Journalof

BANKING &
FINANCE
ELSEVIER Journal of Banking & Finance 20 (1996) 1651- 1671

The composite cost function and efficiency in

giant Japanese banks
D o n a l G. M c K i l l o p a, j. Colin Glass b, * , Y u k i o M o r i k a w a c
a Queen's University, Belfast, UK
b University of Ulster at Coleraine, Banking and Commerce Department. Coleraine, BT52 ISA. UK
c Waseda University, Tok3,o, Japan

Received 26 January 1995; accepted 20 November 1995

Abstract
This study provides further empirical investigation, in the context of giant Japanese
banks, of the recent claim by Pulley and Braunstein (1992, A composite cost function for
multiproduct firms with an application to economies of scope in banking, Review of
Economics and Statistics 74, 221-230), that their new composite model for the multiprod-
uct cost function has important advantages over the separable quadratic, generalized
translog and standard translog models. In addition to assessing the composite model's
relative ability in measuring global scope and scale economies, the study also extends the
P-B analysis to assess measurement of product-specific scope and scale economies,
pairwise cost complementarities between outputs, changes in the marginal costs of outputs
and technological change. The results appear to confirm P-B's claim. The persistent finding
of scale economies for large Japanese banks is also investigated and confirmed.

JEL class!fhTation: G21; D21; C23

Keywords: Japanese banks; Cost functions; Economies of scale: Economies of scope

* Corresponding author. Tel. (+44) 1265 44141; Fax. (+44) 01265 324910.

0378-4266/96/$15.00 Copyright © 1996 Elsevier Science B.V. All rights reserved.

PII S 0 3 7 8 - 4 2 6 6 ( 9 6 ) 0 0 0 19-2
1652 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

1. Introduction

In a recent paper, Pulley and Braunstein (1992) proposed a new composite cost
function model for the analysis of cost efficiency in multiproduct firms. Pulley and
Braunstein (denoted P-B for short) also compared this new model with a set of
alternative functional forms and assessed its advantages in the context of measur-
ing economies of scope in US banking.
The P-B composite cost function combines the log-quadratic input price
structure of the popular translog specification with a quadratic structure for
multiple outputs. As such, the P-B composite model constitutes an empirically
tractable cost function with important advantages. In particular, its quadratic
output structure accords with the Baumol et al. (1982) recommendation that such
structures are more suitable for measuring economies of scope, product-specific
economies of scale and subadditivity. Also, by not imposing strong separability
between outputs and inputs, the P-B composite cost function improves upon the
separable quadratic cost function models suggested by Braunstein and Pulley
(1984) and RiSller (1990). This improvement means that, in contrast to the latter
two models which are multiplicatively separable in outputs and input prices, the
P-B composite model does not require all input demands to vary with outputs in
the same fashion. Consequently, the P-B composite model has the important
advantage that it does not impose the restrictive assumption that input ratios and
cost shares are independent of output levels. In addition, it does not impose the
restrictive assumption that input demand elasticities with respect to each output are
equal and independent of input prices. Lastly, the P-B composite model also has
the advantage that its log-quadratic input price structure is easily restricted to be
linearly homogeneous.
The aim of this study is to provide further investigation of the advantages of the
composite model by utilizing it in the context of giant Japanese banks. As a first
step, the study follows the P-B procedure in estimating the composite, generalized
translog, standard translog and separable quadratic cost functions, and then
compares the results with respect to statistical fit, satisfaction of regularity
conditions and measurement of global economies of scale and scope, i The P-B
analysis is then extended by an assessment of the models with respect to
product-specific economies of scope, pairwise cost complementarities between
outputs, product-specific economies of scale and changes in the marginal costs of
outputs.
The choice of giant Japanese banks for the empirical analysis enables us to also
investigate the persistent, and somewhat surprising, finding of increasing returns to
scale for all sizes of Japanese banks. (It should be noted that this finding relates

i In evaluatingthe global scopeestimates, we also incorporatethe Noulas et al. (1993) suggestionof

checking the satisfactionof regularityconditionsat the point of evaluating scope effects.
 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671 1653

not just to parametric econometric studies such as Kasuya (1986), and Tachibanaki
et al. (1991), which utilize the translog functional form. 2 It also relates to
nonparametric programming studies such as Fukuyama (1993).) In contrast to this
finding, most empirical studies of scale efficiency in banking have found increas-
ing returns to scale only among relatively small banks and decreasing returns to
scale among larger banks.
However, as McAllister and McManus (1993) have recently demonstrated, the
fitting of a single translog cost function over a population of banks that varies
widely in terms of size and output mix (as is the case in the above Japanese
banking studies) results in a specification bias. In particular, they found that this
bias contributes substantially to the traditional conclusion of decreasing returns to
scale among larger banks. Consequently, when using restricted samples of large
US banks, McAllister and McManus (1993) found evidence of increasing returns
to scale for banks fight up to about $7000 million in total assets.
The McAllister and McManus (1993) results, however, only serve to make the
persistent finding of increasing returns to scale for all sizes of Japanese banks
more intriguing - especially as the major studies have fitted a single transtog cost
function over all sizes of banks. Is it the case that large Japanese banks are indeed
characterized by very substantial increasing returns to scale, so that the toregoing
translog studies still yield this finding despite the possibility of the downward
specification bias mentioned above? In order to investigate this, we have utilized a
restricted sample of giant Japanese banks (all of which had assets greater than
$400,000 million in 1991 at the end of our sample period), both to see whether or
not the increasing returns to scale result is still found and to compare our results
with other studies of Japanese banking. Also, by using different functional forms,
we can check whether the results of the scale efficiency investigation are sensitive
to the functional form estimated.
It should be noted that the results of the scale efficiency investigation have an
intrinsic interest beyond that of Japanese banking, since the particular banks
selected are currently the largest banks in the world. It should also be noted that,
since the study utilizes a restricted sample of giant banks, the whole investigation
is made in the context of pooled cross-section time-series estimation (as opposed
to the P-B cross-section estimation). In doing this, each of the four models is
adjusted so as to incorporate technological change and to permit the intercept of
the output structure to vary between banks.

~-Both Kasuya (1986) and Tachibanaki et al. (1991) use a two-output, three-input translog model to
investigate overall scale economies and pairwise cost complementarities. The two outputs are revenue
from loans and revenue from securities business; the three inputs being labour, capital and funds (the
intermediation approach being employed). Kasuya (1986) uses cross-section data for 77 banks,
repeating the analysis for each year over 1975-85. Tachibanaki et al. (1991) uses cross-section data for
61 banks, repeating the analysis for each year over 1985-87.
1654 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

Table 1
Total assets, output mix and input price data for sample banks at the end of March 1991 (millions of
yen) a
Dia-ichi Fuji Sakura Sanwa Sumitomo
Total assets 62573 290 61437697 61 279666 58 126523 62573000

Outputs
1. Loans and bills 36082800 32597900 38243 100 33990200 33063 100
discounted
2. Cash and due 12932800 13088400 7430341 13407600 13782300
from banks +
all loans
3. Securities+ 6610583 6 112 322 7 002 299 6 466478 7597390
trading account
securities

Input prices
1. Price of labour 11.310 10.084 11.063 9.900 9.699
2. Price of capital 0.264 0.167 0.212 0.265 0.211
3. Price of funds 0.070 0.068 0.062 0.070 0.074

Total cost 3 600116 3051632 3 196249 3 174887 3527533

a At the end of March 1991, the exchange rate was ¥141 = $1.

The five giant Japanese banks 3 selected for empirical analysis are all catego-
rized as city banks in the Japanese banking system. Unlike many Western banking
systems, Japanese banking is structured in terms of segregated banking systems
(see Suzuki (1987) for further discussion). As such, there are five distinct
categories of private banks: city banks, regional banks, the former sogo banks
(which are now treated as ordinary commercial banks), trust banks and long-term
credit banks. Given that the last two categories of banks specialize mainly in the
supply of long-term funds for industry, most empirical studies of Japanese banking
relate to the first three categories.
The city banks are characteristically large financial institutions with a national
and international coverage and a predominant, though not exclusive, emphasis on
short-term activities. Regional banks are essentially concerned with banking within
a specific district, their activities being related both to small and medium-sized
businesses and to local governments. Although most regional banks are much
smaller than city banks, some large regional banks are comparable to certain
smaller city banks in terms of asset size. The former sogo banks, which are
generally smaller than regional banks, were previously restricted by regulation to

3 The five giant banks are the Dai-ichi Kangyo Bank, the Fuji Bank, the Sakura Bank, the Sanwa
Bank and the Sumitomo Bank. In 1991 each of these banks had assets greater than $400,000 million.
(Note: prior to April 1, 1992, and during our sample period, the Sakura Bank was known as the Mitsui
Taiyo Kobe Bank. For ease of recognition, we have used the Sakura name in Tables 1 and 3.)
 D.G. McKillop et al./Journal of Banking & Finance 20 (1996) 1651-1671 1655

supplying credit to small firms only (though, unlike regional banks, they were
permitted to engage in instalment financing activities). Recently, these former
sogo banks were permitted to convert into ordinary banks so that their current
activities are essentially the same as those of regional banks.
Given the additional objective of investigating the scale efficiency of very large
Japanese banks, we have concentrated on city banks. Also, since the McAllister
and McManus (1993) analysis highlights the need to try to ensure a reasonably
homogeneous sample of very large banks, in terms of assets size and output mix,
we consider it reasonable to utilise the particular five city banks that are named in
Table 1. An examination of the data for 1991 in Table I indicates that the five
selected banks are quite similar in both size and output mix. By concentrating on
these banks, we were also able to ensure as long a run of consistent (annual)
observations as possible on the variables involved.

2. The composite cost function model

In presenting their composite cost function, P-B first develop a general specifi-
cation for the indirect multiproduct cost function. They then demonstrate that, for
certain parameter restrictions, this general model nests not only the composite cost
function, but also the standard translog, the generalized translog and a separable
quadratic specification. Utilizing superscripts in parentheses to represent the
Box-Cox transformation as follows:
= - for 6 * 0
and thus
v( + ) = y - I for 4 ) = 1 and y(~)--*ln v f o r O ~ 0 ,
the P-B General Specification, adjusted so as to incorporate technological change
and dummy variables, is given by

C (+1= exp a o+E~,D~+Ea~q} ~+y,t (A~+2EEa~Jq~ qi

1
+ + 2n ,q} )t + 2r., ql 1,1 r k

)(')] [ 1
+~p~kt(A)ln r k "exp /3o + F~13k In r k + - ~ 1 3 k l In r k In r /

1
+ ~,~t2,ikq} 7r) In r k + ~O,kt(A)ln rk]

= , f ( 6 ) ( q, In r, t) (1)
where qi, i = l . . . . . m refers to outputs; r k, k = l, .. ,n refers to input prices; t is
an index of time as a proxy for the level of bank technology and D~ refers to
1656 D.G. McKillopet al. / Journal of Banking & Finance 20 (1996) 1651-1671

dummy variables permitting the intercept of the output structure in (1) to vary
between banks. Following P-B, the empirical analysis will retain only one
intercept set (involving c% and the fl~ parameters), and the set of (input)
price-output (Sik) and (input) price-technology (Ptk) interaction terms entering
via the output structure ( flo, ~lbik and Otk thus being deleted).
With the above deletions implemented and with 4) ~ 0, A ~ 0 and ~-= 1 in (1),
the nested Generalized Translog Specification is
1
l n C = a'o + E fls D + Eotiq} ~r) -{- "Yt In t + ~ Y'.EOtijqi(Tr)qj(~)

1
d- ~ Ttt{ln t} 2 + ~'t'litq} ~r) In t + E~_.t~ikq}~r) In r k

1
+ EPtk In t l n r k + E/3~ In r k + ~EE/3klIn rk In r l (2)

where a 'o = a o - 1 and where ~- is empirically determined. The generalized

translog's associated input cost-share equations are
Sk : Ee~ikq} ~r) + Ptk In t + flk + E flkt In r t. (3)
Maintaining the restrictions ~b --* O, A ~ 0 and ~-= 1 in (2), but now requiring
7r-4 0, the nested Standard Translog Specification is readily obtained with all
output terms taking the familiar In qi form. Similarly, setting 7r -* 0 in (3) yields
the standard translog's corresponding input cost-share equations.
Returning to (1), with the above deletions still implemented, the restrictions
7r = 1, A = 1 and ~---* 0 yield the P-B nested Composite Specification as given by
([ 1 l
C (4')= Oto+~,~sDs+~Oliq'i+ Tit'+ -~Y'ff'.olijq;q)+ ~Ytt{t'} 2

+ Erlitq~t' + EY'.rikq'i In r k + 2 p,kt' In rk[

• exp Y'.flk In r k + ~ Y'.~/3kt In r k In r t (4)

where ql = q~ - 1 and t' = t - 1. P-B suggest that three suitable empirical repre-
sentations of the composite structure can be obtained by estimating (4) with
~b --* 0, or with 4) = 1, or with the value of q5 being empirically determined. The
composite function's associated input cost-share equations are
[ ,
S k = (~rikq' i q- Ptkt') a o q- ~flsDs + Y~otiqI q- ytt' + ~ o t i j q ; q ~

1 ,, , , ]-)
+ 2 Ytt{t'}2 + ~rlitqit + ~ S i k q i In r k + F. Ptkt In r k

+ ~k + ~" flkt In r 1. (5)

 D.G. McKillopet al./ Journal of Banking & Finance20 (1996) 1651-1671 1657

Maintaining the restrictions Ir = 1, h = 1 and z ~ 0 in (4), but now setting 6i~,

Ptk = 0 for all i, k in (4), P-B generate a nested Separable Quadratic Specification
similar to that used by Rrller (1990). Similarly, setting 6ik, Ptk = 0 for all i, k in
(5) yields the separable quadratic's input cost-share equations.
To ensure linear homogeneity in input prices we require
Efik = 1 a n d ~ l f l k l = O (k=l . . . . . n)
for all models and, in addition to these restrictions, we require
)Zk 6ik = 0 for all i and ~k Ptk = 0
in the generalized translog, the translog, the general and the composite specifica-
tions. Symmetry requires aij = aji and /3kl = /31k.
P-B estimated the general specification (1) and each of the four functional
forms that are nested within (1), with the cost function and input cost-share
equations being jointly estimated in each case. In doing this, P-B found a trade-off
between statistical fit and the satisfaction of regularity conditions. Thus while the
most flexible general model (1) provided the best fit, it was least effective in
satisfying the output regularity (marginal cost) conditions and the input price
regularity (concavity) conditions. Given this situation, P-B proceeded to assess the
relative advantages of the four nested forms. From this assessment, P-B concluded
that although the composite model also experienced a trade-off between fit and
regularity conditions, it could be strongly recommended on the grounds that it had
the highest log-likelihood value of the four models nested in (1), and it provided
more reasonable measures of global economies of scope.
In order to provide further assessment of the P-B conclusion, this study applies
the P-B analysis to the case of giant Japanese banks. This application also extends
the P-B analysis in that it assesses the four functional forms with respect to
additional measures of cost efficiency in multiproduct banking. The various cost
measures that will be computed are detailed in the following four paragraphs.
Adopting the intermediation approach, a three-output, three-input cost function
model will be estimated and global economies of scope computed via
SC~= [ C ( q , , 0 , 0 ) +C(O, q2,0 ) + C(0,0, q3 ) - C ( q ) ] / C ( q ) (6)
with SC~ > 0( < 0) denoting global economies (diseconomies) of scope. Similarly,
product-specific economies of scope are computed via
SC k = [C(qk) + C ( q , _ , ) - C ( q ) ] / C ( q ) (7)
where C(qk) is the cost of producing only product k and C ( q , _ , ) is the cost of
producing all of the n products except the kth one. Also, as Baumol et al. (1982)
have shown that a multiproduct cost function which exhibits weak cost comple-
mentarities over the full set of outputs up to the observed level of output (as given
by CCkm = 0 2 C/OqkOqm < 0) exhibits economies of scope, pairwise cost comple-
mentarities are computed as an additional check.
1658 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

The degree of scale economies over the production of the three outputs at
q = (ql, q2, q3 ) is computed via
3
e = C( q ) / Y ' ~ qkCk(q) (8)
k=l

where C(q) is the total cost of producing the three outputs (qk) and Ck(q) =
OC(q)/Oqk, with overall returns to scale said to be increasing, constant or
decreasing as • is greater than, equal to, or less than unity, respectively.
Product-specific economies of scale are computed via
PS k : AICk/ ( OCl Oqk) = [ C ( q ) - C( q , - k ) ] / q k ( O C / O q ~ ) (9)
where AIC~ is the average incremental cost of product k. Returns specific to the
scale of product k at q are said to be increasing, decreasing or constant as PS k is
greater than, less than or equal to unity, respectively. Also, as the estimates for
PSk are the result of extrapolating the cost function into the region of zero outputs,
the direction of change of the marginal cost of product k with respect to qk (as
given by AMC k = o Z C / O q 2 ) is also computed.
In those cases where the respective concept's measurement involves zero
outputs, the translog's computational problem is overcome by utilizing the general-
ized translog form with the B o x - C o x parameter on outputs set at the 7r = 0.001
value. 4 Also, technological change is computed via e t = - 0 In ClOt.

3. Data description

We use yearly data obtained from the annual reports of five very large Japanese
city banks over the period 1978-91.
Employing the intermediation approach, 5 the three output categories (measured
in real yen terms) are: loans and bills discounted; cash and due from banks plus
call loans; and securities plus trading account securities. Following Kasuya (1986),
the three inputs are specified as labour (defined as the number of employees in the
respective bank); capital (defined as the real yen value of movable and immovable
capital); and funds (defined as the outstanding real yen balance of funds raised
from customers). The price of labour is given by the ratio of (real yen) personnel
expenses to the number of employees in the respective bank; the price of capital in

4 Although we chose ~r = 0.001 to simulate ~" ---, 0, the statistical fit, the degree of satisfaction of the
regularity conditions, and the computed values of the various concepts are effectively unchanged as 7r
is reduced to a lower value such as ~" = 0.000001 (the only exception being that any unreasonable
concept values become even more unreasonable).
5 The intermediation approach was chosen so as to enable comparison with earlier studies of
Japanese banking. P-B use the value-added approach which considers both output and input character-
istics of deposits.
 D. G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671 [659

each bank is given by the ratio of (real yen) nonpersonnel expenses to the (real
yen) value of movable and immovable capital; and the price of funds in each bank
is given by the ratio of (real yen) fund-raising expenses to the (real yen) value of
raised funds. Consistent with the intermediation approach utilized, total cost in
each bank includes fund-raising expenses in addition to capital and labour
operating expenses. A description of the relevant data for 1991 is presented in
Table 1.
It should be noted that the current study employs 'real yen balance-sheet values
of earning assets' to measure the three outputs. This differs from the (two-output,
cross-section) studies of Japanese banking by Kasuya (1986) and Tachibanaki et
al. (1991), which both use 'revenue produced by earning assets' to measure their
two outputs (with this choice being determined by data limitations on outputs
across all city, regional, and former sogo banks). The latter studies were, of
course, well aware that the existence of market power in setting output prices may
influence their revenue-based measures of output. However based on observations
of the market, Kasuya (1986) argued that such market power was negligible.
Clearly, while this use of revenue-based measures of output may be reasonable for
such cross-section analysis, given data availability, it is not suitable for the current
analysis which utilizes annual observations over the sample period 1978-91.
Consequently, in common with other banking studies using time-series observa-
tions, we have deliberately avoided the use of revenue figures and we have instead
used balance-sheet values of the earning assets.

4. Estimation and empirical results

Each of the four cost functions nested within Eq. (1) was estimated jointly with
its respective input cost-share equations, with the linear homogeneity in input
prices and symmetry restrictions imposed across equations and error terms added
for empirical implementation. 6 The iterative Zellner estimation procedure was
utilized to obtain estimates, with the funds cost-share equation being omitted from
the estimated model in each case.
In estimating the composite function (and the separable quadratic function
which is nested within it), P-B suggest transforming both sides of the cost
function, as shown in Eq. (4), so as to permit the simultaneous examination of the

6 To be an adequate representation of the underlying production technology, the estimated cost

function must be concave in input prices and monotonically non-decreasing in input prices and outputs.
Monotonicity is satisfied if the fitted cost shares and marginal costs of outputs are positive, while
concavity is satisfied if the Hessian matrix [02C/OrkOre], computed using the relevant parameter
estimates, is negative semidefinite. For further details regarding the assumption of normally distributed
errors in (1) and the scaling involved in the P-B procedure see Pulley and Braunstein (1992. pp.
225-226).
1660 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

empirical specifications corresponding to the three cases of ~b ~ 0, 4' = 1 and the

value of ~b being empirically determined. The parameter 4' and the other cost
function parameters can then be estimated either (i) by searching over a grid of
given ~b values and choosing the estimates which yield the smallest sum of
squared errors (SSE), or (ii) by estimating th jointly with the cost function
parameters.
In our estimation of the composite cost function, we first utilized procedure (i)
above. However, before proceeding to the estimation details, it is important to note
that P-B have deliberately chosen to emphasize statistical fit in choosing among
alternative flexible forms, without making any attempt to impose regularity
conditions on these forms. In doing this, they recognise that other researchers have
suggested satisfaction of regularity conditions as a means of choosing among
flexible alternatives. For example, Barnett and Lee (1985) would select the
flexible cost function model that satisfies the regularity conditions over the largest
region. P-B are also aware that others such as Diewert and Wales (1987) and
R~Sller (1990) have emphasized the imposition of regularity conditions in estima-
tion situations. However, by reporting both the statistical fit details and the extent
of the violations of regularity conditions, P-B highlight that regularity conditions
and statistical fit may not be compatible as criteria for choosing among cost
function alternatives. Thus, for example, they found that while the translog model
did a much better job than the other models of satisfying regularity conditions, it
also manifested a much poorer statistical fit than the other models. Consequently,
as we are following the P-B approach, we first utilized procedure (i), in the
preceding paragraph, so as to examine the trade-off between statistical fit and the
satisfaction of regularity conditions.
The search over a grid of given 4' values revealed the useful information that
the highest degree of satisfaction of the regularity conditions (across all observa-
tions) was to be found for given values of ~b in the region around unity
(essentially for 0.7 < 4' < 1.3). Also, as ~b was increased through this region,
although the overall fit improved slightly the degree of satisfaction of the
regularity conditions was effectively unchanged. Moreover, while a better overall
fit could be obtained by a (higher) 4' value outside this region (utilizing either of
the procedures (i) or (ii) noted above), the improved fit could only be achieved at
the cost of an increasingly marked deterioration in the degree to which the
regularity conditions were satisfied. For example, the highest log-likelihood value
(of 645.8) was found at 4' = 3.96, with 25% output (marginal cost) regularity
violations and 31% input-price (concavity) regularity violations. In contrast, when
th = 1.0, the log-likelihood value was 639.9, with 9% output regularity violations
and 13% input-price regularity violations. (A similar situation was found for the
separable quadratic.) This result reinforces the P-B point regarding the trade-off
between increased flexibility for fit and satisfaction of regularity conditions. It also
indicates that this trade-off will vary with the data set employed, P-B utilizing
procedure (ii) above to find 4' = - 0 . 1 4 for the composite model without such
 D.G. McKillop et al. /Journal of Banking & Finance 20 (1996) 165l 1671 1661

Table 2
Estimation of the four cost function models a
Composite Separable Generalized Translog
quadratic translog
Parameter ¢5=Tr=A=l ~b=~r=A=l ¢5~0: A~O q~-~O
r=l 7r ~ (): A-*ll
Restrictions ~- ~ 0 ~- ~ 0: (Tr = 0.675) r =: I
~ik = /gtk = 0
No. of restrictions 4 12 3 4
Log likelihood 639.90 562.34 635.02 554.07

Summary statistics
Cost function
SSE 0.065 0.018 0.034 0.022
R2 0.997 0.999 0.998 0.999
Labour share
SSE 0.012 0.175 0.029 0.028
R2 0.947 0.182 0.862 0.863
Capital share
SSE 0.0024 0.0034 0.0020 0.0200
R2 0.673 0.554 0.730 0.335

Output regularity 9 8 10 7
violations (%)
Price regularity 13 14 6 7
violations (%)

" SSE = Sum of squared errors. No. of restrictions relative to adjusted form of (1) - see text.

extreme violation of the regularity conditions. Given the relative flatness of the
likelihood surface, we decided to concentrate on the case of ~ = 1.0, utilizing
procedure (i) above to estimate the composite model (and the separable quadratic
nested within it) via the iterative Zellner estimation procedure. Note, in doing so,
we therefore deliberately accepted a poorer overall fit so as to ensure better
satisfaction of the regularity conditions. 7 Having made this decision with regard
to the composite and separable quadratic models, the choice between the four cost
models will be made on the basis of statistical fit.
The log-likelihood values and the summary statistics for each of the four
estimated cost function models are given in Table 2. As indicated in the P-B
analysis (pp. 225-228), the SSE values are comparable across models. Table 2
also lists the number of parameter restrictions required for each model relative to

7 In writing up our results, we received the recent paper by Noulas et al. (1993) and were fascinated
to find that they were advocating a B o x - C o x parameter value of 1.0 (for 7r in the generalized translog
model), and thus a semi-log form of cost function for the case of large-sized US banks, so as to secure
the satisfaction of regularity conditions at the point of scope evaluation in output space (but at the cost
of a poorer overall fit). This led us to incorporate their check, as recorded in footnote I above.
1662 D.G. McKillop et al. /Journal of Banking & Finance 20 (1996) 1651-1671

the general model (1) within which it is nested (with the number of restrictions
being relative to (1) after linear homogeneity and symmetry have been imposed on
(1) and after/3 o, /xik and Ot~ have been deleted as noted above).
The results in Table 2 indicate that, even after ~b is restricted to unity, the
composite function achieves the highest log-likelihood value of the four nested
models. Also, utilizing the likelihood ratio test, the separable quadratic model is
rejected relative to the composite model within which it is nested (critical
2
X005 = 15.51; computed X 2 = 155.12). Likewise, the translog model is rejected
relative to the generalized translog in which it is nested (critical X 0.05 2 = 3.84;
computed X 2 = 161.90). Analogous to the P-B study, these comparisons of
statistical fit led us to select the composite model in preference to the other
models.
In relation to the satisfaction of regularity conditions, the fitted cost shares were
positive and less than unity for 99% or more of the possible cases for each of the
estimated functions. Also, although certain violations of the output and price
regularity conditions occurred (with Table 2 indicating the violations as a percent-
age of the possible violations), these regularity conditions were satisfied at the
means of the banks for all the estimated functions.
Having chosen the composite model on the basis of statistical fit, P-B pro-
ceeded to investigate the four cost functions relative to their ability in measuring
global economies of scope. The P-B study also reported measures of overall
economies of scale. Consequently, Table 3 presents the results of a similar
investigation for the five giant Japanese banks. Table 3 thus displays estimates of
the global economies of scope (SC G) and overall economies of scale ( e ) , at the
respective mean values of the variables that pertains for each of the five banks, for
each of the four cost function models. Table 3 also records the average annual
percentage rate of technological change over the 1978-91 period.

4.1. Global scope estimates

The global scope estimates in Table 3 indicate neither economies nor disec-
onomies of scope (SC G not being statistically different from zero) for the
composite, separable quadratic and generalized translog models, but statistically
significant diseconomies of scope for the translog model. However, the translog
estimates for SC6 are suspect in that they are all very close to minus one in
value. 8 This unreasonableness of the translog scope values accords with P - B ' s

8 It is rather unreasonable to accept that the first three terms in the numerator of Eq. (6) effectively
collapse to zero for every observation. The unreasonable values embodied in the component parts of
the computation of these terms reinforces this conclusion. (Moreover, the latter unreasonable values
become even more unreasonable as 7r is reduced in value from 7r = 0.001 to 1r = 0.000001, even
though this reduction leaves the statistical fit and satisfaction of regularity conditions effectively
unchanged. A similar situation prevails with regard to the unreasonable SCk values generated via the
translog model in Table 4.)
 D. G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671 1663

Table 3
Estimates of global economies of scope and scale a

Composite Separable quadratic Generalized translog Translog

Global scope
Dai-ichi - (I.44 - 0.57 0.22 1.00
(0.32) (0.36) (0.58) (0.21)
Fuji - 0.48 - 0.58 - 0.20 1.00
(0.32) (0.38) (0.6t~) (0.42)
Sakura - 0,45 - 0.46 0.27 1.00
(0.33) (0.38) (0.67) (0.41)
Sanwa - 0.48 - 0.49 0.19 - 1.00
(0.33) (0.36) (0.68) (0.42)
Sumitomo - 0.41 - 0.62 - 0.22 -- 1.00
(0.31) (0.37) (0.6'~) (0.41)

Overall scale
Dai-ichi 1.18 1.09 1.24 1.22
(0.07) (0.10) (0.10) (0.08)
Fuji 1.15 1.11 1.27 1.22
(0.06) (0.09) (0.08) (0.08)
Sakura 1.22 1.09 1.26 1.17
(0.07) (0.10) (0.15) (0.05)
Sanwa 1.15 1. I 1 1.27 1.23
(0.05) (0.09) (0.07) (0.08)
Sumitomo 1.19 1.08 1.23 1.23
(0.08) (0.09) (0.12) (0.07)

Technological change (%) 0.53 0.40 0.65 0.26

(0.03) (0.02) (0.12) (0.01)

a Standard errors in parentheses.

findings. With regard to the year by year results, the pattern depicted in Table 3
generally holds, except for the composite and separable quadratic models which
suggest that statistically significant diseconomies of scope set in from the late
1980s. 9
In contrast to the P-B study, the scope estimates in Table 3 for the generalized
translog appear reasonable. This suggests that where the B o x - C o x parameter (Tr)
on outputs is not close to zero, and thus the generalized translog is not a close
approximation to the translog form, the generalized model may provide reasonable
estimates of global economies of scope. In this study, the estimated value of 7r
was 0.675, whereas in the P-B study (where unreasonable S C G estimates were
obtained for both the generalized and the standard translog forms) it was 0.20.
Given this result, it is by no means clear that the composite model provides more

9 W e wish to thank an anonymous referee for an insightful c o m m e n t which helped us locate and
correct a computational error in our calculations of the scope measures for the latter models.
1664 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

reasonable S C c estimates, than does the generalized translog or separable quadratic

models, for the giant Japanese banks investigated in this study. However, we must
be cautious in using this result as a basis for generally recommending the
generalized translog model. The very fact that it is possible for 7r to be estimated
in the range that makes the generalized translog suffer from the same difficulties
as the translog, inevitably raises a question mark over the use of this model.
In a recent paper, Noulas et al. (1993) argue that since scope estimation
necessitates the evaluation of the estimated cost function in portions of output
space well outside the region of the original estimation (namely, where outputs are
set equal to zero in various combinations), the satisfaction of regularity conditions
should also be checked at the points of evaluation of the scope estimates.
Consequently, as we are essentially concerned with the choice between the
composite and generalized translog models, we checked the regularity conditions
for marginal costs of outputs and input cost-shares at the points of calculation of
S C G. In doing so, we also made this check for the generalized translog model with
~-= 1.0 instead of rr = 0.675 (as Noulas et al. (1993) found this to be the only ~-
value satisfying the regularity conditions at the point of scope calculation).
We first checked for positive marginal costs at the mean values of the relevant
non-zero outputs, in the various output triples used to calculate the scope effects.
This revealed that positive marginal costs held in the composite model case for
outputs 1 and 2, but not for output 3. In the generalized translog model case
positive marginal costs only held for output 1. When this check was made across
all observations of the relevant non-zero outputs, the composite model experienced
less violations than did the generalized model with 7r = 0.675. Similar to the
finding of Noulas et al. (1993), the generalized translog's satisfaction of output
regularity conditions, across all observations of the relevant non-zero outputs,
improved when 7r = 1.0. However, despite this improvement, the generalized
translog model was still inferior to the composite model in terms of output
regularity violations.
A check was also made for input cost-shares (S k) which are positive and less
than unity, at the mean values of the relevant non-zero outputs in the various
output triples used to calculate the scope effects. This revealed that the regularity
conditions on S~ held in the composite model case for outputs 1 and 2, but not for
output 3. In the generalized translog model case, the regularity conditions on Sk
held for all three outputs. When this check was made across all observations of the
relevant non-zero outputs, the composite model experienced more violations than
did the generalized translog model. Moreover, when ~" was changed from
7r= 0.675 to 7r= 1.0, the generalized translog's satisfaction of the regularity
conditions on Sk were unchanged.
These results for the Noulas et al. (1993) analysis of regularity conditions at the
point of evaluation of scope effects clearly do not provide criteria for choosing the
composite model in preference to the generalized translog model. They do,
however, indicate the need for caution in utilizing estimates of global scope
 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671 1665

economies based upon these models. Taken as a whole, the foregoing empirical
analysis of global scope economies for Japanese banks does not enable us to
conclude, as P-B could for US banking, that the composite model can be strongly
recommended (in preference to the other three models) for its ability to measure
global economies of scope. In particular, when we consider the mix of statistical
fit, satisfaction of regularity conditions and ability to measure global economies of
scope, taken together, there is not a lot to choose between the composite and the
generalized translog models. However, as noted earlier, it must be remembered
that this conclusion relates only to our particular case where 7r = 0.675 in the
generalized translog model. Clearly it is possible to find estimated values of 7r
that are much closer to zero. When this happens, the generalized translog cost
function is a close approximation to the translog form and thus will encounter
similar problems in estimating scope economies. This, in turn, means that the
generalized translog model cannot be generally recommended for the estimation of
scope effects.

4.2. Ol,erall economies o f scale estimates

The overall economies of scale (e) estimates in Table 3 indicate statistically

significant increasing returns to scale for all models except the separable quadratic
model which indicates constant returns to scale. The estimated values of E in
Table 3 (in range 1.08-1.28) are very similar to those found by the Kasuya (1986)
and Tachibanaki et al. (1991) studies (in each case, increasing returns to scale
being found with the estimated values of e being in the range 1.03-1.24). With
regard to the year by year results, the pattern depicted in Table 3 generally holds,
except from the late 1980s onwards where the results suggest that constant returns
pertain for all models.
As noted in the introduction, a subsidiary aim of this study was to investigate
whether or not the somewhat surprising finding of increasing returns to scale for
all sizes of Japanese banks still held for a restricted, reasonably homogeneous,
sample of giant Japanese banks. As also noted, since the particular five giant
banks used in the study currently constitute the five largest banks in the world,
there is an intrinsic interest in finding whether they manifest scale economies. The
above results indicate that this is indeed the case for most of the sample period.
Moreover, this general finding of scale economies (or of constant returns to scale
from the late 1980s) was essentially confirmed by all four models. Thus, in
comparison to the P-B study which used data from large US banks (having total
assets greater than $1000 million in 1988) and found evidence of only very small
scale economies, the current study uses data from five giant Japanese banks
(having total assets greater than $400,000 million in 1991) to find evidence of
appreciable scale economies.
A possible explanation for the lower scale economies in large US banks, as
compared to large Japanese banks, may be found in the US banking regulations
1666 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

which have hindered the development of interstate banking and branching. For
example, in McCloy (1994), the chairman of Banc One International (one of the
most profitable banks in the US and ranked sixth among the world's 100 best
banks, as ranked by Euromoney) has argued that legislation to encourage interstate
banking would secure cost savings as a result of the more efficient banking
operations that would ensue.
Also, with regard to the substantial scale economies in large Japanese banks, a
possible explanation may be found in the fact that, in addition to holding various
bonds, Japanese banks are allowed to hold shares of other corporate firms within a
certain limit, whereas US banks do not hold them in principle. Tachibanaki et at.
(1991) indicate that this feature of Japanese banking has produced a healthy
correlation between the amount of lending and the amount of shareholding for
inter-related banks and firms. In addition, they argue that the long run nature of
this interrelationship, which is so characteristic of Japanese business, has reduced
the information costs associated with negotiating and monitoring new lending
arrangements over time so that Japanese banks enjoy lower costs. Clearly,
however, both of these possible explanations need further investigation.
Table 3 also indicates that, except for the translog model, there are no dramatic
differences in the estimated average annual percentage rate of technological
change over the sample period. Apart from the usual cautions with regard to using
time as a technological index (see Hunter and Timme, 1991), the comparison of
the four models suggests the need for caution in drawing conclusions as to the rate
of technological change in sub-periods of the sample under consideration. This
caution arises from the different year by year results for the models: the composite
and separable quadratic models showing a lower average rate in the second half of
the period, as compared to the first half of the period, whereas the generalized and
standard translog models show a reverse pattern to this over the sample period.

4.3. Product-specific scope and scale estimates

In an attempt to shed further light on the relative advantages of the four models,
we will now extend the P-B analysis by examining the models' estimates of
additional measures of cost efficiency in multiproduct banking. In particular, we
examine the four models with regard to their respective estimates of product-
specific scope and scale economies.
The estimates of product-specific scope (SC k) in Table 4 (based on the mean
values of all the observations for each of the variables over the whole sample
period), immediately highlight once more the unreasonable nature of the translog
estimates. However, the other three models all show statistically significant
product-specific diseconomies of scope (SC k < 0) for k = 1,2, 3 (except for SC 2,
for the generalized translog, which is not statistically different from zero). The
year by year pattern of the SC~ estimates, for each model, is essentially similar to
the SC k values reported in Table 4.
 D.G. McKillop et al./ Journal of Banking & Finance 20 (1996) 1651-1671 1667

Table 4
Product-specific scope and cost complementaritiesestimates a
Composite Separablequadratic Generalized
translog Translog
Product-specific scope (SCk)
SC~ 0.50 - 0.59 - 0.47 - 1.00
(0.13) (0.07) (0.03) (0.0001 )
SC~ - 0.35 0.30 0.20 1.88D+ 07
((l.04) (0.004) (0.11 ) (0.37D + (17)
SC~ 0.41 - 0.57 - 0.57 - 1.00
(0.10) (0.09) (0.(14) (0.0001 )

Cost complementarities(CCkm)
CCI2 0.29 0.08 0.06 - 0.09
(0.02) (0.01) (0.02) (0.01)
CCI~~ 0.80 1.69 0.86 0.87
(0.01) (0.04) (0.02) (0.05)
CC23 0.11 - 0.08 - 0.01 0.12
(o.oi) (0.002) (o.oi) (o.oi)

a Standard errors in parentheses. In the SCk and CCk.,, the 1, 2 and 3 refer to the outputs 'loans and
bills discounted', 'cash and due from banks pins call loans', and 'sccurities plus trading account
securities', respectively.

All of the pairwise cost complementarities estimates in Table 4 (based on the

mean values of all the observations for each of the variables over the whole
sample period), are statistically different from zero, except for CC23 for the
generalized translog. The CCkm pattern is also not inconsistent with the earlier
global scope results (except in the translog case, where the SC G estimates are
rather unreasonable). Apart from the translog model, the C C t , , results for the
other three models all indicate anticomplementarity (CCk,,, > 0) for CC~2 and Ct3
and complementarity (CCkm < 0) for CC23. The year by year pattern of the CCk,,,
estimates is very similar to the CCkm values given in Table 4.
Given the very different CCkm pattern for the translog model, together with its
unreasonable SC G and SC k estimates, one cannot help but conclude that great care
be taken in utilizing translog estimates for assessing cost efficiency in Japanese
banking. For example, although the Kasuya (1986) and Tachibanaki et al. (1991)
translog models are not strictly comparable to the one above, in that they both
utilize a two-output model with revenue definitions of outputs, they found values
of CCI3 alternating in sign for various years during the period 1975-87 whereas
we found CC13 to be consistently positive over the same period.
The estimates of product-specific economies of scale ( P S i ) in Table 5 (based
on the mean values of all the observations for each of the variables over the whole
sample period), indicate considerable differences in the estimated PS k values for
the four models. Whereas all the PS k estimates generated via the composite and
separable quadratic models are statistically different from zero, only the PS~
1668 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

Table 5
Product-specific scale and change in marginal cost estimates a
Composite Separable quadratic Generalized translog Translog
Product-specific scale ( P S k)
PS 1 2.38 2.02 1.46 1.60
(0.14) (0.06) (0.03) (0.03)
PS 2 0.99 0.58 0.15 - 1.66D + 07
(0.01 ) (0.03) (0.10) (2.72D + 07)
PS 3 4.01 6.70 2.51 5.59
(0.67) (0.58) (2.69) (5.32)

Change in marginal cost ( A M C k)

AMC 1 - 1.44 - 1.26 - 0.92 - 1.13
(0.05) (0.01) (0.04) (0.07)
AMC 2 0.19 0.30 0.21 0.22
(0.02) (0.03) (0.01) (0.06)
AMC 3 - 0.99 - 1.49 -- 0.90 - 0.76
(0.03) (0.02) (0.04) (0.05)
a Standard errors in parentheses. The k = 1,2, 3 refers to the three outputs, as defined in the note at the
bottom of Table 4.

estimates for the generalized and standard translog models are statistically differ-
ent from zero (PS 2 for the translog being particularly unreasonable).
For all four models, statistically significant increasing returns to scale specific
to output 1 (loans and bills discounted) were found. This result also held year by
year over the sample period. For output 2 (cash and due from banks plus call
loans), the composite model suggests constant returns to scale, whereas the
separable quadratic model suggests decreasing returns to scale (with both these
results holding year by year over the period). In regard to output 3 (securities plus
trading account securities), both the composite and the separable quadratic models
indicate increasing returns to scale (with this holding year by year over most of the
period in both cases). These P S 3 values, however, appear to be rather high.
Finally, for all models, the estimates of the direction of change in the marginal
cost of product k with respect to qk (computed at the same point as the P S k
estimates), denoted by A M C k in Table 5, are not inconsistent with the P S k
estimates.
It is interesting to record, with respect to the year by year values of the set of
P S k estimates, that the composite and separable quadratic models manifest less
variation in the P S ~ values than is the case for the generalized and standard
translog models. This situation is summarized in Table 6 (which presents the
average values of the P S k estimates for the sample banks, by model, for selected
sub-periods over the sample period). As we are essentially concerned with the
choice between the composite and generalized translog models, these P S k results
provide further evidence in favour of concluding, along with P-B, that the
 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671 1669

Table 6
Product-specific scale estimates for selected sub-periods a
PSI PS 2 PS 3
Composite
1978-79 1.54 0.996 - 2.32
1980-82 1.56 0.993 2.4 l
1983-85 1.98 0.990 4.88
1986-88 2.72 0.988 6.06
1989-91 3.82 0.987 6.81

Separable quadratic
1978-79 1.44 0.92 5.8
1980-82 1.61 0.80 6.5
1983-85 2.02 0.62 8.4
1986 88 2.42 0.45 7.4
1989-91 2.44 0.23 4.8

Generalized translog
1978-79 1.43 1.20 - 0.82
1980-82 1.40 0.87 - 5.53
1983-85 1.40 0.50 10.36
1986-88 1.50 0.50 9.28
1989 91 1.72 12.60 -1.85

Translog
1978-79 1.31 - 7.57D + 05 5.86
1980-82 1.49 - 4.70D + 06 16.20
1983-85 1.79 - 3.33D + 07 10.41
1986-88 1.62 - 4.46D + 08 18.54
1989-91 1.69 - 1.07D + 09 - 15.16

~' The three outputs are defined as in the note at the bottom of Table 4.

c o m p o s i t e m o d e l d o e s h a v e a d v a n t a g e s o v e r the g e n e r a l i z e d t r a n s l o g m o d e l in
e m p i r i c a l i n v e s t i g a t i o n s o f m u l t i p r o d u c t firms. In o t h e r words, e v e n in a d i f f e r e n t
data c o n t e x t , w h e r e t h e c o m p o s i t e m o d e l d o e s not a p p e a r to h a v e an a d v a n t a g e
o v e r the g e n e r a l i z e d t r a n s l o g m o d e l in m e a s u r i n g g l o b a l a n d p r o d u c t - s p e c i f i c
s c o p e e c o n o m i e s , or p a i r w i s e cost c o m p l e m e n t a r i t i e s , the c o m p o s i t e m o d e l d o e s
a p p e a r to p r o v i d e m o r e r e a s o n a b l e e s t i m a t e s o f p r o d u c t - s p e c i f i c scale e c o n o m i e s
t h a n d o e s the g e n e r a l i z e d t r a n s l o g m o d e l .

5. Summary and conclusions

T h e f o r e g o i n g s t u d y p r o v i d e s f u r t h e r e m p i r i c a l i n v e s t i g a t i o n o f the c l a i m b y
P u l l e y a n d B r a u n s t e i n ( 1 9 9 2 ) that t h e i r n e w c o m p o s i t e m o d e l for the m u l t i p r o d u c t
cost f u n c t i o n h a s i m p o r t a n t a d v a n t a g e s o v e r the s e p a r a b l e q u a d r a t i c , g e n e r a l i z e d
t r a n s l o g a n d s t a n d a r d t r a n s l o g m o d e l s . P - B illustrate the a d v a n t a g e s o f the
c o m p o s i t e m o d e l in the c o n t e x t o f m e a s u r i n g g l o b a l e c o n o m i e s o f s c o p e a n d scale
1670 D.G. McKillop et al. / Journal of Banking & Finance 20 (1996) 1651-1671

in US banking. Our study not only applies the P-B empirical analysis to giant
Japanese banks but also extends it in order to assess the four models with respect
to additional measures of cost efficiency in multiproduct banking. In particular, the
extended analysis assesses the models with respect to product-specific economies
of scope, pairwise cost complementarities between outputs, product-specific
economies of scale and changes in the marginal costs of outputs. Also, by utilizing
pooled cross-section time-series estimation, the analysis incorporates technological
change.
The empirical results for global scope and scale economies generally indicate
(a) neither economies nor diseconomies of scope, and (b) increasing returns to
scale. For the best two models (namely, the composite and the generalized
translog), the results suggest that technological change lowers the real cost of bank
production by 0.53-0.65% per year over the 1978-91 period.
A subsidiary aim of this study was to investigate further the persistent finding
of scale economies for all sizes of Japanese banks. Using a restricted sample of
very large Japanese banks, rather than a sample containing all sizes of banks, we
found that these giant banks did indeed manifest appreciable scale economies. This
finding of increasing, rather than decreasing, returns to scale for very large banks
was essentially confirmed by all four models. A possible explanation, for this
result, may be that Japanese banks enjoy lower monitoring costs with respect to
lending due to their ability to hold shares and the characteristic long-run interrela-
tionship between Japanese banks and business.
In contrast to the P-B study, where the composite model provided more
reasonable global scope estimates than either of the translog models, the current
study found that the generalized translog and the composite models provided
equally reasonable scope estimates. This result was largely unchanged with respect
to measuring product-specific scope economies and pairwise cost complementari-
ties. It should, however, be noted that this result emerged from an empirical
situation where (unlike the P-B situation) the estimated generalized translog cost
function is not a close approximation to the translog form. Moreover, when we
come to the measurement of product-specific scale economies, the results suggest
that the composite model provides more reasonable measures than does the
generalized translog model. This suggests that the foregoing empirical analysis of
giant Japanese banks provides further evidence that the P-B claim, as to the
relative advantages of the composite model, is justified.
Although the above study appears to support the P-B claim, the trade-offs
between statistical fit, satisfaction of regularity conditions and ability to measure
cost concepts, for each functional form, highlights the need for caution in utilizing
the results of such models. 10 This is particularly so with regard to the global

10Caution is also necessary in that the current study implicitly assumes that the multiproduct banks
are operating on the production frontier. Clearly, additional assessment of the possible production
inefficiency of banks via the parametricand non-parametricapproachesto frontier analysis is desirable.
 D. G. McKillop et al. /Journal of Banking & Finance 20 (1996) 1651-1671 1671

s c o p e e s t i m a t e s , in light o f the v i o l a t i o n s o f regularity c o n d i t i o n s at the p o i n t o f

e v a l u a t i o n o f s c o p e effects. M o r e o v e r , w h i l e our c h o i c e o f the i n t e r m e d i a t i o n
a p p r o a c h w a s m o t i v a t e d b y a d e s i r e for c o m p a r a b i l i t y with earlier studies o f
J a p a n e s e b a n k i n g , the use o f a v a l u e - a d d e d a p p r o a c h ( c o n s i d e r i n g both output and
input c h a r a c t e r i s t i c s o f d e p o s i t s as d o e s the P-B study) m a y give rise to d i f f e r e n t
results.

Acknowledgements

T h e authors w i s h to a c k n o w l e d g e h e l p f u l c o m m e n t s f r o m t w o a n o n y m o u s
referees.

References

Barnett, W.A., and Y.W. Lee, 1985, The global properties of the miniflex Laurent, generalized
Leontief and translog fexible functional forms, Econometrica 53, 1421-1437.
Baumol, W.J., J.C. Panzar and R.D. Willig, 1982, Contestable markets and the theory of industry
structure (Harcourt Brace Jovanovich, New York).
Braunstein, Y.M. and L.B. Pulley, 1984, Scope and scale augmenting technological change: An
application in the information sector, in: M. Jussawalla and H. Ebenfield, eds., Communication and
information economics (Elsevier Scientific Publishers, New York) 105-I18.
Diewert, W.E. and T.J. Wales, 1987, Flexible functional forms and global curvature conditions.
Econometrica 55, 43-68.
Fukuyama, H., 1993, Technical and scale efficiency of Japanese commercial banks: A non-parametric
approach, Applied Economics 25, 1101-1112.
Hunter, W.C. and S.G. Timme, 1991, Technological change in large U.S. commercial banks, Journal of
Business 64, 339-362.
Kasuya, M., 1986, Economies of scope: Theory and application to banking, BOJ Monetary and
Economic Studies 4, 59-104.
McAIlister, P.H. and D. McManus, 1993, Resolving the scale efficiency puzzle in banking, Journal of
Banking and Finance 17, 389-405.
McCIoy, J.B., 1994, Interstate banking legislation from a banker's perspective, The Wall Street
Journal, March 9, p. A12.
Noulas, A.G., S.M. Miller and S.C. Ray, 1993, Regularity conditions and scope estimates: The case of
large-sized U.S. banks, Journal of Financial Services Research 7, no. 3, 235-248.
Pulley, L.B. and Y.M. Braunstein, 1992, A composite cost function lor multiproduct firms with an
application to economies of scope in banking, Review of Economics and Statistics 74, 221 230.
Rtiller, LH., 1990, Proper quadratic cost functions with an application to the Bell system, Review of
Economics and Statistics 72, 202-210.
Suzuki, Y., ed., 1987, The Japanese financial system (Oxford University Press, New York).
Tachibanaki, T., K. Mitsui and H. Kitagawa, 1991, Economies of scope and shareholding of banks in
Japan, Journal of the Japanese and International Economies 5, 261 281.