Journal of Monetary Economics 47 (2001) 653–674

New evidence on returns to scale and product

mix among U.S. commercial banks$
David C. Wheelocka,*, Paul W. Wilsonb
a
Research Department, Federal Reserve Bank of St. Louis, P.O. Box 442, St. Louis,
MO 63166, USA
b
Department of Economics, University of Texas, Austin, TX 78712, USA

Received 28 February 1997; received in revised form 4 January 1999; accepted 6 July 2000

Abstract

This paper presents new estimates of scale and product mix economies for U.S.
commercial banks. We compare estimates derived from fitting a translog function to
bank costs with estimates derived from nonparametric methods. We refine measures of
scale and product mix economies introduced by Berger et al. (J. Monet. Econ. 20 (1987)
501) to accommodate nonparametric estimation, and estimate confidence intervals to
assess the statistical significance of returns to scale. Broadly, we find evidence that
potential economies have increased since 1985, with scale economies not exhausted until

$
We are grateful to Robert DeYoung, David Humphrey, James McIntosh, and an anonymous
referee for comments on an earlier draft. Versions of this paper have been presented at the 1997
Texas Econometrics Camp, the 1998 North American summer meetings of the Econometric Society
in Montreal, and the 1998 Georgia Productivity Workshop. We are also grateful to the Texas
Advanced Computing Center (TACC) for a grant of computational time on their SGI=Cray T3E
parallel supercomputer, and especially to Robert Harkness of the TACC staff for help in porting
code to the T3E environment. Any remaining errors, of course, are solely our responsibility. The
views expressed herein do not necessarily reflect official positions of the Federal Reserve Bank of
St. Louis or the Federal Reserve System.
*Corresponding author. Tel.: +1-314-444-8570; fax: +1-314-444-8731.
E-mail address: wheelock@stls.frb.org (D.C. Wheelock).

0304-3932/01/$ - see front matter r 2001 Elsevier Science B.V. All rights reserved.
PII: S 0 3 0 4 - 3 9 3 2 ( 0 1 ) 0 0 0 5 9 - 9
654 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

banks have $300–$500 million of assets. We generally fail to reject constant returns for
larger banks. r 2001 Elsevier Science B.V. All rights reserved.

JEL classification: C14; E59; G21; L11

Keywords: Returns to scale; Economies of scope; Nonparametric; Local linear smoother; Kernel
regression

1. Introduction

The ongoing merger wave in the U.S. banking industry has helped to reduce
the number of American commercial banks by over one-third since 1984 (from
14,419 banks in 1984 to 9103 banks at the end of 1997). A large proportion of
those eliminated have been small banks, and the disappearance of many small
banks through acquisition and failure suggests that they may not be viable in
today’s environment.1
Bankers often justify mergers as attempts to achieve economies of scale. But
conventional wisdom, on the basis of numerous studies, holds that banks
exhaust potential scale economies at very modest levels of output, on the order
of $100–$200 million of assets (McAllister and McManus, 1993). Moreover,
the evidence suggests that ‘‘megamergers’’ among large banks have not
produced significant cost savings (e.g., Berger and Humphrey, 1992; Boyd and
Graham, 1991).2
Conventional wisdom might, however, be mistaken. Recent studies find
potential scale economies for banks at much higher levels of output than
previous studies had found. McAllister and McManus (1993), for example, find
that banks face increasing returns to scale to about $500 million of assets, and
Berger and Mester (1997), who compare banks within size ranges, conclude
that in all ranges the mean bank operates at less than efficient scale. McAllister
and McManus (1993) attribute earlier findings of minimal scale economies to
bias introduced by fitting a single parametric cost function across all banks.
Parameter instability, they argue, can bias estimates of scale economies derived
from global fitting of parametric cost functions. Using less-than-parametric
methods (kernel regression and orthogonal series estimation), McAllister and
McManus (1993) estimate that banks exhaust scale economies at a much larger

1
Between 1984 and 1997, the number of banks with less than $300 million of assets fell from
13,676 to 8082, while the number of banks with more than $300 million of assets increased from 739
to 1061.
2
Akhavein et al. (1997) find that mergers of large banks tend to enhance profit efficiency,
however, because of revenue gains when merged banks adjust their mix of outputs toward higher-
value assets, such as loans.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 655

output level than suggested by global estimation of a translog cost function.

When fitting a translog function locally, however, McManus (1994) estimates
that scale economies are exhausted at smaller output than suggested by global
estimation.
Berger and Mester (1997) attribute differences in estimates of scale
economies not to differences in estimation methods, but to a fundamental
shift in bank costs over time, associated, perhaps, with regulatory or
technological changes. Using data for 1990–1995, Berger and Mester (1997)
find that banks operate at smaller than efficient scale, regardless of whether a
global translog function or a semi-parametric orthogonal series estimator is
used to estimate the bank cost function. Using the same orthogonal series
estimator (the Fourier flexible form), however, Mitchell and Onvural (1996)
find that, if anything, large banks faced greater opportunities to exploit
economies of scale in 1986 than in 1990.
The nature of scale economies in banking, and whether there has been an
increase in potential economies over time, thus remain unsettled questions. The
rapid pace of consolidation within the banking industry poses a challenge for
regulators who must consider questions of competition and market service in
the approval process. It also raises questions about the impact of technological
and regulatory change on potential scale economies and market structure in
general (e.g., see Berger et al., 1995).
In this paper, we present estimates of scale and product mix economies based
on the parametric translog model and three alternative, less-than parametric
techniques: semi-nonparametric orthogonal series estimation (the Fourier
flexible functional form), fully nonparametric kernel regression, and local
polynomial smoothing (LPS). Our sample includes all U.S. commercial banks
(with usable data), and we present estimates for three yearsF1985, 1989, and
1994Fto allow for possible shifts in bank costs or technology over time.3
Despite their popularity in banking studies (e.g., Berger and DeYoung, 1997;
Berger et al., 1997), orthogonal series estimators such as the Fourier flexible
functional form present several unresolved statistical problems, including
whether to augment the underlying translog function with trigonometric terms
or orthogonal polynomials, and the number of such terms to include for
estimation. Kernel regression and LPS involve analogous choices, but here a
substantial body of theory exists to guide researchers. Moreover, kernel and

3
Over this period, both the mean and median bank size increased, and a Kolmogorov–Smirnov
two-sample test of the null hypothesis of no difference in the distribution of total assets rejects the
null at 99 percent significance levels for each pair of years 1985=1989; 1989=1994, and 1985=1994.
Formal analysis also suggests that bank technology has shifted over time (Wheelock and Wilson,
1999). Thus, in contrast to McAllister and McManus (1993) and Berger and Mester (1997), we elect
to investigate explicitly whether scale economies shifted over time. Moreover, by including all
banks in our sample, we avoid possible selection biases associated with using only those banks that
were present throughout the entire sample period, as done in the pooled-data studies.
656 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

LPS estimators are both local in the sense that when estimating the cost
function at a particular point, nearby observations receive more weight than
distant observations. This is particularly advantageous here because the
distribution of bank sizes is heavily skewed toward smaller banks. Observa-
tions on large banks thus act as leverage points, which can have dispropor-
tionate impact on parameter estimates in the translog model, and cause similar
distortions in orthogonal series estimators.
We refine the scale and product mix measures suggested by Berger et al.
(1987) to estimate economies over the range of data and to accommodate
nonparametric approaches. In so doing, we are able to examine ‘‘expansion
path’’ economies, which account for differences in product mix across banks of
different sizes, as well as ray-scale economies. Since previous studies have
found that banks suffer from considerable managerial inefficiency (e.g., Berger
and Humphrey, 1991), which might confound estimates of scale economies, we
estimate scale and product mix economies for a ‘‘thick frontier’’ consisting of
the most x-efficient banks, as well as for all banks. Finally, we provide
estimated confidence intervals to assess the statistical significance of our
estimates of scale and product mix economies.4
Section 2 presents our modification of the Berger et al. (1987) measures of
scale and product mix economies. In Section 3 we describe our model of bank
cost. Section 4 describes our estimation method, and Section 5 presents our
empirical findings.

2. Measuring returns to scale and product mix

Consider a multiple-output cost function CðyÞ, where y ¼ ½y1 y yq 0

denotes a vector of outputs. Berger et al. (1987) note that a firm producing
outputs y is competitively viable if the cost of producing y by that firm is no
greater than the scale-adjusted cost of jointly producing output bundle y by
set of firms. That is, for any and all output vectors yc X0 and y > 0
any other P
such that c yc ¼ yy,

X
CðyÞpy@1 Cðyc Þ; ð2:1Þ
c

4
Berger and Mester (1997) and McAllister and McManus (1993), by contrast, estimate ray-scale
economies only, and do not test for statistical significance. Mitchell and Onvural (1996) examine
expansion-path scale economies and test formally for statistical significance, but focus exclusively
on large banks. Moreover, their significance test requires that the error term of the cost function be
normally distributed; if this assumption were correct, then it should be used in the original
estimation to improve statistical efficiency. We make no parametric assumptions in this regard.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 657

Fig. 1. Graphical representaion of comptetive viability measures.

where c indexes specific output vectors which may be summed to form y.

Unfortunately, there are no simple necessary and sufficient conditions for
competitive viability; a complete examination of viability for a given firm
would require comparing the costs of hypothetical firms producing an infinite
variety of output vectors.5 Of course this is not feasible, and so typically
researchers compare hypothetical, representative firms producing output
vectors at the sample means of outputs within various firm size classes (e.g.,
Berger et al., 1987; Clark, 1996; Mitchell and Onvural, 1996).
For illustration, consider two banks A and B producing two outputs in
quantities ya ¼ ½ya1 ya2  and yb ¼ ½yb1 yb2 , respectively, as shown in Fig. 1 (which
we have adapted from Berger et al., 1987). As Berger et al. (1987) note, ray-
scale economies (RSCE) is the multiproduct analog of marginal cost divided by
average cost on a ray from the origin. Holding product mix constant, RSCE
can be measured along a ray yy by defining
CðyyÞ
SðyjyÞ  : ð2:2Þ
yCðyÞ
It is straightforward to show that returns to scale are increasing (constant,
decreasing) along the ray yy if SðyjyÞ is decreasing (constant, increasing) in y.
# in (2.1), and allowing y to vary, one can
By replacing Cð Þ with estimates Cð Þ
5
Note that we are ignoring demand-side considerations here and throughout.
658 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

examine returns to scale over the entire ray from the origin. Using bootstrap
methods, we estimate simultaneous confidence intervals for SðyjyÞ # to
determine whether any indications of increasing or decreasing returns to scale
are statistically significant.
Because the output mix of banks tends to vary with size, however, RSCE can
yield a misleading impression about scale economies. Berger et al. (1987) thus
propose two alternative measures of returns that commingle scale and product
mix economies. They define expansion path scale economies (EPSCE) as the
elasticity of incremental cost with respect to incremental output along a
nonradial ray such as the one between points A and B in Fig. 1. For a given
pair of output vectors ðya ; yb Þ, we define
Cðya þ yðyb @ya ÞÞ@Cðya Þ
Eðyjya ; yb Þ  : ð2:3Þ
y½Cðyb Þ@Cðya Þ
Straightforward algebra reveals that if Eðyjya ; yb Þ is decreasing (constant,
increasing) in y, then returns to scale along the line from ya to yb are increasing
(constant, decreasing). EPSCE can be evaluated in the neighborhood of bank
A by estimating Eðyjya ; yb Þ for small values of y (by definition, Eðyjya ; yb Þ ¼ 1
when y ¼ 1). For yo1, the first term in the numerator of (2.3) gives the cost of
a hypothetical firm producing at an intermediate point along the segment AB.
If Eðyjya ; yb Þ > 1, then the cost of this hypothetical firm is greater than the
weighted costs of firms A and B, given by the numerator in (2.3). This implies
that the cost surface is concave from below along the path AB, which in turn
implies that total cost is increasing at a decreasing rate as we move from point
A to point B in Fig. 1. Hence returns to scale are increasing along the
expansion path AB. Similar reasoning demonstrates that if Eðyjya ; yb Þo1 for
values yo1, decreasing returns to scale prevail along the expansion path AB.6
The same bootstrap methods used to examine RSCE can be applied here to
determine whether Eðyjya ; yb Þ is significantly greater or less than unity for small
values of y.
As an alternative to EPSCE, Berger et al. (1987) also propose expansion
path subadditivity (EPSUB), which they define as
Cðya Þ þ Cðyb @ya Þ@Cðyb Þ
Aðya ; yb Þ  : ð2:4Þ
Cðyb Þ
The numerator term Cðyb @ya Þ in (2.4) gives the cost of firm D in Fig. 1.
Collectively, firms A and D produce the same output as firm B. If Aðya ; yb Þo0,
then firm B is not competitively viable; i.e., two firms producing output vectors
ya and yb @ya collectively produce the same output as firm B, but at lower
6
Conceivably, Eðyjya ; yb Þ could oscillate around unity for values yo1, which would suggest both
increasing and decreasing returns along different parts of the expansion path. We find no evidence
of this, however, for our sample.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 659

combined total cost. Alternatively, if Aðya ; yb Þ > 0, then (smaller) firm A

should adjust its output vector toward that of (larger) firm B. Although both
EPSCE and EPSUB are composite measures of scale and scope economies,
EPSUB is closer in spirit to a measure of scope economies than EPSCE
because EPSUB compares the cost of production at a given firm B with the cost
of producing an identical level of output in two separate firms with different
output mixes. EPSCE, on the other hand, measures the incremental cost of
incremental output along the expansion path between two different-sized firms.

3. A model of bank cost

To estimate the measures of scale and product mix economies described in

the previous section, a model of bank cost must be specified. Typically, banks
are viewed as transforming various financial resources, as well as labor and
physical plant, into loans, other investments and, sometimes, deposits. We
adopt the ‘‘intermediation’’ approach, which measures outputs in terms of the
dollar amounts of loans and deposits.
Our specific input=output mapping borrows from Kaparakis et al. (1994),
which is somewhat representative. We use annual data for 1985, 1989, and 1994
to examine whether returns to scale and other aspects of bank costs have
changed over recent history. For each bank i ¼ 1; y; n in a given cross-
sectional sample, we define four outputs (loans to individuals for household,
family, and other personal expenses, yi1 ; real estate loans, yi2 ; commercial and
industrial loans, yi3 ; and federal funds sold, securities purchased under
agreements to resell, plus total securities held in trading accounts, yi4 ), four
variable inputs (interest-bearing deposits except certificates of deposit greater
than $100,000, xi1 ; purchased funds (certificates of deposit greater than
$100,000, federal funds purchased, and securities sold plus demand notes) and
other borrowed money, xi2 ; number of employees, xi3 ; and book value of
premises and fixed assets, xi4 ), and one quasi-fixed input (noninterest-bearing
deposits, xi5 ).7
7
Here and elsewhere, we denote the number of observations in a given cross-section as n,
although the number of observations varies across the three cross-sections. Kaparakis et al. (1994)
argue that since, by definition, banks cannot attract more noninterest-bearing deposits by offering
interest, they should be regarded as exogenously determined as a first approximation. Although
banks might offer various services or other incentives to attract noninterest-bearing deposits, we
assume that banks take the quantity of these deposits as given. With the recent proliferation of
‘‘sweep’’ accounts, in which noninterest-bearing transactions deposits are automatically moved into
interest-earning accounts until needed, the treatment of noninterest-bearing accounts as a quasi-
fixed input may be less tenable. Prior to 1995, however, few banks offered such accounts and so our
treatment of noninterest bearing accounts as quasi-fixed seems reasonable. See Hunter and Timme
(1995) for further discussion of quasi-fixed inputs and an investigation of the empirical significance
of taking core deposits as a quasi-fixed input for a sample of large banks.
660 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

We compute input prices in the typical way: pi1 ¼ average interest cost per
dollar of xi1 ; pi2 ¼ average interest cost per dollar of xi2 ; pi3 ¼ average wage
per employee; pi4 ¼ average cost of premises and fixed assets. Finally, the total
cost of the variable inputs defines the Pdependent variable Ci to be used in
estimating the cost function; i.e., Ci ¼ 4j¼1 pij xij : For estimation purposes, we
normalize total variable costs and input prices with respect to the fourth input
price (pi4 ) to ensure linear homogeneity.8 We convert all dollar values to 1992
prices using the GDP deflator.
Because the Call Reports include some firms that have bank charters, but
which do not function as traditional banks (e.g., credit-card subsidiaries of
bank holding companies), we employ several selection criteria to limit the
sample to a group of relatively homogeneous banks. In particular, we omit
banks reporting negative values for inputs, outputs, or prices. Since some
remaining observations contain values for pi1 and pi2 that are suspect, we omit
observations when either of these variables exceed 0.25.9 After omitting
observations with missing or implausible values, we have 13,168, 11,786, and
9819 observations for 1985, 1989, and 1994, respectively, consistent with the
decline in the number of U.S. commercial banks by about one-third between
1985 and 1994.

4. Estimation method

To estimate the measures of scale and product mix economies defined in

(2.2)–(2.4), we must first estimate the relation between the various outputs and
input prices, and bank costs. The usual approach involves estimating the
conditional expectation function mðxÞ ¼ EðCi* jX i ¼ xÞ, where X is an ðN  8Þ
matrix of explanatory variables, with the ith row of X containing the elements
logðyi1 þ 1Þ, logðyi2 þ 1Þ, logðyi3 þ 1Þ, logðyi4 þ 1Þ, logðpi1 =pi4 Þ, logðpi2 =pi4 Þ,
logðpi3 =pi4 Þ, and logðxi5 þ 1Þ, and where C * is an ðN  1Þ column vector whose
elements Ci* equal the log of normalized variable cost, i.e. logðCi =pi4 Þ, for all
i ¼ 1; y; n. Costs and input prices are normalized with respect to the fourth
8
Stocks used to define inputs and outputs (as opposed to flows used to define price variables) are
mean values for the calendar year. For example, to compute outputs for 1985, we average the
values of each stock from the end-of-year Call Reports of 1984 and 1985. All values are book
values, except in the case of total securities held in trading accounts, which are reported in terms of
market value beginning in 1994. Unfortunately, there are no periods for which both book and
market values of these securities are available. However, we believe the effects of this discrepancy
are small since this item represents a small proportion of the fourth output for most banks.
Omitting this item in computing the fourth output had only minor quantitative and no qualitative
effects on our results.
9
The distributions of the price variables are somewhat continuous up to some point below 0.25,
with a few (clearly implausible) large outliers in the right tail.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 661

input price to ensure homogeneity of the cost function with respect to input
prices.
Assume
Ci* ¼ mðX i Þ þ ei ; i ¼ 1; y; N; ð4:1Þ
where ei is an independent stochastic error term, Eðei jX i ¼ xÞ ¼ 0, and
VARðei jX i ¼ xÞ ¼ s2 ðxÞ. In addition, we assume the observations ðX i ; Ci* Þ are
multivariate identically and independently distributed across i.
Parametric estimation involves specifying functional forms both for mðX i Þ
and the distribution of ei in (4.1). The translog specification is given by
mðX i Þ ¼ a þ X i b þ X i PX 0i ; ð4:2Þ

where X i represents the ith row of X, a is a scalar parameter, b is an ð8  1Þ

vector of parameters, and P ¼ ½pij  is an ð8  8Þ matrix of parameters and pjk ¼
0 8j > k. For our translog model, we also assume ei is normally distributed.10
We estimated the translog specification via feasible generalized least squares
using the corresponding share equations with cross-equation restrictions. A
Chow test indicated that the coefficients of the translog function are not stable
across four bank size groups: less than $100 million of assets; $100–$300
million; $300 million–$1 billion; and greater than $1 billion. Similarly, using a
Wald test, we rejected the null hypotheses of no difference in pairwise
comparisons of coefficients across the four subsets. All tests rejected the null at
better than 99.9 percent significance.11
Previous studies have similarly rejected stability of the translog function, and
frequently turn to orthogonal series estimation as an alternative methodology.
We thus also estimate (4.2) using the Fourier flexible form, following the
approach of Berger and Mester (1997).12 Unfortunately, orthogonal series
10
In some studies, the error term is assumed to be composed of a normal component reflecting
white noise, and a one-sided component reflecting inefficiency. To the extent that cost inefficiency is
present here, the distribution of the error terms in (4.2) might be skewed.
11
Detailed descriptions of the tests are available on request. The assumption of normally
distributed errors in (4.1) may not be justified if banks are cost-inefficient. However, in each
instance our tests lead to rejection of the null hypotheses at far greater than 99.9 percent
significance. While it would be straightforward to bootstrap the distributions of our test statistics to
allow for violation of our normality assumption, experience suggests that this would not make a
qualitative difference in our conclusions.
12
Like Berger and Mester (1997) and other studies, for each cross-section, we restrict the number
of parameters to be estimated to n2=3 . Our approach differs from Berger and Mester’s (1997)
however, in that we do not normalize the variables in our cost function with respect to total assets,
since we wish to estimate costs at mean input prices and arbitrary output levels. Although
normalizing by total assets may reduce heteroskedasticity in the error term, and hence increase
statistical efficiency, in our case we would have to estimate total assets corresponding to various
output levels not actually observed in the data, introducing an additional source of estimation
error.
662 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

estimators such as the Fourier flexible form present some yet-unanswered

statistical questions with potentially serious implications for estimation
results.13 We thus also employ kernel estimation and local polynomial
smoothing (LPS). While these nonparametric approaches also raise various
statistical questions, several well-understood, feasible methods have
been developed to address them.14 For our kernel estimation, we use the
Nadarya–Watson estimator (Nadarya, 1964; Watson, 1964) with a multi-
product standard Gaussian kernel function.15 Bandwidth parameters are
selected by least-squares cross-validation as described by Silverman (1986),
which involves minimizing an approximation to the mean integrated square
error. To minimize the mean-square error effects of the curse of dimensionality,
we base our estimates on the first five principal components of X, which
we found to contain more than 95 percent of the independent linear
information in X.16
The Nadarya–Watson estimator is perhaps the most commonly
used nonparametric regression estimator. Yet, as discussed by Fan and
Gijbels (1996), the more recently developed LPS estimator offers several
advantages over the Nadarya–Watson estimator; e.g., the local linear
version of the LPS estimator has the same variance but less bias than the

13
Perhaps most serious concerns the choice of how many series terms to retain for estimation.
In finite samples, the infinite Fourier series must necessarily be truncated to fewer than
n terms. Increasing the number of terms reduces bias, but increases variance. Little
theory or Monte Carlo evidence exists to guide this choice. A second question concerns
the choice of basis functions in the Fourier expansion. Banking studies have used sine
and cosine terms to represent the basis functions, but other choices are possible (e.g., orthogonal
Hermite, Jacobi, Laguerre, or Legendre polynomials), and there is little or no evidence on the
implications of selecting different representations. Finally, Barnett et al. (1991) note that
expansions other than the Fourier may be used, and suggest the Muntz–Satz expansion
instead.
14
With kernel regression and LPS, a bias=variance tradeoff arises in choosing the smoothing
parameter. But, with either method, one can use least-squares cross-validation, which amounts to
choosing the smoothing parameter that minimizes an approximation to mean integrated square
error. There are no corresponding methods for choosing the number of terms in series estimation.
Gallant (1981, 1982) demonstrates that setting the number of terms in the Fourier flexible form at
n2=3 minimizes bias, but this says nothing about variance.
15
Aside from the Nadarya–Watson estimator, one could also use the Priestly and Chao
.
(1972) or Gasser and Muller (1979) estimators, and there are perhaps an infinite number of kernel
functions that one might choose from. See Silverman (1986) and H.ardle (1990) for further
discussion.
16
This is a standard dimension-reduction technique for nonparametric regression; see Scott
(1992) for details. Additional details are also available from the authors on request.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 663

Nadarya–Watson estimator. Use of LPS for estimation in (4.1)

involves choosing both the order of the polynomial as well as a kernel
function and a bandwidth parameter, as noted previously.17 We use a
local linear estimator with the same multiproduct standard Gaussian kernel
function as in the Nadarya–Watson estimator, and nearest-neighbor band-
widths chosen by least-squares cross-validation. In addition, we use the same
dimension-reduction techniques as with our Nadarya–Watson kernel
estimator.
#
Each of our estimation strategies yields estimates mðxÞ of the conditional
mean function in (4.1). Taking exponentials of these estimates yields cost
estimates for arbitrary values x that can be substituted into (2.2)–(2.4) to
obtain estimates of those quantities used to evaluate returns to scale. Inference,
however, requires estimation of confidence intervals. To accomplish this, we
use the so-called ‘‘wild bootstrap’’ proposed by H.ardle and Mammen (1993).18
It is then straightforward to estimate confidence intervals for the quantities
defined in (2.2)–(2.4).

5. Empirical analysis

We first examine ray-scale economies by estimating SðyjyÞ defined in (2.2).

We accomplish this by setting the output vector y in (2.2) equal to the median
output vector for each sample and using each of our estimation methods to
produce estimated costs at yy and at y.19 The estimated costs are substituted
#
into (2.2) to produce estimates SðyjyÞ of SðyjyÞ for various values of y.
Because banks tend to suffer from considerable managerial, or x-, inefficiency,
we estimate scale economies using both the full sample for each year we
consider and a 25 percent, ‘‘thick frontier’’ subsample in each year. These
subsamples consist of the 25 percent of banks with lowest total variable costs in
17
Increasing the order of the polynomial adds more local parameters, which has the effect of
reducing bias but increasing variance. Fan and Gijbels (1996) show that the bias of the regression
estimator decreases for every increase in the order of polynomial, but variance increases only when
passing from an odd-order to the consecutive even-order polynomial. Thus odd-order polynomials
are necessarily preferable to even-order polynomials. McManus (1994) uses a localized version of
the translog, which amounts to a local quadratic, which is of order two. Either a local linear or local
cubic fit would be preferable.
18
The wild bootstrap involves a two-point approximation to the distribution of each residual
error and, in the case of the Nadarya–Watson estimator and LPS, avoids an inconsistency problem
that would result if we used the more conventional bootstrap approach of resampling from the
empirical distribution of the data. While this inconsistency problem does not arise in the case of the
translog model or with the Fourier series estimator, we use the wild bootstrap there as well to
maintain constant treatment across the four estimation approaches.
19
Normalized input prices and the quasi-fixed input are also set at the median values for each
sample. We use medians rather than means because of the skewed distribution of the data.
664 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

Table 1
Minimum points for SðyÞa

Full sample Thick frontier

y Assets y Assets

1994 LPS >36 1964 >36 1859

Kernel 15 818 >36 1859
Fourier >36 1964 >36 1859
Translog 16 873 20 1033

1989 LPS >36 1765 >36 1293

Kernel 16 785 18 647
Fourier >36 1765 23 826
Translog 15 736 >36 1293

1985 LPS >36 1695 >36 1170

Kernel 25 1177 17 552
Fourier 16 753 >36 1170
Translog 11 518 20 650
a
Values of y correspond to point at which SðyÞ reaches a minimum as y increases from 0.1;
corresponding level of total assets are given in millions of 1992 dollars.

each of 10 asset-size groups.20 We compute estimates SðyjyÞ # for yAf0:1; 0:15;

0:20; y; 0:95; 1:0; 2:0; y; 36:0g using each of our six samples. These values of y
correspond to 98 percent of the range of bank sizes in the sense that for 1994,
the 1st and 99th percentiles of the asset-size distribution were approximately
equal to 0.1 and 36.0 times the median level of assets, respectively. In every
#
case, SðyjyÞ is initially decreasing in y, indicating increasing returns to scale.
Table 1 summarizes our findings for ray-scale economies, presenting values
#
of y at which returns to scale are exhausted (i.e., points at which SðyjyÞ reaches
its minimum), and the corresponding level of assets. For smaller values of y,
#
SðyjyÞ is generally falling, indicating increasing returns to scale; for larger
#
values, SðyjyÞ is generally increasing in y, indicating decreasing returns to
scale. Without exception, SðyjyÞ# declines monotonically for yo1, i.e., for
banks less than the median-sized bank. For y > 1, there are local minima for
#
SðyjyÞ in some cases, suggesting movement from increasing to decreasing, and
then back to increasing returns to scale for larger values of y. As we discuss
below, however, all such local minima are in regions where we cannot reject the
hypothesis of constant returns to scale.

20
The 10 asset-size groups are the same as those that are used to examine expansion-path scale
economies, discussed below. See Berger and Humphrey (1997) for a survey of efficiency studies and
evidence, and Berger and Humphrey (1991) for discussion of the thick frontier methodology.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 665

In general, the estimates of ray-scale economies appear to be more similar

across years than across estimation methods. For the full sample in 1985, e.g.,
#
estimation of the translog cost function yields a minimum for SðyjyÞ at y ¼ 11
(approximately $518 million of assets). That is, this point estimate indicates
that banks experience increasing returns to scale up to about $518 million,
constant returns at that point, and decreasing returns above $518 million. The
Fourier flexible functional form yields a somewhat higher point of constant
returns, approximately $753 million, while the kernel regression and LPS
technique yield considerably larger points of constant returns, suggesting that
banks could achieve increasing returns up to $1,177 million and $1,695 million,
respectively. Similar differences in returns to scale estimates occur for 1989 and
1994. But, in those years, returns to scale estimates from the kernel regression
are similar to those obtained from the translog cost function, while the Fourier
flexible form and LPS technique yield considerably larger estimates of the
#
minimum SðyjyÞ.
The results for the thick frontier suggest somewhat different conclusions
about returns to scale. For 1985, the kernel regression method yields a
minimum value of SðyjyÞ # at y ¼ 17 (approximately $552 million), the
translog cost function a minimum at y ¼ 20 (approximately $650 million),
and the Fourier flexible functional form and LPS estimator yield minimums at
y > 36 (i.e., exceeding $1,170 million). For 1989, the kernel regression
again yields a minimum for SðyjyÞ # at the lowest level of output, y ¼ 18
($647 million). For 1994, however, only the translog function yields a minimum
#
for SðyjyÞ in the interior of the range we consider, at y ¼ 20 ($1033
million).
#
Our point estimates SðyjyÞ suggest that banks face increasing returns to
scale over a somewhat larger range than previous studies have found, most of
which conclude that scale economies are exhausted once banks reach sizes of
$100–$200 million.21 Consistent with McAllister and McManus (1993),
however, we find that nonparametric estimation techniques tend to yield
higher estimates of RSCE than the translog function does, though not
dramatically so in comparison with kernel regression in 1989 and 1994.
Consistent with Berger and Mester (1997), we find some evidence that the size
at which banks exhaust returns to scale increased between the 1980s and
the 1990s. It is important to note, however, that the values shown in Table 1
are merely point estimates, and estimates of confidence intervals are necessary
to make any inferences regarding returns to scale. Since the data are sparse in
#
the regions where SðyjyÞ is minimized, the confidence intervals may be
wide. Consequently, the ranges over which we can reject constant returns

21
An exception is Mitchell and Onvural (1996), who look only at banks with at least $500 million
of assets, and find evidence of ray-scale economies for banks of up to $2 billion of assets.
666 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

to scale in favor of increasing returns are more similar across years and
estimating methodologies than the summary results reported in Table 1 would
suggest.
Additional information about returns to scale and the extent of scale
inefficiency for banks of different sizes is given in Table 2.22 Here we report
point estimates and corresponding estimated 95 percent simultaneous
confidence intervals for SðyjyÞ for various values of y from the four different
estimation methods and three different sample years.23 The results indicate the
extent of scale inefficiency for banks of different sizes, relative to banks of the
median size. For example, for 1994, the LPS estimator suggests that
Sðy ¼ 0:2jyÞ falls between 1.2670 and 1.3051. In other words, five banks,
each with size equal to 20 percent of the size of the median bank (in 1994, the
median size was $54.6 million), could have produced the same output as one
median-size bank, but at a combined cost 26.7–30.5 percent higher than
the single median-size bank. The point estimates derived from the alternative
estimation techniques differ somewhat, but all indicate that such small banks
are decidedly scale inefficient.
Where estimated confidence intervals around SðyjyÞ for a particular value of
y do not overlap with corresponding confidence intervals for the next value of
y, we conclude that SðyjyÞ is decreasing or increasing in y in a statistically
significant sense. For each year, using each of our estimation methods we find
#
that SðyjyÞ significantly decreases up to y ¼ 6, i.e., banks up to six times the
median bank sizeFmore than 90 percent of all banks in each yearFoperate
under increasing returns to scale. Beyond y ¼ 6, estimated confidence intervals
from the kernel, Fourier flexible form, and parametric translog methods do not
allow rejection of the null hypothesis of constant returns to scale. However, the
#
LPS estimator suggests that SðyjyÞ significantly decreases in y up to y ¼ 12 for
1985 and up to y ¼ 18 for 1989 and 1994. Moreover, the estimated confidence
interval for Sðy ¼ 36jyÞ obtained with the LPS estimator does not overlap the
confidence intervals at y ¼ 12 or 18, in any of the years we consider, suggesting
that returns to scale are increasing up to the very largest banks in each year,
although the effect becomes less pronounced beyond the 90th percentile. One
should perhaps not make too much of results for the largest banks, however,
since the data are very sparse in this region.

22
We present only the full-sample results. Results for the ‘‘thick-frontier’’ sample are similar and
are available on request from the authors.
23
Simultaneous confidence intervals, as opposed to the more typical individual confidence
intervals, are required to make inferences regarding the slope of SðyjyÞ. We estimate these
confidence intervals by first estimating individual confidence intervals, and then adjusting using the
Bonferonni inequality. This method produces conservative estimates (i.e., confidence intervals that
are too wide), making it harder to find significant differences.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 667

Table 2
Ray-scale economy estimatesa

y LPS Kernel Fourier Translog

1994
0.2 1.2858 1.5529 1.3769 1.3627
(1.2670,1.3051) (1.4171,1.6872) (1.3339,1.4236) (1.3348,1.3935)
0.6 1.0832 1.1739 1.0861 1.0865
(1.0781,1.0883) (1.1351,1.2114) (1.0748,1.0973) (1.0811,1.0925)
6.0 0.7557 0.7401 0.8426 0.8359
(0.7432,0.7678) (0.7082,0.7764) (0.8181,0.8682) (0.8245,0.8478)
12.0 0.6787 0.7190 0.8054 0.8171
(0.6632,0.6937) (0.6709,0.7797) (0.7720,0.8438) (0.8007,0.8361)
18.0 0.6376 0.7143 0.7879 0.8160
(0.6206,0.6539) (0.6622,0.7769) (0.7449,0.8376) (0.7947,0.8404)
24.0 0.6101 0.7440 0.7788 0.8197
(0.5922,0.6273) (0.6878,0.8070) (0.7247,0.8380) (0.7942,0.8484)
30.0 0.5896 0.7569 0.7744 0.8251
(0.5711,0.6076) (0.6918,0.8543) (0.7143,0.8407) (0.7956,0.8594)
36.0 0.5735 0.7631 0.7729 0.8312
(0.5545,0.5919) (0.6923,0.8933) (0.7095,0.8536) (0.7981,0.8705)

1989
0.2 1.3440 1.4643 1.3711 1.3784
(1.3258,1.3651) (1.2784,1.6104) (1.3250,1.4128) (1.3781,1.4253)
0.6 1.0983 1.1210 1.0954 1.0940
(1.0936,1.1037) (1.0874,1.1539) (1.0839,1.1053) (1.0898,1.0986)
6.0 0.7204 0.7110 0.8131 0.8253
(0.7080,0.7314) (0.6787,0.7468) (0.7893,0.8345) (0.8164,0.8348)
12.0 0.6351 0.6372 0.7767 0.8066
(0.6197,0.6485) (0.6008,0.6871) (0.7486,0.8050) (0.7911,0.8229)
18.0 0.5901 0.6226 0.7590 0.8065
(0.5733,0.6046) (0.5846,0.6767) (0.7293,0.7877) (0.7863,0.8298)
24.0 0.5602 0.6740 0.7488 0.8113
(0.5426,0.5753) (0.6260,0.7461) (0.7134,0.7838) (0.7876,0.8403)
30.0 0.5381 0.6842 0.7427 0.8177
(0.5200,0.5536) (0.6337,0.7674) (0.7023,0.7847) (0.7904,0.8519)
36.0 0.5207 0.6654 0.7391 0.8249
(0.5023,0.5366) (0.6183,0.7392) (0.6971,0.7888) (0.7947,0.8635)

1985
0.2 1.3520 1.5534 1.3801 1.4036
(1.3336,1.3703) (1.4009,1.7110) (1.3269,1.4421) (1.3815,1.4260)
0.6 1.1003 1.1186 1.1042 1.0933
(1.0955,1.1049) (1.0738,1.1568) (1.1092,1.1154) (1.0890,1.0973)
6.0 0.7176 0.6901 0.8103 0.8371
(0.7068,0.7281) (0.6574,0.7271) (0.7905,0.8340) (0.8268,0.8464)
12.0 0.6323 0.6643 0.7926 0.8267
(0.6192,0.6451) (0.6233,0.7249) (0.7650,0.8229) (0.8106,0.8423)
18.0 0.5874 0.6304 0.7918 0.8327
(0.5733,0.6013) (0.5899,0.6891) (0.7532,0.8279) (0.8131,0.8530)
668 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

Table 2 (Continued )

y LPS Kernel Fourier Translog

24.0 0.5577 0.5943 0.7921 0.8424

(0.5430,0.5722) (0.5561,0.6482) (0.7478,0.8335) (0.8194,0.8666)
30.0 0.5359 0.6084 0.7920 0.8532
(0.5208,0.5509) (0.5726,0.6558) (0.7426,0.8404) (0.8264,0.8808)
36.0 0.5188 0.6275 0.7918 0.8643
(0.5035,0.5342) (0.5886,0.6843) (0.7365,0.8467) (0.8337,0.8949)
a
Bootstrapped 95 percent Bonferroni simultaneous confidence interval estimates in parentheses.
In 1985, 1989, and 1994, median assets (corresponding to y ¼ 1) were $47.088, $49.037, and
$54.560 million (1992 dollars), respectively.

5.1. Expansion path scale economies

Because banks of different sizes have varying output mixes, ray-scale

economies, which measure returns to scale along a ray emanating from the
origin, might give a misleading picture of scale economies. Thus, we also
examine expansion-path measures, which allow for variation in output mix
across banks of different sizes. Expansion path scale economies (EPSCE) are
estimated by replacing cost Cð Þ in (2.3) with estimated costs Cð Þ# from each of
the four estimation methods. For ya and yb , we use median output vectors for
10 asset-size groups.
Estimates are shown in Table 3 for y ¼ 0:1. The first row, for example, shows
estimates of scale economies for banks operating one-tenth of the distance
along the expansion path from the median output vector of banks with assets
of $0–$10 million to the median output vector of banks with $10–$25 million of
assets. One can examine EPSCE at any point along the expansion path, but by
estimating Eðyjya ; yb Þ at small values of y, e.g., 0.1, we can infer the extent of
scale economies near the median of the smaller size group.24 The estimates of
Eðyjya ; yb Þ in the first row of Table 3 indicate that in 1994, banks producing the
median output vector in the $0–$10 million size range could reduce their scale
inefficiency by expanding toward the output vector of the median bank in the
next larger asset size group. Moreover, with the exception of the estimate
derived from kernel regression, these estimates are statistically significant.
The qualitative results for EPSCE are broadly similar across years and
estimation methodologies. The LPS estimator indicates significantly increasing
returns to scale up to the $100–$200 million group in each year, with
insignificant results beyond this size group. The other estimators also indicate
24 #
By definition, Eðyjy a b
; y Þ ¼ 1 for y ¼ 1. We estimated Eðyjya ; yb Þ at y ¼ 0:1; :2; y; 0:9 and
#
found that Eðyjya b
; y Þ monotonically approached unity in every case where the estimates for y ¼ 0:1
were significantly different from unity, as well as in most other cases.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 669

Table 3
Expansion path scale economy estimatesa

Asset size group LPS Kernel Fourier Translog

(millions of $)

1994
0–10=10–25 1.3313b 1.1155 1.0871b 1.0731b
10–25=25–50 1.1919b 1.0375 1.0449b 1.0548b
25–50=50–75 1.2971b 1.1230b 1.0170b 1.0272b
50–75=75–100 1.1195b 0.9886 1.0026b 1.0137b
75–100=100–200 1.6561b 1.0117 1.0108b 1.0141b
100–200=200–300 1.0103 1.0767b 1.0157b 1.0097b
200–300=300–500 0.6533 0.9428 1.0212b 1.0067b
300–500=500–1000 @5.5494 0.7769b 1.0133 0.9932b
500–1000= > 1000 5.6315 1.4356b 0.9961 0.9802b

1989
0–10=10–25 1.2382b 1.1813 1.0056b 1.0737b
10–25=25–50 1.3156b 0.9572 1.0516b 1.0529b
25–50=50–75 1.3685b 1.1729b 1.0371b 1.0276b
50–75=75–100 1.2885b 1.0125 1.0065b 1.0117b
75–100=100–200 2.4789b 1.0644b 1.0103b 1.0142b
100–200=200–300 0.8033 0.9026b 1.0109b 1.0088b
200–300=300–500 0.4939 1.0546 1.0143b 1.0022
300–500=500–1000 1.3152 1.0262 1.0125 0.9892b
500–1000= > 1000 0.8418 1.1182 0.9676 0.9661b

1985
0–10=10–25 1.3075b 1.0610 1.0533b 1.0718b
10–25=25–50 1.1257b 1.1302b 1.0700b 1.0487b
25–50=50–75 1.4330b 1.0280 1.0274b 1.0236b
50–75=75–100 1.3299b 1.0300 1.0085b 1.0092b
75–100=100–200 1.8670b 1.1559b 1.0062 1.0087b
100–200=200–300 0.7775 0.9440 0.9977 1.0033b
200–300=300–500 0.1340 1.0157 0.9934 0.9435b
300–500=500–1000 1.4084 1.0167 0.9992 0.9881b
500–1000= > 1000 @0.1214 0.8745 0.9978 0.9618b
a
Significance was determined by bootstrap estimation. Values greater than unity indicate
increasing returns to scale, while values less than unity indicate decreasing returns to scale.
b
Significant difference from unity at 95 percent.

increasing returns to this level of output, or possibly larger. Because the 90th
percentiles of the bank size distributions are $196, $220, and $238 million of
assets in 1985, 1989, and 1994, respectively (constant 1992 dollars), however,
the data are too sparse among large banks to produce reliable estimates. The
extraordinarily large absolute values for the last two comparisons in 1994
obtained with the LPS estimator also reflect the sparseness of the data, as do
670 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

similar results for the largest groups in 1985 and 1989. None of these estimates
are statistically significant.25

5.2. Expansion-path subadditivity

Table 4 reports our estimates of expansion-path subadditivity (EPSUB),

AðyÞ, obtained by replacing true costs in (2.4) with estimated costs. The point
#
estimates AðyÞ indicate the potential cost savings associated with producing
the scale and output mix of the median bank in the larger-size group in a single
bank, rather than in two smaller banks. Estimates for the translog specification
are significantly negative when the two largest size groups are compared in
each year, suggesting decreasing returns, but all other statistically significant
estimates are positive, indicating increasing returns to scale. As discussed
previously, however, results for larger banks should be taken cautiously due to
increasing sparseness of data beyond about $200 million of assets. This is
particularly true for estimates of EPSUB, since cost estimates for hypothetical
banks (corresponding to yb @ya in (2.4)) are required.
In qualitative terms, the estimates of EPSUB are again similar across
estimation methods. Quantitatively, however, there are some differences.
Values obtained with the LPS estimator are typically considerably larger than
corresponding values obtained from the kernel estimator, which in turn are
frequently twice or more larger than corresponding estimates from the Fourier
flexible form and the parametric translog specification. That the latter two
methods should give quantitatively similar estimates is perhaps not surprising,
since the translog specification is nested within the Fourier flexible form
specification, and both are fit globally.
Finally, it should not be surprising that the nonparametric LPS and kernel
estimators yield quantitatively different results. Nonparametric estimators
avoid any risk of specification error, which is always a problem for parametric
methods, but at the expense of increased variability in estimates. In addition,
recall that EPSUB is defined in (2.4) as a ratio of costs, and that for technical
reasons, we actually estimate the log of cost. To obtain estimates of EPSUB,
we must take exponentials of estimated log cost, and then take ratios of these
values. Thus, a small difference in estimates from the LPS and kernel methods
is magnified exponentially, then perhaps magnified further when ratios are
taken to obtain estimates of EPSUB. This effect is more pronounced in the case
of EPSUB estimates than with the RSCE and EPSCE estimates because we
25
The parametric translog estimates, by contrast, are statistically significant in all but one case.
This apparent significance, however, is illusory, and is a result of the parametric structure that has
been imposed in the case of the translog model. Because the translog model is fit by least squares,
the few large banks in the sample act as leverage points, perhaps exerting undue influence on
parameter estimates in the model.
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 671

Table 4
Expansion-path subadditivity estimatesa

Asset size group LPS Kernel Fourier Translog

(millions of $)

1994
0–10=10–25 1.0081b 0.1637b 0.1848b 0.1755b
10–25=25–50 1.0030b 0.2222b 0.1496b 0.1395b
25–50=50–75 0.7215b 0.2060b 0.1137b 0.1118b
50–75=75–100 0.6206b 0.1752b 0.0881b 0.0931b
75–100=100–200 0.6863b 0.1423b 0.0756b 0.0789b
100–200=200–300 0.8969b 0.1245b 0.0582b 0.0577b
200–300=300–500 0.8417b 0.0680b 0.0516b 0.0393b
300–500=500–1000 0.8852b @0.0319 0.0440b 0.0174b
500–1000= > 1000 0.8836b @0.0224 0.0193 @0.0176b

1989
0–10=10–25 1.0158b 0.2529b 1.1828b 0.1906b
10–25=25–50 0.9848b 0.1026b 0.1403b 0.1469b
25–50=50–75 0.8064b 0.1841b 0.1234b 0.1189b
50–75=75–100 0.6579b 0.1638b 0.0972b 0.0936b
75–100=100–200 0.7935b 0.1807b 0.0828b 0.0740b
100–200=200–300 0.9091b 0.0731b 0.0562b 0.0527b
200–300=300–500 0.8499b 0.0278 0.0483b 0.0339b
300–500=500–1000 0.8794b 0.0844b 0.0431b 0.0109b
500–1000= > 1000 1.1432b 0.0190 0.0037 @0.0279b

1985
0–10=10–25 1.0311b 0.2103b 0.1554b 0.1910b
10–25=25–50 0.9851b 0.2182b 0.1539b 0.1460b
25–50=50–75 0.8084b 0.1655b 0.1306b 0.1156b
50–75=75–100 0.6485b 0.1346b 0.1013b 0.0900b
75–100=100–200 0.7909b 0.1772b 0.0803b 0.0701b
100–200=200–300 0.9141b 0.1516b 0.0518b 0.0440b
200–300=300–500 0.9008b 0.0939b 0.0295b 0.0213b
300–500=500–1000 0.8850b 0.0535 0.0083 0.0062b
500–1000= > 1000 0.9890b @0.0116 0.0026 @0.0388b
a
Significance was determined by bootstrap estimation. Values greater than zero indicate
increasing returns to scale, while values less than zero indicate decreasing returns to scale.
b
Significant difference from zero at 95 percent.

must rely on estimates of cost for hypothetical, nonexistent banks in the case of
EPSUB which may be far removed from actual banks in the data space.
Despite quantitative differences between the results produced by the different
estimators, qualitatively, our estimates are robust across estimation methods.
Moreover, our results for EPSUB are consistent with our findings for RSCE
and EPSCE. For each year, the results in Table 4 indicate potentially large, and
672 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

statistically significant, cost savings associated with combining the operations

of two (or more) smaller banks into a single bank operating at the median bank
output vector up to the 300–500 million size range, and possibly larger. The
potential savings, however, appear to diminish somewhat with increasing size,
as suggested also by our estimates of RSCE and EPSCE, but the results
certainly seem consistent with the rapid pace of mergers and consolidations
observed in the banking industry in recent years.

6. Summary and conclusion

Previous studies of scale and product mix economies in banking have

produced a myriad of conflicting results. Some of these differences reflect
differences in sample banks and periods studied. Recent work by Berger and
Mester (1997) suggests that bank cost relationships have changed since the
1980s, and they find evidence of larger-scale economies in the 1990s than others
estimated for banks in earlier periods. Differences in estimating methodology
could also account for some of the differences in estimates of scale and product
mix economies across studies. McAllister and McManus (1993) note that
estimates of scale economies derived from fitting a translog function to bank
costs can be biased and, in their sample, they estimate that scale economies are
exhausted at a much lower output level when the translog is fit to bank costs
than when semi- and nonparametric estimators are used to fit bank cost data.
Because of the potential biases associated with using the translog function, the
Fourier flexible functional form has become a popular alternative method of
fitting bank cost data. This method suffers from a number of unresolved
statistical problems, however, as described in Section 3.
In this paper, we present new estimates of scale and product mix economies
for the universe of U.S. commercial banks with usable data for 1985, 1989, and
1994. By using a large sample and including banks of all sizes, we avoid
potentially spurious findings associated with estimating scale economies with
restricted samples. Because neither the translog function nor the Fourier
flexible functional form are free from estimation problems or the potential to
introduce biases, we also estimate scale and product mix economies using
kernel regression and local polynomial smoothing. Although no estimation
methodology is free of problems, kernel regression and local polynomial
smoothing seem especially appropriate techniques to estimate scale and mix
economies in banking due both to their local nature and because statistical
properties of these estimators are well established. Moreover, the additional
results obtained from these techniques provide a robustness check on estimates
derived from the translog and Fourier flexible functional form methods. We
modify measures of expansion path scale economies and subadditivity
suggested by Berger et al. (1987) for use with nonparametric estimation
 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674 673

methods and, unlike many previous studies, we estimate and report robust
confidence intervals for our point estimates of scale and mix economies.
In sum, our estimates of ray-scale economies and expansion path
subadditivity are broadly consistent with one another in indicating that banks
could achieve potential economies by expanding the size of their output and
adjusting their output mix toward those of banks with at least $300–$500
million of assets. Although we find some evidence of scale economies for banks
as large as $1 billion, our point estimates are not estimated precisely across all
methodologies, and, hence, we do not draw firm conclusions. Our ray-scale and
expansion path estimates also suggest that, if anything, the size at which scale
economies are exhausted may have increased between the mid-1980s and mid-
1990s, consistent with a conjecture by Berger and Mester (1997). Our results
thus appear broadly consistent with the ongoing consolidation of the banking
industry and increasing average bank size. The wide range over which we
cannot reject constant returns to scale suggests, however, that banks of many
sizes could be competitively viable, though firm conclusions are difficult to
draw because the density of banks exceeding $1 billion of assets is low.

References

Akhavein, J.D., Berger, A.N., Humphrey, D.B., 1997. The effects of megamergers on efficiency and
prices: evidence from a bank profit function. Review of Industrial Organization 12, 95–139.
Barnett, W.A., Geweke, J., Wolfe, M., 1991. Seminonparametric Bayesian estimation of the
asymptotically ideal production model. Journal of Econometrics 49, 5–50.
Berger, A.N., DeYoung, R., 1997. Problem loans and cost efficiency in commercial banks. Journal
of Banking and Finance 21, 849–870.
Berger, A.N., Hanweck, G.A., Humphrey, D.B., 1987. Competitive viability in banking: scale,
scope, and product mix economies. Journal of Monetary Economics 20, 501–520.
Berger, A.N., Humphrey, D.B., 1991. The dominance of inefficiencies over scale and product mix
economies in banking. Journal of Monetary Economics 28, 117–148.
Berger, A.N., Humphrey, D.B., 1992. Megamergers in banking and the use of cost efficiency as an
antitrust defense. The Antitrust Bulletin 37, 541–600.
Berger, A.N., Humphrey, D.B., 1997. Efficiency of financial institutions: international survey and
directions for future research. European Journal of Operational Research 98, 175–212.
Berger, A.N., Kashyap, A.K., Scalise, J.M., 1995. The transformation of the U.S. banking
industry: what a long, strange trip it’s been. Brookings Papers on Economic Activity 2, 55–218.
Berger, A.N., Leusner, J.H., Mingo, J.J., 1997. The efficiency of bank branches. Journal of
Monetary Economics 30, 141–162.
Berger, A.N., Mester, L.J., 1997. Inside the black box: what explains differences in the efficiencies of
financial institutions? Journal of Banking and Finance 21, 895–947.
Boyd, J.H., Graham, S.L., 1991. Investigating the banking consolidation trend. Quarterly Review,
Federal Reserve Bank of Minneapolis (Spring), pp. 3–15.
Clark, J.A., 1996. Economic cost, scale efficiency, and competitive viability in banking. Journal of
Money, Credit, and Banking 28, 342–364.
Fan, J., Gijbels, I., 1996. Local Polynomial Modelling and Its Applications.. Chapman and Hall,
London.
674 D.C. Wheelock, P.W. Wilson / Journal of Monetary Economics 47 (2001) 653–674

Gallant, A.R., 1981. On the bias in flexible functional forms and an essentially unbiased form.
Journal of Econometrics 15, 211–245.
Gallant, A.R., 1982. Unbiased determination of production technologies. Journal of Econometrics
20, 285–323.
.
Gasser, T., Muller, H.G., 1979. Kernel estimation of regression functions.. In: Gasser, T.,
Rosenblatt, M. (Eds.), Smoothing Techniques for Curve Estimation. Springer, Berlin, pp.
23–68.
H.ardle, W., 1990. Applied Nonparametric Regression.. Cambridge University Press, Cambridge.
H.ardle, W., Mammen, E., 1993. Comparing nonparametric versus parametric regression fits.
Annals of Statistics 21, 1926–1947.
Hunter, W.C., Timme, S.G., 1995. Core deposits and physical capital: a reexamination of bank
scale economies and efficiency with quasi-fixed inputs. Journal of Money, Credit, and Banking
27, 165–185.
Kaparakis, E.I., Miller, S.M., Noulas, A., 1994. Short-run cost inefficiency of commercial banks: a
flexible stochastic frontier approach. Journal of Money, Credit and Banking 26, 875–893.
McAllister, P.H., McManus, D., 1993. Resolving the scale efficiency puzzle in banking. Journal of
Banking and Finance 17, 389–405.
McManus, D.A., 1994. Measuring scale and scope economies in banking using the localized
translog functional form. Finance and Economics Discussion Series working paper #94-30,
Board of Governors of the Federal Reserve System, Washington, DC.
Mitchell, K., Onvural, N.M., 1996. Economies of scale and scope at large commercial banks:
evidence from the Fourier flexible functional form. Journal of Money, Credit, and Banking 28,
178–199.
Nadarya, E.A., 1964. On estimating regression. Theory of Probability and its Applications 10,
186–190.
Priestly, M.B., Chao, M.T., 1972. Nonparametric function fitting. Journal of the Royal Statistical
Society, Series B 34, 385–392.
Scott, D.W., 1992. Multivariate Density Estimation: Theory, Practice, and Visualization. Wiley,
New York.
Silverman, B.W., 1986. Density Estimation for Statistics and Data Analysis. Chapman a Hall,
London.
Watson, G., 1964. Smooth regression analysis. Sankhya Series A 26, 359–372.
Wheelock, D.C., Wilson, P.W., 1999. Technical progress, inefficiency, and productivity change in
US banking, 1984–1993. Journal of Money, Credit, and Banking 31, 212–234.