro-
pean Journal of Operational Research, 74, 61-69.

[278] Sengupta, J.K. (1995), Dynamics of Data Envelopment Analysis. Theory of Systems
Efficiency, Kluwer Academic Publishers, Dordrecht.

[279] Sengupta, J.K. (2000), Dynamic and Stochastic Efficiency Analysis, Economics of
Data Envelopment Analysis, World Scientific, Singapore.

[280] Sheather S.J., and Jones M.C. (1991), “A relyable data-based bandwidth selection
method for kernel density estimation”, Journal of the Royal Statistical Society, Series
B, 53:3, pp. 683-690.

[281] Shephard, R.W. (1953). Cost and Production Functions. Princeton, NJ: Princeton
University Press.

[282] Shephard, R.W. (1970). Theory of Cost and Production Function. Princeton, NJ:
Princeton University Press.

[283] Shephard, R.W. (1974). Indirect Production Functions. Princeton, NJ: Princeton Uni-
versity Press.

[284] Shukla, R. and Trzcinca C. (1992), “Performance Measurement of Managed Portfo-

lios”, Financial Markets, Institutions and Instruments, 1, 1-59.

[285] Sickles R. C. (2005), “Panel estimators and the identification of firm-specific effi-
ciency levels in parametric, semiparametric and nonparametric settings”, Journal of
Econometrics, 126, 305-334.

[286] Silverman, B.W. (1986). Density Estimation for Statistics and Data Analysis, Chapman
and Hall, London.

[287] Silverman B.W., and Young G.A. (1987), “The Bootstrap: Smooth or Not to Smooth?”,
Biometrika, 74, 469-479.

[288] Simar L., (1992), “Estimating Efficiencies from Frontier Models with Panel Data: A
comparison of Parametric, Nonparametric and Semi-parametric Methods with Boot-
strapping”, The Journal of Productivity Analysis, 3, 167-203.

[289] Simar L., (1996) “Aspects of statistical Analysis in DEA-type frontier models”, The
Journal of Productivity Analysis, 7, 177-185.
REFERENCES 239

[290] Simar, L. (2003a), “Detecting outliers in frontier models: a simple approach”, Journal
of Productivity Analysis, 20, 391-424.

[291] Simar, L. (2003b), “How to Improve the Performance of DEA/FDH Estimators in the
Presence of Noise?, Discussion Paper 0323, Institut de Statistique, UCL, Belgium.

[292] Simar, L. and Wilson, P.W. (1998), “Sensitivity analysis of efficiency scores: how to
bootstrap in nonparametric frontier models”, Management Science, vol. 44, 1, 49-61.

[293] Simar, L. and Wilson, P.W. (1999a), “Some problems with the Ferrier/Hirschberg
Bootstrap Idea”, The Journal of Productivity Analysis, 11, 67-80.

[294] Simar, L. and Wilson, P.W. (1999b), “Of Course we Can Bootstrap DEA scores!
But does it mean anything? Logic Trumps and Wishful Thinking”, The Journal of
Productivity Analysis, 11, 67-80.

[295] Simar, L. and Wilson, P.W. (1999c), “Estimating and Bootstrapping Malmquist In-
dices”, European Journal of Operational Research, 115, 459-471.

[296] Simar, L. and Wilson, P.W. (2000a), “Statistical Inference in Nonparametric Frontier
Models: The State of the Art”, The Journal of Productivity Analysis, 13, 49-78.

[297] Simar, L. and Wilson, P.W. (2000b), “A general methodology for bootstrapping in
non-parametric frontier models”, Journal of Applied Statistics, vol.27, 6, 779-802.

[298] Simar, L. and Wilson, P.W. (2001), “Testing restrictions in nonparametric efficiency
models”, Communications in Statistics, 30(1), 159-184.

[299] Simar, L. and Wilson, P.W. (2002), “Nonparametric tests of returns to scale”, European
Journal of Operational Research, 139, 115-132.

[300] Simar, L. and Wilson, P.W. (2003), “Estimation and Inference in Two-stage, Semi-
parametric Models of Production Processes”, Discussion Paper 0307, Institut de Sta-
tistique, UCL, Belgium, forthcoming in Journal of Econometrics.

[301] Simar, L. and P.W. Wilson (2005), “Estimation and Inference in Cross-Sectional Sto-
chastic Frontier Models”, Discussion paper 0524, Institut de Statistique, UCL.

[302] Simar, L. and P.W. Wilson (2006a), “Statistical Inference in Nonparametric Frontier
Models: recent Developments and Perspectives”, forthcoming in The Measurement
of Productive Efficiency, 2nd Edition, Harold Fried, C.A.Knox Lovell and Shelton
Schmidt, editors, Oxford University Press, 2006.

[303] Simar, L. and Wilson, P.W. (2006b), “Efficiency Analysis: The Statistical Approach”,
Manuscript, Institute of Statistics, UCL, Belgium.

[304] Simar, L. and V. Zelenyuk (2003), “Statistical Inference for Aggregates of Farrell-type
Efficiencies”, Discussion paper 0324, Institute of Statistics, UCL, Belgium, forthcom-
ing in Journal of Applied Econometrics.

[305] Simar, L. and V. Zelenyuk (2004), “On Testing Equality of Distributions of Technical
Efficiency Scores”, Discussion paper 0434, Institute of Statistics, UCL, Belgium.

[306] Simon, H. A. (1955), “A behavioral model of rational choice”, Quarterly Journal of

Economics, 69(1), 99-118.
240 REFERENCES

[307] Simon, H. A. (1957), Models of Man, John Wiley and Sons, NY.

[308] Simonoff J.S., (1996), Smoothing methods in statistics, Springer series in Statistics,
NY.

[309] Sitorius, B.L. (1966), “Productive Efficiency and Redundant Factors of Production
in Traditional Agricolture of Underdeveloped Countries”, Proceedings of the Thirty
Ninth Annual Meeting of the Western Farm Economics Associations, 153-158.

[310] Schmidt, P, and R.C. Sickles (1984), Production frontier and panel data, Journal of
Business and Economic Statistics,3, 171-203.

[311] Stevenson, R.E. (1980), “Likelihood Functions for Generalized Stochastic Frontier
Estimation”, Journal of Econometrics, 13(1), 57-66.

[312] Stigler, G.J. (1976), “The Xistence of X-Efficiency”, American Economic Review, 66
(1), 213-216.

[313] Swiss Re, (1996), “Deregulation and Liberalization of Market Access: The European
Insurance Industry on the Threshold of a New Era in Competition,” Sigma, no. 7 of
1996.

[314] Swiss Re, (2000a), “Japan’s Insurance Markets - A Sea Change,” Sigma, no. 8 of 2000.

[315] Swiss Re, (2000b), “Europe in Focus: Non-life Markets Undergoing Structural
Change,” Sigma, no. 3 of 2000.

[316] Taubes G. (1993), “Measure for measure in science”, Science, 14/05/93, Vol. 260, n.
5110, pp. 884-886.

[317] Taveres G. (2002), “A bibliography of Data Envelopment Analysis (1978-2001)”.

RUTCOR Research Report RRR 01-02, Rutgers University, Piscatway, NJ.

[318] Thanassoulis, E. (2001) Introduction to the Theory and Application of Data Envelop-
ment Analysis, Kluwer Academic Publishers, Boston.

[319] Thrall, R.M. (1999), “What is the economic meaning of FDH?”, Journal of Produc-
tivity Analysis, 11, 243-250.

[320] Thursby J.G., and Kemp S. (2002), “Growth and productive efficiency of university
intellectual property licensing”, Research Policy, 31(1), 109-124.

[321] Timmer, C.P. (1971), “Using a Probabilistic Frontier Production Function to Measure
Technical Efficiency”, Journal of Political Economy, 79 (4), 776-794.

[322] Tornqvist, L. (1936), “The Bank of Finland’s Consumption Price Index”, Bank of
Finland Montly Bullettin, 10, 1-8.

[323] Treynor, J.L. (1965), “How to Rate Management of Investment funds”, Harvard Busi-
ness Review, 43, 63-75.

[324] Triantis, K. and Girod, O. (1998), “A Mathematical Programming Approach for Mea-
suring Technical Efficiency in a Fuzzy Environment”, Journal of Productivity Analysis,
10, 85-102.
REFERENCES 241

[325] Triantis, K and Vanden Eeckaut P. (2000), “Fuzzy Pair-wise Dominance and Impli-
cations for Technical Efficiency Performance Assessment”, Journal of Productivity
Analysis, 13, 207-230.

[326] Tulkens, H. (1993), “On FDH Efficiency Analysis: Some methodological Issues and
Applications to Retail Banking, Courts, and Urban Transit”, Journal of Productivity
Analysis, 4 (1/2), 183-210.

[327] Tulkens, H. and Vanden Eeckaut P. (1995a), “Non-frontier measures of efficiency,

progress and regress for time series data”, International Journal of Production Eco-
nomics, 39, 83-97.

[328] Tulkens, H. and Vanden Eeckaut P. (1995b), “Non-parametric efficiency, progress

and regress measures for panel data: methodological aspects”, European Journal of
Operational Research, 80, 474-499.

[329] Turchetti G. and Daraio, C. (2004), “How Deregulation Shapes Market Structure and
Industry Efficiency: The case of the Italian Motor Insurance Industry”, Geneva Papers
on Risk and Insurance, 29 (2), 202-218.

[330] van Raan, A.F.J. (1993), “Advanced bibliometric methods to assess research perfor-
mance and scientific development: basic principles and recent practical applications”,
Research Evaluation, 3:151.

[331] van Raan, A.F.J. (1997), “Scientometrics: state of the art”, Scientometrics, Vol. 38,
205-218.

[332] van den Broeck J., Koop G., Osiewalski J., and Steel M.F.J. (1994), “Stochastic frontier
models: a Bayesian perspective”, Journal of Econometrics, 61, 273-303.

[333] Vanden Eeckaut, P. (1997), Free Disposal Hull and Measurement of efficiency: Theory,
Application and Software, PhD Thesis, Faculté des Sciences Economiques, Sociales
et Politiques, Nouvelle série (229), Université Catholique de Louvain, Louvain-la-
Neuve, Belgium.

[334] Varian, H.R. (1984), “The Non-Parametric Approach to Production Analysis”, Journal
of Productivity Analysis, 52, 279-297.

[335] Varian, H.R. (1985), “Nonparametric Analysis of Optimizing Behaviour with Mea-
surement Error”, Journal of Econometrics, 30 (1/2), 445-458.

[336] Varian, H.R. (1990), “Goodness-of-Fit in Optimizing Models”,in Lewin, A.Y., and
Lovell, C.A.K., (eds.) Frontier Analysis: Parametric and Nonparametric Approaches,
Journal of Econometrics, 46 (1/2).

[337] Varian H.R. (1992) Microeconomic Analysis, 3rd edition, W. W. Norton & Company.

[338] Vincent, A. (1968), La mesure de la productivité, Dunod, Paris.

[339] Watson, G.S. (1964), “Smooth regression analysis”, Sankhya Series A, 26, 359-372.

[340] Wilkens, K., J. Zhu (2001), “Portfolio evaluation and benchmark selection: A mathe-
matical programming approach”, Journal of Alternative investments, 4 (1), 9-20.
242 REFERENCES

[341] Williamson, O.E. (1964), The Economics of Discretionary Behavior: Managerial Ob-
jectives in a Theory of the Firm, Englewood Cliffs, NJ: Prentice-Hall.

[342] Wilson, P.W. (1995), “Detecting Influential Observations in Data Envelopment Analy-
sis”, Journal of Economics and Business, 6, 27-46.

[343] Wilson, P. W. (2005a), “FEAR 1.0: A Software Package for Frontier Efficiency
Analysis with R”, unpublished working paper, Department of Economics, Uni-
versity of Texas, Austin, Texas. Software and working paper downloadable at
http://www.eco.utexas.edu/faculty/Wilson/Software/FEAR/ .

[344] Wilson, P. W. (2005b), “FEAR 0.9 Command Reference”, Department of Economics,

University of Texas, Austin, Texas.

[345] Wilson, P. W. (2005c), “FEAR 0.9 User’s Guide”, Department of Economics, Univer-
sity of Texas, Austin, Texas.

[346] Zhu, J. (1996), “Data envelopment analysis with preference structure”, Journal of the
Operational Research Society, 47, 136-150.

[347] Zucker L.G., Darby M.R., Armstrong J. (1998), “Geographically localized knowledge:
Spillovers or markets?” Economic Inquiry, XXXVI, 65-86.
Topic Index

Activity analysis, 21 Deterministic frontier approach, 45

Age effects, 174 Deterministic models, 28
Aggregation of inputs and outputs, 148 DGP, 50
Asymptotic results, 47 Economic axioms, 21
consistency, 47 Economies of agglomeration, 173
sampling distributions, 49 Economies of experience, 163
Bandwidth selection, 109 Economies of scale, 157, 171
k-Nearest Neighbor method, 109 Economies of scope, 160
empirical rules, 111 Efficiency, 14
likelihood cross-validation, 109 allocative, 15
Bootstrap, 50 coefficient of resource utilization, 14
basic, 56 directional, 41
bias, 55 radial measures, 14
bootstrap principle, 54 scale, 15
bootstrap world, 52 structural, 15
confidence intervals, 55 technical, 14, 20
consistent, 56 Explaining efficiency, 96
empirical distribution, 51 External-environmental factors, 95
Monte-Carlo approximation, 51 FDH, 33
sampling distribution, 51–52 FEAR, 63
smoothing parameter, 59 Free disposability, 21
true world, 52 Full frontiers, 67
Chance-constrained programming, 39 Fund age, 208
Conditional efficiency, 96 Input oriented, 30
decomposition, 113 Input requirement set, 20
externality index, 120 Kernel, 101
individual index, 120 multivariate, 111
internal or managerial, 120 truncated, 112
full frontiers, 101 Malmquist productivity indices, 28
order-α frontiers, 103 Manager tenure, 199
conditional α efficiency, 104 Multiple objective approach, 41
order-m frontiers, 101 Nonparametric frontier approach, 30
Convexity, 22 Nonparametric models, 26
Cross-evaluation, 41 Normal reference rule, 60
Cross-sectional models, 28 Outliers, 79
Cross-validation, 61 Output correspondence set, 20
Curse of dimensionality, 48, 65, 78 Output oriented, 30
DEA, 31 Panel data models, 28
DEAsoft, 63 Parametric approaches, 85
244 TOPIC INDEX

Parametric approximations, 86 order-m efficiency measures, 70

bootstrap, 89 partial frontiers, 67
Generalized Cobb-Douglas, 91 probabilistic formulation, 66
multivariate, 90 properties, 77
translog, 93, 186 advantages, 78
Parametric models, 26 estimators full frontier, 77
Plug-in principle, 34 outliers, 79
Preference structure models, 40 statistical, 77
Principal component analysis, 145 Robust normal reference rule, 60
Production set, 20 Scale elasticity, 189
Productivity, 13 Shephard distance functions, 24
partial, 13 Slacks, 32
total factor, 13 Statistical model, 43
Reflection method, 58 statistical assumptions, 44
Returns to scale, 22 stochastic elements, 45
testing, 151 Stochastic frontier approach, 30
Robust frontiers, 65 Stochastic frontier models, 46
order-α frontiers, 72 Stochastic models, 28
α efficiency, 75 Subsampling techniques, 57
computational aspects, 74 Super efficiency, 41
order-α efficiency scores, 73 Taxonomy of efficient frontier models, 25
order-m frontiers, 68 Two stage methods, 87
computational aspects, 72 Value efficiency analysis, 40
definition, 68 Weight restrictions, 40
economic meaning, 71
Author Index

Acs, Z., 173 Cazals, C., 65–66, 68, 72, 75, 77, 96, 102
Adams, J.D., 168 Cesari, R., 193
Afriat, S.N., 17–18, 29 Chambers, R.G., 41
Aigner, D.J., 4, 29, 46 Chang, H., 19
Alchian, A., 18 Charnes, A., 2, 14–15, 17, 29, 31, 168
Allen, R., 40 Chavas, J.P., 18
Allison, P.D., 174 Cherchye, L., 37
Amel, D., 139 Chevalier, J., 195
Andersen, P., 41 Choi, B., 193
Angulo-Meza, L., 40 Christensen, L.R., 16, 28, 186
Annaert, J., 193 Chu, S.F., 29
Aragon, Y., 72, 74–75 Chung, Y., 41
Armstrong, J., 174 Clark, G.L., 173
Athanassopoulos, A., 40 Coelli, T., 26, 90, 92, 168, 186, 188
Audretsch, D.B., 174 Collins, P.M.D., 169
Avveduto, S., 175
Cooke, P., 173
Banker, R.D., 15, 18–19, 31, 47, 98
Cooper, W.W., 2, 14–15, 17, 19, 25, 29, 31, 39, 168
Barnes, C., 139
Cornwell, C., 29
Barnett, V., 81
Cox, T.L., 18
Bartelsman, E.J., 97–98
Cummins, J.D., 139, 142–143
Barth, J.R., 138
Dan Brumbaugh, R.Jr., 138
Basso, A., 193
Battese, G.E., 26 Daniel, H.D., 169
Becattini, G., 173 Dantzig, G.B., 17
Berger, A.N., 138–139, 143 Daouia, A., 5, 72, 74–75, 77, 86–88, 91, 94, 96,
Bergson, A., 17 103–104, 186, 192
Bertoletti, P., 191 Daraio, C., 6, 65–66, 70–71, 77, 96, 100–102, 113,
Bessent, A., 168 122, 144, 168–170, 176, 184, 193–194
Bhadra, D., 18 Darby, M.R., 174
Bickel, P.J., 57 David, P.A., 175
Bjurek, H., 16 De Alessi, L., 19
Bogetoft, P., 41 Debreu, G., 2, 14, 16–17, 24
Boles, J.N., 17 Deckle, R., 138
Bonaccorsi, A., 168–170, 176, 184 Deprins, D., 2, 17, 30, 33–34, 86, 145
Braun, T., 169, 178 Desai, P., 193
Bressler, R.G., 17 Detzel, L.F., 196, 200, 216
Briec, W., 41, 193–194 Diewert, W.E., 16, 18, 28
Cantarelli, D., 191 Doms, M., 97–98
Castells, M., 173 Dorfman, R., 2
Caves, D.W., 16, 28 Dorof, W.A., 169
246 AUTHOR INDEX

Dyson, R.G., 40 Kao, C., 40

Efron, B., 50, 55, 57, 62 Kashayap, A.K., 138
Ellison, G., 195 Kemp, S., 168
Emrouznejad, A., 25 Kennington, J., 168
Farrell, M.J., 2, 14–17, 21, 24, 26, 31, 33 Kerstens, K., 41, 194
Feldman, M.P., 173–174 Kessel, R., 18
Ferrier, G.D., 63 Khorana, A., 196
Fieldhouse, M., 17 King, D., 169
Fisch, R., 169 Kneip, A., 4, 29, 45–46, 48–49, 57
Fisher, I., 28 Koopmans, T.C., 2, 14, 16–17
Florens, J.P., 5, 65–66, 68, 72, 75, 77, 85, 87–92, Korhonen, P., 40, 168
94, 96, 102, 186, 188, 191–192 Korostelev, A., 48
Forsund, F.R., 16–17 Kostoff, R.N., 169
Fortin, R., 196, 200, 216 Krugman, P., 173
Freedman, D.A., 57, 59 Kumbhakar, S.C., 4, 16, 29, 46, 86, 98
Fried, H.O., 6, 29–30, 39, 99 Kuosmanen, T.P., 37
Funari, S., 193 Land, K.C., 39
Färe, R., 5, 15–17, 19, 28–29, 32, 41, 98, 168, 188 Laredo, P., 168
Garfield, E., 169 Lau, L.J., 186
Gattoufi, S., 25 Leibenstein, H., 19
Gertler, M.S., 173 Lemak, D., 195
Gijbels, I., 49, 56 Leontief, W.W., 17
Girod, O., 40 Lesourd, J.B., 193–194
Glanzel, W., 169 Levin, S.G., 170, 175
Golec, J., 195, 200 Lewin, A.Y., 29, 98
Goto, I., 138 Lewis, T., 81
Green, J.R., 191 Lewison, G., 169
Greene, W.H., 29, 46, 86 Li, S., 39
Griliches, Z., 168 Lindgren, B., 29
Grinblatt, M., 193 Lindsay, C.M., 19
Grosskopf, S., 5, 16–17, 19, 28–29, 32, 41, 43, 90, Liu, S.T., 40
92, 98, 168, 186, 191 Lovell, C.A.K., 4–6, 13–17, 29–30, 32, 39, 46, 86,
Hall, P., 4, 29, 46, 173 97–99
Halme, M., 40 Luenberger, D.G., 41
Hamilton, K.S., 169 Luwel, M., 168
Hanoch, G., 18 Maindiratta, A., 18
Hansmann, H., 19 Malmquist, S., 16, 28
Harker, P.T., 139 Mammen, E., 49, 56
Hayes, K., 90, 92, 168, 186 Margaritis, D., 191
Henderson, D.J., 29, 47 Markowitz, H., 193
Henderson, R., 174 Martin, B.R., 169
Hess, T., 136, 138 Martin, S., 172
Hicks, J.R., 18 Mas-Colell, A., 191
Hirschberg, J.G., 63 May, R., 169
Hjalmarsson, L., 16 Meeusen, W., 29, 46
Hogan, A.M.B., 138 Merton, R.K., 174
Holbrook, J.A.D., 169 Michelson, S., 196, 200, 216
Holmstrom, B.R., 19 Milgrom, P., 172
Humphrey, D.B., 139 Minton, J.W., 98
Härdle, W., 114, 145, 218 Moed, H.F., 169, 179
Ippolito, R.A., 193 Moorsteen, R.H., 17
Jaffe, A.B., 174 Morey, M.R., 193
Jeong, S.O., 64, 101 Morey, R.C., 98, 193
Johnston, R., 168 Morgan, K., 173
Jones, M.C., 60, 113 Morroni, M., 97
Jorgenson, D.W., 186 Mouchart, M., 148
Joro, J., 40 Moutray, C., 168
AUTHOR INDEX 247
Mullins, N., 169 Schuster, E.F., 58
Murthi, A., 193 Scott, A.J., 173
Mustar, P., 168 Scott, D.W., 57, 110
Nadaraya, E.A., 114 Seaver, B., 40
Narin, F., 169 Seglen, P.O., 179
Niskanen, W.A.Jr., 19 Seiford, L.M., 14, 25, 29, 39
Nolle, D.E., 138 Seitz, J.K., 17, 97
Oehler, K., 169 Sengenberger, W., 173
Okubo, Y., 169 Sengupta, J.K., 14–15, 40, 193–194
Olesen, O.B., 39 Sheather, S.J., 60, 113
Olivastro, D., 169 Shephard, R.W., 2, 16–17, 21, 24, 52
Oral, M., 25 Shukla, R., 193
Panetta, F., 139, 193 Sickles, R.C., 29
Park, B.U., 4, 29, 45–46, 48–49, 56, 77, 100–101 Silverman, B.W., 58–61, 110
Park, H.S., 194 Simar, L., 2, 4–6, 17, 19, 22, 25–26, 29–30, 33–34,
Parkan, C., 18 38, 43–46, 48–49, 52, 56–61, 63–66, 68,
Pasurka, C., 98 70–72, 74–75, 77, 79, 85–92, 94, 96,
Pedraja-Chaparro, R., 40 99–104, 113, 122, 145, 148, 151–153, 170,
Pereira Estellita Lins, M., 40 186, 188, 191–194, 218
Perelman, S., 188 Simon, H.A., 19, 174
Petersen, N.C., 39, 41 Simonoff, J.S., 57, 110
Porrini, D., 138 Sitorius, B.L., 17
Porter, G., 196, 200, 216 Siverman, B.W., 57
Post, G.T., 37 Smith, J., 40
Primont, D., 188 Smith, P., 40
Pyke, F., 173 Snizek, W., 169
Quiggin, J., 41 Solow, R., 2
Ramsden, P., 168 Stephan, P.E., 170, 175
Rao, D.S.P., 26 Stevens, K.A., 169
Ray, S.C., 3, 18, 29 Stevenson, R.E., 46
Reagan, B., 168 Stewart, J.A., 174
Reisman, A., 25 Stigler, G.J., 19
Rhodes, E., 2, 15, 17, 31, 168 Tainio, R., 168
Rice, T.N., 138 Taubes, G., 169
Richmond, J., 29 Taveres, G., 25
Ritter, C., 46 Taylor, L., 90, 92, 168, 186
Roberts, J., 172 Thanassoulis, E., 26, 40
Roos, P., 29 Thomas-Agnan, C., 72, 74–75
Rosenberg, N., 169 Thore, S., 39
Rothschild, M., 18 Thrall, R.M., 37, 39
Rousseau, R., 168 Thursby, J.G., 168
Rousseau, S., 168 Tibshirani, R.J., 50, 55, 57, 62
Russell, R.R., 15, 28 Timmer, C.P., 17, 97
Salinas-Jimenes, J., 40 Tirole, J., 19
Salleo, C., 139 Titman, S., 193
Salo, S., 40 Tone, K., 14, 25, 29, 39
Samuelson, P.A., 2, 18 Tornqvist, L., 16, 28
Santin, D., 188 Trajtenberg, M., 174
Sarafoglou, N., 17 Trauth, T., 136, 138
Satish, P., 195 Triantis, K., 40
Saxenian, A., 173 Trifts, J., 196, 200, 216
Scalise, J.M., 138 Trzcinca, C., 193
Scherer, F.M., 172 Tsionas, E.G., 4, 29, 46
Schmidt, P., 4, 29, 46 Tsybakov, A.B., 48
Schmidt, S.S., 6, 29–30, 39, 99 Tulkens, H., 2, 17, 29, 33–35
Schmoch, U., 169 Turchetti, G., 143–144
Schubert, A., 169, 178 Valdmanis, V., 191
248 AUTHOR INDEX

Van den Broeck, J., 29, 46, 193 Whinston, M.D., 191
Van Leeuwen, T.N., 179 Wilcox, J.A., 138
Van Raan, A.F.J., 169 Wilkens, K., 193
Vanden Eeckaut, P., 29, 37, 40–41 Williamson, O.E., 18
Varian, H.R., 18, 191 Wilson, P.W., 4, 6, 19, 22, 25–26, 30, 43–46, 49, 52,
Vennet, R.V., 193 57–61, 63–64, 79, 99, 113, 151–153, 218
Vincent, A., 13 Yaisawarng, S., 6, 99
Wagner, J., 196, 200, 216
Young, G.A., 57
Wallenius, J., 40, 168
Zelenyuk, V., 16, 63, 100
Watson, G.S., 114
Weber, W., 90, 92, 168, 186 Zenios, S.A., 139
Weigand, R., 196, 200, 216 Zhu, J., 39–40, 193
Weiner, C., 45, 48–49, 77 Zi, H., 143
Weiss, M.A., 139, 142–143 Zucker, L.G., 174