V. A Thick Frontier Analysis for 1996
The preceding analysis implicitly assumed that all thrifts in each year operated on the same cost frontier with the same underlying efficiency. Several authors, e.g., Bauer et al.
1993, Berger and Humphrey 1992, 1991 and Berger and Mester 1997a, 1997b, however, have questioned this assumption and used a thick cost frontier to address the
issue of varying efficiency levels between groups of financial institutions. By recognizing that there are observed cost differences inefficiency which do not reflect random error,
the thick frontier methodology reduces distributional assumptions on the error term in cost function regressions and provides more precise empirical measures.
12
A more general cost function regression decomposes the stochastic error into two parts as:
lnC 5 C~ p,q 1 p; 13
lnC 5 C~ p,q 1 m 1 n, where C p,q is a general cost function as above; p is the composite error term; m
represents firm-specific inefficiencies both allocative and technical which raise costs above best practice, and n is a true random error which represents measurement error or
luck.
The thick cost frontier approach decomposes the composite error and isolates the inefficiency component, m, by comparing the average efficiency of large subsamples of
the available data. The complete sample is divided into quartiles based on average costs and it is then assumed that variation from expected values within each quartile represents
random error, n, while variation between quartiles represents more fundamental ineffi- ciencies or exogenous differences, m. By estimating a separate regression for each
quartile, therefore, the error terms are more likely to meet the statistical assumptions, e.g., n is mean zero and finite variance.
13
Following Berger and Mester 1997b, the 1,202 thrifts in 1996 were divided into four quartiles based on the residual that was obtained from the cost function estimation
described in Section IV.
14
Thrifts with the smallest residuals were assumed to be cost efficient low m, while large residuals were indicative of cost inefficiency large m.
Table 4 shows descriptive statistics for the thrifts in each quartile and the total of 1,202 thrifts in 1996. The lowest cost quartile shows the smallest average cost residual by
definition, and there is a steady increase in average costs total cost divided by total assets across the four quartiles. The lowest cost quartile also shows the highest return on
average assets ROA, return on average equity ROE, and Tier 1 Risk-Based Capital Ratio. These 300 thrifts appear to be the most cost efficient and the most successful, as one
would expect. There is not, however, any definite trend in average size across the quartiles.
15
The system for the SUR analysis, equations 7 to 9, was then reestimated separately for each quartile in 1996. To compare behavioral differences across thrifts with varying
12
See Berger and Humphrey 1992, especially page 257, for a discussion of the econometric advantages.
13
The data envelopment analysis DEA approach is another alternative, which assumes that the random error component is zero, so that all unexplained variation represents inefficiency.
14
As an alternative, Bauer et al. 1993 created quartiles based on total costs per dollar of total assets.
15
To conserve space, the thick frontier analysis is reported only for 1996. The results for 1991 are similar and are available upon request.
154 K. J. Stiroh
levels of efficiency, both the AES and the MES for the lowest cost quartile and the highest cost quartile were calculated using the estimated parameters not shown and mean values
from each quartile, respectively. This approach compares behavioral differences between the two quartiles that reflect both technology as determined by the estimated parameters
and behavioral variation as determined by the mean values. Table 5 reports the results.
The estimated substitution patterns show some similarity between the two groups—all own-price AES were negative as expected, substitution was most responsive to a change
Table 4. 1996 Summary Statistics by Efficiency Quartile
Variables Cost Quartiles
Total Lowest
Second Third
Highest Total assets
576,379.9 451,667.5
523,137.9 611,622.6
540,746.4 2,198,571.3
1,658,846.8 2,629,598.0
1,833,243.8 2,111,602.8
ROA 0.669
0.467 0.434
0.266 0.459
0.593 0.466
0.573 1.644
0.959 ROE
5.793 4.814
4.697 3.686
4.747 6.394
5.677 7.342
9.827 7.506
Tier 1 ratio 24.878
19.403 18.881
19.799 20.738
14.167 9.096
10.496 15.868
12.915 Average cost
0.056 0.059
0.061 0.065
0.060 0.007
0.007 0.008
0.021 0.013
Cost residual 20.268
20.066 0.052
0.285 0.001
0.144 0.038
0.037 0.177
0.232 No. obs.
300 300
301 301
1202
Standard deviations are in parentheses. Quartiles are based on a residual from equation 7, estimated for 1,202 thrifts in 1996.
Table 5. Elasticities of Substitution by Cost Quartile in 1996
Allen Elasticities of Substitution AES - s
jk
Lowest Cost Highest Cost
Capital Labor
IBL Capital
Labor IBL
Capital 212.289
Capital 27.384
30.344 27.431
Labor 0.728
22.957 Labor
1.839 21.452
2.838 12.824
9.159 8.251
IBL 0.880
0.699 20.257
IBL 0.393
0.340 20.214
10.402 9.974
10.974 3.633
3.673 3.963
Morishima Elasticities of Substitution MES - M
jk
Lowest Cost Highest Cost
Capital Labor
IBL Capital
Labor IBL
Capital 0.841
0.851 Capital
0.938 0.791
32.475 28.269
31.477 22.472
Labor 0.707
0.701 Labor
0.905 0.493
23.401 12.257
27.485 6.763
IBL 0.845
0.711 IBL
0.379 0.346
10.894 10.273
3.902 3.805
t statistics are in parentheses. Elasticities are defined in equations 10–12.
Measuring Input Substitution in Thrifts 155
in the price of capital, and the MES appeared asymmetric— but there were also several interesting differences between the two quartiles.
The MES asymmetry appeared larger for the highest cost thrifts, which is consistent with their relatively high cost structure. If these thrifts cannot respond as quickly or easily
to a change in a particular price, one would expect them to be less efficient and incur higher costs. Second, the own-price AES were generally larger for the lowest cost quartile.
Third, the AES overstated the elasticity of substitution between capital and labor for the entire industry, but not for the lowest cost quartile separately. Both of these results imply
more flexible substitution patterns for the lowest cost thrifts. Finally, the lowest cost quartile shows larger MES in terms of capitalIBL and laborIBL, but smaller MES in
terms of capitallabor substitution. This suggests that the highest cost thrifts are the most restricted with regard to interest-bearing liabilities.
VI. Conclusions