EXPENSES IN DISTRICTS I AND II 205 
year average level tend to increase in the following year, the 
net percentage amounts of decrease or of increase being positively 
correlated with the percentage amounts of dispersion from type. 
It is now shown that similar tendencies for the individual banks 
in the Boston district are in operation, the ratios of the indi- 
vidual banks being compared in the successive pairs of years 
with the yearly averages for all banks. 
Are these regressions to type due to the method of expressing 
the deviations and to the fact that the averages change from 
year to year? They do not appear to be. It is true that the 
averages change, but the amounts of variation, except as between 
1922 and 1923, are small. Moreover, as shown in Tables 161 and 
162, the averages are functionally related to volumes of earning 
assets and to sizes of city of location. Accordingly, let us treat 
the data in a manner similar to that used to secure the detail 
in Table 168, but compute the deviations, not in terms of the 
averages for all banks, but as compared with (1) those in the 
different city groups, and (2) those in the various volume groups. 
In measuring regression, we shall later discard all averages as 
points of departure for computing deviations. It will be found 
that whatever methods are followed, the results are substantially 
the same. These appear to be due to the properties of the data 
rather than to the methods of manipulating them. 
Table 169 shows the nature and amount of regression to type 
in the ratios of total expense to earning assets when the indi- 
vidual bank ratios are expressed as deviations in the first and 
in the second of each pair of years, 1922-1925 (1) from the 
respective city-group averages, and (2) from the respective 
volume-group (earning assets) averages, the results being those 
secured from aggregating and averaging the amounts found in the 
respective pairs of years. For purposes of comparison, the total 
section of Table 168 is also included. 
As is indicated in Table 169, the three methods of computing 
the deviations give substantially the same results, both as to the 
fact and as to the amounts of regression. That is, according to 
all methods, ratios deviating from type in a given year deviate 
less widely the following year, the changes in the deviations in 
the second year varying directly with the percentage amounts of 
deviation in the first year. This condition obtains not only for