EUROPEAN

JOURNAL
OF OPERATIONAL
RESEARCH
ELSEVIER European Joumal of Operational Research 98 (1997) 175-212

Efficiency of financial institutions: International survey and

directions for future research
A l l e n N. B e r g e r a,b, *, D a v i d B. H u m p h r e y c
a Board o f Governors of the Federal Reserve System, Washington, DC 20551, USA
b Wharton Financial Institutions Center, University o f Pennsylvania, Philadelphia, PA 19104, USA
c Department o f Finance, Florida State University, Tallahassee, FL 32306-1042, USA

Abstract

This paper surveys 130 studies that apply frontier efficiency analysis to financial institutions in 21 countries. The primary
goals are to summarize and critically review empirical estimates of financial institution efficiency and to attempt to arrive at
a consensus view. We find that the various efficiency methods do not necessarily yield consistent results and suggest some
ways that these methods might be improved to bring about findings that are more consistent, accurate, and useful. Secondary
goals are to address the implications of efficiency results for financial institutions in the areas of government policy,
research, and managerial performance. Areas needing additional research are also outlined.

1. Introduction ployed; or (3) to improve managerial performance by

identifying 'best practices' and ' worst practices' as-
The first task in evaluating the performance of sociated with high and low measured efficiency,
financial institutions is to separate those production respectively, and encouraging the former practices
units that by some standard perform well from those while discouraging the latter.
that perform poorly. This is done by applying non- At its heart, frontier analysis is essentially a so-
parametric or parametric frontier analysis to firms phisticated way to 'benchmark' the relative perfor-
within the financial industry or to branches within a mance of production units. Most financial institu-
financial firm. The information obtained can be used tions, with varying degrees of success, benchmark
either: (1) to inform government policy by assessing themselves a n d / o r use industry consultants to per-
the effects of deregulation, mergers, or market struc- form this task. The power of frontier analysis is
ture on efficiency; (2) to address research issues by twofold. First, it permits individuals with very little
describing the efficiency of an industry, ranking its institutional knowledge or experience to select 'best
firms, or checking how measured efficiency may be practice' firms within the industry (or 'best practice'
related to the different efficiency techniques em- branches within the firm), assign numerical effi-
ciency values, broadly identify areas of input overuse
a n d / o r output underproduction, and relate these re-
sults to questions of government policy or academic
* Corresponding author. Emaih [email protected].
The opinions expressed do not necessarily reflect those of the research interest. Second, in the hands of individuals
Board of Governors or its staff. with sufficient institutional background, frontier

0377-2217/97/$17.00 Published by Elsevier Science B.V.

PII S0377- 2 2 1 7 ( 9 6 ) 0 0 3 4 2 - 6
176 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

analysis permits management to objectively identify more fully informed regarding these techniques are
areas of best practice within complex service opera- referred to the numerous comprehensive method-
tions, a determination not always possible with tradi- ological surveys which exist (Banker et al., 1989;
tional benchmarking techniques due to a lack of a Bauer, 1990; Seiford and Thrall, 1990; Aly and
powerful optimizing methodology such as linear pro- Seiford, 1993; Greene, 1993; Grosskopf, 1993;
gramming. Lovell, 1993; Charnes et al., 1994).
As practiced by academics, frontier analysis will In Section 3, the average efficiency and disper-
generally tell informed industry participants little sion of efficiency for US commercial banks - the
they do not already know in a general, qualitative most studied class of financial institutions - is dis-
way. While the qualitative 'news' may not be new, played. These data are used to illustrate the differ-
the quantification of it is. Frontier analysis provides ences in efficiency estimates between nonparametric
an overall, objectively determined, numerical effi- and parametric frontier techniques. As some investi-
ciency value and ranking of firms (also called X-ef- gators have already hinted at, the central tendency of
ficiency in the economics literature) that is not other- the distribution of estimates of average efficiency
wise available. This attribute makes frontier analysis derived from either type of technique is similar but
particularly valuable in assessing and informing gov- the degree of dispersion differs. The similarity that
ernment policy regarding financial institutions, such exists for average efficiency within an industry across
as determining the efficiency effects of mergers and frontier techniques is weaker when rankings of firms
acquisitions for possible use in antitrust policy. When by their efficiency value are being compared.
frontier analysis is more narrowly focused on propri- In Section 4, we discuss the similarity of average
etary transactions data and detailed input use across efficiency estimates across countries and by type of
branches of a financial institution, a firm's internal financial institution. We compare the results for 21
performance can often be enhanced beyond that pos- nations and four types of financial institutions -
sible with its own benchmarking procedures. banks, S &Ls, credit unions, and insurance firms.
There are now enough frontier efficiency studies Applications of efficiency analysis are reviewed
of financial institutions to make some tentative com- in Sections 5-7, segmented according to the main
parisons of average efficiency levels both across purpose of the research. Section 5 reviews studies
measurement techniques and across countries, as well which provide valuable information for government
as outline the primary results of the many applica- policy, such as the effects of deregulation, financial
tions of efficiency analysis to policy and research institution failure, market structure, and mergers.
issues. Toward this end, we survey and contrast the Section 6 reviews studies that are chiefly concerned
results of 130 financial institution efficiency studies. with research issues, such as the measurement of
This literature has employed at least five major efficiency, comparisons of efficiency across interna-
different efficiency techniques, which have been ap- tional borders, issues of corporate control, risk, and
plied to financial institutions in at least 21 countries. the stability over time of firm-level efficiency. Sec-
We also cover studies of several different types of tion 7 analyzes studies that are primarily associated
depository institutions - commercial banks, savings with improving managerial performance, most of
and loans, and credit unions - as well as firms in the which measure the relative efficiencies of individual
insurance industry. We include this large number of branches within the s a m e firm.
nations and wide array of types of financial institu- We recognize the somewhat artificial nature of
tions because the financial markets of the future are this division of issues into government policy, re-
likely to become more globalized, and have more search, and managerial performance. For example,
universal-type institutions offering greater selections studies which advance the efficiency research agenda
of financial services within a single institution. will eventually be useful for studying policy, man-
Section 2 critiques the main nonparametric and agement, or any other efficiency issue.
parametric efficiency estimation methods. A reason- Finally, Section 8 concludes, assessing the results
able familiarity with the various frontier measure- of applications of efficiency analysis to financial
ment techniques is assumed. Readers wishing to be institutions, and suggesting some new directions for
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 177

future research. Most of the important suggestions Nonparametric frontiers. Nonparametric ap-
concern finding explanations of efficiency that may proaches, such as much of the work in data envelop-
help inform government policy, identify the eco- ment analysis (DEA) and Free Disposal Hull (FDH),
nomic conditions that create inefficiency, and im- put relatively little structure on the specification of
prove managerial performance. the best-practice frontier. DEA is a linear program-
ming technique where the set of best-practice or
frontier observations are those for which no other
2. Nonparametric and parametric approaches to decision making unit or linear combination of units
measuring efficiency has as much or more of every output (given inputs)
or as little or less of every input (given outputs). 3
Our focus in this article is on frontier efficiency, The DEA frontier is formed as the piecewise linear
or how close financial institutions are to a 'best-prac- combinations that connect the set of these best-prac-
tice' frontier. Since engineering information on the tice observations, yielding a convex production pos-
technology of financial institutions is not available, sibilities set. As such, DEA does not require the
studies of frontier efficiency rely on accounting mea- explicit specification of the form of the underlying
sures of costs, outputs, inputs, revenues, profits, etc. production relationship. The free disposal hull ap-
to impute efficiency relative to the best practice proach (FDH) is a special case of the DEA model
within the available sample. There is a virtual con- where the points on lines connecting the DEA ver-
sensus in the literature that differences in frontier tices are not included in the frontier. Instead, the
efficiency among financial institutions exceed ineffi- FDH production possibilities set is composed only of
ciencies attributable to incorrect scale or scope of the DEA vertices and the free disposal hull points
output. 2 However, there is really no consensus on interior to these vertices. 4 Because the FDH frontier
the preferred method for determining the best-prac- is either congruent with or interior to the DEA
tice frontier against which relative efficiencies are frontier, FDH will typically generate larger estimates
measured. of average efficiency than DEA (Tulkens, 1993).
At least five different types of approaches have Either approach permits efficiency to vary over time
been employed in evaluating the efficiency of finan- and makes no prior assumption regarding the form of
cial institutions and branches. These methods differ the distribution of inefficiencies across observations
primarily in the assumptions imposed on the data in except that undominated observations are 100% effi-
terms of (a) the functional form of the best-practice cient.
frontier (a more restrictive parametric functional form However, a key drawback to these nonparametric
vs. a less restrictive nonparametric form), (b) whether approaches is that they generally assume that there is
or not account is taken of random error that may no random error. There is assumed to be: (a) no
temporarily give some production units high or low measurement error in constructing the frontier; (b)
outputs, inputs, costs, or profits, and (c) if there is
random error, the probability distribution assumed
for the inefficiencies (e.g., half-normal, truncated
normal) used to disentangle the inefficiencies from 3 Developed by Charnes et al. (1978), DEA was originally
the random error. Thus, the established approaches intended for use in public sector and not-for-profit settings where
typical economic behavioral objectives, such as cost minimization
to efficiency measurement differ primarily in how
or profit maximization, may not apply. Thus, DEA could be used
much shape is imposed on the frontier and the even when conventional cost and profit functions that depend on
distributional assumptions imposed on the random optimizing reactions to prices could not be justified.
error and inefficiency. 4 From the perspective of input requirements to produce a given
output, DEA presumes that linear substitution is possible between
observed input combinations on an isoquant (which is generated
from the observations in piccewise linear forms). In contrast, FDH
2 See Berger et al. (1993b) for a review of studies of scale and presumes that no substitution is possible so the isoquant looks like
scope efficiencies of financial institutions and how these compare a step function formed by the intersection of lines drawn from
to frontier efficiencies. observed (local) Leontief-type input combinations.
178 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

no luck that temporarily gives a decision making unit posed-error framework, since the truncated normal
better measured performance one year from the next, and gamma distributions may be close to the sym-
and (c) no inaccuracies created by accounting rules metric normal distribution assumed for the random
that would make measured outputs and inputs devi- error.
ate from economic outputs and inputs. Any of these The distribution-free approach (DFA) also speci-
errors that did appear in an inefficient unit's data fies a functional form for the frontier, but separates
may be reflected as a change in its measured effi- the inefficiencies from random error in a different
ciency. What may be more problematical is that any way. Unlike SFA, DFA makes no strong assump-
of these errors in one of the units on the efficient tions regarding the specific distributions of the inef-
frontier may alter the measured efficiency of all the ficiencies or random errors. Instead, DFA assumes
units that are compared to this unit or linear combi- that the efficiency of each firm is stable over time,
nations involving this unit. whereas random error tends to average out to zero
Parametric frontiers. There are three main para- over time. The estimate of inefficiency for each firm
metric frontier approaches. The stochastic frontier in a panel data set is then determined as the differ-
approach (SFA) - sometimes also referred to as the ence between its average residual and the average
econometric frontier approach - specifies a func- residual of the firm on the frontier, with some trunca-
tional form for the cost, profit, or production rela- tion performed to account for the failure of the
tionship among inputs, outputs, and environmental random error to average out to zero fully. 5 With
factors, and allows for random error. SFA posits a DFA, inefficiencies can follow almost any distribu-
composed error model where inefficiencies'are as- tion, even one that is fairly close to symmetric, as
sumed to follow an asymmetric distribution, usually long as the inefficiencies are nonnegative. 6 How-
the half-normal, while random errors follow a sym- ever, if efficiency is shifting over time due to techni-
metric distribution, usually the standard normal. The cal change, regulatory reform, the interest rate cycle,
logic is that the inefficiencies must have a truncated or other influences, then DFA describes the average
distribution because inefficiencies cannot be nega- deviation of each firm from the best average-practice
tive. Both the inefficiencies and the errors are as- frontier, rather than the efficiency at any one point in
sumed to be orthogonal to the input, output, or time.
environmental variables specified in the estimating Lastly, the thick frontier approach (TFA) specifies
equation. The estimated inefficiency for any firm is a functional form and assumes that deviations from
taken as the conditional mean or mode of the distri- predicted performance values within the highest and
bution of the inefficiency term, given the observation lowest performance quartiles of observations (strati-
of the composed error term. fied by size class) represent random error, while
The half-normal assumption for the distribution of deviations in predicted performance between the
inefficiencies is relatively inflexible and presumes highest and lowest quartiles represent inefficiencies.
that most firms are clustered near full efficiency. In This approach imposes no distributional assumptions
practice, however, other distributions may be more
appropriate (Greene, 1990). Some financial institu-
tion studies have found that specifying the more 5 A n alternative way to apply DFA is to use a fixed effects
general truncated normal distribution for inefficiency model. In a fixed effects model, a dummy variable is specified for
yields minor, but statistically significant, different each firm in a panel data set. Differences in the fixed effects
estimated across firms represent firm inefficiencies (e.g., Lang and
results from the special case of the half-normal
Welzel, 1996). However, Berger (1993) found that the fixed
(Berger and DeYoung, 1997). A similar result using effects were confounded by the differences in scale, which are
life insurance data occurred when a gamma distribu- several thousand times larger in magnitude than differences in
tion, which is also more flexible than the half-nor- efficiency in typical banking data sets.
mal, was used (Yuengert, 1993). However, this 6 A plot of an unrestricted distribution of inefficiencies implied
by the data in one DFA study determined that the resulting
method of allowing for flexibility in the assumed
frequency distribution was closer to the shape of a symmetric
distribution of inefficiency may make it difficult to normal rather than an asymmetric half-normal distribution (Berger,
separate inefficiency from random error in a corn- 1993).
 A.N. Berger, D.B. Humphrey~European Journal of Operational Research 98 (1997) 175-212 179

on either inefficiency or random error except to To date, this has focused on specifying a Fourier-
assume that inefficiencies differ between the highest flexible functional form which adds Fourier trigono-
and lowest quartiles and that random error exists metric terms to a standard translog function (Berger
within these quartiles. TFA itself does not provide and DeYoung, 1997; Berger and Mester, 1997;
point estimates of efficiency for individual firms but Berger et al., 1996a, 1997). This greatly increases
is intended instead to provide an estimate of the the flexibility of the frontier by allowing for many
general level of overall efficiency. The TFA reduces inflection points and by including essentially orthog-
the effect of extreme points in the data, as can DFA onal trigonometric terms that help fit the frontier to
when the extreme average residuals are truncated. the data wherever it is most needed. 7
Is there a 'best' frontier method? The lack of In the nonparametric approaches, two research
agreement among researchers regarding a preferred agendas are being pursued. 8 One is analytical, and
frontier model at present boils down to a difference seeks to provide a statistical foundation for DEA.
of opinion regarding the lesser of evils. The paramet- The other is empirical, and seeks to develop and
ric approaches commit the sin of imposing a particu- implement a stochastic version of DEA. The analyti-
lar functional form (and associated behavioral as- cal research has demonstrated that, given certain
sumptions) that presupposes the shape of the frontier. plausible assumptions concerning the structure of
If the functional form is misspecified, measured technology and the distribution of the 'true' efficien-
efficiency may be confounded with the specification cies, (a) the empirical efficiencies calculated from a
errors. Usually a local approximation such as the DEA model provide consistent estimators for the
translog is specified, which has been shown to pro- true efficiencies, (b) the DEA estimators can be
vide poor approximations for banking data that are interpreted as maximum likelihood estimators, and
not near the mean scale and product mix (see McAI- (c) the asymptotic empirical distribution recovers the
lister and McManus, 1993; Mitchell and Onvural, true distribution under the maintained assumptions.
1996). The translog also forces the frontier average This work thus provides a theoretical foundation for
cost curve to have a symmetric U-shape in logs. statistical hypothesis testing in a DEA environment
The nonparametric studies impose less structure (see Banker, 1996, for a summary). However, the
on the frontier but commit the sin of not allowing for fundamental problem is one of specifying the distri-
random error owing to luck, data problems, or other bution of efficiency across observations (Kneip and
measurement errors. If random error exists, mea- Simar, 1996; Simar, 1996). Hypothesis testing can
sured efficiency may be confounded with these ran- be conducted only after the data generating process
dom deviations from the true efficiency frontier. As has been specified, and in a multidimensional non-
seen below, the conflict between the nonparametric parametric setting in which the inefficiencies are
and parametric approaches is important because the one-sided, this is a statistically non-trivial matter.
two types of methods tend to have different degrees Moreover, the sampling distribution of the DEA
of dispersion and rank the same financial institutions efficiency estimators remains unknown, and this ob-
somewhat differently. servation motivates the second line of research.
It is not possible to determine which of the two
major approaches dominates the other since the true
level of efficiency is unknown. The solution, in our
opinion, lies in adding more flexibility to the para-
metric approaches and introducing a degree of ran- 7 The use of the Fourier-flexible form in place of the translog in
dom error into the nonparametric approaches. By one case reduced the amount of measured inefficiency by about
addressing the main limitation of each approach, the half - from 10% to 5% of costs - since the more flexible frontier
efficiency results will presumably yield efficiency was able to be closer to more of the data (Berger and DeYoung,
1997). Globally flexible functional forms have also been applied
estimates which are more consistent across the ap-
to banking data in non-frontier models of scale economies (McAI-
proaches. These processes have already begun. In the lister and McManus, 1993; Mitchell and Onvural, 1996).
parametric approaches, some studies have experi- s We thank Knox Lovell for his gracious assistance with this
mented with specifying more globally flexible forms. and the following paragraph.
180 A.N. Berger, D.B. Humphrey~European Journal of Operational Research 98 (1997) 175-212

A resampling technique, such as bootstrapping, is depository financial institutions and notes which of
one way of obtaining an empirical approximation to the five frontier methods were used. 10 The 8 studies
the underlying sampling distribution of D E A effi- that apply frontier analysis to insurance firms are
ciency estimates. Once the underlying distribution is shown in another table and are discussed later. Table
approximated, statistical inference can be conducted. 1 also shows the country the analysis was applied to,
This computer-intensive approach to hypothesis test- the author(s) of the study, the average yearly effi-
ing, however, requires a careful specification of the ciency estimates reported, and the type of institution
data generating process (Simar and Wilson, 1995). A covered. Overall, there were 69 applications of non-
different approach is to apply the techniques of parametric techniques and 60 using parametric ap-
chance-constrained programming to the D E A model proaches (some papers used more than one ap-
(Land et al., 1993; Olesen and Petersen, 1995). Here proach). 11 Studies focusing on US financial institu-
inequality constraints describing the structure of the tions were the most numerous, accounting for 66 of
nonparametric DEA technology are converted to the 116 single country studies in Table 1. 12
'chance constraints' which, due to noise in the data, A frequency distribution of 188 nonparametric
are allowed to be violated by a certain proportion of and parametric annual average efficiency estimates
the observations. If probability distributions are spec- for US banks from Table 1 (excluding profit effi-
ified for these violations (the data generating process ciency and branch efficiency studies) is shown in
again), the constraints can be converted into certainty Fig. 1. 13 The 188 annual estimates exceeds the 50
equivalents, and a chance-constrained DEA model US bank efficiency studies because many of these
emerges as a nonlinear programming problem. Al- studies report values for multiple years, techniques,
though the chance-constrained D E A model remains a n d / o r classes of banks, and each is treated as a
deterministic, it incorporates noise in the data (see single observation here. 14 The distribution combines
Grosskopf, 1996, for a survey of both empirical
approaches). 9

l0 This feature issue also contains a novel application of DEA

3. S u m m a r y o f efficiency findings by m e a s u r e - efficiency analysis to the performanceof mutual funds (Murthi et
ment method al., 1997), but is not listed in our tables.
t z Of the 69 nonparametricapplications,62 were DEA, 5 were
We now turn to the results of studies of financial FDH, and 2 were other approaches noted in Table 1. The 60
parametric applicationswere 24 SFA, 20 DFA, and 16 TFA.
institution efficiency. Along the way we will take
12Although we have tried hard to be comprehensive, there are
note of how the efficiency estimates vary by the undoubtedly some studies we have missed, and we apologize to
efficiency approach specified (DEA, FDH, SFA, the authors of those articles. Some that we know we have missed
DFA, TFA) and a number of other facts about the were not written in English or were in journals to which we did
method and sample. Table 1 lists the 122 frontier not have access.
13Estimates of profit efficiency and branch efficiency are ex-
studies we found that apply efficiency analysis to
cluded from the display because they are difficult to compare to
cost and production efficiencies. Profit efficiency is measured in
terms of best-practice profits, which are typically much smaller
than the costs, inputs, or output levels used in conventional
9An earlier effort to combine parametric and nonparametric studies. Branch efficiencyis measured relative to the best-practice
approaches has involved using FDH (or DEA) to first 'screen the branch within a firm, which is a very different target than the best
data' in order to identify the set of efficient observationsand then firm in a sample.
use only these observations in a regression-basedestimate of a 14For example, efficiency estimates obtained by making differ-
cost frontier (Thiry and Tulkens, 1992; Bauer and Hancock, 1993) ent assumptionsregarding the distributionof inefficiencyin SFA
or identify these observationswith a dummy variable and use all composed error models were treated as separate estimates in both
the observations in the regression, circumventingthe problem of Table 1 and Fig. 1. This treatment was also applied to efficiency
having too few observationsfor a large regression(Bardhan et al., estimates obtained from banks with different organizationalforms
1996). This approach is similar to that of the thick frontier or separate samples of banks in states with different branching
approach except that the criterion used to screen the data with is laws. If semi-annual estimates of efficiency were made, these
different. were averaged into annual figures.
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 181

Table 1
Studies of the efficiency of depository financial institutions
Country Method a Author (date) Average annual efficiency estimate Institution type
Belgium FDH Tulkens (1993) 0.97, 0.93 Branch
Belgium FDH Tulkens and Malnero (1994) 0.93 Branch

Canada DEA Parkan (1987) 0.98 Branch

Canada DEA Schaffnit et al. (1997) 0.87 Branch
Cyprus DEA Zenios et al. (1996) 0.89, 0.92, 0.88 Branch
Denmark DEA Bukh (1994) 0.80, 0.85 Bank
Finland DEA Kuussaari (1993) 0.80, 0.86 Bank
Finland DEA Kuussaari and Vesala (1995) 0.86 Bank
France DFA Chaffai and Dietsch (1995) 0.24, 0.33 Bank
France DFA Dietsch (1994) 0.72, 0.71, 0.68, 0.71, 0.69 Bank
Germany SFA Altunbas and Molyneux (n.d.) 0.8 l, 0.77, 0.77 Bank
Germany TFA Lang and Welzel (1995) 0.93 Bank
Germany DFA Lang and Welzel (1996) 0.54, 0.61 Bank

Greece DEA, Giokas ( 1991) 0.87 Branch

and SFA 0.72 Branch
Greece DEA Vassiloglou and Giokas (1990) 0.91 Branch
India DEA Bhattacharyya et al. (1997) 0.86, 0.75, 0.79 Bank
Italy DEA Favero and Papi (1995) 0.88, 0.91, 0.79, 0.84 Bank
Italy DEA Ferrier and Hirschberg (1994) 0.98 Bank
Italy DEA, Resti (1995) 0.74, 0.76, 0.74, 0.75, 0.73 Bank
and SFA 0.69, 0.70, 0.70, 0.70, 0.70 Bank
Japan DEA Fukuyama (1993) 0.86 Bank
Japan DEA Fukuyama (1995) 0.46, 0.46, 0.44 Bank
Mexico DEA Taylor et al. (1997) 0.75, 0.72, 0.69 Bank
Norway DEA Berg (1992) 0.62, 0.51, 0.57, 0.47, 0.49, 0.68, 0.57 Bank
Norway DEA Berg et al. (1991) 0.81 Bank
Norway DEA Berg et al. (1992) n.a. Bank
Norway TFA Berg and Kim (1994) 0.81, 0.81 Bank
Norway TFA Berg and Kim (1996) 0.89, 0.74 Bank
Saudi Arabia DEA Al-Faraj et al. (1993) 0.87 Branch
Spain DEA Grifell-Tatj6 and Lovell (1994) n.a. S&L
Spain DEA Grifell-Tatj6 and Lovell (1996) n.a. Bank, S&L, 7r
Spain DEA Grifell-Tatj6 and Lovell (1997a) 0.81, 0.85, 0.85, 0.84, 0.83, 0.84, Bank
0.83, 0.87, 0.84, 0.85, 0.84, 0.83, S&L
0.80, 0.82, 0.81, 0.77
Spain DEA Grifell-Tatj6 and Lovell (1997b) 0.76, 0.75, 0.75, 0.80, 0.78, 0.80 S&L
Spain DEA Lovell and Pastor (1997) 0.92, 0.90 Branch
Spain TFA Lozano (1995) 0.90, 0.88, 0.89, 0.88, 0.87, 0.87, 0.87 Bank
Spain TFA Lozano (1997) 0.68, 0.67, 0.66, 0.73, 0.78, 0.81 S&L, 7r
Spain DEA Pastor (1995) 0.87, 0.80 Branch
Spain DEA Perez and Quesada (1994) 0.83 Bank
Spain SFA Maudos (1996a) 0.81, 0.83, 0.82, 0.81, 0.81, 0.81, Bank
0.81, 0.79, 0.80, 0.82, 0.83, 0.82,
0.80, 0.82, 0.82, 0.81, 0.80, 0.81
0.85, 0.87, 0.85, 0.81, 0.84, 0.85,
0.84, 0.82, 0.82
182 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175-212

Table 1 (continued)
Country Method Author (date) Average annual efficiency estimate Institution type
Spain SFA Maudos (1996b) 0.82, 0.83, 0.83 Bank
Sweden DEA Hartman and Storbeck (1995) 0.85, 0.78 S&L
Switzerland DEA Sheldon and Haegler (1993) 0.56 Bank
Tunisia SFA Chaffai (1993) 0.66, 0.65, 0.65, 0.64, 0.63, 0.63 Bank
0.62, 0.62, 0.62, 0.61, 0.61, 0.61
Tunisia SFA Chaffai (1997) n.a. Bank
Turkey DEA Oral and Yolalan (1990) 0.87, 0.53 Branch
Turkey DEA Zaim (1995) 0.83, 0.94 Bank
UK DEA Athanassopoulos (1995) 0.85 Branch
UK DEA Athanassopoulos (1997) 0.90 Branch
UK DEA Drake and Howcroft (1997) 0.93, 0.97 Branch
UK DEA Drake and Weyman-Jones (1992) 0.98 S&L
UK DEA Field (1990) 0.93 S&L
US DFA Adams et al. (1995) 0.64, 0.61, 0.69, 0.64, 0.70, 0.77 Bank
US DFA Akhavein et al. (1997a) 0.24 0.34 Bank, 7r
US DFA Akhavein et al. (1997b) 0.99 0.44 Bank, ~r
US DEA Aly et al. (1990) 0.75 0.81 Bank
US DEA Barr et al. (1994) 0.81 0.83 Bank
US SFA, Bauer et al. (1993) 0.87 0.86, 0.86, 0.87, 0.86, 0.88 Bank
0.89 0.87, 0.87, 0.87, 0.85, 0.85
DFA, 0.86 0.85, 0.86, 0.86, 0.85, 0.86 Bank
0.87 0.86, 0.86, 0.86, 0.85, 0.85
and TFA 0.86, 0.90, 0.81, 0.80, 0.84, 0.84 Bank
0.86, 0.83, 0.81, 0.82, 0.83, 0.79
US DEA Bauer et al. (1995) 0.73, 0.71,0.71, 0,73, 0.75, 0.76 Bank
0.74, 0.73, 0.73, 0.70, 0.67, 0.67
US DFA Berger (1993) 0.85, 0.84, 0.75 Bank
US DFA Berger (1995) n.a. Bank
US SFA Berger and DeYoung (1997) 0.92, 0.94, 0.95, 0.91, 0.93, 0.91 Bank
0.91, 0.91, 0.93, 0.95
US DFA Berger et al. (1993a) 0.52, 0.65, 0.66 Bank, ~"
US DFA Berger and Hannan (1997) n.a. Bank
US TFA Berger and Humphrey (1991) 0.81, 0.84 Bank
US TFA Berger and Humphrey (1992a) 0.85, 0.81, 0.71, 0.80, 0.84, 0.80 Bank
US DFA Berger and Humphrey (1992b) n.a. Bank
US DFA Berger et al. (1997) 0.94, 0.79 Branch
US DFA Berger and Mester (1997) 0.87 Bank
0.55, 0.46 Bank, w
US SFA Cebenoyan et al. (1993a) 0.77, 0.83 S&L
US SFA Cebenoyan et al. (1993b) 0.87, 0.86 S&L
US SFA Chang et al. (1993) 0.81 Bank
US DEA Charnes et al. (1990) n.a, Bank
US TFA Clark (1996) 0.73, 0.90 Bank
US FDH DeBorger et al. (1995) 0.94, 0.88, 0.89, 0.80, 0.95, 0.88, 0.89 Bank
0.80, 0.97, 0.95, 0.89, 0.89, 0.77
US DEA Devaney and Weber (1995) 0.75, 0.75, 0.71 Bank
US TFA DeYoung (1994) u.a. Bank
US DFA DeYoung (1997a) 0.80 Bank
US TFA DeYoung (1997b) 0.82 Bank
US TFA DeYoung (1997c) 0.84, 0.89 Bank
US DFA DeYoung and Nolle (1996) 0.56, 0.73 Bank, ~-
 A.N. Berger, D.B. Humphrey/ European Journal of Operational Research 98 (I 997) 175-212 183

Table 1 (continued)
Country Method Author (date) Average annual efficiency estimate Institution type
US DEA, Eisenbeis et al. (1996) 0.72, 0.73, 0.73, 0.78 Bank
and SFA 0.84, 0.87, 0.89, 0.93
US SFA Ellinger et al. (1997) n.a. Bank
US DEA Elyasiani and Mehdian (1990a) 0.90, 0.78 Bank
US SFA Elyasiani and Mehdian (1990b) 0.88 Bank
US DEA Elyasiani and Mehdian (1992) 0.89 Bank
US DEA Elyasiani and Mehdian (1995) 0.97, 0.95, 0.95, 0.96 Bank
US DEA Elyasiani et al. (1994) 0.86, 0.83 Bank
US DEA English et al. (1993) 0.75, 0.76 Bank
US DEA Ferrier et al. (1993) 0.69, 0.60 Bank
US DEA Fender et al. (1994) 0.37, 0.33 Bank
US DEA, Fender and Lovell (1990) 0.83 Bank
and SFA 0.79 Bank
US IN Fixler and Zieschang (1993) n.a. Bank
US FDH Fried and Lovell (1994) 0.93 CU
US FDH Fried et al. (1993) 0.83 CU
US DEA Grabowski et al. (1993) 0.72 Bank
US SFA, Hasan and Hunter (forthcoming) 0.82, 0.79 Bank
and TFA 0.64, 0.70 Bank, 7r
US DEA Hermalin and Wallace (1994) 0.75, 0.73 S&L
US TFA Humphrey and Pulley (1997) 0.81,0.82, 0.85 Bank, 7r
US DFA Hunter and Timme (1995) 0.84, 0.77, 0.78 Bank
US SFA Kaparakis et al. (1994) 0.90 Bank
US SFA Kwan and Eisenbeis (1994) 0.88, 0.85, 0.84, 0.84, 0.88, 0.88 Bank
US TFA Mahajan et al. (1996) 0.77, 0.88 Bank
US SFA Mester (1993) 0.92, 0.87 S&L
US SFA Mester (1996) 0.86 Bank
US SFA Mester (1997) 0.93, 0.92, 0.85, 0.87, 0.89, 0.88, Bank
0.86, 0.85
US DEA Miller and Noulas (1996) 0.97 Bank, 7r
US DFA Newman and Shrieves (1993) n.a. Bank
US DFA Peristiani (1997) 0.79, 0.79, 0.77, 0.81, 0.81, 0.77 Bank
US SFA Pi and Timme (1993) 0.87 Bank
US DEA Rangan et al. (1988) 0.70 Bank
US DEA Ray and Mukherjee (1994) 0.88 Bank
US DEA Sherman and Gold (1985) 0.96 Branch
US DEA Sherman and Ladino (1995) 0.80 Branch
US DEA Thompson et al. (1997) 0.81, 0.69, 0.59, 0.59, 0.54, 0.62 Bank
US DEA Thompson et al. (1996b) 0.53, 0.51, 0.45, 0.39, 0.35, 0.31, Bank
0.46, 0.44, 0.53
US DEA Wheelock and Wilson (1994) 0.84, 0.77, 0.69, 0.59, 0.59, 0.46, Bank
0.51,0.42
US SFA Zhu et al. (1997) 0.88, 0.86, 0.82 Bank

Multiple countries
Norway DEA Berg et ai. (1993) 0.57 Bank
Sweden 0.78 Bank
Finland 0.53 Bank
Norway MOS Bergendahl (1995) 0.09-1.00; Average = 0.51 Bank
Sweden 0.05-1.00; Average = 0.64 Bank
Finland
Denmark
Norway DEA Bukh et al. (1995) 0.54 Bank
Sweden 0.85 Bank
Finland 0.52 Bank
Denmark 0.78 Bank
184 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175-212

Table 1 (continued)
Country Method Author (date) Average annual efficiency estimate Institution type
11 OECD DFA Fecher and Pestieau (1993) 0.71-0.98; Average = 0.82 Financial services
countries
8 developed DEA Pastor et al. (1997) 0.55-0.95; Average = 0.86 Bank
countries
15 developed TFA Ruthenberg and Elias (1996) 0.55-0.94; Average = 0.70 Bank
countries

Notes: n.a. indicates either not reported, not comparable, or duplicates earlier estimates.
• r indicates a profit efficiency measure. The profit efficiency ratios employ a substantially different denominator (maximum or optimal
profits), and therefore are not comparable to the other ratios.
In order to make the reported efficiencies as comparable as possible, we try to report only technical efficiency ratios, and exclude scale,
scope, and allocative inefficiencies, which are not measured in most studies. In some cases, these other types of inefficiencies could not be
separated out. For example, some of the profit efficiency ratios incorporate scale and scope inefficiencies which create deviations from the
optimal output point.
a IN refers to a nonparametric index number approach. MOS is a mixed optimal strategy where the most efficient 'parts' of different banks
are combined and used as a frontier, in contrast to DEA and FDH where all parts of an individual bank define the frontier.
Key:
Nonparametric: DEA Data Envelopment Analysis Parametric: SFA Stochastic Frontier Approach (composed error)
FDH Free Disposal Hull DFA Distribution Free Approach (different composed error)
IN Index Numbers TFA Thick Frontier Approach
MOS Mixed Optimal Strategy

[3 Paremetdc (SFA, DFA, TFA)

• Nonparametdc (DEA, FDH)
F
r 4o
e Mean: ,79
q Median: .83
u Standard deviation: .13
e Interquartile range: .13
3o
n
c
Y

20

,20 .25 .30 .35 .40 .45 .50 .55 .60 .65 .70 .75 .80 .85 90 .95

.24 .29 .34 .39 .44 .49 .54 .59 .64 .69 .74 .79 .84 ,89 +94 .99

Annual Average Efficency Values

Fig. 1. Nonparametric and parametric annual average efficiency estimates for US banks (cost and productive efficiency values).
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 185

TabLe 2 80%, with the nonparametric techniques generally

Average efficiency of US banks by frontier technique giving lower efficiency estimates. The mean and
Nonparametric Parametric median efficiencies for the nonparametric techniques
techniques techniques
are 0.72 and 0.74, respectively; the parametric tech-
DEA and FDH a SFA, DFA, and TFA niques have a mean of 0.84 and median of 0.85. A
(78 observations) ( 110 observations)
greater difference between the approaches is that the
Mean 0.72 0.84 nonparametric studies suggest a greater dispersion in
Median 0.74 0.85
estimated efficiency ratios. The standard deviation,
Standard 0.17 0.06
deviation
range, and interquartile range of the nonparametric
Interquartile 0.24 0.07 studies are 0.17, [0.31,0.97], and 0.24, respectively,
range which is more dispersed than the 0.06, [0.61,0.95],
Range 0.31-0.97 0.61-0.95 and 0.07 values, respectively, for the parametric
studies. 16 The dispersion shown in Fig. 1 and Table
a Two nonparametric studies (noted as IN and MOS in Table 1)
have not been included. 2 suggests that the standard errors associated with
individual average efficiency estimates may be rela-
tively large, particularly for the nonparametric esti-
average efficiency estimates of US banks from dif-
mates. As discussed below, this also appears to be
ferent time periods, size classes, input-output speci-
the case so far for those studies that have determined
fications, and frontier techniques. For DEA-type
confidence intervals for nonparametric bank effi-
models, variable returns to scale estimates (if re-
ciency estimates using bootstrapping procedures.
ported) were chosen over efficiency values based on
Efficiency rankings for nonparametric and para-
constant returns.
metric models. Although there is a good deal of
The mean of Fig. 1 using both nonparametric and
information regarding the average efficiency of de-
parametric techniques is 0.79 with median of 0.83,
pository financial institutions by frontier technique,
standard deviation of 0.13, range of 0.31 to 0.97, and
there is only limited information comparing the effi-
interquartile range of 0.13. The mean of 0.79 implies
ciency rankings of firms across techniques. Based on
an average inefficiency of 27% [(1 - 0.79)/0.79]. 15
the few studies that exist, it appears that the similar-
The interval formed by the mean plus and minus one
ity of the central tendency of average efficiency
standard deviation would cover efficiency values
estimates evident in Fig. 1 between nonparametric
from 0.66 to 0.92, and capture 82% of the observa-
and parametric techniques does not consistently carry
tions.
over to the rankings of firms within the banking
The distribution of average efficiency from non-
industry. Some studies support a strong relationship
parametric studies of US banks is shown in the dark
between the findings of different techniques, while
(bottom) portion of each bar in Fig. 1 while the light
others find only weak relationships.
(top) portion indicates the distribution of the para-
Only two studies have compared the efficiency
metric results. These separate results are also sum-
ranking of banking firms between nonparametric and
marized; see Table 2. As seen, the central tendencies
parametric techniques. The Spearman rank correla-
of efficiency using these two broad classes of fron-
tion coefficient (RRANS) between DEA and SFA
tier techniques give similar ballpark figures near
technical efficiency rankings in one study of smaller
US banks for one year was R~ANK= 0.02 and not
significantly different from zero (Ferrier and Lovell,
t5 Efficiency results are typically reported in either of two ways.
The 0.79 efficiency figure means that if the average firm were
producing on the frontier instead of at its current location, then
only 79% of the resources currently being used would be neces-
sary to produce the same output (or meet the same objectives). 16 If the extremely low efficiency estimates (0.31 to 0.39 values
The 27% inefficiency figure means that the average firm requires in Fig. 1) from two nonparametric studies are deleted, the mean of
27% more resources to produce the same output (or meet the same the nonparametric studies rises from 0.72 to 0.74 and the standard
objectives) as an efficient firm on the frontier (the relationship is deviation is reduced from 0.17 to 0.14, slightly closer to the
0.79 = 1/( 1 + 0.27) or 0.27 = (1 - 0.79)/0.79). summary statistics for the parametric studies.
186 A.N. Berger, D.B. Humphrey/ European Journal of Operational Research 98 (1997) 175-212

1990). 17 In another study, using averages over 6 dial technical efficiency measures using both input-
years, RaANK between DEA and SFA varied from based and output-based FDH reference technology
0.44 to 0.59 across four size classes of larger US and found a wider range of similarity in efficiency
banks (Eisenbeis et al., 1996). These bank results are rankings, with correlations ranging between 0.32 and
weaker than those obtained for efficiency rankings 0.96 (DeBorger et al., 1995). A third study reported
across nonparametric and parametric frontier tech- rank correlation values between 0.86 and 0.99 for
niques for Federal Reserve check processing offices SFA efficiency estimates using assumed half-normal,
with RRANK values on the order of 0.70 (Bauer and truncated-normal, and exponential distributions of
Hancock, 1993) or for insurance firms which yielded inefficiency (Maudos, 1996a). 18
RRnNK values above 0.50 (Cummins and Zi, 1995) Overall, it seems clear that the estimates of mean
or above 0.72 (Fecher et al., 1993). This is one area or median efficiency for an industry may be a more
where further research would prove useful - deter- consistently reliable guide for policy and research
mining how the different frontier techniques affect purposes than are rankings of firms by their effi-
the relative efficiency rankings of individual finan- ciency value, especially between nonparametric and
cial institutions. parametric approaches. Because the consistency in
There is greater similarity in bank efficiency rank- rankings of individual firms by their efficiency value
ings when, instead of comparing nonparametric with can differ across frontier techniques, it follows that
parametric techniques, the comparison is between statistical results from the numerous ex post analyses
different techniques within one of these categories. correlating firm-level efficiency estimates with vari-
Two parametric techniques - SFA and TFA - were ous sets of explanatory variables should be viewed
compared when both methods were used to sepa- with caution. The use of a different method for
rately identify quartiles of US banks that were, re- determining efficiency may affect the qualitative re-
spectively, most or least efficient over a 12 year sults when searching for explanations of what makes
period. The degree of correspondence was 38% for some firms more efficient. Indeed, SFA efficiency
banks identified by each technique as being in the values in one study were significantly associated
most efficient quartile (Bauer et al., 1993). A some- with differences in market and accounting measures
what higher correspondence, at 46%, was found of bank risk and seem to strongly affect bank stock
across techniques for banks in the least efficient returns while DEA efficiency values were much less
quartiles. This is compared to an expected 25% informative in this regard (Eisenbeis et al., 1996).
correspondence due to chance alone, suggesting a This result occurred even though the rankings of
moderate positive relationship between the rankings banks by their SFA and DEA efficiency values were
of the two techniques. similar (with rank correlation values of 0.44 to 0.59).
Finally, there are three studies that compared Therefore, policy and research issues that rely upon
efficiency rankings of banks when different assump- firm-level efficiency estimates (as opposed to indus-
tions were applied within a given efficiency ap- try-wide averages) may be more convincingly ad-
proach. One study found that correlation coefficients dressed if more than one frontier technique is applied
for efficiency rankings of US banks using four dif- to the same set of data to demonstrate the robustness
ferent radial and nonradial technical efficiency mea- of the explanatory results obtained.
sures with a (variable returns to scale) DEA refer-
ence technology were relatively large and ranged
from 0.87 to 0.99 (Ferrier et al., 1994). A second 4. A v e r a g e efficiency across countries and by type
study undertook a comparison of radial and nonra- of financial institution

Average efficiency estimates across countries.

17Giokas (1991) compared average efficiency results between Five studies that have compared efficiency levels
DEA and a Cobb-Douglas (frontier) econometric estimation for
branches of a single bank, not across banks. Although differences
in efficiencyresults between these two techniques were discussed, ~8Rank correlations of these three sets of estimates with a fixed
no rank correlationwas reported. effects and a random effects model ranged from 0.56 to 0.90.
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 187

across countries are noted at the end of Table 1. In noted where Swedish banks were the most efficient.
one study, a D E A analysis of banks in Norway, In another study, D E A was applied to a cross-section
Sweden, and Finland was first performed with sepa- of 427 banks in 8 developed countries (Pastor et al.,
rate frontiers for each country and then with a 'com- 1997). The mean efficiency value was 0.86 with a
mon' frontier. In both the variable and constant range of 0.55 (UK) to 0.95 (France). US banks had
returns to scale cases, Sweden was found to be the the second lowest efficiency value (0.81) in the
more efficient of the three (Berg et al., 1993). The cross-section, which is consistent with the finding in
robustness of the common frontier results were the previous study that US banks were relatively
demonstrated by deleting all banks on the frontier, inefficient. We note that these cross-country compar-
recomputing efficiency values, and then correlating isons are difficult to interpret because the regulatory
the new efficiency ranking with the ranking prior to and economic environments faced by financial insti-
deleting any banks. Even after all the original fron- tutions are likely to differ importantly across nations
tier banks were deleted, the RRAN~ for the remaining and because the level and quality of service associ-
institutions was 0.96, attesting to the robustness of ated with deposits and loans in different countries
the original D E A rankings with a common frontier may differ in ways that are difficult to measure. Such
and the conclusion that Swedish banks are more cross-country differences were not specified when a
efficient. A follow-up analysis, adding Denmark, ' c o m m o n ' frontier was being estimated and this may
found broadly similar results (Bukh et al., 1995). ~9 affect the cross-country results. Difficult as they may
Two other cross-country studies applied DFA and be to perform and interpret, however, cross-country
D E A analysis to, respectively, 11 OECD countries studies can provide valuable information regarding
and 8 developed countries. 20 In both cases, the the competitiveness of banks in different countries, a
cross-country data are pooled and used to define a concern of particular importance in the increasingly
common frontier. In the first study, the average harmonized European market for banking services
efficiency of financial services (banking and insur- and the perhaps more globalized financial markets of
ance) is determined for 11 O E C D countries using the future.
national accounts data over 1971-1986 (Fecher and Fig. 2 shows a frequency distribution for the 131
Pestieau, 1993). Using a DFA-based fixed effects average efficiency values for banks from 14 non-US
model, the mean average efficiency value was 0.82 countries. 21 The comparability of efficiency esti-
with a range of 0.67 (Denmark) to 0.98 (Japan). mates for specific countries is limited by the fact that
Among other results, average efficiency in Sweden each country's efficiency estimate is determined rela-
(0.76) is found to be lower than that for Norway tive only to the frontier for that country. Since
(0.90) and the US (0.71) had the second lowest frontiers may differ across countries, our comparison
efficiency of the 11 countries studied. This result for here can only illustrate (a) the average dispersion of
Sweden is the opposite of that found in the more banks in each country away from that country's own
focused study of Norway, Sweden, and Finland just measured best-practice frontier, rather than (b) bank
efficiency measured relative to any global best-prac-
tice frontier. The advantage of (a) is that banks are
measured against a frontier that embodies similar
19The same four countries were covered in another study levels of service, regulatory treatment, and economic
(Bergendahl, 1995) which sought to develop a composite 'refer- environment. The advantage of (b) would be that a
ence bank' composed of the most efficient parts of the banks in
the sample. This generates higher benchmarks than does DEA and frontier formed from the complete data set across
indicates what may be possible rather than only what has been nations would allow for a better comparison across
achieved by any one bank alone.
20 A different approach would be to contrast individual banks in
all countries with only two other banks - one bank with the
lowest and another bank with the highest predicted average cost 21 Efficiency comparisons among S&Ls or branches of a single
from a standard non-frontier cost function model (Ruthenberg and bank, the only information available for 6 other countries (Bel-
Elias, 1996). In this analysis the average efficiency was 0.70 for gium, Canada, Cyprus, Greece, Saudi Arabia, and the UK). is
individual banks in 15 (mostly European) countries. discussed below.
188 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

nations, since the banks in each country would be try frontiers would lie close to the same global
compared against the same standard. Since frontiers frontier. We expect this to be an unlikely event and
likely differ across countries, efficiency measured so we cannot draw the conclusion from these figures
relative to single-nation frontiers will be overstated that US banks are more efficient than banks in other
relative to what would be measured with a common countries. Indeed, the opposite result was found in
or global frontier, so (a) will likely show greater two of the multiple country studies noted above
overall efficiency than would (b). (where US banks were among the least efficient).
With these caveats in mind, the mean annual Clearly, this is an area where more work is needed,
average efficiency value in Fig. 2 is 0.75, with especially the proper specification of country-specific
median of 0.81, standard deviation of 0.13, range of environmental influences that will justify using a
0.24 to 0.98, and interquartile range of 0.15. An common frontier for cross-country comparisons of
interval formed by the mean plus and minus one efficiency.
standard deviation would cover efficiency values Average efficiency of thrift institutions. The 14
from 0.62 to 0.88, covering 84 percent of the obser- studies that have focused on savings and loan associ-
vations in Fig. 2. The mean average efficiency de- ations (S&Ls) and credit unions cover the US, UK,
rived from nonparametric (0.75) and parametric Spain, and Sweden. These are listed in Table 1. The
(0.76) models is very similar but the standard devia- mean average efficiency level for US S&Ls is 0.83,
tion of the nonparametric model results (0.14) is which is higher than, but close to, the value reported
slightly larger than that for parametric models (0.12). for US banks (0.79) in Fig. 1. The average efficiency
Strictly speaking, the results of Figs. 1 and 2 are for credit unions, at 0.88, is higher still. The average
comparable only if all or most of the separate coun- efficiency of Spanish savings banks, at 0.80, is higher

5o
F
r
45
e
q
u 40
e
n
c 35 rn Parametric (SFA, DFA, T F A )

Y • Nonparametdc (DEA, FDH)

30-

25 -p Mean: .75
Median: .81
Standard deviation: .13
2o- Interquartile range: .15

15-

o
.20 .25 .30 .35 .4o .45 .50 .55 ,6o .85 .7o .75 ,80 .85 .9o .g5

.24 .29 .34 .39 .44 .49 .54 .59 .64 .69 .74 .79 .84 .8~ .94 .99
Annual Average Efficiency Values

Fig. 2. Nonparametric and parametric annual average efficiency estimates for non-US banks (cost and productive efficiency values).
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 189

Table 3
Studies of the efficiency of insurance firms
Country Method Author (date) Average annual efficiency estimate Institution type
France DEA, Fecher et al. (1993) 0.50, 0.33 Life and non-life
and SFA 0.41, 0.24 Life and non-life
Italy DEA Cummins et al. (1995a) 0.71, 0.71, 0.72, 0.76, 0.72, 0.78, 0.77, 0.74 Life and non-life

US SFA Cummins and Weiss (1993) 0.90, 0.79, 0.88 Property liability
US DFA Gardner and Grace (1993) 0.42 Life
US SFA, Yuengert (1993) 0.75 Life
and TFA 0.63 Life
US DFA Berger et al. (1996a) 0.74, 0.70 Property liability
0.63, 0.51, 0.68, 0.58 Property liability, ~r
US DEA Cummins et al. (1995b) 0.88, 0.88, 0.85, 0.86, 0.86, 0.85, 0.85, 0.85, 0.85, 0.85 Property liability
0.90, 0.91, 0.88, 0.88, 0.88, 0.87, 0.87, 0.89, 0.88, 0.87 Property liability
US SFA, Cummins and Zi (1995) 0.58, 0.63, 0.61, 0.61, 0.63 Life
and DFA, 0.47 Life
and DEA, 0.56, 0.58, 0.56, 0.61, 0.60 Life
and FDH 0.98, 0.98, 0.98, 0.98, 0.98 Life

than the mean for banks in other countries in Fig. 2 DFA) techniques. Thus, consistent with the bank
(0.75). 22 Similar to the international case above, it results, there is greater similarity in firm level effi-
is difficult to compare results across industries be- ciency estimates among techniques within the non-
cause no common frontier has been established. parametric or parametric category than there is among
Nonetheless, a tentative conclusion is that there is no techniques between these categories.
significant evidence to suggest that there is much of A different issue was addressed by Yuengert
a difference among the average efficiencies of these (1993). This concerned the problem of disentangling
types of depository financial institutions - banks, the effects of scale inefficiencies from frontier ineffi-
S&Ls, and credit unions. A more definitive result ciencies in the presence of heteroscedasticity. The
will have to await further study. problem arises in cross-section data on US insurance
Average efficiency of insurance firms. The aver- firms and banks where there is a greater dispersion
age efficiency for different types of insurance firms of average costs for smaller firms than for larger
is shown in Table 3. The mean average annual ones, but the envelope of lowest-cost firms across all
efficiency for the US insurance firms shown is 0.79 size classes is relatively flat. 23 In such a situation,
with a standard deviation of 0.15 (profit efficiency the average of firm average costs for smaller institu-
excluded). As a central tendency, US insurance firms tions will tend to be higher for smaller institutions
seem to have an average efficiency close to that for than for larger ones. One interpretation is that there
US banks (Fig. 1). The insurance studies are notable are economies of scale and that the greater disper-
in another respect: rankings of firms by their effi- sion in costs for smaller firms is due to heteroscedas-
ciency level between nonparametric and parametric ticity in the random error. An alternative interpreta-
techniques yielded (as noted earlier) RRANK values tion is that there are no scale economies, but rather a
above 0.50 (Cummins and Zi, 1995) or above 0.72 greater dispersion in efficiency levels for smaller
(Fecher et ai., 1993), showing greater consistency in firms than larger ones (reflecting 'heteroscedasticity'
firm rankings than similar evidence for banks. Cum- of efficiency across firms). Standard composed error
mins and Zi (1995) found even higher correlations models cannot distinguish between these cases or
for rankings of firms by different nonparametric determine which interpretation is more correct.
(DEA and FDH) or different parametric (EFA and

22 The remaining S&L studies for Sweden and the UK are 23 This result can be seen in scatter diagrams in Yuengert (1993)
branch analyses, which are discussed below. for insurance firms and Humphrey (1987) for banks.
190 A.N. Berger, D.B. Humphrey~European Journal of Operational Research 98 (1997) 175-212

Yuengert's 'solution' to this problem is to permit policy toward depository financial institutions is di-
both the random error term and the inefficiency term vided into four subcategories: (1) deregulation and
to be heteroscedastic and let the data determine the financial disruption, (2) institution failure, risk, prob-
outcome. While this is a useful idea from a theoreti- lem loans, and management quality, (3) market struc-
cal perspective, and could work in very large data ture and concentration, and (4) the effects of mergers
sets, it may not be a practical solution when data sets and acquisitions.
are relatively small as they typically are for most Deregulation, financial disruption. Deregulation
countries. That is, it is very difficult to estimate two is typically undertaken to improve the performance
types of heteroscedasticity from a single composed of the industry being deregulated. If efficiency is
error under the best of situations, but it is likely to be raised, the improvement in resource allocation will
even more difficult when data are relatively limited. benefit society and may lead to price reductions
In practice, it may be best to note the potential for and/or service expansion for consumers if competi-
confounding scale with efficiency effects and at- tion is sufficient. However, in many cases deregula-
tempt to judge the potential bias by comparing scale tion is initiated less by a desire to benefit consumers
estimates obtained in non-frontier models with fron- than by a need to improve the competitive viability
tier scale estimates. If these two sets of scale esti- of the industry. One such example was the removal
mates are very similar, then the bias from 'efficiency of interest rate ceilings on deposits paid by US banks
heteroscedasticity' is likely small, and the measured in the 1980s, which permitted banks to compete
differences in frontier efficiency across size classes better with money market mutual funds in acquiring
of firms may be relatively accurately estimated. For- funds. Another example is the harmonization and
tunately, most estimates of average and frontier scale unification of banking markets in Europe - remov-
economies in banking are fairly similar, suggesting ing restrictions that have limited the ability of banks
that this problem of confounding scale economies in one country from aggressively entering markets in
and inefficiencies due to heteroscedasticity is not other countries.
substantial. Given that a primary goal of deregulation has
been to improve efficiency, the results have been
mixed. Norwegian banks experienced improved effi-
5. Informing government policy toward financial ciency and productivity after deregulation (Berg et
institutions al., 1992) as did Turkish institutions in a more
liberalized banking environment (Zaim, 1995). In
In order to summarize the main findings of effi- contrast, banking efficiency in the US was relatively
ciency studies of financial institutions, the studies unchanged by the deregulation of the early 1980s
listed in Table 1 have been rearranged into three (Bauer et al., 1993; Elyasiani and Mehdian, 1995).
broad categories based upon whether a study's pri- Although measured bank productivity fell
mary contribution was to inform government policy, (Humphrey, 1993; Humphrey and Pulley, 1997), this
to address general research issues, or to improve was largely because interest rate deregulation in-
managerial performance. These studies are shown in duced a competitive scramble to pay higher interest
Table 4. While many studies have contributed di- rates on consumer deposits without a corresponding
rectly to more than one area, and most can be viewed reduction in either banking services or an immediate
as contributing indirectly to policy makers, re- and fully offsetting increase in deposit fees. Thus
searchers, and managers, each study is listed only productivity benefits which otherwise would have
once. 24 The discussion of informing government been captured by banks was instead passed onto
consumers. Spain experienced deregulation results
similar to the US (Grifell-Tatj6 and Lovell, 199To;
24 For example, Giokas (1991) is listed under 'Address research
Lozano, 1995). Lastly, the bursting of the specula-
issues' because it compares efficiency measurement techniques,
but it could have easily been listed under 'Improving managerial tive bubble in Japan seemed to have little effect
performance' because it estimates the efficiency of individual overall on the efficiency of Japanese banks
bank branches. (Fukuyama, 1995), although the bad loans it created
 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175-212 191

clearly had a significant adverse effect on the finan- lation - such as existing excess loan demand in
cial conditions of Japanese banks. Norway, a desire to rapidly expand market share in
Depending on industry conditions prior to deregu- Spain, or competition to pay higher deposit interest

Table 4
Applications of efficiency analysis of f'mancial institutions
Application Country Method a Author (date)
Inform government policy:
Deregulation, financial disruption Norway DEA Berg et al. (1992)
US DEA Elyasiani and Mehdian (1995)
Japan DEA Fukuyama (1995)
Spain TFA Lozano (1995)
Turkey DEA Zaim (1995)
US TFA Humphrey and Pulley (1997)
Spain DEA Grifell-Tatj~ and Lovell (1997b)

Institution failure, US TFA Berger and Humphrey (1992a)

risk, problem loans, US SFA Cebenoyan et al. (1993a)
and management quality US DEA Ban" et al. (1994)
US DEA Elyasiani et al. (1994)
US DEA Hermalin and Wallace (1994)
US SFA Berger and DeYoung (1997)
US SFA Mester (1996)
US SFA Mester (1997)
US TFA DeYoung (1997c)

Market structure and concentration Norway TFA Berg and Kim (1994)
US DFA Berger (1995)
US DEA Devaney and Weber (1995)
Norway TFA Berg and Kim (1996)
Spain SFA Mandos (1996b)
US DFA Berger and Hannah (1997)

Me~e~ Norway DEA Berg (1992)

US DFA Berger and Humphrey (1992b)
US IN Fixler and Zieschang (I 993)
US DFA Akhavein et al. (1997a)
US TFA DeYoung (1997b)
US DFA Peristiani (1997)

Address research issues:

Confidence intervals Italy DEA Ferrier and Hirschberg (1994)
US DEA Wheelock and Wilson (1994)

Comparing different US DEA,SFA Ferrier and Lovell (1990)

efficiency techniques Greece DEA,SFA Giokas (1991)
or assumptions US SFA,DFA ,TFA Bauer et al. (1993)
US DEA,SFA Eisenbeis et al. (1996)
Spain SFA Maudos (1996a)
Germany SFA Altanbas and Molyneux (n.d.)
US SFA Zhu et al. (1997)

Comparing different Norway DEA Berg et al. (1991)

output measures Finland DEA Kuussaari (1993)
Italy DEA Favero and Papi (1995)
US DFA Hunter and Timme (1995)
Finland DEA Kuussaari and Vesala (1995)
192 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175-212

Table 4 (continued)
Application Country Method Author (date)
Organizational form, US DEA Rangan et al. (1988)
corporate control issues US DEA Aly et al. (1990)
US DEA Elyasiani and Mehdian (1992)
US SFA Cebenoyan et al. (1993b)
US SFA Chang et al. (1993)
US DEA Grabowski et al. (1993)
US SFA Mester (1993)
US DFA Newman and Shrieves (1993)
US SFA Pi and Timme (1993)
US DFA DeYoung and Nolle (1996)
US TFA Mahajan et al. (1996)
India DEA Bhattacharyya et al. (1997)
US SFA,TFA Hasan and Hunter (forthcoming)
General level US DEA Eiyasiani and Mehdian (1990a)
of efficiency UK DEA Field (1990)
UK DEA Drake and Weyman-Jones (1992)
Tunisia SFA Chaffai (1993)
Japan DEA Fukuyama (1993)
Switzerland DEA Sheldon and Haegler (1993)
Denmark DEA Bukh (1994)
US SFA Kaparakis et al. (1994)
Spain DEA Perez and Quesada (1994)
Germany TFA Lang and Welzei (1995)
Italy DEA,SFA Resti (1995)
Germany DFA Lang and Welzel (1996)
US DEA Miller and Noulas (1996)
Intercountry comparisons Norway DEA Berg et al. (1993)
Sweden
Finland
11 OECD SFA Fecher and Pestiean (1993)
countries
8 developed DEA Pastor et al. (1997)
countries
Norway DEA Bukh et al. (1995)
Sweden
Finland
Denmark
15 developed TFA Ruthenberg and Elias (1996)
countries
Methodology issues US DEA Charnes et al. (1990)
US TFA Berger and Humphrey (1991)
US DFA Berger (1993)
Belgium FDH Tulkens (1993)
US DEA Ferrier et al. (1994)
Spain DEA Grifell-Tatj6 and Lovell (1994)
Norway MOS Bergendahl (1995)
Sweden
Finland
Denmark
US DFA Adams et al. (1995)
US DFA Akhavein et al. (1997b)
US FDH DeBorger et al. (1995)
Spain DEA Pastor (1995)
US DEA Thompson et al. (1996b)
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 193

Table 4 (continued)
Application Country Method Author (date)
US DFA Berger and Mester (1997)
Tunisia SFA Chaffai (1997)
US DFA DeYoung (1997a)
Spain DEA Grifell-Tatj6 and Lovell (1997a)
Spain DEA Lovell and Pastor (1997)
Mexico DEA Taylor et al. (1997)
US DEA Thompson et al. (1997)

Opportunity cost, US DEA Ferrier et al. (1993)

output diversification US TFA DeYonng (1994)
France DFA Dietsch (1994)
France DFA Chaffai and Dietsch (1995)
US TFA Clark (1996)
Profit, revenue US SFA Elyasiani and Mehdian (1990b)
US DFA Berger et al. (1993a)
US DEA English et al. (1993)
Spain TFA Lozano (1997)
Spain DEA Grifell-Tatj6 and Lovell (1996)
US SFA Ellinger et al. (1997)
Stability over time, US SFA Kwan and Eisenbeis (1994)
institution size US DEA Ray and Mukherjee (1994)
US DEA Bauer et al. (1995)

Improve managerialperformance:
Credit unions US FDH Fried et al. (1993)
US FDH Fried and Lovell (1994)

Bank branch US DEA Sherman and Gold (1985)

Canada DEA Parkan (1987)
Turkey DEA Oral and Yolalan (1990)
Greece DEA Vassiloglou and Giokas (1990)
Saudi Arabia DEA AI-Faraj et al. (1993)
Belgium FDH Tulkens and Malnero (1994)
UK DEA Athanassopoulos (1995)
US DEA Sherman and Ladino (1995)
US DFA Berger et al. (1997)
Cyprus DEA Zenios et al. (1996)
UK DEA Athanassopoulos (1997)
Canada DEA Schaffnit et al. (1997)
UK DEA Drake and Howcroft (1997)

S & L branch Sweden DEA Hartman and Storbeck (1995)

a See notes to Table I.

rates in the US - the consequences of deregulation incorrect. Industry conditions prior to deregulation
may differ across countries. Indeed, in some cases, and other incentives may intervene. Measurement
deregulation appears to have led to a reduction in over longer time periods may eventually show a net
measured productivity rather than an improvement. improvement in both efficiency and productivity but
The implication for government policy is that the this has not yet been demonstrated.
conventional wisdom which holds that deregulation Institution failure, risk, problem loans, and man-
always improves efficiency and productivity may be agement quality. A key role of a country's financial
194 A.N. Berger, D.B. Humphrey~European Journal of Operational Research 98 (1997) 175-212

institution regulators is to limit systemic risk - the regional downtums, then measured cost efficiency
risk that the problems of a few institutions spread to may be artificially low because of the expenses
many other institutions that are otherwise solvent associated with dealing with these loans (e.g., extra
and liquid. This protects the money supply and the monitoring, negotiating workout arrangements, etc.).
payment system from being severely disrupted and Alternatively, problem loans may be related to mea-
involves the management of bank failures. Most sured efficiency because 'bad management' is poor
bank failures are directly related to having a large at controlling both costs and risks. If 'bad luck'
number of problem loans, a low capital position, a dominates, then problem loans are mostly exogenous
weak or negative cash flow, and poor management and should be controlled for in efficiency models. If
quality. It might be expected that institutions would 'bad management' dominates, then problems loans
display low efficiency prior to failure and that man- are essentially endogenous to financial institution
agement quality would be positively related to effi- efficiency and should not be controlled for in the
ciency. Both of these priors are supported in studies analysis of efficiency, To this point, the evidence is
that have looked at these issues. mixed, yielding some support for both hypotheses
Banks and S&Ls with low efficiency failed at (Berger and DeYoung, 1997). A potential solution to
greater rates than institutions with higher efficiency this problem is to control for the problem loan ratio
levels (Berger and Humphrey, 1992a; Cebenoyan et for the state or region of the bank, which should
al., 1993a; Hermalin and Wallace, 1994) and this primarily reflect the 'bad luck' facing the bank,
relationship was evident a number of years ahead of rather than its own 'bad management' (see Berger
eventual failure (Barr et al., 1994). Management and Mester, 1997).
quality, as measured by regulatory agency assess- Market structure and concentration. An important
ments, is positively related to cost efficiency (De- area of government policy concerns antitrust issues.
Young, 1997c) which, in turn, Granger-causes reduc- Many studies of financial institutions and other firms
tions in problem loans (past due and nonaccrual, have found a positive statistical relationship between
Berger and DeYoung, 1997). As a result, efficiency market concentration and profitability. This may be
measures have been shown to improve the predictive due to market-power explanations in which firms in
accuracy of failure prediction models and thus may concentrated markets exercise market power in pric-
represent a useful addition to current modeling ef- ing and earn supernormal profits. Altematively, the
forts by regulatory agencies (as shown by Barr et al., efficient-structure paradigm links concentration to
1994, for banks and Kramer, 1997, for insurance high profitability through efficiency (Demsetz, 1973).
firms). 25 Under efficient-structure, relatively efficient finns
Problem loans have been included as explanatory compete more aggressively for and gain dominant
variables in some efficiency studies (e.g., Hughes market shares and also have high profits because of
and Mester, 1993; Mester, 1996, 1997) with the their low costs of production. These different expla-
result that slight measured scale diseconomies for nations of differences in profitability across firms -
larger institutions are altered to economies and effi- market power vs. efficiency - have directly oppos-
ciency is increased. Whether or not it is appropriate ing implications for antitrust policy. If high profits
to control for problem loans depends on which is the are created by market power, then antitrust actions
dominating explanation for the observed negative are likely to be socially beneficial, moving prices
relationship between measured efficiency and prob- toward competitive levels and allocating resources
lem loans. If problem loans are generally caused by more effectively. However, if high efficiency is the
'bad luck' events exogenous to the bank, such as explanation for high profits, then breaking up effi-
cient finns that have gained large market shares or
disallowing efficient firms to acquire other firms is
likely to raise costs and may lead to prices less
25 This result is not surprising since it has been shown that
information contained in a efficiency measure closely corresponds
favorable to consumers. Regulatory agencies have
to that contained in standard financial ratios (Elyasiani et al., typically followed the market-power paradigm in
1994). their antitrust policies.
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 195

The evidence comparing market power and effi- whether the relationships among efficiency, concen-
ciency effects is limited, but it suggests that cost tration, prices, and profitability found in local US
efficiency is somewhat more important than market markets obtain in other nations, where banking mar-
power in explaining profitability. However, as mea- kets are typically more national in scope and are
sured by R 2,s, neither efficiency nor market power generally much more highly concentrated.
explains much of the observed variance of profitabil- An alternative approach to examining market
ity (Berger, 1995; Maudos, 1996b). structure questions is to rely on direct estimates of
Although concentration is not significantly posi- the degree of oligopolistic output interdependence
tively related to profitability after controlling for among suppliers of financial services. Adopting this
efficiency, higher concentration is significantly asso- approach, a frontier conjectural variations model has
ciated with lower deposit interest rates and higher been estimated for Norwegian banks (Berg and Kim,
loan rates even after accounting for efficiency differ- 1994). From the view of efficiency analysis, the
ences (Berger and Hannan, 1997). One explanation innovation is that market structure effects are di-
seems to be that financial institutions with more rectly accounted for when estimating scale economies
market power charge higher prices but, instead of and frontier efficiency. This is done by specifying
enjoying higher than average profits, experience re- that each firm's cost is a function of its own output
duced cost efficiency as managers pursue other goals as well as the output level of other firms in the same
and a 'quiet life' (in Hicks' words). 26 The extra market. While the average of firm frontier cost effi-
costs from 'quiet life' inefficiency have been esti- ciency estimates is only little affected by including
mated to several times larger than the traditional or excluding conjectural variation effects, the scale
welfare triangle costs from the exercise of market economy measure moves from indicating constant
power (Berger and Hannan, 1995). costs in a standard cost frontier framework to indi-
Unfortunately, most of the research on this topic cating decreasing costs when conjectural variations
has been on the US banking industry, where the are added. In a profit function context, however,
structure of the industry is quite different from the adding conjectural variations significantly improves
rest of the world. In the US, many financial products measured profit efficiency, suggesting that this as-
such as retail deposits and small business loans are pect of market behavior is likely an important factor
essentially only competed for on a local basis so in efficiency measurement (Berg and Kim, 1996).
prices can differ significantly among these markets. Mergers and acquisitions. Relative to historical
Most of the studies have focused on the relationship trends, banking industries in a number of countries
between local market concentration and measures of have been subject to an increased number of mergers
bank performance. Although some financial prod- and acquisitions. In the US, much of the activity has
ucts, such as large certificates of deposit and large been spawned by liberalizations of state rules regard-
wholesale loans are competed for on a nationwide ing bank and bank holding company expansion both
basis, the US national market is extremely unconcen- within and between states. In the early 1980s, there
trated by world standards. For example, it would was almost no interstate banking activity, but by the
take over 2000 banking organizations to account for end of 1994, 28% of US banking assets were con-
90% of deposits in the US, while in most other trolled by out-of-state banking organizations, primar-
developed countries 90% of deposits would be ac- ily through regional compacts among nearby states
counted for by fewer than 10 organizations. It would (Berger et al., 1995).
be of research and policy importance to discover The conventional wisdom among bank consul-
tants and the popular press is that mergers can be
and have been successful in improving cost ratios
and cost efficiency, at least for a number of firms.
However, academic studies usually find no such
26 Some support is found in Devaney and Weber (1995) who
improvement on average. This holds whether simple
determined that rural US banking markets that experience a
decrease in concentration appear to have greater efficiency and accounting ratios are compared pre- and post-merger,
productivity growth. holding industry effects constant, or in more sophis-
196 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

ticated econometric analyses using frontier cost func- 6. Address research issues related to financial
tions (Berg, 1992; Berger and Humphrey, 1992b; institutions
Rhoades, 1993; Peristiani, 1997; DeYoung, 1997b).
Although many individual mergers have been Much of the work in efficiency analysis has been
quite successful in improving cost performance, many focused on methodology and measurement issues.
others have worsened their cost ratios or cost effi- Research issues include the study of: (1) the similar-
ciency, so that on average there is no significant ity of efficiency results derived from different fron-
improvement. This would suggest that government tier models, (2) the sensitivity of efficiency results
merger policy should not as a rule be influenced by when different output measures are applied, (3) the
claims of expected cost efficiency benefits from association between efficiency and firm organiza-
mergers. However, an exception could occur if there tional structure, (4) different ways to measure effi-
existed a reliable precondition that could be used to ciency, (5) the effects of incorporating opportunity
identify mergers that are very likely to improve cost cost and product diversification in the analysis, (6)
efficiency. Two plausible preconditions were found, the consistency among cost, profit, and production
upon testing, to provide little in the way of signifi- efficiency measures, and (7) the variability of effi-
cant additional information. First, it was expected ciency estimates over time. The general level of
that a successful merger might be one with a high efficiency, along with broad comparisons of effi-
degree of local market overlap between merging ciency levels across different frontier techniques and
institutions because of the greater potential for elimi- countries, has been discussed above. The survey that
nating duplicate expenditures on branches and back- follows will thus focus on the remaining research
office operations. Second, it was thought that merg- issues noted in Table 4.
ers would be more successful when the acquiring Confidence intervals and comparing different effi-
firm is more cost efficient than the firm being ac- ciency techniques or assumptions. The effect that
quired, because the superior management team would different frontier approaches can have on estimates
gain control and use its (apparently) demonstrated of industry average and individual firm efficiency
ability to improve the less efficient firm. Upon inves- estimates has been noted above. 27 In almost all of
tigation, neither of these expectations were realized these analyses, conclusions have been drawn from
(Berger and Humphrey, 1992b). only point estimates of efficiency. Thus it is of
The effects of mergers on profit efficiency have interest to derive confidence intervals for efficiency
been less intensively investigated. However, initial estimates in order to determine if the efficiency
results suggest that profit efficiency improves signif- comparisons being made are meaningful in a statisti-
icantly from mergers of large banks (Akhavein et al., cal sense. Fortunately, bootstrapping methods have
1997a). The different results experienced for cost become more widely known and available (Efron
and profit efficiency appear to occur because mea- and Tibshirani, 1993; Hall et al., 1993; Mooney and
sured cost efficiency changes do not take into ac- Duvai, 1993; Atldnson and Wilson, 1995). Thus new
count the effects of the changes in output that occur research in the efficiency area should try to make it a
after the merger. Merging banks tend to shift their practice to provide confidence intervals for the effi-
output mixes away from securities toward loans, ciency estimates they generate. Somewhat similar
which raises profit efficiency because issuing loans information may be obtained through sensitivity
creates more value (and usually more risk) than analysis to examine the robustness of efficiency esti-
purchasing securities. This shift in mix may occur mates (Brockett et al., 1995; Thompson et al., 1996a).
because merging banks are better able to diversify
these risks than the previous management, allowing a
higher loan/asset ratio to be held with same amount
of capital (see also Benston et al., 1995; Hughes et 27 Additional studies, not mentioned earlier, have examined the
stability of efficiency estimates from an SFA model when differ-
al., 1996). Further investigation of the profit effi-
ent efficiency distributions are specified (Altunbas and Molyneux,
ciency vs. cost efficiency effects of mergers repre- n.d.) or when different functional forms are specified (Zhu et al.,
sents an area for fruitful additional research. 1997).
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 197

When confidence intervals of efficiency estimates the alternative 'intermediation' approach, financial
have been provided, these intervals appear to be institutions are thought of as primarily intermediat-
quite large (Simar, 1992; Ferrier and Hirschberg, ing funds between savers and investors. With this
1994; Mester, 1996, 1997). 28 When confidence in- approach, since service flow data are not usually
tervals are large, comparisons of efficiency among available, the flows are typically assumed to be
firms in an industry, or branches within a firm, may proportional to the stock of financial value in the
be more meaningful when groups of observations, accounts, such as the numbers of dollars of loans,
rather than individual values, are being compared. deposits, or insurance in force (e.g., Berger and
Thus the common practice of regressing point esti- Humphrey, 1991).
mates of firm efficiency, or rankings of firms by These approaches also have implications for which
their efficiency value, on sets of explanatory vari- inputs or costs should be included in the analysis.
ables might be improved upon or augmented with a Under the production approach, only physical inputs
subset analysis. This would involve an additional such as labor and capital and their costs should be
examination of the data where only subsets of finns included, since only physical inputs are needed to
with relatively high and relatively low efficiency perform transactions and process financial docu-
values are used in a regression (with the middle ments. Under the intermediation approach, the input
group of firms excluded) to explore the robustness of of funds and their interest cost should also be in-
the posited relationships. This could also be used to cluded in the analysis, since funds are the main 'raw
determine if the correspondence of firm-level effi- material' which is transformed in the financial inter-
ciency estimates among different frontier methods mediation process.
could be improved if only the most important subsets Neither of these two approaches is perfect be-
of observations, rather than all the observations, cause neither fully captures the dual roles of finan-
were used. 29 cial institutions as (i) providing transactions/docu-
Comparing different output measures. There are ment processing services and (ii) being financial
two main approaches to the choice of how to mea- intermediaries that transfer funds from savers to
sure the flow of services provided by financial insti- investors. While it would probably be best to employ
tutions. Under the 'production' approach, financial both approaches to determine whether the results
institutions are thought of as primarily producing were qualitatively affected by the choice of output
services for account holders. The financial institu- metric, sufficient data to implement such a research
tions perform transactions and process documents for design are not usually available. Nevertheless, each
customers, such as loan applications, credit reports, of the approaches has some advantages. The produc-
checks or other payment instruments, and insurance tion approach may be somewhat better for evaluating
policy or claim forms. Under this approach, output is the efficiencies of branches of financial institutions,
best measured by the number and type of transac- because branches primarily process customer docu-
tions or documents processed over a given time ments for the institution as a whole and branch
period (e.g., Kuussaari and Vesala, 1995). Unfortu- managers typically have little influence over bank
nately, such detailed transaction flow data is typi- funding and investment decisions. The intermedia-
cally proprietary and not generally available. As a tion approach may be more appropriate for evaluat-
result, data on the stock of the number of deposit or ing entire financial institutions because this approach
loan accounts serviced or the number of insurance is inclusive of interest expenses, which (depending
policies outstanding are sometimes used instead (e.g., on the phase of the interest rate cycle) often accounts
Ferrier and Lovell, 1990; Ferrier et al., 1993). Under for one-half to two-thirds of total costs. As well, the
intermediation approach may be superior for evaluat-
ing the importance of frontier efficiency to the prof-
2s Confidence intervals are computed for year-to-year changes in itability of the financial institution, since minimiza-
efficiency in Wheelock and Wilson (1994).
29 One study where this was done suggested that the improve-
tion of total costs, not just production costs, is
ment may be slight (Bauer et al., 1993). Even so, it may be useful needed to maximize profits.
to see if this result holds up in additional analyses. One study compared the production and interme-
198 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

diation approaches by applying both to the same data proportionate to the output of depositor services
set of bank branches using the same functional form, provided (Berger and Humphrey, 1991; Bauer et al.,
finding correlations above 0.40 between the frontier 1993).
efficiency rankings of the two approaches (Berger et Other efficiency studies have first treated deposits
al., 1997). Other studies have also compared effi- as an input and then as an output. These investiga-
ciency results obtained with outputs measured by tions find that efficiency is somewhat higher when
numbers of accounts vs. the financial values in these deposits are specified as an output. In a DEA model,
accounts. In one case, little difference was found in the RRANK between these two specifications aver-
the distribution of efficiency estimates when these aged 0.77 (Favero and Papi, 1995), while in a DFA
two stock indicators of financial firm output were model RRAN~= 0.16 (Hunter and Timme, 1995).
used (Berg et al., 1991) while in another case, a Since the treatment of deposits in efficiency models
similar distribution was found but mean efficiency can affect the efficiency estimates, this aspect of
was higher when financial values were specified model specification may be of some importance to
(Kuussaari, 1993). Although the efficiency estimates the outcome.
had a similar distribution, the rankings of firms Organizational form and corporate control issues.
within these distributions differed. In Berg et al. Financial institutions are organized in a number of
(1991), the average RRANK for the two comparisons different ways. Relying on agency theory, some
made was 0.64 while RRANK was 0.32 in Kuussaari studies have investigated whether organizational form
(1993). Overall, it appears that inferences regarding is associated with differences in frontier efficiency.
efficiency may be importantly affected by how out- Firms owned by stockholders might be expected to
put is measured, a result which is usually less depen- face stronger incentives to control costs a n d / o r en-
dent upon investigator choice than availability of hance profits compared to mutual organizations
data. where depositors or policyholders own the firm. The
Despite the many other differences in assumptions evidence is mixed, One study found that stock S &Ls
involved in measuring efficiency, there is reasonable were less efficient than mutual S &Ls (Mester, 1993)
agreement about the specification of most of the while another found that efficiency was not signifi-
important inputs and outputs for financial institu- cantly related to this difference in ownership
tions. The asset, user cost, and value-added methods (Cebenoyan et al., 1993b). This issue might be
of assigning financial goods to input and output somewhat confounded by the fact that so many
categories all agree that loans and other major assets S &Ls have switched status, possibly creating a sam-
of financial institutions should count as outputs. ple selection bias if either inefficient or efficient
However, there is a longstanding controversy whether firms switched at a greater pace. Study of frontier
deposits should count as inputs or outputs. Deposits efficiency in the US life insurance industry (Gardner
have input characteristics because they are paid for and Grace, 1993) and in the US property-liability
in part by interest payments and the funds raised insurance industry (Berger et al., 1996a) found no
provide the institution with the raw material of in- significant differences between stocks and mutuals in
vestible funds. However, deposits also have output cost efficiency, but stock firms providing property-li-
characteristics because they are associated with a ability insurance were sometimes statistically signifi-
substantial amount of liquidity, safekeeping, and cantly more profit efficient than mutuals, all else
payments services provided to depositors. held equal.
Some studies resolve this issue with a dual ap- In the US banking industry, the primary organiza-
proach that captures both the input and output char- tional trade off for large organizations is between a
acteristics of deposits. The interest paid on deposits multibank holding company (MBHC) arrangement,
is counted as part of costs and the rate paid is where a commonly-owned group of banks have sep-
included as an input price, both consistent with the arate charters and financial books, vs. an extensive
input of the raw material of investible funds. These branch banking arrangement where banks have been
same studies specify the quantities of deposits as merged under a single charter within a larger branch-
outputs because these quantities are assumed to be ing network with a consolidated operation. This will
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 199

likely be an important issue over the next several found no difference between one-bank and multi-
years in the US, as the Riegle-Neal Interstate Bank- bank holding company affiliation (Elyasiani and
ing and Branching Efficiency Act of 1994 allows Mehdian, 1990b). All of these issues will require
widespread interstate branch banking for the first substantially more future research to resolve.
time in many decades. The results of one study Methodology issues. Since our focus in this article
suggest that branch banking may lead to greater is on the application of frontier efficiency techniques
efficiency than keeping banks separate within an to financial institutions, the methodology studies
MBHC (Grabowski et al., 1993). 30 noted here are mostly limited to those in which there
A related issue concerns possible efficiency dif- has been an application of new methodology to
ferences associated with foreign vs. domestic owner- financial institutions. There have been a number of
ship. Four studies have found that foreign-owned attempts to improve both nonparametric and para-
banks in the US were significantly less efficient than metric frontier models and estimation in this field.
US-owned banks (Chang et al., 1993; DeYoung and Improvements a n d / o r alternatives to the standard
Nolle, 1996; Mahajan et al., 1996; Hasan and Hunter, DEA nonparametric approach concern the develop-
1996). In contrast, foreign owned banks in India ment and application of FDH (Fried et al., 1993;
were found to be somewhat more efficient than Tulkens, 1993; Fried and Lovell, 1994), the polyhe-
privately owned domestic banks but government dral cone-ratio DEA model (Chames et al., 1990;
owned banks were more efficient that both (Bhat- Brockett et al., 1997), and the assurance region DEA
tacharyya et al., 1997). It has been suggested that model (Thompson et al., 1997; Taylor et al., 1997).
foreign-owned banks in the US have in effect traded In addition, the nonparametric Malmquist Index ap-
current profits for rapid expansion of market share. proach to efficiency measurement has been general-
The rapid growth was made possible by relying on ized (Grifell-Tatj6 and Lovell, 1994), goal program-
purchased funds, which are more expensive than ming is being applied (Cooper et al., 1997), and the
core deposits raised through a network of branches, sensitivity of DEA and FDH efficiency models to
which takes time to establish. different radial and non-radial measurement tech-
The evidence is also quite limited on the links niques is being tested (Ferrier et al., 1994; Pastor,
between other aspects of corporate governance and 1995; DeBorger et al., 1995). The general conclusion
frontier efficiency. When the CEO is also the chair- of this work is that the standard DEA model, along
man of the board, efficiency has been measured to be with the radial measurement of efficiency, may not
lower in one study, and this effect is not offset by be as well-suited to distinguishing efficient from
having a higher proportion of outside directors on inefficient observations as the newer approaches cited
the board (Pi and Timme, 1993). In another study, here.
minority ownership was investigated, but no signifi- From another perspective, two recent additions to
cant differences were found (Elyasiani and Mehdian, the DEA literature promise to extend the analysis in
1992). A different ownership issue concerns possible important new directions. First, it has been suggested
efficiency differences among banks depending upon (Bergendahl, 1995) that perhaps the DEA frontier
holding company status. One study found that being should be composed of the most efficient parts of
in a holding company seemed to confer some cost banks within the sample - forming a composite or
advantages compared to remaining independent representative firm, rather than being composed of
(Newman and Shrieves, 1993), but another study separate and individual firms as is now the case. A
'composite frontier' would serve to indicate the effi-
ciency that had been achieved within the sample,
30However,no significantdifferencesin efficiencywere found although not necessarily all at a single institution.
between banks locatedin branching vs. non-branching(unit bank- Such a frontier more accurately represents what is
ing) states in the US (Rangan et al., 1988; Aly et al., 1990). This possible and does not confound efficient results in
result is essentially a weaker test of the branching/separatebank
relationship since the separate bank arrangementwas importantin one specified area of interest with inefficient results
states that restrictedbranching while both branching and separate from other areas. This 'composite frontier' is theoret-
bank arrangementsexisted in states that permitted branching. ically similar to the 'true' best-practice frontier dis-
200 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

cussed elsewhere, which would be made up only of dures have also been attempted. This work has fo-
branches (and other financial institution units) that cused on replacing the translog functional form with
are fully efficient (Berger et al., 1997). Both con- the more flexible Fourier form (e.g., Berger and
cepts seek to set higher standards for the frontier Mester, 1997), the use of random coefficient estima-
than any firm in the sample has achieved by looking tion which also provides greater flexibility (Akhavein
at the best-practice segments of firms. et al., 1997b), and correcting for situations where the
A second analysis (Lovell and Pastor, 1997) im- regressors and error are correlated (Adams et al.,
plements a statistical test of the effect of sequentially 1995). In general, greater flexibility has resulted in
reducing the number of constraints in a DEA model. higher estimates of efficiency. At present, the choice
The goal is to provide a method whereby the con- between the various parametric models and estima-
straints in the DEA model can be collapsed down to tion procedures is based primarily on ease of use
only those that are important to the results obtained. a n d / o r the apparent reasonableness of underlying
With this approach, extraneous constraints can be assumptions, rather than on any strong theoretical
discarded and attention focused on only influential foundation.
constraints. This work, as well as that on cone-ratio To date, parametric efficiency analysis has essen-
and assurance region DEA models which both spec- tially assigned all deviations from an estimated effi-
ify additional a priori information, also address the cient frontier to a dependent variable such as total
problem where individual bank observations may be costs or profits. Importantly, the resulting ineffi-
100% efficient by default (due to non-comparability ciency value can be made more informative by addi-
among observations when 'too many' constraints are tionally decomposing it into its technical and alloca-
specified). The DEA assurance region model, for tive components. Further information is obtained
example, has consistently reduced the number of when inefficiency can be directly related to specific
bank 'self-identified' observations (Taylor et al., inputs. This has been done by Kumbhakar (1988)
1997; Thompson et al., 1996b, 1997). These tech- and Chaffai (1997).
niques should go a long way toward ensuring that Opportunity cost, output diversification. Nonpara-
extraneous constraints do not 'contaminate' the DEA metric and parametric studies can underestimate effi-
results and thus may generate more consistent effi- ciency when important cost influences have not been
ciency estimates across different studies. 31 included in the analysis. Two such influences rou-
Similar efforts to improve the standard parametric tinely neglected in earlier studies have been the
SFA frontier model include the development of two opportunity cost of equity and the expenses under-
alternatives - the thick frontier approach TFA taken to reduce risk. An expected result would be
(Berger and Humphrey, 1991), and distribution-free that including these additional costs may improve
approach DFA (Berger, 1993), the latter being a efficiency estimates as the cost or profit function
modification of earlier work by Schmidt and Sickles would fit the data more closely and less specification
(1984). More general parametric estimation proce- error might be counted as inefficiency. Studies incor-
porating these factors include Dietsch (1994), Clark
(1996), Mester (1996, 1997), and Berger and Mester
sl In some cases, the choice of which constraints to discard may (1997). Clark (1996) and Berger and Mester (1997)
be fairly straightforward, such as when some of the constraints compared efficiency estimates that did and did not
essentially contain the same economic information. An example is account for the effects of equity capital and Clark
a DEA model that specifies as bank inputs transactions deposits,
nontransactions deposits, total noninterest expenses, and total
found that accounting for equity raised measured
interest expenses (e.g., Miller and Nonlas, 1996, and other papers cost efficiency, and Berger and Mester found that it
referenced there). The goal of trying to capture the effects of raised estimated profit efficiency substantially. As
funding mix (transactions vs. nontransactions deposits) along with noted above, there also can be a problem with con-
the interest rates paid (interest vs. noninterest expenses), while trolling for problem loans or other variables that may
laudable, may be problematic since in cross-section data nontrans-
actions deposits are virtually the sole source of interest expenses
be endogenous to the decisions of the firm being
while transactions deposits make up the largest segment of nonin- studied.
terest expenses. Extending efficiency analysis in a different direc-
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 201

tion, some research has been done on the effects on A number of the studies cited in Table 1 measure
efficiency from output diversification and product profit efficiency. The mean profit efficiency from
diversity. The 'optimal scope economies' concept studies of US depository institutions is 0.64, so these
based on the profit function rather than the cost firms were earning about 64% of their potential
function includes all the revenue effects of output profits on average (Berger et al., 1993a; DeYoung
choices as well the cost effects of input choices and Nolle, 1996; Miller and Noulas, 1996; Akhavein
(Berger et al., 1993a). On the cost side, a measure of et al., 1997a; Akhavein et al., 1997b; Berger and
diversification more general than the traditional scope Mester, 1997; Humphrey and Pulley, 1997; Hasan
concept was applied to US banks (Ferrier et al., and Hunter, 1996). Similarly, a study of Spanish
1993). It was found that greater diversification tended depositories found average profit efficiency of 0.72
to reduce cost efficiency. Similarly, 'universal' banks (Lozano, 1997). 33 Much lower profit efficiency was
in Europe (who provide a broader mix of services) found for large merging US banking organizations
were found to experience lower cost efficiency than using the D F A method, 0.24 before merger, 0.34
more specialized banks (Chaffai and Dietsch, 1995). after merger (Akhavein et al., 1997a). This contrasts
A seemingly contrary result was found in an analysis with a D E A study of large US banks which found
of the effects of shifting from making bank loans to profit efficiency of 0.97, with 42% of the firms being
providing a broader mix of services by expanding 100% technically efficient (Miller and Noulas, 1996).
fee-based services, since the shift was associated A profit efficiency study of US banks using random
with higher (not lower) banking efficiency (De- coefficients found the average efficiency to be highly
Young, 1994). dependent upon the choice of subsample (Akhavein
Profit and revenue efficiency. Most of the para- et al., 1997b).
metric models applied to financial institutions have A study of insurance companies found average
focused on cost efficiency while nonparametric mod- profit efficiencies on the order of about 60% effi-
els have concentrated on the relationship between cient (Berger et al., 1996a). When profit efficiency
inputs and outputs directly. An area only recently and cost efficiency results are made comparable by
attracting interest has been the estimation of profit expressing the quantities of inefficiency in terms of a
and revenue efficiency. The techniques are essen- c o m m o n denominator, cost inefficiency was found to
tially the same but the data are different. Profit be larger than profit inefficiency, suggesting that cost
efficiency is concerned with both cost and revenue inefficiency may be overstated because of differ-
efficiency but only under certain conditions would it ences in service quality or other variables not ac-
be likely that the former will equal the sum of the counted for in the analysis. 34
latter. This is because cost (revenue) efficiency pre- Some of these studies employ an alternative profit
sumes that the observed level of output (input use) is function in which the firm maximizes profits given
already profit maximizing, which may or may not be
the case in practice. In addition, there may be differ-
ences in the quality of some financial services that 33 Profits (and productivity) in the Spanish banking industry
have been decomposed into a productivityeffect (technical change
are not captured in the output measures. This may and operating efficiency), an activity effect (product mix and
make high-quality producers appear to be cost ineffi- scale), and a price effect (Grifell-Tatj6 and Lovell, 1996, 1997a).
cient because of the extra expenses associated with Less comprehensively, profits in the US banking industry have
producing the higher quality output. Such a problem been split into an endogenous or management determined compo-
may be ameliorated by the use of a profit function or nent and a exogenous or external 'business conditions' component
(Humphrey and Pulley, 1997).
profit programming orientation because high quality 34 In another study of banks, the reverse was found: expressed
should be rewarded in the marketplace by extra as a percent of total assets, profit inefficiency was over 10% of
revenues that offset the extra expenses. 32 asset value while cost inefficiency was between 1% and 3% of
asset value (Ellinger et al., 1997). The cost result is possible, since
total costs as a percent of assets averages around 7% for banks.
32 The alternative of directly specifying service-level or quality However, since profits as a percent of assets are usually only
constraints or variables directly in a cost model is usually not around 1%, it is bard to believe that average inefficiency is 10
possible due to limited and proprietary data. times the level of profits.
202 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

output quantities, rather than taking output prices as Humphrey, 1991); and yet another examined the
exogenous (Berger et al., 1996a; Humphrey and stability of frontier banks over time (Berg, 1992).
Pulley, 1997; Akhavein et al., 1997a; Berger and Another study tried to determine for DFA the
Mester, 1997; Hasan and Hunter, 1996). In most number of years that may be needed to strike a
cases, the alternative profit function provides qualita- balance between the benefits and costs of the extra
tively similar results to the standard profit function. information from adding another year of data. The
The alternative profit function may be useful when benefits come from having another residual to help
one or more of the assumptions underlying the stan- average the random error toward zero to get more
dard cost and profit efficiency models are violated precise estimates of the inefficiency term, whereas
by the data (e.g., competitive imperfections, unmea- the costs come from the increasing likelihood that
sured differences in product quality). 35 the efficiency in the extra year has drifted further
Revenue efficiency is essentially the mirror image away from its level at point being measured. The
of cost inefficiency, incorporating errors in the choice study found that the benefits and costs balance out at
of output mix, having too little output, etc. Although about 6 years for US banking data (DeYoung, 1997a).
few revenue frontier analyses have been undertaken, However, the apparent persistence of relative effi-
revenue efficiency estimates (as measured by an ciencies across firms over time does not necessarily
output distance function) appear to be similar to carry over to changes in the overall level or distribu-
those for cost efficiency (English et al., 1993). 36 tion of efficiency. Advances and declines in year-to-
Stability over time. A final research issue con- year efficiency affect banks over time (Bauer et al.,
cerns the stability of efficiency over time. This refers 1995). Finally, although numerous studies have com-
both to average efficiency levels for an industry and mented on how efficiency seems to differ or not
for rankings of firms by their efficiency level. This is differ across size classes of banks (e.g., Ray and
an important issue for the DFA frontier model since Mukherjee, 1994), our view is that these simple
this efficiency measure is based on the assumption contrasts remain unreliable until the issue discussed
that firm efficiency is stable over time and that above of possibly confounding scale economies with
random error, when averaged, will be close to zero. inefficiency due to heteroscedasticity in the data is
Several studies have found that efficiency is reason- more completely resolved.
ably persistent over time: two studies computed a
series of correlations among firms ranked by their
estimated SFA and DEA efficiency level over time 7. Improving managerial performance at financial
(Kwan and Eisenbeis, 1994; Eisenbeis et al., 1996); institutions
another looked for consistency in groups of high and
low cost firms over a number of years (Berger and In principle, virtually any efficiency study of
financial institutions can be used as a tool by man-
agers to improve performance, as long at there is
information in the study on the characteristics or
identities of the relatively efficient and inefficient
institutions. Management practices or characteristics
35 See Berger and Mester (1997) for further discussion. that are found to be relatively common among finan-
36 Revenue efficiency is expressed as a percent of revenue and,
cial institutions on or near the efficient frontier may
since revenues are typically only a bit larger than costs, revenue
efficiency estimates are essentially comparable to those for cost
be identified as 'best practices', which should be
efficiency. An ad hoc revenue frontier approach-essentially a adopted if possible. Managers can also adjust their
one-year DFA model-found a similar result (Elyasiani and policies and procedures to avoid 'worst practices'
Mehdian, 1990b). Berger et al. (1996b) specified an alternative that are relatively common among institutions that
revenue function, similar to the alternative profit function and
are far from the efficient frontier. In addition, owners
found no evidence of revenue economies of scope, suggesting that
customers do not value 'one-stop banking' or that banks do not
and managers of financial institutions may pay par-
have sufficient market power to extract the value that consumers ticular attention to the relationship between mea-
place on this convenience. sured frontier efficiency and organizational form,
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 203

which may suggest managerial arrangements which relatively simple comparisons or rankings of offices
are more conducive to high performance. according to a small (but sometimes, an overly large)
Many frontier efficiency Studies perform ex post set of partial performance ratios (Colwell and Davis,
analyses to identify the most important determinants 1992; Sherman and Ladino, 1995; Lovell and Pastor,
of firm efficiency. However, to date, the results of 1997). Although informative, such comparisons are
these analyses have not been very informative be- not as broadly-based as frontier analysis and typi-
cause of a lack of detailed data. Exceptions have cally lack a powerful and comprehensive optimizing
been studies that compared and contrasted the per- methodology. As well, the use of simple ratios typi-
formance of individual credit unions (Fried et al., cally does not account for differences in output mix
1993; Fried and Lovell, 1994). The incorporation of and input prices faced by the different branches.
price and service variety components into the output However, frontier analysis may not always indi-
of credit unions resulted in more accurate bench- cate the remedy for inefficient observations. Internal
marking of these firms and yielded higher average audits or intensive reviews of procedures are often
efficiency values being measured, because certain also needed to uncover the source and nature of the
'high cost' credit unions were found to incur higher operating and other changes that will likely improve
costs in order to improve the services they provided. efficiency at less efficient branches.
This can be important for mutual and cooperative As shown in Table 4, there have been a number
types of organizations in which the customers are the of frontier analyses focusing on branch performance
owners. The customer/owner may prefer an increase within a single banking firm. Only one of these
in costs which would lower conventionally-measured studies used parametric methods, applying DFA
efficiency if the higher costs were in the form of (Berger et al., 1997), while all the other studies have
higher interest paid or additional services provided. relied on nonparametric approaches, DEA or FDH.
Perhaps the best potential use of frontier effi- As proprietary data is often available for these stud-
ciency methods in improving managerial perfor- ies, many inputs and outputs can be expressed in
mance, largely due to the availability of detailed physical flow terms (e.g., hours worked by type of
proprietary data, comes from efficiency analysis of labor, numbers of transactions processed) and more
the branches of an individual financial institution. A accurate measures of stock inputs may be specified
financial institution may use frontier efficiency rank- (e.g., square footage of office space used). This has
ings, along with its own internal measures of perfor- been the case for the branch operations of a large
mance, to determine where problems lay and help Canadian bank (Schaffnit et al., 1997) which permits
solve them. In the hands of a researcher who has a a detailed and comprehensive efficiency analysis.
good institutional understanding of a given industry, Regional or seasonal influences, differences in mar-
frontier analysis can assist management to determine ket location or operating environments, office size,
objectively those procedures or branches that may be or even management style and organization may also
classified as best practice and worst practice within a be considered (Parkan, 1987; Oral and Yolalan, 1990;
firm. The best and worse practices that are discov- Tulkens and Malnero, 1994; Athanassopoulos, 1995,
ered can be used to rewrite the policies and proce- 1997; Sherman and Ladino, 1995; Zenios et al.,
dures book for the branches. In addition, manage- 1996; Drake and Howcroft, 1997). Less comprehen-
ment may use frontier efficiency rankings to deter- sive studies have to rely on more indirect indicators
mine which branches are in most need of reform, such as the stock of accounts serviced or the values
local management replacement, or closure. The mea- within various accounts, with little or no information
surement and use of frontier efficiency for these on important customer characteristics or other envi-
purposes may work particularly well in analyzing ronmental influences that can importantly affect the
branches which effectively have the same production outcome.
function and produce a similar output mix but may When detailed transactions and service data are
differ importantly in productivity and efficiency. available, it often is grouped into a smaller set of
While many firms have their own intemal bench- similar categories to be made operational, such as
marking procedures, they often are composed of aggregating 60 banking operations into only eight
204 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175-212

service areas (Tulkens, 1993) or constructing a purposes than are the estimated efficiency rankings
weighted measure of 4 service categories from 17 of of firms, and that analyses of the causes or correlates
the most common services offered at the branch of efficiency should be viewed with caution. The
level (Sherman and Gold, 1985). One reason for standard deviation of the efficiency estimates, at 13
partially aggregating the data is that it reduces the percentage points, is relatively large. This suggests,
number of constraints that have to be specified, and and some initial studies confirm, that the confidence
so reduces the number of observations that are deter- intervals surrounding individual firm or branch effi-
mined to be 100% efficient by virtue of having no ciency estimates may be substantial.
other observations with which to be compared (self- Applications of efficiency analysis. In terms of
identifiers). A more appropriate way to do this is applications, research on financial institution effi-
through a statistical test which can discriminate be- ciency has largely focused on using institution effi-
tween informative and extraneous constraints (Lovell ciency estimates: (1) to inform government policy
and Pastor, 1997) or applying a cone-ratio or assur- (e.g., by assessing the effects of deregulation, merg-
ance region DEA model (Schaffnit et al., 1997). ers, and market structure on industry efficiency); (2)
The one parametric study finds frontier efficiency to address research issues (e.g., by determining how
of about 0.90 to 0.95 for total branching costs (in- efficiency varies with different frontier approaches,
cluding interest expenses) or about 0.75 to 0.80 of output definitions, and time periods); and (3) to
branch operating costs, consistent with studies of improve managerial performance (e.g., by identify-
financial institutions generally. In contrast, the non- ing best-practice and worst-practice branches within
parametric frontier analyses of branches tend to find a single firm).
a relatively large proportion of branches to be 100% Results from these applications suggest the fol-
efficient. This may occur in some cases because the lowing sets of conclusions. First, the govemment
number of inputs and outputs is large relative to the policy-efficiency literature finds that deregulation of
number of observations available, making it difficult financial institutions can either improve or worsen
to find other branches or linear combination of efficiency, depending upon industry conditions prior
branches that dominate in every input and output. to deregulation. In a number of countries, deregula-
tion has led to rapid branch expansion, excessive
asset growth, a run-up in bank failures, and reduced
8. Conclusions and directions for future research efficiency. Although one goal of deregulation has
been to improve efficiency, other incentives may
We have outlined the results of 130 studies of intervene.
financial institution efficiency covering 21 countries A similar result applies to mergers and acquisi-
that apply five different frontier approaches. The tions: some consolidations improve cost efficiency,
efficiency estimates from nonparametric (DEA and whereas others worsen the performance of the com-
FDH) studies are similar to those from parametric bined institution relative to the separate institutions.
frontier models (SFA, DFA, and TFA), but the non- On average, there appears to be no significant cost
parametric methods generally yield slightly lower improvement. However, profit efficiency may im-
mean efficiency estimates and seem to have greater prove with mergers and acquisitions due to altering
dispersion than the results of the parametric models. output mix toward more profitable products (e.g.,
Overall, depository financial institutions (banks, S & from securities to loans), rather than improved cost
Ls, credit unions) in these studies experience an efficiency.
average efficiency of around 77% (median 82%). The application of frontier efficiency analysis to
The similarity in average efficiency values for firms the market-power vs. efficient-structure debate about
across different frontier models, however, does not the determinants of profitability also yields mixed
strongly carry over to rankings of individual firms by results. Cost efficiency is found to be more important
their efficiency values across models. This suggests than market concentration in explaining financial
that estimates of mean efficiency for an industry may institution profitability, but both influences together
be a more reliable guide for policy and research only weakly explain performance variation. Market
 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175 -212 205

power does seem to affect the prices of some types output mix and diversification, and the development
of local deposits and loans, but has little apparent of profit efficiency.
effect on profits. One reason may be that the man- The management performance-efficiency litera-
agers of financial institutions with market power ture on financial institutions is perhaps the least
appear to take some of the benefits of charging developed of the three types of applications. Some of
higher prices as a 'quiet life' in which they pursue this research has focused on alternative goals for
goals other than maximizing efficiency. managers, particularly when the firm is organized on
The research-efficiency literature on financial in- a mutual or cooperative basis, rather than as a
stitutions generally finds that efficiency rankings dif- value-maximizing enterprise. The burgeoning litera-
fer depending on which frontier approach is used (as ture on bank branch efficiency offers an opportunity
noted above) and by how financial institution output for researchers to provide managers with information
is measured - as a transaction-based flow, a stock of that may help to identify troubled branches and to
numbers of accounts, or a stock of value in these help rewrite operational policies and procedures
accounts. Once a frontier approach is adopted and an books based upon practices that are common among
output specification is selected, however, efficiency branches with the highest or lowest measured effi-
estimates are fairly stable from year to year, showing ciency. Unfortunately, few of these studies have
persistence. The limited evidence also suggests that noted in any detail the specific changes implemented
the confidence intervals around efficiency estimates to improve performance at inefficient branches.
may be quite large. Directions for future research. Finally, it is im-
Much of this literature is also concerned with the portant to point out shortcomings in existing research
determinants of efficiency. Firm efficiency appears that should be addressed, suggest ways in which the
to be greater for some forms of corporate organiza- existing research may be refocused to fill gaps in the
tion or control than others. However, most of these literature, and outline potential areas for future re-
effects are slight and may not always be economi- search. Existing research has shown us that financial
cally important, even if they are statistically signifi- institutions are less than fully efficient and have
cant. quantified the apparent extent of this deficiency.
There are a number of important methodological However, little has been offered in terms of the
developments under way that may help resolve some significance of the measured efficiency differences,
of the conflicts among methods, make efficiency in determining the specific causes of these differ-
estimation more accurate, and help find the determi- ences, and in explaining why they seem to persist in
nants of efficiency. For the nonparametric tech- market-based economies.
niques, these developments include non-radial mea- One problem of frontier analysis is that although
sures, the use of 'composite' frontiers which embody the central tendency of average efficiency values for
the best parts of different financial institutions, the financial institutions is generally similar across fron-
use of output distance functions, measurement of tier techniques, rankings of firms by their measured
confidence intervals, optimization of the number of efficiencies can differ. Since rankings differ depend-
constraints, finding a statistical basis for the non- ing on the frontier technique used, the common
stochastic approaches, and resampling to take ac- practice of regressing firm efficiency values (or
count of some of the random error in the data. For ranks) on other variables of interest may lead to
the parametric techniques, the new developments misleading results. If these ex post regressions are to
include the specification of more globally flexible be informative, they should be demonstrated to be
functional forms, the use of less restrictive assump- robust to efficiency estimates from more than just
tions on the distributions of inefficiencies, the al- one class of frontier technique.
lowance for heteroscedasticity in the distributions of There are also shortcomings in applying both
both inefficiencies and random errors, the measure- nonparametric and parametric frontier methods. The
ment of confidence intervals, random coefficient es- parametric approaches impose functional forms that
timation, allowance for correlations between regres- restrict the shape of the frontier, and the nonparamet-
sors and inefficiencies, measurement of the effects of ric approaches do not allow for random error that
206 A.N. Berger, D.B. Humphrey/ European Journal of Operational Research 98 (1997) 175-212

may affect measured performance. Attempts to rem- high efficiency values can be contrasted with at-
edy these situations by specifying more globally tributes associated with the group with relatively low
flexible functional forms in the parametric ap- values (with the middle group excluded entirely). In
proaches and trying to implement stochastic versions this context, it would be interesting to see if the
of the nonparametric approaches should continue. By imperfect correspondence found for firm-level effi-
generalizing both types of approaches, the data will ciency estimates among different frontier methods is
presumably have a better chance to yield results that markedly improved if groups of observations, rather
are more accurate and more consistent across ap- than individual observations, were used.
proaches. An area of research also deserving additional
Other shortcomings in the two types of ap- attention concerns efficiency comparisons among
proaches are clear as well. For example, the choice countries. With so few cross-country comparative
among the various parametric models is typically efficiency studies to draw upon, the results obtained
based more on ease of use a n d / o r the apparent so far should be taken with caution unless the robust-
reasonableness of the assumptions that underlie the ness of an intercountry comparison is demonstrated
different approaches than on any strong theoretical by finding the same result using different frontier
or empirical foundation. This gap in the literature is techniques on the same data set. As well, most
being filled for nonparametric models with an at- financial institution efficiency studies have been ap-
tempt to demonstrate that a stronger theoretical foun- plied to the US banking industry, which has distinct
dation exists for FDH than for DEA and that both local markets for many products and is quite uncon-
approaches have a valid statistical foundation. Even centrated by world standards. It is important for
so, nonparametric models are often specified in such research and policy purposes to see if the US results
a manner that many observations turn out to be carry over into other nations with banking markets
100% efficient, and this has been particularly so in that are more national in scope with much higher
the case of bank branches. Financial institutions or levels of concentration.
branches may be found to be fully efficient either Finally, there is a considerable lack of informa-
because there truly are no other units that dominate tion on what the main determinants of efficiency are
them (even when a small set of important core both across firms within the financial industry and
variables/constraints are specified). Alternatively, across branches within a single firm. Almost all of
these units may he found to be efficient because too the studies which estimate efficiency and then regress
many constraints have been specified, leading to it on sets of explanatory variables have been unable
excessive numbers of self-identifiers - units which to explain more than just a small portion of its total
neither dominated any other unit nor were dominated variation. While some differences have been found,
by any other unit or combination of units in every little published information exists regarding those
dimension. While this problem is well-known, there influences that are under direct management control,
have been few attempts to solve it. The statistical test such as the choice of funding sources, wholesale vs.
applied by Lovell and Pastor (1997) to identify retail orientation, etc. In sum, while there have been
extraneous constraints, however, may finally address improvements made in applying efficiency analysis
this issue. to financial institutions, there are many areas which
In addition, efficiency studies should try to pro- deserve further research.
vide confidence intervals for the estimates they gen-
erate, as some very recent studies have done. These
intervals, when they have been provided, appear to
Acknowledgements
be large relative to the range of efficiency estimates The authors thank Sigbjom Berg, Bill Cooper,
provided. As a result, comparisons of efficiency Gary Ferrier, Joaquin Maudos, and Jesus Pastor for
estimates across observations may be more meaning- insightful comments on earlier drafts and Knox
ful if groups of observations, rather than individual Lovell for bringing us up to date on stochastic DEA.
observations, were being compared. Attributes asso- We also thank Seth Bonime and Emilia Bonaccorsi
ciated with the group of observations with relatively for outstanding research assistance.
 A.N. Berger, D.B. Humphrey~European Journal of Operational Research 98 (1997) 175-212 207

References Bauer, P.W. (1990), "Recent developments in the econometric

estimation of frontiers", Journal of Econometrics 46, 39-56.
Bauer, P.W., and Hancock, D. (1993), "The efficiency of the
Adams, R., Berger, A.N., and Sickles, R. (1995), "Semiparamet- Federal Reserve in providing check processing services",
ric approaches to stochastic panel frontiers with applications in Journal of Banking and Finance 17, 287-311.
the banking industry", Working Paper, Rice University, USA. Bauer, P.W., Berger, A.N., and Humphrey, D.B. (1993), "Ef-
Akhavein, J.D., Berger, A.N., and Humphrey, D.B. (1997a), "The ficiency and productivity growth in US banking", in: H.O.
effects of megamergers on efficiency and prices: Evidence Fried, C.A.K. Lovell and S.S. Schmidt (eds.), The Measure-
from a bank profit function", Review of Industrial Organiza- ment of Productive Efficiency: Techniques and Applications,
tion 12, 95-130. Oxford University Press, Oxford, 386-413.
Akhavein, J.D., Swamy, P.A.V.B., and Taubman, S.B. (1997b), Bauer, P.W., Berger, A.N., Ferrier, G.D., and Humphrey, D.B.
" A general method of deriving the efficiencies of banks from (1995), "An empirical investigation of productivity change in
a profit function", Journal of Productivity Analysis 8. US banking, 1977-1988", Working Paper, University of
Al-Faraj, T.N., Alidi, A.S., and Bu-Bshait, K.A. (1993), "Evalua- Arkansas, Fayetteville, AR.
tion of bank branches by means of Data Envelopment Analy- Benston, G.J., Hunter, W.C., and Wall, L.D. (1995), "Motiva-
sis", International Journal of Operations and Production tions for bank mergers and acquisitions: Enhancing the deposit
Management 13, 45-52. insurance put option versus earnings diversification", Journal
Altunbas, Y., and Molyneux, P. (n.d.), "Stochastic estimators, of Money, Credit, and Banking 27, 777-788.
technical efficiency and bank size", Working Paper, Univer- Berg, S.A., (1992), "Mergers, efficiency and productivity growth
sity College of North Wales, UK. in banking: The Norwegian experience 1984-1990", Working
Aly, A.I., and Seiford, L.M. (1993), "The Mathematical Program- Paper, Norges Bank, Oslo, Norway.
ruing approach to efficiency analysis", in: H.O. Fried, C.A.K. Berg, S.A., and Kim, M. (1994), "Oligopolistic interdependence
Lovell and S.S. Schmidt (eds.), The Measurement of Produc- and the structure of production in banking: An empirical
tive Efficiency: Techniques and Applications, Oxford Univer- evaluation", Journal of Money, Credit, and Banking 26,
sity Press, Oxford, 120-159. 309-322.
Aly, H.Y., Grnhowski, R., Pasurka, C., and Rangan, N. (1990), Berg, S.A., and Kim, M. (1996), "Banks as multioutput
"Technical, scale, and allocative efficiencies in US banking: oligopolies: An empirical evaluation of the retail and corporate
An empirical investigation", Review of Economics and Statis- banking markets", Working Paper, Norges Bank, Oslo, Nor-
tics 72, 211-218. way.
Athanassopoulos, A.D. (1995), "Multivariate and frontier analysis Berg, S.A., Forsund, F., and Jansen, E. (1991), "Technical effi-
for assessing the market and cost efficiency of large scale ciency of Norwegian banks: A nonparametric approach to
bank branch networks", Working Paper, University of War- efficiency measurement", Journal of Productivity Analysis 2,
wick, Coventry, UK. 127-142.
Athanassoponlos, A.D. (1997), "Service quality and operating Berg, S.A., Forsund, F., and Jansen, E. (1992), "Malmquist
efficiency synergies for management control in the provision indices of productivity growth during the deregulation of
of financial services: Evidence from Greek bank branches", Norwegian banking, 1980-89", Scandinavian Journal of Eco-
European Journal of Operational Research 98, 301-314 (this nomics 94 (Supplemen0, $211-$228.
issue). Berg, S.A., Forsuod, F., Hjalmarsson, L., and Suominen, M.
Atkinson, S.E., and Wilson, P.W. (1995), "Comparing mean (1993), "Banking efficiency in the Nordic countries", Jour-
efficiency and productivity scores from small samples: A nal of Banking and Finance 17, 371-388.
bootstrap methodology", Journal of Productivity Analysis 6, Bergendahl, G., (1995), "DEA and benchmarks for Nordic
137-152. banks", Working Paper, Gothenburg University, Gothenburg,
Banker, R.D. (1996), "Hypothesis tests using Data Envelopment Sweden.
Analysis", Journal of Productivity Analysis 7, 139-159. Berger, A.N. (1993), "'Distribution-free' estimates of efficiency
Banker, R.D., Charnes, A., Cooper, W.W., Swarts, J., and Thomas, in the US banking industry and tests of the standard distribu-
D.A. (1989), "An introduction to Data Envelopment Analysis tional assumptions", Journal of Productivity Analysis 4, 261-
with some of its models and their uses", in: J.L. Chan and 292.
J.M. Patton (eds.), Research in Governmental and Nonprofit Berger, A.N. (1995), "The profit-structure relationship in bank-
Accounting, Vol. 5, JAI Press, Greenwich, CT, 125-163. ing-tests of market-power and efficient-structure hypotheses",
Bardhan, I.R., Cooper, W.W., Kozmetsky, G., and Kumbhakar, Journal of Money, Credit, and Banking 27, 404-431.
S.C. (1996), "A simulation study of joint uses of Data Envel- Berger, A.N., and DeYonng, R. (1997), "Problem loans and cost
opment Analysis and statistical regressions for production efficiency in commercial banks", Journal of Banking and
function estimation and efficiency evaluation", Working Pa- Finance 21 (forthcoming).
per, University of Texas, Austin, TX. Berger, A.N., and Hannan, T.H. (1995), "The efficiency cost of
Ban', R., Seiford, L., and Siems, T. (1994), "Forecasting bank market power in the banking industry: A test of the 'Quiet
failure: A non-parametric approach", Recherches Econo- Life' and related hypotheses", Working Paper, Board of Gov-
miques de Louvain 60, 411-429. ernors of the Federal Reserve System, Washington, IX:.
208 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

Berger, A.N., and Hannan, T.H. (1997), "Using measures of finn efficiency in the Nordic countries: A four-country Malmquist
efficiency to distinguish among alternative explanations of the index analysis", Working Paper, University of Aarhus, Den-
structure-performance relationship", Managerial Finance 23. mark.
Berger, A.N., and Humphrey, D.B. (1991), "The dominance of Cebenoyan, A.S., Cooperman, E.S., and Register, C.A. (1993a),
inefficiencies over scale and product mix economies in bank- "Firm inefficiency and the regulatory closure of S&Ls: An
ing", Journal of Monetary Economics 28, 117-148. empirical investigation", Review of Economics and Statistics
Berger, A.N., and Humphrey, D.B. (1992a), "Measurement and 75, 540-545.
efficiency issues in commercial banking", in: Z. Griliches Cebenoyan, A.S., Cooperman, E.S., Register, C.A., and Hudgins,
(ed.), Measurement Issues in the Service Sectors, National S. (1993b), "' The relative efficiency of stock vs. mutual S&Ls:
Bureau of Economic Research, University of Chicago Press, A stochastic cost frontier approach", Journal of Financial
Chicago, IL, 245-279. Services Research 7, 151-170.
Berger, A.N., and Humphrey, D.B. (1992b), "Megamergers in Chaffai, M., (1993), "Technical and time variant allocative ineffi-
banking and the use of cost efficiency as an antitrust defense", ciency of Tunisian commercial banks: A shadow cost frontier
Antitrust Bulletin 33, 541-600. approach using panel data", Working Paper, Facult~ des
Berger, A.N., and Mester, L.J. (1997), "Inside the black box: Sciences Economiques, Tunisia.
What explains differences in the efficiencies of fmancial insti- Chaffai, M.E., (1997), "Estimating input-specific technical ineffi-
tutions?", Journal of Banking and Finance 21 (forthcoming). ciency: The case of the Tunisian banking industry", European
Berger, A.N., Hancock, D., and Humphrey, D.B. (1993a), "Bank Journal of Operational Research 98, 315-332 (this issue).
efficiency derived from the profit function", Journal of Bank- Chaffai, M., and Dietsch, M. (1995), "Should banks be 'univer-
ing and Finance 17, 317-347. sal'? The relationship between economies of scope and effi-
Berger, A.N., Hunter, W.C., and Timme, S.G. (1993b), "The ciency in the French banking industry' ', Working Paper, Uni-
efficiency of financial institutions: A review and preview of versity Robert Schuman of Strasbourg, France.
research past, present, and future", Journal of Banking and Chang, C.E., Hasan, I., and Hunter, W.C. (1993), "Efficiency of
Finance 17, 221-249. multinational banks: An empirical investigation", Working
Berger, A.N., Kashyap, A.K., and Scalise, J.M. (1995), "The Paper, New Jersey Institute of Technology.
transformation of the US banking industry: What a long, Charnes, A., Cooper, W.W., and Rhodes, E. (1978), "Measuring
strange trip it's been", Brookings Papers on Economic Activ- the efficiency of decision making units", European Journal of
ity 2, 55-218. Operational Research 2, 429-444.
Berger, A.N., Cummins, D., and Weiss, M. (1996a), "The coexis- Charnes, A., Cooper, W.W., Huang, Z.M., and Sun, D.B. (1990),
tence of multiple distribution systems for financial services: "Polyhedral cone-ratio DEA models with an illustrative appli-
The case of Property-Liability Insurance", Working Paper, cation to large commercial banks", Journal of Econometrics
Wharton Financial Institutions Center, University of Pennsyl- 46, 73-91.
vania. Charnes, A., Cooper, W.W., Lewin, A., and Seiford, L. (eds.)
Berger, A.N., Humphrey, D.B., and Pulley, L.B. (1996b), "Do (1994), Data Envelopment Analysis: Theory, Methodology
consumers pay for one-stop banking? Evidence from an alter- and Applications, Kluwer Academic Publishers, Boston, MA.
native revenue function", Journal of Banking and Finance Clark, J. (1996), "Economic cost, scale efficiency and competi-
20, 1601-1621. tive viability in banking", Journal of Money, Credit, and
Berger, A.N., Leusner, J., and Mingo, J. (1997), "The efficiency Banking 28, 342-364.
of bank branches", Journal of Monetary Economics (forth- Colwell, R.J., and Davis, E.P. (1992), "Output and productivity in
coming). banking", Scandinavian Journal of Economics 94 (Supple-
Bhattacharyya, A., Lovell, C.A.K., and Sahay, P. (1997), "The ment), SI 11-S129.
impact of liberalization on the productive efficiency of Indian Cooper, W.W., Lelas, V., and Sueyoshi, T. (1997), "Goal pro-
commercial banks", European Journal of Operational Re- gramming models and duality relations for use in evaluating
search 98, 333-346 (this issue). security portfolio and regression relations", European Journal
Brockett, P.L., Rousseau, J., and Wang, Y. (1995), "An investiga- of Operational Research 98, 432-444 (this issue).
tion in the active army recruiting environment post 'Desert Cummins, D., and Weiss, M. (1993), "Measuring cost efficiency
Storm"', CCS Research Report No. 777, Graduate School of in the property-liability insurance industry", Journal of Bank-
Business, University of Texas, Austin, TX. ing and Finance 17, 463-481.
Brockett, P.L., Charnes, A., Cooper, W.W., Huang, Z.M., and Cummins, D., and Zi, H. (1995), "Measuring economic efficiency
Sun, D.B. (1997), "Data transformations in DEA cone-ratio of the US life insurance industry: Econometric and mathemati-
envelopment approaches for monitoring bank performances", cal programming techniques", Working Paper, Wharton Fi-
European Journal of Operational Research 98, 251-269 (this nancial Institutions Center, University of Pennsylvania.
issue). Cummins, D., Turchetti, G., and Weiss, M. (1995a), "Productiv-
Bukh, P.N.D., (1994), "Efficiency loss in the Danish banking ity and technical efficiency in the Italian insurance industry",
sector. A Data Envelopment approach", Working Paper, Uni- Working Paper, Wharton Financial Institutions Center, Uni-
versity of Aarhus, Denmark. versity of Pennsylvania.
Bukh, P.N.D., Berg, S.A., and Forsund, F.R. (1995), "Banking Cummins, D., Weiss, M., and Zi, H. (1995b), "Organizational
 A.N. Berger, D.B. Humphrey / European Journal of Operational Research 98 (1997) 175-212 209

form and efficiency: An analysis of stock and mutual prop- nonparametric approach", Journal of Banking and Finance
erty-liability insurers", Working Paper, Wharton Financial 16, 933-948.
Institutions Center, University of Pennsylvania. Elyasiani, E., and Mehdian, S.M. (1995), "The comparative
DeBorger, B., Ferrier, G., and Kerstens, K. (1995), "The choice efficiency performance of small and large US commercial
of a technical efficiency measure on the free disposal hull banks in the pre- and post-deregulation eras", Applied Eco-
reference technology: A comparison using US banking data", nomics 27, 1069-1079.
Working Paper, University of Arkansas. Elyasiani, E., Mehdian, S.M., and Rezvanian, R. (1994), "An
Demsetz, H. (1973), "Industry structure, market rivalry, and empirical test of association between production and financial
public policy", Journal of Law and Economics 16, 1-9. performance", Applied Financial Economics 4, 55-59.
Devaney, M., and Weber, W. (1995), "Rural bank efficiency and English, M., Grosskopf, S., Hayes, K., and Yaisawamg, S. (1993),
contestable markets", Working Paper, Southeast Missouri State "Output allocative and technical efficiency of banks", Jour-
University. nal o f Banking and Finance 17, 349-366.
DeYoung, R. (1994), "Fee-based services and cost efficiency in Favero, C., and Papi, L. (1995), "Technical efficiency and scale
commercial banks", in: Proceedings: Conference on Bank efficiency in the Italian banking sector: A non-parametric
Structure and Competition, Federal Reserve Bank of Chicago, approach", Applied Economics 27, 385-395.
May 1994. Fecher, F., and Pestieau, P. (1993), "Efficiency and competition
DeYoung, R. (1997a), " A diagnostic test for the distribution-free in OECD Financial Services", in: H.O. Fried, C.A.K. LoveU
efficiency estimator. An example using US commercial bank and S.S. Schmidt (eds.), The Measurement of Productive
data", European Journal of Operational Research 98, 244- Efficiency: Techniques and Applications, Oxford University
250 (this issue). Press, Oxford, 374-385.
DeYoung, R. (1997b), "Bank mergers, X-efficiency, and the Fecher, F., Kessler, D., Perelman, S., and Pestieau, P. (1993),
market for corporate control", Managerial Finance 23. "Productive performance of the French insurance industry",
DeYoung, R. (1997c), "Management quality and X-efficiency in Journal of Productivity Analysis 4, 77-93.
national banks", Journal of Financial Services Research 11. Ferrier, G., and Hirschberg, J. (1994), "Bootstrapping confidence
DeYoung, R., and Nolle, D. (1996), "Foreign-owned banks in the intervals for linear programming efficiency scores: With an
US: Earning market share or buying it?", Journal of Money, illustration using Italian banking data", Working Paper, Uni-
Credit, and Banking 28, 622-636. versity of Arkansas.
Dietsch, M. (1994), "Risk-taking and cost efficiency in French Ferrier, G., and Lovell, C.A.K. (1990), "Measuring cost effi-
banking industry", Working Paper, Robert Schuman Univer- ciency in banking: Econometric and linear programming evi-
sity of Strasbourg, France. dence", Journal of Econometrics 46, 229-245.
Drake, L., and Howcroft, B. (1997), " A study of the relative Ferrier, G., Grosskopf, S., Hayes, K., and Yaisawamg, S. (1993),
efficiency of UK bank branches", Journal of Banking and "Economies of diversification in the banking industry: A
Finance, forthcoming. frontier approach' ', Journal of Monetary Economics 31,229-
Drake, L., and Weyman-Jones, T. (1992), "Technical and scale 249.
efficiency in UK building societies", Applied Financial Eco- Ferrier, G., Kerstens, K., and Vanden Eeckaut, P. (1994), "Radial
nomics 2, 1-9. and nonradial technical efficiency measures on a DEA refer-
Efron, B., and Tibshirani, R.J. (1993), An Introduction to the ence technology: A comparison using banking data",
Bootstrap, Chapman & Hall, New York. Recherches Economiques de Louvain 60, 449-479.
Eisenbeis, R.A., Ferrier, G.D., and Kwan, S.H. (1996), "An Field, K. (1990), "Production efficiency of British building soci-
empirical analysis of the informativeness of programming and eties", Applied Econom&s 22, 415-426.
SFA efficiency scores: Efficiency and bank performance", Fixler, D., and Zieschang, K. (1993), "An index number approach
Working Paper, University of North Carolina, Chapel Hill, to measuring bank efficiency: An application to mergers",
NC. Journal of Banking and Finance 17, 437-450.
Ellinger, P., Zhu, S., Shumway, R., and Neff, D. (1997), "Speci- Fried, H.O., and Lovell, C.A.K. (1994), "Enhancing the perfor-
fication of inefficiency in banking: A comparison of cost and mance of credit unions: The evolution of a methodology",
profit function approaches", Journal of Financial Services Recherches Economiques de Louvain 60, 431-4,47.
Research, forthcoming. Fried, H.O., Lovell, C.A.K., and Vanden Eeckaut, P. (1993),
Elyasiani, E., and Mehdian, S.M. (1990a), " A nonparametric "Evaluating the performance of US credit unions", Journal
approach to measurement of efficiency and technological of Banking and Finance 17, 251-265.
change: The case of large US commercial banks", Journal of Fukuyama, H. (1993), "Technical and scale efficiency of Japanese
Financial Services Research 4, 157-168. commercial banks: A non-parametric approach", Applied
Elyasiani, E., and Mehdian, S.M. (1990b), "Efficiency in the Economics 25, 1101-1112.
commercial banking industry. A production frontier approach' ', Fukuyama, H. (1995), "Measuring efficiency and productivity
Applied Economics 22, 539-551. growth in Japanese banking: A nonparametric frontier ap-
Elyasiani, E., and Mehdian, S.M. (1992), "Productive efficiency proach", Applied Financial Economics 5, 95-117.
performance of minority and nonminority-owned banks: A Gardner, L., and Grace, M. (1993), "X-efficiency in the US life
210 A.N. Berger, D.B. Humphrey~European Journal of Operational Research 98 (1997) 175-212

insurance industry", Journal of Banking and Finance 17, Humphrey, D.B., and Pulley, L.B. (1997), "Banks' responses to
497-510. deregulation: Profits, technology, and efficiency", Journal of
Giokas, D. (1990, "Bank branch operating efficiency: A compar- Money, Credit, and Banking 73-93.
ative application of DEA and the loglinear model", OMEGA Hunter, W.C., and Timme, S. (1995), "Core deposits and physical
International Journal of Management Science 19, 549-557. capital: A reexamination of bank scale economies and effi-
Grabowski, R., Rangan, N., and Rezvanian, R. (1993), "Organi- ciency with quasi-fixed inputs", Journal of Money, Credit,
zational forms in banking: An empirical investigation of cost and Banking 27, 165-185.
efficiency", Journal of Banldng and Finance 17, 531-538. Kaparakis, E., Miller, S., and Noulas, A. (1994), "Short-run cost
Greene, W.H. (1990), " A gamma-distributed stochastic frontier inefficiency of commercial banks" A flexible stochastic fron-
model", Journal of Econometrics 46, 141-163. tier approach", Journal of Money, Credit, and Banking 26,
Greene, W.H. (1993), "The econometric approach to efficiency 875-893.
analysis", in: H.O. Fried, C.A.K. Lovell and S.S. Schmidt Kneip, A., and Simar, L. (1996), " A general framework for
(eds.), The Measurement of Productive Efficiency: Techniques frontier estimation with panel data", Journal of Productivity
and Applications, Oxford University Press, Oxford, 68-119. Analysis 7, 187-212.
Grifell-Tatjt, E., and Lovell, C.A.K. (1994), " A generalized Kraraer, B. (1997), "N.E.W.S.: A model for the evaluation of
Mahnquist productivity index", Working Paper, University of non-life insurance companies", European Journal of Opera-
Georgia, Athens, GA. tional Research 98, 420-431 (this issue).
Grifell-Tatjt, E., and Loveil, C.A.K. (1996), "Profits and produc- Kumbhakar, S.C. (1988), "Estimation of input-specific technical
tivity: A theoretical analysis and an empirical application to and allocative inefficiency in stochastic frontier models",
Spanish banking", Working Paper, University of Barcelona. Oxford Economic Papers 40, 535-549.
Grifell-Tatjt, E., and Loveil, C.A.K. (1997a), "The sources of Kuussaari, H. (1993), "Productive efficiency in Finnish local
productivity change in Spanish banking", European Journal banking during 1985-1990", Working Paper, Bank of Fin-
of Operational Research 98, 365-381 (this issue). land.
GrifelI-Tatjt, E., and Lovell, C.A.K. (1997b), "Dereguiation and Kuussaari, H., and Vesala, J. (1995), "The efficiency of Finnish
productivity decline: The case of Spanish savings banks", banks in producing payment and account transactions", Work-
European Economic Reoiew, forthcoming. hag Paper, Bank of Finland.
Grosskopf, S. (1993), "Efficiency and productivity", in: H.O. Kwan, S.H., and Eisenbeis, R.A. (1994), "An analysis of ineffi-
Fried, C.A.K. Lovell and S.S. Schmidt (eds.), The Measure- ciencies in banking: A stochastic cost frontier approach",
ment of Productive Efficiency: Techniques and Applications, Working Paper, Federal Reserve Bank of San Francisco.
Oxford University Press, Oxford, 160-194. Land, K., LoveU, C.A.K., and Thole, S. (1993), "Chance-con-
Grosskopf, S. (1996), "Statistical inference and nonparametric slxained Data Envelopment Analysis", Managerial and Deci-
efficiency: A selective survey", Journal of Productivity Anal- sion Economics 14, 541-554.
ysis 7, 161-176. Lang, G., and Welzel, P. (1995), "Technology and cost efficiency
Hall, P., Hardle, W., and Simar, L. (1993), "On the inconsistency in universal banking", Working Paper, University of Augs-
of bootstrap estimators", Computational Statistics and Data burg, Germany.
Analysis 16, 11-18. Lang, G., and Welzel, P. (1996), "Efficiency and technical
Hartman, T., and Storbeck, J.E. (1995), "Measuring managerial progress in banking: Empirical results for a panel of German
and program efficiencies in a Swedish savings and loan", banks", Journal of Banking and Finance 20, 1003-1023.
Working Paper, Kecle University, UK. Lovell, C.A.K. (1993), "Production frontiers and productive effi-
Hasan, I., and Hunter, W.C. (1996), "Efficiency of Japanese ciency", in: H.O. Fried, C.A.K. Lovell and S.S. Schmidt
multinational banks in the US", Research in Finance, forth- (eds.), The Measurement of Productive Efficiency: Techniques
coming. and Applications, Oxford University Press, Oxford, 3-67.
Hermalin, B.E., and Wallace, N.E. (1994), "The determinants of Lovell, C.A.K., and Pastor, J.T. (1997), "Target setting: An
efficiency and solvency in savings and loans", Rand Journal application to a hank branch network", European Journal of
of Economics 25, 361-381. Operational Research 98, 291-300 (this issue).
Hughes, J.P., and Mester, L.J. (1993), " A quality and risk-ad- Lozano, A. (1995), "Efficiency and technical change for Spanish
justed cost function for banks: Evidence on the 'too-big-to-fail' banks", Working Paper, University of Malaga, Spain.
doctrine", Journal of Productivity Analysis 4, 293-315. Lozano, A. (1997), "Profit efficiency for Spanish savings banks' ',
Hughes, J.P., Lang, W., Mester, L.J., and Moon, C.-G. (1996), European Journal of Operational Research 98, 382-395 (this
"Efficient banking under interstate branching", Journal of issue).
Money, Credit, and Banking 28, 1045-1071. Mahajan, A., Rangan, N., and Zardkoohi, A. (1996), "Cost
Humphrey, D.B. (1987), "Cost dispersion and the measurement structures in multinational and domestic banking", Journal of
of economies in banking", Economic Review, May/June, Banking and Finance 20, 238-306.
24-38. Mandos, J. (1996a), "A comparison of different stochastic fron-
Humphrey, D.B. (1993), "Cost and technical change: Effects tier techniques with panel data: An application for efficiency
from bank deregulation", Journal of Productivity Analysis 4, of Spanish banks", Working Paper, University of Valencia,
5-34. Spain.
 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212 211

Maudos, J. (1996b), "Market structure and performance in Span- "The technical efficiency of US banks", Economics Letters
ish banking using a direct measure of efficiency", Working 28, 169-175.
Paper, University of Valencia, Spain. Ray, S., and Mukherjee, K. (1994), "Identifying banks that are
McAllister, P.H., and McManus, D.A. (1993), "Resolving the too large: A study of size efficiency in US banking", Working
scale efficiency puzzle in banking", Journal of Banking and Paper, University of Connecticut.
Finance 17, 389-405. Resti, A. (1995), "Linear programming and econometric methods
Mester, L.J. (1993), "Efficiency in the savings and loan industry", for bank efficiency evaluation: An empirical comparison based
Journal of Banking and Finance 17, 267-286. on a panel of Italian banks", Working Paper, University of
Mester, L.J. (1996), " A study of bank efficiency taking into Bergamo, Italy.
account risk-preferences", Journal of Banking and Finance Rhoades, S.A. (1993), "The efficiency effects of horizontal bank
20, 1025-1045. mergers", Journal of Banking and Finance 17, 411-422.
Mester, L.J. (1997), "Measuring efficiency at US banks: Account- Ruthenberg, D., and Elias, R. (1996), "Cost economies and
ing for heterogeneity is important", European Journal of interest rate margins in a unified European banking market",
Operational Research 98, 230-243 (this issue). Journal of Economics and Business 48, 231-249.
Miller, S.M., and Noulas, A.G. (1996), "The technical efficiency Schaffuit, C., Rosen, D., and Paradi, J.C. (1997), "Best practice
of large bank production", Journal of Banking and Finance analysis of bank branches: An application of DEA in a large
20, 495-509. Canadian bank' ', European Journal of Operational Research
Mitchell, K., and Onvural, N.M. (1996), "Economies of scale and 98, 270-290 (this issue).
scope at large commercial banks: Evidence from the Fourier Schmidt, P., and Sickles, R.C. (1984), "Production frontiers and
flexible functional form", Journal of Money, Credit, and panel data", Journal of Business and Economic Statistics 2,
Banking 28, 178-199. 367-374.
Mooney, C.Z., and Duval, R.D. (1993), Bootstrapping: A Non- Seiford, L.M., and Thrall, R.M. (1990), "Recent developments in
parametric Approach to Statistical Inference, Sage, Newbury DEA: The mathematical programming approach to frontier
Park, CA. analysis", Journal of Econometrics 46, 7-38.
Murthi, B.P.S., Choi, Y.K., and Desai, P. (1997), "Efficiency of Sheldon, G., and Haegler, U. (1993), "Economies of scale and
mutual funds and portfolio performance measurement: A non- scope and inefficiencies in Swiss banking", in: N. Blatmer, H.
parametric approach", European Journal of Operational Re- Genberg and A. Swoboda (eds.), Banking in Switzerland,
search 98, 409-419 (this issue). Springer-Verlag, New York, 103-140.
Newman, J., and Shrieves, R. (1993), "Multibank holding com- Sherman, D., and Gold, F. (1985), "Branch operating efficiency:
pany effect on cost efficiency in banking", Journal of Bank- Evaluation with Data Envelopment Analysis", Journal of
ing and Finance 17, 709-732. Banking and Finance 9, 297-315.
Olesen, O.B., and Petersen, N.C. (1995), "Chance constrained Sherman, D., and Ladino, G. (1995), "Managing bank productiv-
efficiency evaluation", Management Science 41,442-457. ity using Data Envelopment Analysis (DEA)", Interfaces 25,
Oral, M., and Yolalan, R. (1990), "An empirical study on measur- 60-73.
ing operating efficiency and profitability of bank branches", Simar, L. (1992), "Estimating efficiencies from frontier models
European Journal of Operational Research 46, 282-294. with panel data: A comparison of parametric, nonparametric
Parkan, C. (1987), "Measuring the efficiency of service opera- and semiparametric methods with bootstrapping", Journal of
tions: An application to bank branches", Engineering Costs Productivity Analysis 3, 171-203.
and Production Economics 12, 237-242. Simar, L. (1996), "Aspects of statistical analysis in DEA-type
Pastor, J. (1995), "How to account for environmental effects in frontier models", Journal of Productivity Analysis 7, 177-
DEA: An application to bank branches", Working Paper, 185.
University of Alicante, Spain. Simar, L., and Wilson, P.W. (1995), "Sensitivity analysis of
Pastor, J., Perez, F., and Quesada, J. (1997), "Efficiency analysis efficiency scores: How to bootstrap in nonparametric frontier
in banking finns: An international comparison", European models", Working Paper, Institute of Statistics, Universit6
Journal of Operational Research 98, 396-408 (this issue). Catholique de Louvain, Belgium.
Perez, F., and Quesada, J. (1994), "Efficiency and banking Taylor, W.M., Thompson, R.G., Thrall, R.M., and Dharmapala,
strategies in Spain", in: D. Fair and R. Raymond (eds.), The P.S. (1997), " D E A / A R efficiency and profitability of Mexi-
Competitiveness of Financial Institutions and Centres in Eu- can banks: A total income model", European Journal of
rope, Financial and Monetary Policy Studies, Vol. 28, Kluwer Operational Research 98, 347-364 (this issue),
Academic Publishers, Dordrecht, 135-149. Thiry, B., and Tulkens, H. (1992), "Allowing for inefficiency in
Peristiani, S. (1997), "Do mergers improve the X-efficiency and parametric estimates of production functions for urban transit
scale efficiency of US banks? Evidence from the 1980s", firms", Journal of Productivity Analysis 3, 45-66.
Journal of Money, Credit, and Banking, 29. Thompson, R.G., Dharmapala, P.S., Diaz, J., Gonzalez-Lima,
Pi, L., and Timme, S. (1993), "Corporate control and bank M.D., and Thrall, R.M. (1996a), "DEA multiplier analytic
efficiency", Journal of Banking and Finance 17, 515-530. center sensitivity with an illustrative application to indepen-
Rangan, N., Grabowski, R., Aly, H., and Pasurka, C. (1988), dent oil companies", Annals of Operations Research 66.
212 A.N. Berger, D.B. Humphrey/European Journal of Operational Research 98 (1997) 175-212

Thompson, R.G., Dharmapala, P.S., Humphrey, D.B., Taylor, ment Analysis", Journal of the Operational Research Society
W.M., and Thrall, R.M. (1996b), "Computing DEA/AR effi- 41,591-597.
ciency and profit ratio measures with an illustrative bank Yuengert, A. (1993), "The measurement of efficiency in life
application", Annals of Operations Research 68. insurance: Estimates of a mixed Normal-Gamma error
Thompson, R.G., Brinkmarm, E.J., Dhannapala, P.S., Gonzalez- model", Journal of Banking and Finance 17, 483-496.
Lima, M.D., and Thrall, R.M. (1997), "DEA/AR profit-ratios Wheelock, D.C., and Wilson, P.W. (1994), "Productivity changes
and sensitivity of 100 large US commercial banks", European in US banking: 1984-93", Working Paper, Federal Reserve
Journal of Operational Research 98, 213-229 (this issue). Bank of St. Louis.
Tulkens, H. (1993), "On FDH efficiency analysis: Some method- Zaim, O. (1995), "The effect of f'mancial liberalization on the
ological issues and applications to retail banking, courts, and efficiency of Turkish commercial banks", Applied Financial
urban transit", Journal of Productivity Analysis 4, 183-210. Economics 5,257-264.
Tulkens, H., and Malnero, A. (1994), "Nonparametric approaches Zenios, C., Zenios, S., Agathocleous, K., and Soteriou, A. (1996),
to the assessment of the relative efficiency of bank branches", "Benchmarks of the efficiency of bank branches", Working
Working Paper, Center for Operations Research and Econo- Paper, University of Cyprus, Nicosia, Cyprus.
metrics, Universit6 Catholique de Louvain, Belgium. Zhu, S., Ellinger, P., and Shumway, R. (1997), "The choice of
Vassiloglou, M., and Giokas, D. (1990), " A study of the relative functional form and estimation of banking efficiency", Ap-
efficiency of bank branches: An application of Data Envelop- plied Economic Issues, forthcoming.