The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
JULY/AUGUST 1991 ECONOMIC PERSPECTIVES Contents Investm ent, G N P , and real exch an g e rate s........................................................................................... 2 Paula W o rthing to n A new study shows that industry investment rates and exchange rates are correlated, suggesting that changes in the value of the dollar affect the international competitiveness of U.S. firms. Prod uctive e fficie n cy in b an kin g ............................................................... 11 Douglas D. E van off and Philip R. Israilevich Studies show that banks are inefficient. The authors discuss why, and what this means for the future of the industry. ECONOMIC PERSPECTIVES Karl A. Scheld, Senior Vice President and Director of Research Editorial direction Carolyn McMullen, editor, David R. Allardice, regional studies, Herbert Baer, financial structure and regulation, Steven Strongin, monetary policy, Anne Weaver, administration Production Nancy Ahlstrom, typesetting coordinator, Rita Molloy, Yvonne Peeples, typesetters, Kathleen Solotroff, graphics coordinator Roger Thryselius, Thomas O’Connell, Lynn Busby-Ward, John Dixon, graphics Kathryn Moran, assistant editor ju ly/a u g u st 1991 Volum e XV , Issue 4 ECONOMIC PERSPECTIVES is published by the Research Department of the Federal Reserve Bank of Chicago. The views expressed are the authors’ and do not necessarily reflect the views of the management of the Federal Reserve Bank. Single-copy subscriptions are available free of charge. Please send requests for single- and multiple-copy subscriptions, back issues, and address changes to Public Information Center, Federal Reserve Bank of Chicago, P.O. Box 834, Chicago, Illinois 60690-0834, or telephone (312) 322-5111. Articles may be reprinted provided source is credited and The Public Information Center is provided with a copy of the published material. ISSN 0164-0682 In v e stm e n t, GIMP, and real e x ch a n g e rates Paula R. W o rth in g to n The value of the U.S. dollar varied widely over the 19631986 time period. Those same years witnessed several cyclical expansions and contractions and even wider swings in aggregate fixed investment rates. One explanation for some of the investment rate swings is the dramatic movements in exchange rates over this period. In this article, I use newly constructed capital stock and investment series for 270 U.S. manufacturing industries to examine investment responsiveness to changes in real exchange rates for 1963-1986. My research shows that invest ment rates are sensitive to real exchange rate movements and that appreciation of the U.S. dollar is associated with a decrease in industry investment rates—particularly in durable goods industries. Analysis of industries for which imports-sales data are available further suggests that investment is more responsive in industries with greater exposure to foreign competition. Finally, I document the existence of substantial interindustry variation in the influence of real exchange rates on investment. My results are broadly consistent with international trade models in which changes in real exchange rates drive changes in the relative competitiveness of domestic and foreign industries. Changes in real exchange rates are often thought to reflect changes in the international competitiveness of domestic and foreign industries. For example, the depreciation of the dollar is said to be correlated with improved competitiveness of U.S. firms, because U.S. and foreign consumers find it relatively cheap to buy 2 U.S. goods. In the long run, being competitive in international markets requires investing in capital equipment that will be used to satisfy current and future market demand. This suggests that real exchange rate movements are correlated with changes in international competi tiveness now and will continue to be in the future. By analyzing the extent to which investment spending of U.S. manufacturing industries has historically varied with changes in the value of the dollar, I indirectly examine how internationally competitive the U.S. manufactur ing sector will be in the future. The article is organized as follows. The next section outlines the expected effects of changes in the value of the dollar on output and input demands of U.S. manufacturing industries. The third section describes the data used in the article, and the fourth reports the results. Conclusions are in the final section. Why should real exchange rates m atter? Movements in the value of the dollar will affect the input and output choices of U.S. manufacturing firms as long as the goods produced are tradeable, that is, as long as output demand is sensitive to the relative price of domestic and foreign goods. Simply put, an The author is visiting assistant professor in the Department of Economics, Northwestern University, and consultant in the Economic Research Depart ment, Federal Reserve Bank of Chicago. The author w ould like to thank Hesna Genay, Jack Hervey, Prakash Loungani, Steve Strongin, and the editor for useful com m ents on earlier drafts and Jack Hervey and W illiam Strauss fo r assistance w ith some of the data. ECONOMIC PERSPECTIVES appreciation of the dollar lowers the relative price of foreign goods to U.S. goods. This causes demand for domestically produced goods to fall and, as a consequence, reduces input demands in the affected sectors. The appropriate measure of the relative price of home and foreign goods is the real exchange rate, which depends on the nominal exchange rate and home and foreign prices. To illustrate this relationship, consider Equation (1), where E is the nominal exchange rate, expressed in terms of units of foreign currency per U.S. dollar, and Pus (PF is the price level of the ) United States (foreign country). Equation (1) shows that the real exchange rate, e, is defined as Pus rates and investment rates. My analysis relies on the assumption that changes in e are exogenous at the individual industry level, that is, that the exchange rate is not affected by the actions of individual industries. This exogeneity assump tion has been exploited by other researchers interested in measuring the impact of exchange rate movements on industry outputs and inputs.3 The present analysis is only a first step towards understanding the relationship among investment spending, exchange rate movements, and international competitiveness. The evidence for the patterns documented here is suggestive, not conclusive, about the nature of this relation ship, and this article lays the groundwork for future analysis. (1) e = E* — . A review o f th e data The idea behind many theories of interna tional trade is that increases in e (appreciation of the dollar) cause decreases in domestic output and derived input demands. According to this view, the size of the output response in any given sector or industry will depend on the relevant demand elasticities and the expected persistence of the exchange rate shock. In turn, technologically determined elasticities of substitution and adjustment costs will determine the size of the input demand response.1 Shocks that are expected to be permanent may be met with changes in inputs that are relatively costly to adjust, such as capital, while more transitory shocks may be met with change in more easily altered inputs, such as labor. Furthermore, firms may alter prices instead of outputs and inputs, so that price-cost margins may also be affected when real exchange rates change. In this article, I do not seek to directly develop and test a model of real effects of exchange rate movements. Instead, I focus on the correlation between changes in the demand for one particular input, capital, and changes in an index of the real value of the dollar.2 Changes in the demand for capital, as measured by investment spending, are of interest because of the strong empirical evidence that investment spending is a large and cyclically sensitive component of U.S. total aggregate spending. Because industries differ widely with respect to their output and input demand elasticities as well as in their exposure to international markets, I expect to observe substantial cross-sectional variation in the relationship between exchange The industry data used in this article are annual figures for a subset of U.S. four-digit Standard Industrial Classification (SIC) manu facturing industries during the years 1963-1986.4 After elimination of industries with missing data, 270 industries remain in the data set. The data are derived from the Census of Manufac tures and the Annual Survey of Manufactures and were originally assembled by Domowitz, Hubbard, and Petersen (DHP) (1987). Data on capital stocks and investment, as well as other variables, are included in the data base, and the original data were used to construct several series used in this article. Capital stock series were computed by applying standard recursion formulas to benchmark stocks. See the Box for details. Table 1 gives the reader some background information on the industries studied here. The Table reports the full sample means and standard deviations for the gross investment rate, the sales to capital ratio, and the price-cost margin, and it also presents the same statistics for durable goods and nondurable goods industries sepa rately.5 The mean gross investment rate in the sample was .132, and the average sales to capital ratio was 5.11, implying a .20 capital-sales ratio. Durable goods industries are characterized by higher levels of capital intensity, higher invest ment rates, and higher price-cost margins than nondurable goods industries.6 Because investment spending is highly procyclical, I need to control for the level of macroeconomic activity in the analysis below. I use the ratio of actual to potential gross national product (GNP) for each year in the sample as my FEDERAL RESERVE RANK OF CHICAGO 3 Data sources and construction Data for the four-digit SIC industries are obtained from the data of Domowitz, Hubbard, and Petersen (DHP) (1987), who assembled the set from various years of the C e n s u s o f M a n u f a c tu r e s and A n n u a l S u r v e y o f M a n u f a c tu r e s . DHP’s original data set was updated and expanded at the Federal Reserve Bank of Chicago. Macroeconomic data are obtained from the National Income and Product Accounts (NIPA). Specifically, the following definitions and proce dures were used in constructing the data set used in this article. Unless otherwise noted, the annual data cover the 1963-1986 time period. In v es tm en t The Census reports total gross investment (dollars spent on new capital goods) in current (nominal) dollars. C apital sto ck The Census contains gross stock figures, but these data are not good measures o f capital for at least two reasons. First, the data embody an assumption of “one-horse-shay” depreciation.* Second, because stocks purchased at different times are added together, it is difficult to correct for changes in the price o f capital goods. Consequently, I construct a current (nominal) dollar capital stock series for each industry by applying a standard capital accumulation relationship to a benchmark capital stock. I use an annual geometric depreciation rate (5) for the total capital stock o f .0926, computed by the Bureau of Economic Analysis (BEA) and cited in Shapiro (1986). The capital accumulation equation embodies the “time-to-build” assumption and applies depreciation only to the current stock, not to the current year’s investment: measure of aggregate economic activity. The mean of this ratio over the 1963-1986 time period is 1.00. The real exchange rate measure used in this article is the real, trade-weighted index of the U.S. dollar developed at the Federal Reserve Bank of Chicago. This index, which is described in detail by Hervey and Strauss (1987a, 1987b, 1987c), was originally developed to measure exchange rate movements over the 1971-1986 time period and has recently been extended as far back as 1960. The index includes 16 countries, uses current consumer price indexes to convert nominal to real ex change rates, and is based to equal 1.0 in the first 4 (1) v 7 K it =1 i t - 1 Ip k J K it—1v ,(1-6),’ 7 Pt-I where K is the capital stock and p* is the implicit price deflator for capital goods, taken from NIPA. I use the 1958 gross stocks as benchmarks. Gross in v es tm en t rate The gross investment rate is defined as the ratio o f gross investment expenditures to the previous year’s capital stock: A I K .t ,, where both f and K jt jare measured in current dollars. N o m in al sales Nominal sales is defined as output minus the value o f final goods inventory changes. Specifically, S = V A D + C M . - T I N T Y + T I N T Y „ where V A D is value added, C M is cost of materials, and T IN T Y is final goods inventories, all taken directly from the Census. The sales-capital ratio, S _ K , is defined as S — K n = S.it IK .it-\ Price-cost m argin The price-cost margin (P C M jt) is defined as ( V A D - P A Y . ) ! ( V A D . + C M .) , where P A Y is total payroll, which is reported directly by the Census. M acro econ o m ic m easures I used the actual and potential gross national product (GNP) figures reported in NIPA, and I defined A _ P G N P t as the ratio o f actual to potential GNP in year t. This measure is identical to the one used by Petersen and Strauss (1989, 1991). *See Hulten and Wykoff (1981) for evidence that depreciation patterns tend to be geometric. quarter of 1973.7 The index series is quarterly; ,8 I use the four-quarter average index for each year in the sample.9 Results Before analyzing industry-level investment sensitivity to GNP and real exchange rate movements, it is instructive to consider invest ment behavior in the aggregate. Let I_Kt be defined as the simple cross-sectional average investment rate in year t. Figure 1 contains a graph which plots the ratio of /_Kt to its mean (.132), the ratio of actual to potential GNP (APGNP), and the real, trade-weighted dollar ECONOMIC PERSPECTIVES value of the dollar (R7GMA) more formally: TABLE 1 Summary statistics, 1963-1986 All industries Durable Nondurable goods goods industries industries Variable name Label Price-cost margin PCMit .274 (.093) .278 (.074) Gross investment rate l_Kit .132 (.069) .137 (.069) Sales-capital ratio S-Kit Number of industries 5.11 (3.29) 4.44 (2.01) 270 140 NOTE: Standard deviations are in parentheses. index, R7GMAt, over the 1963-1986 time period. The investment rate clearly varies procyclically, and investment’s variability appears to exceed that of output. The relationship between investment and the value of the dollar appears to be negative, at least after 1971 or so. This Figure suggests that, in the aggregate, invest ment does indeed vary procyclically and does increase when the dollar depreciates. The remainder of the article examines the data at the four-digit level. Table 2 presents the results of estimating the relationship between investment rates (I K.), actual to potential GNP (APG N P), and the real FEDERAL RESERVE BANK OF CHICAGO (2) I_K, = P0 + P,*A_PGNP,+ P,*R7GMAt + £it7 , • 2 .269 (.110) where /' denotes industry i, t denotes year t, and e7is an .128 econometric error term. Because (.068) preliminary analysis suggested that 5.82 the error term was serially corre (4.15) lated, I present both ordinary least squares (OLS) estimates and least 130 squares estimates corrected for first order serial correlation, which are denoted as PW, for PraisWinsten.1 I present results for the 0 full sample, for durable and nondurable goods industries separately, and for producer and consumer goods industries separately. The OLS and PW results are qualitatively similar; I will discuss only the PW results." The positive and significant coefficients on APG NP are interpreted as measuring the sensitivity of investment rates to changes in the strength of the macroeconomy. For the sample as a whole, a 1 percent increase in A PGNP from its mean of 1.00 implies an increase of .00407 in the investment rate, or a 3.1 percent increase relative to the rate’s mean of .132. These results conform with previous work by Petersen and 5 associated with a decrease in the investment rate of .00075, or a .57 Investment rates, GNP, and real exchange rates percent decrease relative to its Dependent variable: I Kit7 1963-1986 . mean. Table 2 also confirms that Coefficients: CONSTANT /!LPGNP, R7GMA, R2 investment patterns in durable All goods industries differ from industries (270) patterns in nondurable goods OLS -.238a .448a -.076a .04 industries. Durable goods invest (.029) (.029) (.004) ment is more cyclical and more PW -,203a ,407a -.072a responsive to changes in real (.029) (.030) (.008) exchange rates than is nondurable Durable goods goods investment. This difference industries (140) is significant at the 1 percent level. OLS -.312a ,568a —117a . .07 An alternative method of (.040) (.041) (.011) distinguishing broad groups of PW -,269a .518a -.1 1 1a industries is to group them on the (.040) (.041) (.011) basis of the buyer’s identity rather Nondurable goods than the good type. Table 2 reports industries (130) the results of estimating Equation OLS -.159a .319a -0 3 3 a .02 (2) separately for producer goods (.041) (.043) (.011) and consumer goods industries.1 2 PW -.131a .285a -,030b Although the coefficient estimates (.042) (.043) (.012) do differ between the groups, an F test at conventional significance Producer goods industries (196) levels fails to reject the hypothesis OLS -.247a .467° -.086a .04 that the coefficients do not differ. (.034) (.036) (.009) It appears, then, that the type of PW -. 184a good produced (the durable goods/ .401a -.085a (.034) (.035) (.010) nondurable goods distinction) matters more than the identity of Consumer goods industries (74) the customer (the producer goods/ OLS -.217a .398a -.049a .03 consumer goods distinction) in (.053) (.056) (.014) explaining investment patterns over this time period. PW -.255a .424a -.039b (.054) (.056) (.015) Previous researchers have documented substantial variation NOTES: l_Kit is the investment rate for industry /' in year t, A_PGNPt is the ratio of actual to potential GNP at time f, and R7GMAt is the real tradein output and input demand weighted dollar index at time t. OLS refers to the ordinary least squares behavior at the two-digit SIC level, estimates, and PW refers to the Prais-Winsten estimates, which correct for first order serial correlation. Standard errors are in parentheses under so in Table 3 I present the results coefficient estimates. Superscripts a, b, and c denote statistical significance at the 1 percent, 5 percent, and 10 percent level, respectively. of reestimating Equation (2) while allowing all coefficients to vary across two-digit groups.1 The 3 Table’s results confirm that investment rates Strauss (1989, 1991), which concludes that vary negatively with the value of the dollar and investment is more cyclical, relative to its that this effect varies across two-digit groups. mean, than output. Consider first the coefficients on A PGNP; most The coefficients on R7GMA are significant are positive, as expected. Thus investment is and have the expected negative signs. Thus procyclical, and the degree of procyclicality increases in the value of the dollar are associat varies across industries. Of the five two-digit ed with declines in investment rates in U.S. groups with negative coefficients, only textiles manufacturing. The magnitude of the effect (SIC 22) and rubber (SIC 30) have significant suggests that investment is fairly responsive to coefficients. changes in the value of the dollar. For the full The coefficients on R7GMA are a bit more sample, a 1 percent increase in R7GMA is TABLE 2 6 ECONOMIC PERSPECTIVES TABLE 3 Industry investment rates, GNP, and real exchange rates Dependent variable: I K t, 1963-1986 Prais-Winsten estimates Industry CONSTANT A_PGNPt R7GMA, 20 Food .158“ (.011) .066b (.032) -.107“ (.028) 21 Tobacco .163“ (.032) .200b (.091) -.236“ (.078) 22 Textiles 151a (.015) 115a (.043) .066° (.037) 23 Clothing .065“ (.015) .075“ (.042) -.005 (.037) 24 Lumber .121“ (.033) .059 (.078) -.055 (.067) 25 Furniture .098“ (.035) .300“ (.074) -.216“ (.057) 26 Paper .142“ (.026) -.035 (.059) .018 (.046) 27 Publishing .126“ (.026) -.055 (.065) .082 (.054) 28 Chemicals .141“ (.022) .056 (.061) -.074 (.048) 29 Petroleum refining .108“ (.035) .458“ (.084) -.388“ (.069) 30 Rubber .240“ (.040) -.300“ (.086) ,142b (.070) 31 Leather .100 (.069) .142 (.166) -.124 (.111) 32 Stone, clay, glass .065“ (.012) .243“ (.029) -.177“ (.025) 33 Primary metals .098“ (.016) .083c (.046) -.055 (.039) 34 Metal products .110“ (.016) .202“ (.043) -.167“ (.036) 35 Industrial equipment .155“ (.012) .127“ (.035) -.146“ (.031) 36 Electronic equipment .122“ (.015) .187“ (.040) -.145“ (.034) .149“ (.020) .101c (.055) -,104b (.049) 38 Instruments .245“ (.040) -.126 (.137) -.007 (.098) 39 Miscellaneous .163“ (.038) .014 (.078) -.050 (.057) 37 Transportation equipment Notes: l_Klt is the investment rate for industry /' in year t, A _ P G N P t is the ratio of actual to potential GNP at time t, and R 7 G M A : is the real trade-weighted dollar index at time t. The table reports the results of the PW regression of l_ K on A _ P G N P and R 7 G M A , while permitting all coefficients to vary over two-digit SIC groups. Reported coefficients are the total effect for the given two-digit group. Standard errors are in parentheses under coefficient estimates. Superscripts a, b, and c denote statistical significance at the 1 percent, 5 percent and 10 percent level, respectively. FEDERAL RESERVE BANK OF CHICAGO varied in sign and magnitude. For 16 of the 20 two-digit groups, R7GMA enters with a negative sign, as expected; 9 of these 16 coefficients differ significantly from 0. The coefficients are largest for two-digit groups 29 (petroleum), 21 (tobacco), 25 (furniture), 32 (stone, clay, and glass), and 34 (metal products).1 4 These results appear generally consistent with those of Branson and Love (1988), who find that the real exchange rate has its greatest effects on employment in the two digit groups 33 (primary metal), 35 (industrial equipment), 34, 29, 32, and 39 (miscellaneous).1 Again, 5 textiles and rubber are the only groups whose coefficients are significant and the wrong sign. The textiles industry enjoyed substantial import protection during the time period covered by this study, so the industry’s investment spending may not have been likely to respond in the expected way to the appreciation of the dollar. Finally, as indicated earlier, it is likely that an industry’s exposure to international markets influences the size of its investment responsiveness to exchange rate changes. One measure of that exposure, the industry import-sales (IMS) ratio, is available for 173 of the sample’s 270 industries over the 1965-1980 time period. Because of this limited availability, I computed each industry’s average IMS over the available time period and then grouped industries into high IMS and low IMS industries, comparing industry averages to the overall average. I then re-estimated Equation (2) over the 173-industry sample and separately over the high and low IMS industries, respectively.1 The 6 results appear in Table 4. In brief, the coefficient on R7GMA is larger in the high IMS industries, and an 7 Sum m ary and conclusions TABLE 4 Investment rates and real exchange rates Dependent variable: I_Kjt, 1963-1986 CONSTANT A_PGNPt R7GMA, R2 OLS -.165a (.036) .373a (.038) -.078a (.010) .03 PW -.1 19a (.037) .326a (.038) o o 00 «o o f — F test rejects the null hypothesis of pooling of high and low IMS industries, showing that this difference is statistically significant. So, higher IMS ratios are associated with larger investment responses to exchange rate fluctuations. This is reasonable, because industries experiencing substantial foreign competition at home are likely to be sensitive to exchange rate fluctua tions. Coefficients: All industries with IMS data (173) High IMS industries (49) ECONOMIC PERSPECTIVES CO CO O 00 o 8 CO In this article, I presented OLS -.094 .322a —102a . .03 (.075) (.079) (.020) evidence that fixed investment rates are sensitive to changes in the value PW -.074 -.101a (.081) (.023) of the dollar. Investment responds more in durable goods industries Low IMS than in nondurable goods industries, industries (124) but there appears to be little differ OLS -.193a .393a -.069a .03 (.041) (.043) (.011) ence between consumer goods and producer goods industries. Further, PW .337a -.138a -.071a investment is more sensitive to (.041) (.041) (.012) exchange rate fluctuations for NOTES: l_Kit is the investment rate for industry / in year f, IMS is the industries experiencing substantial import-sales ratio, A_PGNPt is the ratio of actual to potential GNP at time t, and R7GMAf is the real trade-weighted dollar index at time t. OLS refers to foreign competition. the ordinary least squares estimates, and PW refers to the Prais-Winsten estimates, which correct for first order serial correlation. Standard errors Some readers may be surprised are in parentheses under coefficient estimates. Superscripts a, b, and c at investment’s responsiveness to denote statistical significance at the 1 percent, 5 percent, and 10 percent level, respectively. relative price changes, given the limited role for relative factor prices in much recent research on invest highly valued dollar of the 1980s may have led ment spending. To what extent might industries to some long run deterioration in the ability of absorb exchange rate fluctuations into their U.S. industries to compete in international price-cost margins (PCMs) instead of their input markets. Between 1978 and 1985, the dollar demands? In fact, in a related, unpublished index rose from .856 to 1.149, an appreciation of analysis of industry PCMs, I found that this 34 percent. Table 2’s coefficient estimates price adjustment effect is present in the data: imply that the average industry’s investment rate when the dollar appreciates, domestic PCMs fall, was .021 lower in 1985 than it would have been especially so in durable goods industries. So it in the absence of the dollar’s appreciation. The appears that as the relative price of domestic raw investment data for the sample show that goods changes, U.S. industries respond by total investment spending in 1985 was $61.8 changing both the price and quantity of output billion. Combining this figure with appropriate (hence inputs like capital). Developing struc capital stock figures and Table 2’s estimates, I tural models that can distinguish these two sets estimate that investment spending in 1985 was of exchange rate effects is an important area for $11.3 billion less than it would have been had future research. the dollar not appreciated. The decline in the Finally, although my results should be dollar in recent years, though not examined viewed as suggestive only, they do indicate the directly in this article, may have reversed this potential importance of exchange rate move trend, thus enabling U.S. industries to effectively ments for the future international competitive compete at home and abroad. ness of U.S. manufacturing industries. The FOOTNOTES 'Note that strict application of the “purchasing power parity” argument implies that e - 1, that is, that changes in E are simultaneously offset by changes in relative prices. Consequently, for these arguments to be correct, some sort of price stickiness must prevent parity from being reached. 2In other words, I estimate a “reduced form” relationship between investment and real exchange rates. For an example of analysis of a structural relationship between input demands and real exchange rates, see Krieger (1989), who argues that real exchange rate changes affect factor demands through two channels. The first is the one discussed above: an increase in the value of the dollar causes an increase in the relative price of U.S. goods, thus a decrease in aggregate derived factor demands. The second channel involves the sectoral reallocation of resources that follows an exchange rate shock, regardless of whether the shock is positive or negative. The two channels are not mutually exclusive, and distinguishing between the two requires a structural model. 9Using fourth quarter values made no qualitative and minor quantitative difference to the analysis. l0See Judge et al. (1985), p. 286. Ordinary least squares (OLS) estimation is appropriate under the following assumptions: (3a) E(ejt) = 0 , (3b) E ($ = a 2, (3c) E(eu = 0 for i * j , and ejt) (3d) E(e.tejs) = 0 for t * s . Preliminary analysis suggested that first order serial correlation was significant and that the autoregressive parameter differed across four-digit industries, so I permitted the parameter to vary in the estimation. This amounts to replacing assumption (3d) by E (t.ji.J = p., so that e is assumed to follow the autoregressive process e i, = P ie i,-, + u i,’ 3 example, see Branson and Love (1988) and Krieger For (1989). 4 The SIC system assigns all manufacturing establishments into categories based on the primary activities at the establishments, and its most often used categories are the two- and four-digit groupings. Two-digit numbers are used to denote major groups, such as SIC 20, which is Food and Kindred Products, while four-digit numbers correspond to more narrowly defined categories, such as SIC 2011, which is Meat Packing Plants. industries in two-digit SIC groups 24, 25, or 32-38 were labeled durable goods industries; others were placed in the nondurables group. See Petersen and Strauss (1991). 6For each of the three variables reported in Table 1 ,1 can reject at the 5 percent level the hypothesis that the mean is the same for durable and nondurable goods industries. 7Using versions of the index based on lagged (as opposed to current) prices led to results similar to those reported below. 8See Hervey and Strauss (1987a) for a discussion of the appropriate price index to use when constructing a real exchange rate. Branson and Love (1988) report that using producer price indexes or more general price indexes made little difference to their ranking of industries in terms of their output and employment elasticities with respect to the real exchange rate. FEDERAL RESERVE BANK OF CHICAGO where I assume that u is a mean zero, variance a 2 random variable with no serial or contemporaneous correlation. 1'The reader will notice the low R 2 values in the Table. Low R 2s are common in pooled time-series cross-sectional analyses. Estimating a pure time series version of (2), so that the dependent variable is /_K , yields an R 2 of .47, with coefficient estimates identical to those in the first line of Table 2. The Durbin-Watson statistic is .98. l2The classification is taken from DHP (1987). l3Only the Prais-Winsten estimates are presented. Specifications that restricted all slope coefficients to be equal across two-digit groups were rejected by F tests at the 1 percent level. Further, specifications that restricted the coefficients on A P G N P (R 7 G M A ) while permitting those on R 7 G M A (A P G N P ) to vary were also rejected. l4It is possible to compute an elasticity of the investment rate with respect to the dollar index, but the measure is difficult to interpret. I choose to focus on the absolute coefficient estimates themselves. l5Branson and Love (1988) obtain similar results when analyzing industrial production’s response to real exchange rate changes. l6This procedure is strictly appropriate only if the variable used to group industries, here the IMS ratio, is exogenous. 9 REFERENCES Branson, W.H., and J.P. Love, “The real exchange rate, employment, and output in manufacturing in the U.S. and Japan,” National Bureau of Economic Research, Working Paper 2491, January 1988. Domowitz, I., R.G. Hubbard, and B.C. Petersen, “Oligopoly supergames: some empirical evidence on prices and margins,” Journal of Industrial Economics, Vol. 35, No. 4, June 1987, pp. 379-398. Hervey, J.L., and W.A. Strauss, “The interna tional value of the dollar: an inflation-adjusted index,” Economic Perspectives, Vol 2, No. 1, January/February 1987a, pp. 17-28. Hervey, J.L., and W.A. Strauss, “Technical correction: the inflation-adjusted index of the dollar,” Economic Perspectives, Vol. 2, No. 2, March/April 1987b, pp. 29-31. Hervey, J.L., and W.A. Strauss, “The new dollar indexes are no different from the old ones,” Economic Perspectives, Vol. 2, No. 4, July/August 1987c, pp. 3-22. Hulten, C.R., and F.C. Wykoff, “The measure ment of economic depreciation,” In Hulten, C. R., ed., Depreciation, Inflation, and the Taxation of Income from Capital, Washington, D. C., Urban Institute Press, 1981. 10 Judge, G.G., W.E. Griffiths, R.C. Hill, H. Lutkepohl, and T.-C. Lee, The theory and practice of econometrics, 2nd ed, New York, John Wiley and Sons, 1985. Krieger, R., “Real exchange rates, sectoral shifts, and aggregate unemployment,” Federal Reserve Board of Governors, Finance and Economics Discussion Series, Working Paper 92, September 1989. Petersen, B.C., and W.A. Strauss, “Investment cyclicality in manufacturing industries,” Economic Perspectives, Vol. 13, No. 6, November/December 1989, pp. 19-28. Petersen, B.C., and W.A. Strauss, “The cyclicality of cash flow and investment in U.S. manufacturing,” Economic Perspectives, Vol. 15, No. 1, January/February 1991, pp. 9-19. Shapiro, M.D., “The dynamic demand for capital and labor,” Quarterly Journal of Eco nomics, Vol. 51, No. 3, 1986, pp. 513-542. U.S. Bureau of the Census, Annual Survey of Manufactures, various years. U.S. Bureau of the Census, Census of Manu factures, various years. ECONOMIC PERSPECTIVES Productive efficiency in banking Douglas D. E van off and Philip R. Israilevich Then a new CEO came in who asked, ... “What do we have to produce by way of results?" Every one of his store manag ers knew the answer, “We have to increase the amount each shopper spends per visit.” Then he asked, “Do any of our stores actually do this?" Three or four— out of 30 or so—did it. “Will you then tell us,” the new CEO asked, “what your people do that gives you the desired results?”' In tro d u c tio n In the above epigraph the managers are attempting to identify, in a particular context, the firms which are doing the best job of ac complishing the company objectives. Such firms are known as the best practice firms. Economists typically make similar inquiries concerning the production process. They address the issue by theoretically defining the best practice firm, empirically identifying it, determining its resource utilization, and then evaluating how others compare to it. More generally, economists, like the new CEO, are concerned with productive efficiency. Because of changes taking place in the banking industry, the importance of efficiency has increased substantially. As geographic and product deregulation occurs, the resulting increase in competition should place banks in a situation where their success will depend on their ability to adapt and operate efficiently in the new environment. Banks unable to do so will have difficulty surviving. FEDERAL RESERVE RANK OF CHICAGO Most studies of bank efficiency have concentrated on cost advantages resulting from the scale of production. In fact, this is probably one of the most researched topics in banking.2 There are, however, other aspects of efficiency which students of the industry have just begun to evaluate. For example, do the producers of banking services effectively combine their productive inputs? Once employed, do they use the inputs effectively? If not, how ineffi cient are they? What allows them to continue to do this and stay in business? Given its im portance in the deregulated environment, it is imperative that the various aspects of bank efficiency be understood and empirically analyzed. In this article we discuss the concept of efficiency in production, define its various aspects and the means to measure it, and review the relevant literature concerning inefficiency in the banking industry. Our major conclusion is that there appears to be significant inefficien cy in banking. Inefficiency resulting from operating at an inappropriate scale of operation is probably in the range of 10-20 percent of costs. However, by emphasizing the role of scale, researchers have essentially overlooked a major portion of bank inefficiency. The evi dence suggests that inefficiencies resulting The authors are economists at the Federal Reserve Bank of Chicago. Helpful com m ents on earlier drafts by Herb Baer, Paul Bauer, Allen Berger, Dave Humphrey, Curt Hunter, Carl Pasurka, and Sherrill Shaffer are gratefully acknowledged. The views expressed, however, are those of the authors and are not necessarily shared by others. 11 from the suboptimal utilization of inputs is larger than that resulting from other factors. According to a majority of studies, banks operate relatively efficiently with respect to the optimal combination of inputs, yet many are very inefficient in converting these inputs into outputs. This inefficient utilization of inputs accounts for an additional 20-30 percent of costs. This is particularly interesting because it implies that, to a great extent, the future viabili ty of an individual bank is under its own con trol. To the extent that bank inefficiency can be accurately measured, it appears that the largest inefficiencies are not the result of regulation or technology, but result directly from an under utilization of factor inputs by bank manage ment. This inefficiency will most likely decline in the future as bankers respond to increased competitive pressures and strive to become more efficient. Failing this, the inefficient firms will become prime merger candidates to be acquired and restructured. The article proceeds as follows. In the next section we define, discuss, and illustrate the components of production efficiency. We then evaluate the alternative means to generate measures of efficiency. A review of the litera ture on bank efficiency is then presented. The final section summarizes and evaluates policy concerns. We have also included an extensive reference list for readers interested in more detailed analysis of productive efficiency. P ro d u ctio n e ffic ie n c y The economic theory of the firm assumes that production takes place in an environment in which managers attempt to maximize profits by operating in the most efficient manner possible. The competitive model suggests that firms which fail to do so will be driven from the market by more efficient ones. However, when natural entry barriers or regulation weak en competitive forces, inefficient firms may continue to prosper. That is, true firm behav ior may vary from that implied by the competi tive model as managers attempt to maximize their own well-being instead of profits, or find that they are not required to operate very efficiently to remain in business. Variations from productive efficiency can be broken down into input and output induced inefficiencies. By input inefficiency we mean that, for a given level of output, the firm is not optimally using the factors of production . 12 Overall input inefficiency resulting from the suboptimal use of inputs can be decomposed into allocative and pure technical inefficiency. Allocative inefficiency occurs when inputs are combined in sub-optimal proportions. Regula tion is typically given as a major reason for this occurrence. Pure technical inefficiency occurs when more of each input is used than should be required to produce a given level of output. This occurrence is more difficult to explain, but is typically attributed to weak competitive forces which allow management to “get away” with slackened productivity. Combining these two notions of inefficiency we get the overall inefficiency resulting from the improper use of inputs.3 The distinction between the two types of inefficiency is important because they may be caused by totally different forces. Productive efficiency requires optimizing behavior with respect to outputs as well as inputs. With respect to outputs, optimal behav ior necessitates production of the level and combination of outputs corresponding to the lowest per unit cost production process. An optimal output level is possible if economies and diseconomies of scale exist at different output levels. Economies of scale exist if, over a given range of output, per unit costs decline as output increases. Increases in per unit cost correspond to decreasing returns to scale. A scale efficient firm will produce where there are constant returns to scale; that is, changes in output result in proportional changes in costs. Because it involves the choice of an inefficient level, scale inefficiency is considered a form of technical inefficiency. Thus total technical inefficiency includes both pure technical and scale inefficiency; that is, inefficient levels of both inputs and outputs. Additional cost advantages may result from producing more than one product. For example, a firm may be able to jointly produce two or more outputs more cheaply than producing them separately. If the cost of joint production is less than the cost resulting from independent production processes, economies of scope are said to exist. Diseconomies of scope exist if the joint production costs are actually higher than specialized or stand-alone production of the individual products. A final point should be mentioned concern ing the various categories of inefficiency. Pure technical inefficiency is entirely under the control of, and results directly because of, the ECONOMIC PERSPECTIVES behavior of the producer. Output inefficiency and allocative inefficiency may, from the perspective of the firm, be unavoidable. For example, a firm optimally using factor inputs may find that per unit cost declines over the entire range of market demand. While increas ing production would generate cost savings or efficiencies, the characteristics of market demand may not justify it. Failure to exploit scope advantages may also result from factors outside of the control of the firm. In banking, the array of allowable activities is obviously constrained by regulation. This may preclude potential gains from the joint production of various financial services. Finally, as men tioned earlier, allocative inefficiency may occur as a direct result of regulation. For example, during the 1970s, banks were restricted with respect to the explicit rates they could pay depositors. As market rates rose above allow able levels, banks frequently substituted implic it interest payments in the form of improved service levels; for example, more offices per capita or per area, see Evanoff (1988). This resulted in an over-utilization of physical capital relative to other factor inputs. In this case, regulation was the driving force behind the resulting allocative inefficiency. The point is that much inefficiency may be beyond the control of the individual firm. In the following sections we illustrate the inefficiencies described above and discuss alternative methods used to empirically capture them. The reader who is most interested in an analysis of efficiency in banking may skip directly to the section entitled “The role of production inefficiency in banking: A survey of the literature.” Illu s tra tin g in p u t e ffic ie n c y The notions of input inefficiencies can be illustrated as shown in Figure 1. Assume that jc, and x2 are two factor inputs required to produce a single output, y. Isoquant /-/’ depicts various efficient combinations of the two inputs which can be used to produce a specific level of output, y r Isoquants further to the right corre spond to higher levels of output, those to the left to lower levels of output. For example, the output level associated with isoquant //-//’ is less than y r For a given set of input prices, the isocost line, P-P\ represents the various combi nations of inputs which generate the same level of expenditures. Isocost lines further to the right FEDERAL RESERVE BANK OF CHICAGO correspond to higher level of expenditures on inputs. The slope of the isocost line is, obvi ously, determined by input prices. If the objective of the producer is to pro duce a particular level of output at minimum cost then the optimal input combination in Figure 1 is at point E. That is, given factor prices, output y ] can be optimally produced by employing x\ units of input x { and xe units of 2 input x2 Any other combination of the inputs . along the P-P’ isocost line would generate less output for the same cost. For example, the input combinations corresponding to points W or Z would result in similar expenditures on inputs, but generate the lower level of output associated with isoquant //-//’. Alternatively, the production of y ] using any combination of inputs other than that corresponding to point E would cost more. Therefore, at point E, input efficiency exists.4 To illustrate input inefficiency, suppose that the observed combination of inputs used by a particular firm to produce y, is at point A in Figure 1. We know that inefficiency exists because E was shown above to correspond to the most efficient combination of inputs to produce y,. Comparing the input utilization at point A to that at E we can derive the level of inefficiency resulting from the suboptimal use of inputs. In order to illustrate allocative and pure technical inefficiency, we have drawn a line from the origin to point A. Along this line different levels of factor inputs are employed but the ratio between the two inputs is fixed at 13 the actual ratio (that is, the ratio at point A). Reference points along this line and on isoquant /-/’ and isocost line P-P' are highlighted. Consider allocative inefficiency first. Point C represents a level of costs equal to that of the efficient production process at point E because it is on line P-P'. Point B corresponds to an output level equal to y, because it is on isoquant IP . Therefore, the distance CB corresponds to additional production expenses resulting from the suboptimal allocation of inputs. That is, allocative inefficiency exists because we are not on the isocost line, P-P'. Formally, OC/OB is a measure of allocative efficiency. Values less than 1.0 reflect inefficiency.5 For this same example, we can also depict pure technical inefficiency resulting from producing at point A. We have seen that producing y, using .v“ and x“ involves allocative inefficiency because point A is to the right of line P-P' and ray 0A does not go through point E. However, there is additional inefficiency because point A is above isoquant IP . That is, the combination of inputs associated with point A should enable the firm to produce a level of output greater than y,. (It should be able to produce output y3 corresponding to isoquant //////’.) Given that the isocost line depicts total expenditures used in production, distance CA constitutes a less than optimal usage of all inputs and corresponds to additional production expenses. Therefore, overall input inefficiency is measured as OC/OA. Because OC/OB is attributed to allocative inefficiency, the remain ing portion, OB/OA, can be attributed to pure technical inefficiency. Since these are radial measures, overall input inefficiency is the product of the two subcomponents, that is, OC/OA = (OC/OB) ■(OB/OA). The pure technical inefficiency shown in Figure 1 can also be illustrated in terms of output, instead of input, using a total output or total product relationship as depicted in Figure 2. The ratio of input usage, x jx 2, is held fixed by assumption in Figure 2 to represent input combinations along the ray OA in Figure 1. Since the fixed input ratio precludes the analy sis of allocative efficiency, we are analyzing only pure technical efficiency. Because changes in inputs result in proportional changes in output (the total product curve is linear) we have constant returns to scale as was assumed in Figure 1. Employing x\ units of input jc, we 14 FIGURE 2 Pure technical efficiency measured in terms of outputs output could produce an output level y, //the inputs were fully utilized. This corresponds to point B in Figure 1. Similarly, using x( units of input jc( we should be able to produce yr Again, this corresponds to point A in Figure 1. However, if inputs are not used effectively, that is if techni cal inefficiency exists, the resulting production point will be below the total product curve. That is, pure technical inefficiency occurs when we operate beneath the total product relation ship. For example, the pure technical ineffi ciency depicted in Figure 1 corresponds to that found at point G in Figure 2, where inputs are under-utilized and x( only generates an output level of yr If we are producing y, at point A in Figure 1 or, equivalently, at point G in Figure 2, pure technical inefficiency is measured with respect to inputs as OB/OA and with respect to outputs as AG/AM. The inefficiency measures are equivalent. This illustration is important because it indicates that technical inefficiency can be measured in terms of either inputs or outputs. Below we drop the constant returns to scale assumption and expand on this output inefficiency measure. Illu s tra tin g o u tp u t e ffic ie n c y Point E in Figure 1 corresponds to the least cost, most efficient means to produce y,. How ever, because of particular characteristics of the production technology, this level of output may not be the optimal one to produce. For exam ple, it may be that over a certain range of outputs, economies of scale exist. Production ECONOMIC PERSPECTIVES efficiency, therefore, requires optimal decisions concerning both input and output levels. In Figure 3 we have dropped the assumption of constant returns to scale. The production process is now characterized by increasing returns up to point R, constant returns at R, and decreasing returns at output levels above R. Now the firm corresponding to point G in Figure 3 is technically inefficient for two reasons. First, there is pure technical ineffi ciency resulting from the under-utilization of inputs; that is, we are beneath the total product curve. If inputs are fully utilized, input jc“ should produce the higher output level corre sponding to point A/, that is, yy Second, we have decreasing returns to scale at the current level of output since the production process is not represented as the linear relationship OH. The output not produced because of scale inefficiency can be measured as HM. This output is what could have been produced if inputs were used efficiently and constant returns to scale existed at this output level. Therefore, for the input usage depicted at point A, the input efficient firm could produce at point M, and the input and scale efficient firm could produce at point H. As explained above, scale inefficiency is generally considered a form of technical inefficiency because it in volves the choice of an inefficient level. Thus, total technical inefficiency includes pure technical and scale inefficiency; that is, ineffi ciency in the use of both inputs and outputs. Figure 4 depicts the reference points just discussed in Figure 3 in terms of production cost. The total product relationship in Figure 3 corresponds to the average cost relationship depicted in Figure 4. Points H and R each correspond to constant returns to scale and, therefore, correspond to minimum points on average cost relationships. Total technical inefficiency can be depicted here as the ratio of the average costs. For the example just dis cussed, total technical inefficiency is equal to ACJACg. The alternative measures of ineffi ciency illustrated in Figures 1 through 4 are equivalent and correspond to the alternative means of calculating inefficiency estimates commonly cited in the literature. In the above discussion we assumed the production of a single output for illustrative purposes. Additional cost advantages may result from multiproduct production. For example, economies may exist for the joint production of two or more outputs, relative to the stand-alone production of the individual products. That is, scope advantages may exist. More formally, economies of scope exist in the joint production of (2, and Q, if (1) [C, + CJ > C|2 where C, and C, are the cost of producing Q] and Q, independently (that is, as stand-alone processes), and Cj2 is the cost of joint produc tion. With multiproduct production, some FEDERAL RESERVE RANK OF CHICAGO 15 fixed cost of production can be spread across the outputs and there may be synergies when the two products are produced jointly. A multiproduct cost relationship which exhibits production synergies between the two outputs, y, and y2, is illustrated in Figure 5. Joint pro duction moves the cost off the “lip” of the relationship onto the inner surface. Potential cost gains obviously exist. M easu rin g p ro d u c tio n in e ffic ie n c ie s The relationships depicted in the above figures, as well as all standard textbook presen tations of the production process, present extreme values; that is, the maximum output that can be produced from a given set of inputs, or the minimum cost required to produce a given level of output. However, when attempts are made to generate estimates of the produc tion process we typically abstract from the extreme values. The traditional approach to evaluating the production process is to assume the standard competitive model is appropriate and to estimate an average production, cost, or profit function.6 Realizing that this restrictive model may not adequately describe the produc tion process (and definitely avoids efficiency 16 issues), methods have been developed which allow for variations in this approach. We discuss these variations in this section. The methodologies differ from each other in a number of ways, not the least of which is a result of differences in assumptions imposed in the analysis. The restrictiveness of these assumptions is determined by the individual data sets. Each of the methods discussed here is superior to the basic competitive model as long as the assumptions employed are correct. More will be said about this later. While the concept of firm efficiency is rather straightforward, various difficulties are encountered when attempting to measure it. Essentially, one needs to derive the best prac tice firm, or the production frontier which depicts the maximum performance possible by firms, and contrast existing firms to this stan dard. Ideally, we would compare firm perfor mance to the true frontier, however, the best that can be achieved is an empirical frontier or best practice firm generated from the observed data. Once the best practice firm is established, input related pure technical and allocative efficiency, and output related scale and scope efficiency, can be analyzed. For example, assuming constant returns to scale in Figure 1, all firms can be compared to one producing at point E. Differences in estimates of firm efficiency typically result from different means of gener ating the best practice firm. There are two general approaches used to model this relation ship. First, the parametric or econometric approach employs statistical methods to esti mate an efficient cost frontier. Second, the nonparametric or deterministic approach is based on the linear programming approach for optimal allocation of resources called data envelopment analysis (DEA). This technique is used to directly generate individual frontiers for each firm. Below we discuss alternative meth odologies within these two broad categories. It should be emphasized that empirical measures of inefficiency are no different from estimated parameters in any economic model. The model may mistakenly reflect measure ment errors or specification errors for produc tive inefficiency. As the literature on banking develops, more comprehensive models should be analyzed. ECONOMIC PERSPECTIVES P a ra m e tric approach: S h a d o w price m odels To generate estimates of allocative effi ciency, one can use the parametric approach developed by Lau and Yotopoulos (1971) and refined by Atkinson and Halvorsen (1980, 1984).7 This method assumes that firms are combining the factor inputs correctly, but that the combination is not necessarily based on observed prices. Rather, there are factors in addition to explicit market prices which enter the firm’s employment decision process. These additional factors are combined with the explic it prices to generate shadow prices which are more comprehensive and which determine factor utilization. These additional factors typically include distortions induced by union ism, regulation, or managerial goals other than profit maximization. These alternative goals may include profit satisficing or expense preference behavior.8 More formally, a basic contention of economic theory is that, in competitive mar kets, the optimal level of employment for each factor of production can be determined by employing additional units until the last dollar spent on each factor yields the same amount of productivity. That is, f f (2) — — , for i * j = 1,..., m, = P> Pj where/ = 8 // 8X. is the marginal product of input j, and P is the price of input /, or f P j (3) (4) J__!l , for i * j = 1,... , m, fJ PJ where P* is the effective or shadow price of input /, and the marginal rate of technical substitution between the inputs is set equal to the ratio of the shadow prices of the inputs. Given competitive markets and the absence of additional binding constraints, shadow and actual prices are equal and the employment decision is not affected. Because the shadow prices of the inputs are not directly observable, Lau and Yotopolous developed a means to estimate them along with other parameters of the cost relationship. Assuming shadow prices are proportional to market prices, shadow prices can be approxi mated by (5) P* = k P , for i = 1,... , m, where k. is input-specific.’ Again, if the addi tional constraints are not binding, all shadow prices equal the respective market prices, that is, k. = 1 for all i. Standard econometric techniques can then be used to generate cost estimates employing the additional information. That is, the stan dard cost structure (6) C = C(P, Q, Z), where C depicts costs, Q outputs, P explicit factor prices, and Z additional pertinent exoge nous variables, is replaced with j , for i * j = 1,... , m, where / / / i s the marginal rate of technical substitution between the inputs. This relation ship corresponds to the tangency of the isoquant and the isocost curve (point E) in Figure 1. Given input prices and the predetermined level of output as the only constraint, the optimal combination of inputs, as in Equation (3), can be derived to minimize cost. However, if additional constraints exist (for example, regulatory constraints), they need to be ac counted for and incorporated into the optimiza tion process. Concerning the employment decision, Equation (3) becomes FEDERAL RESERVE BANK OF CHICAGO (7) Cs = Cs(kP, Q, Z), where kP denotes shadow factor prices, and k is estimated along with the other parameters in the cost function.1 0 The shadow price model also allows one to calculate the optimal (unobserved) input combi nation given observed prices, P. This combina tion is relevant for measuring the cost differ ences resulting from production under competi tive conditions and those when additional binding constraints exist. In the banking industry, these additional constraints are typi cally thought to be regulatory induced. The cost differences can be determined by contrast ing costs when market prices equal shadow 17 prices (k = l ) to that found using the estimated shadow prices (k = k where k denotes the estimated value for k). The difference between the two cost values will be the result of combin ing inputs in a suboptimal manner. Estimation of the cost function will yield k values which can be considered to reflect the effect of binding constraints on average. Ideal ly, the k. measure would be firm specific. However, statistical problems typically make this prohibitive in terms of the degrees of freedom required for the estimation procedure. All the parametric approaches cited below have this same shortcoming. Some progress toward resolving this shortcoming has recently been made; see Evanoff and Israilevich (1991, 1990a). One of the advantages of the shadow price model approach is that it allows for the estima tion of returns to scale and scope along with allocative efficiency. However, pure technical efficiencies can not be measured by this ap proach although, as discussed later, this short coming can also be partially resolved. P a ra m e tric approach: S to ch a stic co s t fro n tie rs Another more comprehensive parametric approach for measuring efficiency is to use stochastic frontier models. With this approach, the cost frontier is empirically estimated and firm specific deviations from the frontier are attributed to productive inefficiencies. A number of alternative parametric techniques can be used to generate the frontier. The major difference between these techniques is in the maintained assumptions which, obviously, can produce significantly different results. The restrictiveness of these assumptions is deter mined by the individual data sets. Here we summarize alternative parametric methods used to develop the frontier. Using a parametric approach, the standard cost structure is typically generated by impos ing a specific functional form on the data and obtaining the “best fit” by minimizing devia tions from the estimated structure. For exam ple, the estimated total cost relationship may be fitted to the data to produce a relationship such as TC in Figure 6. However, when evaluating efficiency, we are interested in the best practice firm or the cost frontier. We are not interested in the average relationship, rather we are looking for a minima in the data. Therefore, 18 FIGURE 6 Total cost relationship TC adjustments to the standard estimation proce dures are required. Typically the standard parametric procedure is adjusted by employing a more complex error structure. A “composed” error can be used which consists of two compo nents: one is the standard statistical noise which is randomly distributed about the rela tionship, and the other consists entirely of positive deviations from the cost structure (that is, a one-sided disturbance term) and represents inefficiency.* Stated crudely, the resulting 1 * 9 8 1 frontier is simply a transformation of TC in Figure 6 (shifted downward) to generate the best cost relationship instead of the average relationship. For example, and more formally, assume a stochastic frontier model which consists of the following cost and share equations: (8) In C* = In q + lnTh+ InA + uh; (9) M* = VF + b.i + u„’ for i = 1,..., m; v 7 th ih ih where In denotes the natural log, and CAand MA h are observed cost and factor shares for firm h. CFis the lowest production cost relationship or h the cost frontier, lnTh reflects additions to cost resulting from pure technical inefficiency, InA reflects additions to cost resulting from alloca tive inefficiency, and uh is a random error. M F is the efficient share equation, /? depicts share distortions resulting from allocative inefficien cy, u.h captures random distortions from effi- ECONOMIC PERSPECTIVES cient shares, and B h is the composed error term. Measures of technical inefficiency are calculated as firm specific deviations from the frontier and are derived from the additional error term discussed above. Since technical inefficiency can result only in increases to total cost, this error structure must consist entirely of non-negative values. That is, this component of the error structure is one-sided relative to the frontier. Choice of a specific one-sided distri bution could obviously influence the empirical results.1 2 As with the shadow price model, allocative inefficiency is computed as an average for the sample and is not firm specific. InA is non negative as deviations from use of the optimal combination of inputs can lead only to addi tions to cost. However, b. can be positive or negative suggesting over- or under-utilization of a particular input. Obviously, InA and b. are related because suboptimal combinations of factor inputs (b ^ 0) result in additions to cost. However, empiri cally modeling this relationship is problematic. One standard means to do it is to impose restrictions on the relationship reflecting prior knowledge. For example, assuming increased costs occur only when mistakes are made (A = 0 only when /? = 0), and that large mistakes cost more than small ones, one can impose a relationship between allocative mistakes and cost increases:1 3 (10) In A = b’ F b; where F is a diagonal matrix with nonnegative elements. Positive elements of F represent weights for each br For example,/j. represents the relative effect of allocative distortions from factor / on the increased production costs. To summarize, the additional cost of allocative inefficiency is a weighted sum of squared mistakes from the misallocation of each input. The (nonnegative) weights are additional parameters to be estimated. An alternative approach to generate a cost frontier is to utilize a cost structure consisting of cost and share equations, but to sever the link between the error terms of the cost and share equations. That is, the share equations are used only for efficiency gains in parameter estimation; not to link suboptimal combinations of inputs to increases in cost. Under this FEDERAL RESERVE BANK OF CHICAGO approach, both allocative and technical ineffi ciencies are depicted as one-sided errors from the cost frontier. Therefore, the estimated system of Equations 8 and 9 becomes (11) lnCA= l n C F+ vh+ uh, (12) MA = MF + u., for i = 1,..., m, where the error term depicting inefficiency, vh, can be decomposed into its two components (that is, InT + InA) using techniques developed by Kopp and Diewert (1982) and refined by Zieschang (1983). This approach essentially ignores information concerning the relationship between disturbances in the cost and share equations, but is easier to work with than the above approach and does not necessarily generate results inferior to more complicated linkage approaches. This is particularly true if the more complicated approach, which is typically based on a set of untested assump tions, incorrectly models the linkage. This attempt to simplify the methodology brings us to the most recent approach intro duced by Berger and Humphrey (1990). These authors take the view that the preceding meth odologies impose rather restrictive ad hoc assumptions concerning the data, the validity of which are questionable. For example, the assumed linkage between the error structure in the share cost equations, discussed above, could be inaccurate as could the assumptions con cerning the one-sided error distribution. To partially remedy these problems the authors developed a “thick frontier” approach. Instead of imposing restrictive characteristics on the cost relationship to generate a true frontier or frontier edge, a thick frontier is estimated from a subsample of the data which, based on a priori information, is considered to be an efficient subgroup. This group is then com pared to another group which, based on a priori information, is considered an inefficient sub group. Therefore the authors are able to relax the restrictive assumptions employed in the methodologies discussed above, but at the cost of using a somewhat ad hoc means to catego rize the data into efficient and inefficient groups. This approach was implemented using banking sector data by assuming subgroups could be delineated based on their average cost 19 per dollar of output. The data were then strati fied by size and divided into quartiles and the lowest and highest cost quartiles were contrast ed. After accounting for differences resulting from market characteristics, the remaining differences between the two groups were assumed to constitute inefficiency. This can be distributed into its allocative and technical components using procedures similar to those of Kopp and Diewert discussed above. Obvi ously, this approach lacks precision and also imposes some rather ad hoc assumptions to develop the subgroups and produce the frontier. However, the assumptions may be less restric tive than those made in the more elaborate models discussed above. In fact, some of the maintained assumptions in these models were statistically tested and rejected by Berger and Humphrey. As a result, the relatively easy-toimplement approach may perform quite well in generating a rough measure of the extent of production inefficiency in an industry.1 4 N o n p a ra m e tric approach While intuitively appealing, and somewhat similar to the procedures commonly used to estimate standard cost relationships, the para metric approaches have been criticized for requiring more information than is typically available for estimation of the cost frontier. In an attempt to decrease the required information, some have chosen to use a nonparametric, linear programming approach known as data envelopment analysis (DEA). Although there are various permutations to the DEA approach, the basic objective is to “envelop” the data by producing a piecewise linear fit via linear programming techniques. That is, instead of using regression techniques to fit a smooth relationship, a piecewise linear surface is produced which borders the observa tions, for example, the broken line qo in -qo Figure 7. The technique identifies observations for which the firm is producing a given level of output with the fewest inputs. These will be observations on the frontier. All other observa tions will be given an efficiency measure based on the distance from the frontier and indicating the extent to which inputs are being effectively utilized. This is comparable to the measure of pure technical inefficiency, OB/OA, for obser vation A in Figure 1. The technique allows for the derivation of a frontier for each firm in the sample based on 20 FIGURE 7 DEA measured efficiency the output and input utilization of all firms in the sample. As a simple example for the two input, one output case, the linear programming problem for technical inefficiency could be written as (13) Min 0 A , subject to qA< |f q1+ (I2 q2+ .... + jnnqn 0 A > in1x| + JJ.2 x!, + .... + fin x1 xA 0 A > ft' x^ + |i2x2+ .... + |lnx x* !J M > 0, .’ where 0* is the fraction of the actual inputs which could be used to optimally produce the given level of output, qA for observation A; , and x, are quantities of the two inputs; |Ts are the weights generated for each observation via the linear programming optimization process to obtain the optimal value for 0; A is the observa tion we are evaluating, and superscripts denote individual firms. Again, 0^ = OBIOA for firm A in Figure 1 or Figure 7. Therefore, we are finding the lowest fraction of the inputs used which would produce an output level at least as great as that actually produced by firm A. Additional linear programs can be solved to derive allocative inefficiency. A more complete description and an example of DEA analysis which has been applied to the banking industry is presented in the Box. ECONOMIC PERSPECTIVES Example of a data envelopment analysis (DEA) program applied to banking technical inefficiency, and then taking the difference between the two. To determine overall inefficiency, take the observed input prices vv4 faced by the bank A and assume cost minimizing behavior: n (2) M i n xA £ wA • xA j i= i J < X (ih • q \, i = 1 , ... , m , h=l Z h=I j = l , - s |T h • X h , J , n, = ’ S *’ r C C H > H < (1) Min J II cz> + Technical inefficiency is measured as the difference between the observed behavior o f bank A to that which would occur if bank A were on the production frontier. Therefore, the unobserved frontier must be projected. This is done via DEA analysis by developing a program which determines the minimum required amount of inputs necessary for bank A to produce as much, or more, of each of the outputs currently being produced. The input vector is chosen based on the observed behavior of the sample firms. Again, this reduces to a linear programming problem. For example, for bank A , the technically efficient combination of inputs is deter mined as Z |i h • z h , s 1 , h=l 0A , ... H s.t. i = 1, ..., m, qA< Z ph- q1, ; hl = ^ Z !th • Z* , ... , S, h=l H 0 A • XA> £ |Ih • x}, h= l j = 1 ,... , n, > 0 , h = 1 ,... , H , Z |T h = 1. h=l H s = 17,... , sr7 , 7 z * < Z M ’ Zs> -h h= l H s = sr+ ... , S, 1, zA> Z M ' Zs> -h h= l The optimization process determines the mini mum input vector, jc for the observed price vector 4* w4. Scalar ve4- jc is the minimum production cost 4* for the vector of outputs q A. Overall inefficiency for any firm, h , is therefore the ratio o f cost of the observed and the best practice bank:2 H ph> 0 , h = 1 ,... ,H , Z ph = 1, hl = where Q4 is our radial measure of technical efficiency for firm A , q . is the output vector, \ i h is a vector of weights assigned to each observation (an intensity vector) which determines the combination of tech nologies of each firm to form the production frontier, Jt*is the observed amount o f input j used by firm h, and z is a vector o f additional exogenous variables.1 There are two types of these exogenous variables; those that need to be maximized, z h for s = 1 ,... , s , s and those that should be minimized, z h for s = s + 1 , ..., S . An example of these exogenous variables in banking would be the number of branch offices. Banks would, c e t e r i s p a r i b u s , want to minimize the number of branch offices required to provide a given level o f output. The output o f each firm in the sample is weighted in such a way that the combina tion o f observed outputs, i, is not less than the output actually produced by firm A . Thus the frontier for firm A is constructed as a weighted technology from the sample. If O'4 = 1, then firm A is as efficient as any firm in the sample, that is, firm A is on the frontier. If Gf4 < 1 then firm A is inefficient. Allocative inefficiency for firm A can be derived by determining overall inefficiency and FEDERAL RESERVE BANK OF CHICAGO (3) Oh = (wh ■xh / ( w h • xh ) - l . ) * The difference between the costs o f technically efficient production and overall efficient production determines the cost resulting from allocative ineffi ciency. That is, A h = [(wA• 0 h* • xf1 / ( w h ■y 1 -1 is ) *)] an index o f allocative inefficiency for firm h , and & * is the optimal value of Q h determined in Equation ( l) .3 FOOTNOTES •The sum of the weights \ i h used in the optimization process is restricted to unity to allow for varying returns to scale. See Afriat (1972). The appropriate number of constraints for exogenous variables is difficult to determine and the estimated inefficiency for a given model typically varies inversely with the number chosen. 2The inequality in the linear program implies free disposability of both inputs and outputs. •Technical inefficiency, determined in Equation 1, ’ obviously is the difference between overall and allocative inefficiency: T h = O h - A h. 21 C o m p ariso n o f th e p a ra m e tric and n o n p a ra m e tric approaches The role o f p ro d u c tio n in e ffic ie n c y in b anking: A survey o f th e lite ra tu re Using either the parametric or DEA ap proach, the goal is to generate an accurate frontier. However the two methods use signifi cantly different approaches to achieve this objective. Because the parametric approach generates a stochastic cost frontier and the DEA approach generates a production frontier, and because the methodologies are fundamentally different, one should expect differences in the efficiency projections. Which methodology is preferable? There are advantages and disadvantages with each of the procedures. The parametric approach for generating cost relationships requires (accurate) information on factor prices and other exogenous variables, knowledge of the proper functional form of the frontier and the one-sided error structure (if used), and an adequate sample size to generate reliable statistical inferences. The DEA approach uses none of this information, therefore, less data is required, fewer assumptions have to be made, and a smaller sample can be utilized.1 Howev 5 er, statistical inferences cannot be made using the nonparametric approach. Another major difference is that the para metric approach includes a random error term around the frontier, while the DEA approach does not. Consequently, the DEA approach will account for the influence of factors such as regional factor price differences, regulatory differences, luck, bad data, extreme observa tions, etc., as inefficiency.1 Therefore, one 6 would expect the nonparametric approach to produce greater measured inefficiency.1 The 7 importance of this difference should not be understated because single outliers can signifi cantly influence the calculated efficiency measure for each firm using the DEA approach. Obviously, one would like to be able to take comfort in the fact that either approach generates similar results. This is more likely to occur if the sample analyzed has homogeneous units which utilize similar production process es. However, similar results have not been found in the literature. In fact, it is common for studies contrasting results produced from the two methodologies to find no correlation between the efficiency estimates. This has also occurred in studies of efficiency for the banking sector. We next review some of that literature. In this section we review the literature on productive efficiency for financial institutions. Most of the studies reviewed, particularly those analyzing input efficiency, were conducted recently and involve flexible functional forms and state of the art research techniques. For a more comprehensive review of much of the earlier literature on output efficiency, which typically utilized somewhat restrictive function al forms and single output measures, the reader is referred to Gilbert (1984). 22 O u tp u t e ffic ie n c y The production process has been one of the most extensively investigated topics in banking. A major purpose of most of these studies has been to obtain estimates of scale elasticities, that is, to evaluate how bank costs change with changes in the level of output.1 More recently, 8 efforts have also been made to estimate econo mies of scope; that is, advantages from the joint production of multiple outputs. Concerning scale economies, if changes in bank costs are proportional to changes in output then the scale elasticity measure equals 1.0 and all cost advantages resulting from the scale of production are being fully exploited. If the changes are not proportional, that is, varying returns to scale exist, then efficiency gains could be obtained by leaving the production process unchanged, but altering the quantity of output produced. Scale elasticities less than one imply that increases in output would produce less than proportional increases in costs. Efficiency gains, therefore, could be obtained by increasing the scale of production. This is typically a justification given for bank merger activity. Efficiency gains could be obtained by reducing production levels if decreasing returns to scale exist; that is, the scale elasticity is greater than 1.0. Although much effort has been spent evaluating scale economies, it is one of the most disagreed upon topics in banking. For example, a number of studies find cost advan tages from size to be fully exhausted at relative ly low levels of output. Even when potential economies exist they appear to be relatively small. Some of these studies are summarized in Table 1 which presents the estimated scale elasticity for the average bank in the sample, the range of the estimates for all banks, and the ECONOMIC PERSPECTIVES TABLE 1 Economies of scale estimates for small banks Author Scale elasticity at sample m ean3 Range of scale elasticity measure Relevant range for significant scale (disecon om ies1 1 Benston, Hanweck and Humphrey (1982) U B 1.09 1.10 0.89- 1.24 0.97- 1.16 Diseconomies above $25 m illion Diseconomies above $25 m illion Berger, Hanweck and Humphrey (1987) U B 1.04 1.03 0.87- 1.21 1.00- 1.03 Diseconomies above $100 m illion No significant (dis)economies Cebenoyan (1988) U B 1.08* 0.97 0.88- 1.39 0.92- 1.03 Diseconomies above $50 m illion Economies above $100 m illion U 0.99 0.98- 1.10 Economies above $10 m illion and diseconomies above $50 m illion U 1.03* 0.93- 1.27 Economies below $25 m illion and diseconomies above $100 m illion B 1.02* 0.94- 1.17 Economies below $25 m illion and diseconomies above $100 m illion B U (n.a.) (n.a.) 0.99- 1.02 0.88- 0 .93 No significant (dis)economies Economies below $100 m illion 0.99 0.91 - 0.99 Economies below $100 m illion Gilligan and Sm irlock (1984)c Gilligan, Sm irlock and Marshall (1984) Kolari and Zardkoohi (1987) Lawrence and Shay (1986) “Calculated as (d InC/d InQ) for single output measures or I (d InC/d InY.) for all /=outputs. Benston, Hanweck and Humphrey (1982) calculated a scale elasticity augmented for output expansion via office expansion. bln these studies the banks are grouped by deposit size for calculation of the scale elasticity measure. The figures presented are for the minimum bound on the group where statistically significant (dis)economies were realized. cGilligan and Smirlock did not use the FCA data, as did the other studies, but did evaluate institutions similar in size to those in the FCA sample. ^Denotes statistically significant difference from 1.0. Note: U and B represent unit and branch bank subsamples, respectively. Many of the studies provided results for a number of years and/or are based on alternative output measures. When multiple sets of findings were provided, the results reported here are for the most recent year, based on earning assets as the output measure, and use the intermediation approach (i.e., dollar value of funds transformed to assets). level of output at which significant advantages or disadvantages from the scale of production occurs. Basically, the results imply that scale advantages are fully exhausted once an institu tion achieves a size of approximately $100-200 million, a relatively small bank in the United States.1 Higher output levels result in either 9 constant or decreasing returns to scale. The implications from these results are that very small banks are inefficient because they operate under increasing returns to scale, and inefficiencies may exist for banks above ap proximately $100-200 million in deposits. The extent of the inefficiency, however, would not appear to be very large: scale elasticities typically range from .95 to 1.05. These find FEDERAL RESERVE BANK OF CHICAGO ings would appear to run counter to the argu ments typically found in the popular banking press which imply that merger activity, desires to expand geographically, and product expan sion are all driven by the desire to reap cost advantages; for example, see Moynihan (1991). However, this may partially result from the fact that, until very recently, most of the bank cost studies excluded large institutions; the very ones which are most interested in expanding. Most of the studies presented in Table 1 uti lized the Federal Reserve’s Functional Cost Analysis (FCA) survey data which typically includes only institutions with less than one billion dollars in assets. Although banks in this size group constitute over 95 percent of all 23 banks in the United States, they TABLE 2 constitute only about 30 percent of Results from large bank cost studies the nation’s banking assets. It Range of Size at which excludes the larger banks which are calculated economies of most active in merger activity Author scale are exhausted0 scale elasticities8 (Rhoades 1985) and most vocal about expanded product and geo Berger and 0.3 billion® graphic expansion powers. Humphrey (1990) 0.98- 1.03e 0.08 billion ' 0.92- 1.06' Table 2 provides a summary of results from recent studies which C la rk(1984) 0.95-0.96 Non-exhausted through $500m illionh have analyzed larger financial institutions; typically in excess of Evanoff and one billion dollars. The evidence 5.5 billion Israilevich (1990)' 0.98 suggests that scale advantages exist Hunter and well beyond the $100-200 million Tim m e (1986)' 1.05 $4.2 billion® range. While typically significant 0.97 $12.5 billion" in a statistical sense, the scale Hunter, Tim m e elasticity measure is close to 1.0. $25.0 billion 0.86-1.14 and Yang (1990) Again, the measures tend to range $6.0 billion 0.97-1.09 from .95 to 1.05. Therefore, the Noulas, et al. (1990) studies employing data for larger 0.949 Non-exhausted Shaffer (1988) banks tend to argue against the through $140 billion" finding that inefficiencies resulting Shaffer (1984) 0.95 Non-exhausted from diseconomies of scale set in at through $50 billion" relatively low levels of output. However, the most typical conclu Shaffer and David 0.92 $37.0 billion (1991)' sion the authors draw from these bank cost studies is that potential a The reported values are based on elasticity calculations for alternative asset size groups (when available). Statistical significance is not taken gains from altering scale via inter into account for figures reported in this column—that is, the calculated nal growth or merger activity are values may not be significantly different from 1.0 in a statistical sense. relatively minor.2 0 b The values should be considered approximations. The authors frequently reported scale elasticity measures for a group of banks It should be emphasized, covering a relatively broad size range, for example, 10-25 billion. If the however, that the scale elasticity calculated value was insignificantly different from 1.0, then banks up to $25 billion were said to have constant returns to scale. Unlike the measure is not a measure of ineffi figures reported in the previous column, whether or not a calculated ciency. This may partially explain scale elasticity is significantly different from 1.0 in a statistical sense is all important for figures in this column. some of the disagreement between cFor one bank holding companies. The value is probably biased past research studies claiming downward. This is the sample mean value at which the calculated scale elasticity was insignificantly different from 1.0. potential savings from growth are “For multibank holding companies. See note c. not very great because scale elastic eBranch bank results for the low cost banks. ity measures are not very different 'Unit bank results for low cost banks. from 1.0, and the popular banking 8For a $10 billion bank. press which typically claims that hNon-exhausted for the entire sample. The values are calculated at the sample mean. significant cost savings could be Note: Many of the studies provided results for a number of years gained by expanding the bank scale and/or are based on alternative output measures. When multiple sets of operation. Relatively minor of findings were provided, the results reported here are for the most recent year, are based on earning assets as the output measure, and scale elasticity deviations from 1.0 use the intermediation approach. can actually result in nontrivial inefficiency.2 To determine poten 1 tial gains from scale advantages, $500 million banks compare to that resulting the relative comparison is the production costs of existing banks to that of the most efficient from the one large bank? The scale elasticity measure is required to estimate the cost differ sized bank. For example, assuming scale advantages are exhausted at a $5 billion bank, ence, however it by itself is not a measure of how does the production cost for ten existing inefficiency.2 2 24 ECONOMIC PERSPECTIVES TABLE 3 Estimated scale inefficiencies in banking Author Calculated scale inefficiency (p e r c e n t) Aly, et al. (1990) 3.3C Berger and Humphrey (1990) 4.2b 12.7U Clark (1984) 18.3a Elyasiani and Mehdian (1990a) 38.9C Evanoff and Israilevich (1990) 16.6 G illigan, Sm irlock and Marshall (1984) Hunter, Tim m e, and Yang (1990) 5.0U 4.3b 26.6 Lawrence and Shay (1986) 5.5 Noulas, et al. (1990) 2.7 Shaffer (1984) 12.0d Shaffer (1988) 10.0d a The scale elasticity for the "efficient" firm was .9637 since scale advantages were not exhausted in the data sample. The calculated inefficiency would be larger if we extrapolated outside the study data sample. bDenotes branch banks. Taken directly from the cited study. T h e inefficiency measure is biased downward because data limitations necessitated using an "inefficient" size bank which was not the most inefficient in the sample. “Denotes unit banks. Note: The reported inefficiencies were derived assuming prices, exogenous variables, and product mix were constant across banks (for example, at the sample mean), and that the cost representation could be approximated by lnC= a + b (InQ) + .5cilnQ)2 (where Q represents output). Evaluating only inefficiency resulting from production in the range of increasing returns to scale, the data were centered about the values of the inefficient bank. Hence, for this bank, the scale elasticity measure is simply the coefficient b. The cost of production for the scale efficient bank is InC = a + b ln(F ■Q) + .5c[/n(F •Q))2where Fis the size of the efficient firm relative to the inefficient one. The scale elasticity for the efficient bank is d InC/d ln{F • Q) = b + c ln{F ■ Q) = 1.0. Scale inefficiency is the difference between cost values of the two banks relative to F, that is, [ F-s- CJCt ] -1, where C, and CEdenote costs of the inefficient and efficient bank, respectively. The same methodology could be used to calculate inefficiency resulting from production in the range of decreasing returns to scale. In the studies considered, scale measures are typically reported for various size ranges. Unless noted, the calculated inefficiency is based on the smallest bank in the size group in which statistically significant economies of scale existed, relative to the largest bank in the size category in which minimum efficient scale existed (that is, the scale measure was not significantly different from 1.0 in a statistical sense). Details are available from the authors. By holding product mix constant we restrict the cost savings to scale effects only; precluding any savings resulting from altering the mix. This implicitly assumes either that the mix is actually invariant over the banks considered or that the scale efficient bank analyzed is equal in size to the scale efficient bank observed in the data. Given the assumptions employed and the relatively broad size categories reported in the studies considered, the reported inefficiencies should be considered rough approximations. FEDERAL RESERVE BANK OF CHICAGO Using the actual scale esti mates and sample data from a number of bank cost studies, measures of scale inefficiency were calculated and are presented in Table 3. The reported ineffi ciencies are for banks producing in the range of increasing returns to scale. They suggest that poten tial gains resulting from scale inefficiency are not trivial. While some of the studies suggest ineffi ciencies in the range of five percent, estimates in the 10-20 percent range are not uncommon, and they range up to nearly 40 percent. The major point is that although their importance is typically played down in the bank cost literature, scale inefficiencies appear to be significant enough to warrant efforts by banks to achieve an efficient scale.2 3 The evidence concerning efficiency gains from economies of scope is not conclusive. Studies to date typically focus on the outputs currently produced and find very slight or no potential for efficiency gains; for example, see Benston, et al. (1982), Cebenoyan (1990), Clark (1988), Hunter, Timme, and Yang (1990), Law rence and Shay (1986), and Mester (1987).2 However, the methodol 4 ogies used to evaluate advantages from joint production have typi cally been criticized on the grounds that most functional forms utilized for bank cost analysis are ill suited for analyzing economies of scope. Additionally, the evaluation of potential effi ciency gains is commonly preclud ed as a result of regulation. Since numerous products cannot be provided by banks, there is no available quantitative means to evaluate the joint cost relationship or potential efficiency gains. In p u t e ffic ie n c y While much research has been conducted evaluating output 25 efficiency, only recently has input efficiency been considered. The evidence suggests that the assumption of input efficiency, common in most studies of bank production, is typically violated. Table 4 presents summary findings for recent studies evaluating input efficiency in banking. While substantially different techniques were used in the studies reviewed, the results are surprisingly similar. Total input inefficiency is commonly in the range of 20-30 percent, and is as high as 50 percent in one of the studies. This implies that significant cost savings could be realized if bank management more efficiently utilized productive inputs. Breaking down the study findings into more detail, allocative inefficiency is typically found to be relatively minor and, with one exception, dominated by technical inefficiency.2 Evanoff 5 and Israilevich (1990a, 1990b, 1991) found that the allocative inefficiency that does exist results from the overuse of physical capital relative to other inputs. As mentioned earlier, this is consistent with expectations since past bank regulation did not allow price competition in the market for deposits. As a result, it appears that banks simply responded by competing using alternative means such as service levels. The introduction of numerous branch offices resulted in brick-and-mortar competition instead of price competition. While the typical ly small allocative inefficiency estimate cannot be ignored as a potential source of future cost savings in banking, it does suggest that the frequent criticism of bank regulation based on the burden it imposes on the bank production process may be somewhat exaggerated. Appar ently the optimal mix of factor inputs is only marginally affected by regulation. The results presented in Table 4 suggest that the major source of input inefficiency is T A B LE 4 Input inefficiency in banking Author Approach Overall input inefficiency ( Allocative inefficiency Pure technical inefficiency ---------------- ------ p e r c e n t ------ ---------------- ) 24.8 20.2 M inim al M inim al Berger and Humphrey (1990)b Berger and Humphrey (1990)u Parametric Parametric Elyasiani and Mehdian (1990b)d Parametric Evanoff and Israilevich (1990)a Parametric Evanoff, Israilevich, and Merris (1990) Parametric Ferrier and Lovell (1990) Parametric 26.0 17.1 8.9 Aly, et al. (1990)® Nonparametric 50.7 14.9 35.8 Elyasiani and Mehdian (1990a) Nonparametric Ferrier and Lovell (1990) Nonparametric Gold and Sherman (1985)c Nonparametric 13.6 22.0 1.0 21.0 1.8 11.7 21.0 5.0 16.0 27.9 aForthe 1972-87 period. Subperiods produced different results. bFor branch banks. cFor the most inefficient decision making unit. dFor 1980. Scale inefficiency was also calculated to be 38.9%. eScale inefficiency was also calculated to be 3.1%. “For unit banks. Note: The figures presented are the level of inefficiency relative to the firm using its inputs efficiently. The studies frequently reported inefficiency relative to the observed firm or efficiency as a percentage of input utilization (see Figure 1 for an illustration of input inefficiency measures). These measures were converted to the measure presented here. Gaps in the results are due to the fact that not all studies considered all components of inefficiency. 26 ECONOMIC PERSPECTIVES pure technical inefficiency. Simply put, firms use too much input per unit of output. Com bined with the finding of relatively small allocative inefficiency, this implies that bank managers do a relatively good job of choosing the proper input mix, but then simply under utilize all factor inputs.2 This inefficiency 6 obviously cannot be sustained over time if the banks are subject to competitive forces. Com paring the findings summarized in Tables 3 and 4, it is apparent that the inefficiencies resulting from the suboptimal use of inputs are somewhat larger than those resulting from producing suboptimal levels of output.2 7 Causes o f in e ffic ie n c y and im p lic a tio n s fo r th e fu tu re There are a number of possible explana tions for the inefficiency in banking. Basically, the expected causes should be the same as those found in any industry. As discussed earlier, economic theory suggests that allocative ineffi ciency is driven by market distortions from factors such as regulation. Pure technical inefficiency may be the result of weak market forces (induced by market structure or regula tion) which allow bank management to become remiss and to continue their inefficient behav ior. Scale and scope induced inefficiency may be the result of either market or regulatory forces which make the optimal level and mix of outputs unachievable. Some analysts would also argue that bank size should be a determi nant of efficiency. According to this argument, larger banks may have more astute manage ment and/or be more cost conscious because of greater pressure from owners concerning bottom-line profits. Additionally, these banks are typically located in the larger, more com petitive markets which may induce a more efficient production process. The evidence suggests that these forces are indeed operative in determining efficiency levels in banking. Analyzing data for large banks over the 1972-87 period, Evanoff and Israilevich (1990a, 1990b, 1991) found alloca tive inefficiency to be related to alternative measures of regulatory stringency. It was also found to be greater in regions characterized by more restrictive state level regulation, and significantly less after industry deregulation occurred in the early 1980s. Allocative effi ciency has not been found to be related to bank size (for example, Aly, et al. 1990). This, FEDERAL RESERVE BANK OF CHICAGO however, should not be surprising since ineffi ciency may occur although the bank is operat ing efficiently in response to shadow market prices (that is, those including market distor tions). The evidence also suggests that pure technical inefficiency is induced by regulation, and some evidence exists suggesting that it results from elements of market structure. Berger and Humphrey (1990) found that the inefficiency was greater, on average, for banks located in the more restrictive unit banking states than those in states allowing branching. Additional analysis of data used in Evanoff and Israilevich (1990b) produced similar findings.2 8 Pure technical inefficiency has also been found to be negatively associated with bank size [Berger and Humphrey (1990), Aly, et al. (1990), Elyasiani and Mehdian (1990a), Rangan, et al. (1988)]. To the extent that small institutions are located in the smaller, least competitive markets, the absence of market pressures could be producing the higher levels of inefficiency. Aly, et al. (1990) tested this contention more directly by relating pure technical inefficiency to bank location. Banks located in large metropolitan areas were found to be significantly more efficient than those in smaller markets, suggesting market structure may influence efficiency levels. The evidence here, however, is also not conclusive. It may be that cost savings realized by urban banks may exist because the increased population density makes possible less costly delivery systems. This cost savings may be interpreted as being driven by greater market competition in the urban markets while actually it is simply a function of demographics. Scale and scope diseconomies are also expected to be partially determined by regula tory forces. Unit banking restrictions force banks to expand at one physical location in stead of allowing them to expand by opening additional offices to serve customers. Disecon omies of scale may set in at the individual office causing higher cost for larger single office institutions. Expansion via new offices has been shown to be more cost effective. Review of the findings presented in Tables 1 and 2 indicates that diseconomies of scale are typically larger in unit banking markets. Anal ysis which combines data for both unit and branch banks usually find that the larger unit banks typically operate under conditions of 27 diseconomies of scale; see for example, Evanoff, Israilevich, and Merris (1990). LeCompte and Smith (1990) also found that inefficiency resulting from not producing the proper mix of outputs, and therefore failing to take advantage of economies from joint produc tion, was greater under conditions of more stringent regulation. What are the implications of these find ings? Given the important role regulation apparently serves in determining efficiency levels, the recent trend toward industry deregu lation should result in improved efficiency. Reductions in entry barriers resulting in less regulatory-created market protection, and fewer regulatory-induced market distortions should significantly increase competitive pressures. The beneficial aspects of increased competition will be accomplished by weeding out the less efficient firms. Obviously, in an environment of deregulation and increased competition, reducing pure technical inefficiency could be a major determinant of firm survival. Merger activity in the financial services industry will probably increase in the future as banks strive to compete in the deregulated market. The deregulation will provide banks with both the desire and the ability to expand acquisition activity. Scale inefficient firms will be absorbed to exploit cost advantages. Firms whose management does an inadequate job of utilizing factor inputs may soon find it difficult to survive in the more competitive market. They will be required to eliminate the ineffi ciency or become prime targets for acquiring firms looking to “trim the fat” from new acqui sitions.2 Given that pure technical inefficiency 9 is so significant in banking, and given that it is the one aspect of efficiency over which the firm has direct control, one would expect significant increases in bank productive efficiency in the coming years. S u m m ary and conclusions The purpose of this study has been to discuss productive efficiency at a conceptual level and to review the relevant literature for the banking industry. We categorize efficiency into input and output related measures. Output inefficiency results from producing suboptimal 28 output levels or a suboptimal combination of outputs. Input inefficiency results from produc tion using a sub-optimal input mix (allocative inefficiency) and not effectively utilizing the inputs employed (pure technical inefficiency). A review of existing bank cost studies suggests that banks have substantial room to increase productive efficiency and, as a result, to signifi cantly lower costs. Although the range of findings in the studies surveyed is relatively broad, it is not uncommon to find 10-20 percent bank scale inefficiency generated by producing at suboptimal output levels. Allocative ineffi ciency is typically found to be relatively minor; usually less than five percent. Pure technical inefficiency, however, is apparently quite significant; in the range of 20-30 percent. Combining these three effects results in sub stantial potential cost savings for banks. What are the major causes of bank ineffi ciency? The evidence suggests that industry regulation is a dominant source. Allocative inefficiency, although relatively minor, is directly induced by regulation. Inefficiencies resulting from not producing the optimal combination of outputs have also been shown to be related to regulation. However, the major source of inefficiency, pure technical ineffi ciency, is managerially induced. That is, the absence of competitive forces, which is also influenced by industry regulation, has allowed banks to continue to operate in spite of the fact that management has not effectively utilized the resources available to them. Given that the industry is undergoing a process of significant deregulation, the findings from these studies have both positive and negative implications for banking. As deregu lation continues, the increased competitive pressures will force banks to operate more efficiently. Those unable to do so by adjusting to the new competitive environment will have difficulty surviving. However, one of the major sources of inefficiency, pure technical ineffi ciency, is directly under the control of the banks themselves. Therefore, they will have control of their own destiny. In light of the recent significant number of branch closings and cost saving campaigns aimed at reducing payrolls, it would appear that efforts to improve bank efficiency are already underway. ECONOMIC PERSPECTIVES FOOTNOTES 'Drucker (1991). 2However, the research process continues because of differences in results from previous studies, methodologies, assumptions, output definitions, etc. For a review of some of these studies see Gilbert (1984), Clark (1988), or Humphrey (1990). 3These definitions of productive inefficiency were introduced by Farrell (1957). They are radial measures and coincide with much of the discussion that follows. For alternative measures of (in)efficiency see Fare and Lovell [1978]. 4At this stage we are ignoring the potential for economies of scale. Farrell assumed a linearly homogeneous production process; that is, constant returns to scale. We assume, as discussed below, that any gains from scale advantages would result in a higher level isoquant for a given efficient combination of inputs. 5For example, if production was allocatively efficient the measure would, obviously, be O E I O E = 1.0; that is, points 6, C , and E would coincide. Typically we assume profit maximization as the objective under competitive conditions, that is, frictionless markets and the absence of monopoly power and regulatory distortions. In this model the production, cost, and profit functions are essentially alternative means of evaluating the same optimization process, that is, the production process. The cost relationship is frequently evaluated instead of the production function because less information is required. When discussing frontier analysis, it is irrelevant whether the cost or production side is considered. However, empirically, the choice of a cost, production, or profit representation may generate different results because the researcher is required to use approximations for the true functional forms of these representations. 7A s a byproduct, the methodology also allows for the estimation of scale and scope induced inefficiency. It does not, however, allow for estimates of pure technical inefficiency. 8However, this is an empirical approach, therefore the true cause of the distortion in factor prices could be generated by a number of things, including data measurement errors, etc. ’Using this methodology, the shadow price approximations can be interpreted as first-order Taylor’s series expansions of arbitrary shadow-price functions. It should be emphasized that there is nothing special about the linear relationship. Alternative specifications can and should be considered; for example, see Evanoff and Israilevich (1991, 1990a). 'Typically, factor share equations are derived from the cost relationship via Shephard’s Lemma and the system of cost and share equations are jointly estimated. The additional FEDERAL RESERVE BANK OF CHICAGO share equations provide increased efficiency of the estimates. The share equations are derived from the s h a d o w cost relationship. "For a lucid description of these models, see Bauer (1990). The foundation for this approach was developed by Aigner, Lovell and Schmidt (1977). See also Fare, Grosskopf, and Lovell (1985). 12One sided distributions which have been used in estimation include the half-normal, exponential, truncated normal and the Gamma distribution. Again, see Bauer (1990) and the sources cited. 'T h is is the linkage employed in Ferrier and Lovell (1990) in their study of bank efficiency. 'T h e approach can also be combined with others to incorporate additional information. For example, Evanoff and Israilevich (1990b) augmented the thick frontier approach by estimating shadow cost functions for high and low cost banks. In this way, estimates of allocative inefficiency could be obtained directly from the model instead of using an auxiliary, somewhat arbitrary, procedure to decompose the total inefficiency into its component parts. 'Technical inefficiency can be calculated ignoring this information. Measures of allocative efficiency require information on factor prices. l6Evanoff and Israilevich (1991) found significant regional differences in bank production techniques a n d levels of efficiency. "Surprisingly, Ferrier and Lovell (1990) find exactly the opposite in their analysis of banks. '"More formally, the scale elasticity measure is the percentage change in cost relative to the percentage change in output, or d ln C / d l n Q . One major issue in bank cost studies is determining what constitutes output. Although defining output is difficult in any service oriented industry, there seems to be more controversy with respect to banking. However, of the measures used to date, the findings tend to be similar regardless of the measure employed; see Humphrey (1990). 19In 1989, over three-quarters of U.S. banks had less than $100 million in assets; see FDIC (1989). However, banks over 1000 times this size also existed. 20It is possible that scale estimates could be biased as a result of misspecifying the cost relationship. For example, the standard assumption of efficient utilization of factor inputs, if incorrect, could produce misleading findings concerning scale (dis)advantages. However, Berger and Humphrey (1990), Evanoff and Israilevich (1990b) and Evanoff, Israilevich, and Merris (1990) found scale estimates were n o t substantially different when input inefficiency was accounted for. 29 2lThe distinction between the scale elasticity and inefficiency measures has been emphasized in Shaffer (1988) and Shaffer and David (1991). Using data from a previous study, Shaffer and David show that a scale elasticity of .99 could result in a 25% cost savings if production was shifted from small to large banks. 22Actually the scale inefficiency will be determined by the difference in average cost between the efficient and inefficient firm. The elasticity at the output level corresponding to the inefficient firm gives us information about cost changes at slightly larger or slightly smaller output levels. Neither of these levels is relevant for determining inefficiency since we never produce at these levels. For efficiency analysis, production takes place either at the efficient or the inefficient firm; therefore only the corresponding two average cost values are relevant. Whereas the elasticity measure gives percentage changes in cost induced by incremental changes in output, in the banking studies analyzed in Table 3 the difference in output between the efficient and inefficient firm is not incremental. 23Caution should be taken in deriving policy implications from findings concerning scale efficiency alone. There may be alternative factors which partially offset these potential gains. In fact, in viewing bank data, Humphrey (1990) finds the average cost across all bank size groups to be amazingly similar. That, combined with the potential efficiency gains from scale economies discussed here, suggest that there may be some factors counteracting these potential efficiencies. However, with respect to scale efficiency alone, there would appear to be significant potential gains for banking. 24Some studies have, however, found significant advantag es resulting from joint production; for example, Gilligan, Smirlock, and Marshall (1984), and Evanoff and Israilevich (1990b). However, the finding of relatively small or no scope economies is most typical. The methodologies utilized to generate estimates of scope economies have been critiqued in Pulley and Humphrey (1990). This is obviously a rich area for future research. 25Although Aly et al.(1990) find evidence of greater allocative inefficiency than most of the studies reviewed, the major exception to the norm is the study by Ferrier and Lovell (1990). Using a parametric approach the authors found significant allocative inefficiency (over 17 percent). However, as mentioned earlier, the reliability of these techniques decreases significantly when non-homogeneous decision making units are considered. The data for this study included mutual savings banks, credit unions, savings and loan associations, and “noncommercial” institutions. Nearly a third of the sample was made up of noncommer cial banks. Given that the technology for these institutions may differ from that of commercial banks, one would expect these observations to have a substantial influence on the error structure of the estimates. The authors themselves even state that some of these observations d o significantly influence their results (Ferrier and Lovell, p. 243). Since the distribution of the errors is the major determinant of the efficiency measure, this may bias the results concerning commercial bank efficiency. The study also found that the allocative efficiency resulted from an over-utilization of labor relative to the other factors. This is precisely the opposite of what one, intuitively, would expect in banking (see Evanoff and Israilevich 1990b). Finally, the measures used for factor prices may bias the results toward finding allocative inefficiency resulting from over-utilization of labor (Berger and Humphrey 1990, p. 21). 26This finding has implications for the bank expense preference literature, for example, Mester (1989). Typically it is assumed that managers of the expensing bank prefer one input to others—usually labor. The results presented here suggest that a more restricted form of expense preference, a preference for all the inputs to the same degree, may best describe the situation in banking. 27However, this excludes any inefficiencies resulting from scope disadvantages which cannot be empirically captured. 28The evidence on this, however, is not conclusive. Aly, et al. (1990) found no significant efficiency difference across unit and branch banks. 29This does not imply that there will no longer be small banks. While most of the bank cost literature has assumed homogeneous outputs, recent research suggests that banks frequently find a market niche in an attempt to differentiate themselves from others. Efficient banks which are able to fill a needed market niche should continue to prosper in a deregulated environment. See Amel and Rhoades (1988). REFERENCES Afriat, Sydney, “Efficiency estimation of production functions,” International Economic Review, 13, 1972, pp. 568-598. Aigner, Dennis, C.A. Knox Lovell, and Peter Schmidt, “Formulation and estimation of stochastic frontier production function models,” Journal of Econometrics, 6, 1977, pp. 21-37. Aly, Hassan Y., Richard Grabowski, Carl Pasurka, and Nanda Rangan, “Technical, scale, and allocative efficiencies in U.S. banking: an 30 empirical investigation,” The Review of Econom ics and Statistics, 72, 1990, pp. 211-218. Amel, Dean F., and Stephen A. Rhoades, “Strategic groups in banking,” The Review of Economics and Statistics, 70, 1988, pp. 685-689. Atkinson, Scott E., and Robert Halvorsen, “Parametric efficiency tests, economies of scale, and input demand in U.S. electric power genera tion,” International Economic Review, 25, 1984, pp. 647-662. ECONOMIC PERSPECTIVES Atkinson, Scott E., and Robert Halvorsen, “A test of relative and absolute price efficiency in regulated utilities,” The Review of Economics and Statistics, 62, 1980, pp. 185-196. Elyasiani, Elyas, and Seyed M. Mehdian, “Efficiency in the commercial banking industry, a production frontier approach,” Applied Econom ics, 22, 1990b, pp. 539-551. Bauer, Paul W., “Recent developments in the econometric estimation of frontiers,” Journal of Econometrics, 46, 1990, pp. 39-56. Evanoff, Douglas D., “Branch banking and service assessibility,” Journal of Money, Credit, and Banking, 1988, pp. 191-202. Benston, George, Gerald A. Hanweck, and David B. Humphrey, “Scale economies in banking,” Journal of Money, Credit, and Banking, 14, 1982, pp. 435-456. Evanoff, Douglas D., and Philip R. Israilevich, “Cost economies and allocative efficiency in large U.S. commercial banks,” Proceedings of a Conference on Bank Structure and Competition, 26, 1990a, pp. 152-169. Berger, Allen N., Gerald A. Hanweck, and David B. Humphrey, “Competitive viability in banking: Scale, scope, and product mix econo mies,” Journal of Monetary Economics, 16, 1987, pp. 501-520. Berger, Allen N., and David B. Humphrey, “The dominance of inefficiencies over scale and product mix economies in banking,” forthcoming in Journal of Monetary Economics, 28, 1991. Also in Finance and Economics Discussion Series, 107, Board of Governors of the Federal Reserve System, 1990. Evanoff, Douglas D., and Philip R. Israilevich, “Deregulation, cost economies and allocative efficiency of large commercial banks,” Issues in Financial Regulation, Federal Reserve Bank of Chicago Working Paper 90-19, 1990b. Evanoff, Douglas D., and Philip R. Israilevich, “Regional differences in bank efficiency and technology,” The Annals of Regional Science, 25, 1991, pp. 41-54. Cebenoyan, A. Sinan, “Multiproduct cost functions and scale economies in banking,” The Financial Review, 23, 1988, pp. 499-512. Evanoff, Douglas D., Philip R. Israilevich, and Randall C. Merris, “Relative efficiency, technical change, and economies of scale for large commercial banks,” Journal of Regulatory Economics, 2, 1990, pp. 281-298. Cebenoyan, A. Sinan, “Scope economies in banking: the hybrid box-cox function,” The Financial Review, 25, 1990, pp. 115-125. Fare, R.J., Shawna Grosskopf, C.A. Knox Lovell, The measurement of efficiency of produc tion, Boston, Kluwer Academic Publishers, 1985. Clark, Jeffrey A., ‘‘Estimation of economies of scale in banking using a generalized functional form,” Journal of Money, Credit, and Banking, 16, 1984, pp. 53-67. Fare, R.J., and C.A. Knox Lovell, “Measuring the technical efficiency of production,” Journal of Economic Theory, 19, 1978, pp. 150-162. Clark, Jeffrey A., “Economies of scale and scope at depository financial institutions: A review of the literature,” Economic Review, Federal Reserve Bank of Kansas City, 1988, p. 16-33. Drucker, Peter F., “Don’t change corporate culture—use it,” The Wall Street Journal, March 28, 1991, p. A 14. Elyasiani, Elyas, and Seyed M. Mehdian, “A nonparametric approach to measurement of efficiency and technological change: The case of large U.S. commercial banks,” Journal of Financial Services Research, A, 1990a, pp. 157168. Farrell, M.J., “The measurement of productive efficiency,” Journal of Royal Statistical Analysis, A, 120, 1957, pp. 253-281. FDIC, “Statistics on banking,” Federal Deposit Insurance Corporation, Washington, GPO, 1989. Ferrier, Gary D., and C.A. Knox Lovell, “Measuring cost efficiency in banking: Economet ric and linear programming evidence,” Journal of Econometrics, 46, 1990, pp. 229-245 Gilbert, R. Alton, “Bank market structure and competition,” Journal of Money, Credit, and Banking, 16, 1984, pp. 617-644. Gilligan, Thomas W., and Michael L. Smirlock, “An empirical study of joint production and scale FEDERAL RESERVE RANK OF CHICAGO 31 economies in commercial banking,” Journal of Banking and Finance, 8, 1984, pp. 67-77. Gilligan, Thomas W., Michael L. Smirlock, and William Marshall, “Scale and scope economies in the multi-product banking firm,” Journal of Monetary Economics, 13, 1984, pp. 393-405. Humphrey, David B., “Why do estimates of bank scale economies differ?,” Economic Review, Federal Reserve Bank of Richmond, 1990, pp. 3850. Hunter, William C., and Stephen G. Timme, “Technical change, organization form, and the structure of bank production,” Journal of Money, Credit, and Banking, 18, 1986, pp. 152-166. Hunter, William C., Stephen G. Timme, and Won Keun Yang, “An examination of cost subadditivity and multiproduct production in large U.S. banks,” Journal of Money, Credit, and Banking, 22, 1990, pp. 504-525. Kolari, James, and Asghar Zardkoohi, “Bank cost, structure, and performance,” Lexington, D.C., Heath Publishers. Kopp, Raymond, and W. Erwin Diewert, “The decomposition of frontier cost function deviations into measures of technical and allocative efficien cy,” Journal of Econometrics, 18, 1982, pp. 319331. Lau, L. J., and P. A. Yotopoulos, “A test for relative efficiency and application to Indian agriculture,” American Economic Review, 61, 1971, pp. 94-109. Lawrence, Colin, and Robert Shay, “Technolo gy and financial intermediation in a multiproduct banking firm: an econometric study of U.S. banks, 1979-82,” in Colin Lawrence and Robert Shay (ed.), Technological Innovation, Regulation, and the Monetary Economy, Cambridge, Ballinger, 1986, pp. 53-92. Lecompte, Richard, L. B., and Stephen D. Smith, “Changes in the cost of intermediation: The case of savings and loans,” Journal of Finance, 45, 1990, pp. 1337-1345. Mester, Loretta J., “A multiproduct cost study of savings and loans,” Journal of Finance, 42, 1987, pp. 423-445. 32 Mester, Loretta J., “Testing for expense preference behavior: Mutual and stock savings and loans,” Rand Journal of Economics, 20, 1989, 483-498. Moynihan, Jonathan P., “Banking in the 90s— where will the profits come from?,” Proceedings of a Conference on Bank Structure and Competi tion, Federal Reserve Bank of Chicago, 27, 1991. Noulas, Athanasios G., Subhash C. Ray, and Stephen M. Miller, “Returns to scale and input substitution for large U.S. banks,” Journal of Money, Credit, and Banking, 22, 1990, pp. 94108. Pulley, Lawrence B., and David B. Humphrey, “Correcting the instability of bank scope econo mies from the translog model: A composite function approach,” paper presented at the Financial Management Association meetings, Orlando Florida, October, 1990. Rangan, Nanda, Richard Grabowski, Hassan Aly, and Carl Pasurka, “The technical efficiency of U.S. banks,” Economic Letters, 28, 1988, pp. 169-75. Rhoades, Stephen A., “Mergers and acquisitions by commercial banks,” Staff Studies, 142, Board of Governors of the Federal Reserve System, 1985. Shaffer, Sherrill, “Scale economies in multiprod uct firms,” Bulletin of Economic Research, 1, 1984, pp. 51-58. Shaffer, Sherrill, “A revenue-restricted cost study of 100 large banks,” Federal Reserve Bank of New York, unpublished research paper, 1988. Shaffer, Sherrill, and Edmond David, “Econo mies of superscale in commercial banking,” Applied Economics, 23, 1991, pp. 283-293. Sherman, H. David, and Franklin Gold, “Bank branch operating efficiency,” Journal of Banking and Finance, 9, 1985, pp. 297-315. Zieschang, Kimberly D., “A note on the decom position of cost efficiency into technical and allocative components,” Journal of Econometrics, 23, 1983, pp. 401-405. ECONOMIC PERSPECTIVES ECONOMIC PERSPECTIVES BULK RATE Public Information Center Federal Reserve Bank of Chicago P.O. Box 834 Chicago, Illinois 60690-0834 U.S. POSTAGE PAID CHICAGO, ILLINOIS PERMIT NO. 1942 D o N o t F orw ard A d d ress C o rrec tio n Requested R eturn Postage G uaranteed FEDERAL RESERVE BANK OF CHICAGO