Markets vs Management 659 a second benefit is that the maximum likeli- As can be seen from table 3. our results hood estimation scheme is an iterative procedure produced a high degree of agreement with the which incorporates the relative magnitudes of the Rumelt(1991) results of the Ftc data base in estimates at each step of the iteration. The alter- percent variance accounted for by both industry native is to use single-step(noniterated) quadratic and SBU effects. The results of the MIVQUEO estimates which do not account for the relative analysis provided the same relative order of vari- magnitudes of the variance components, and such ance accounted for, but as expected due to estimates lose efficiency relative to the iterated reasons discussed earlier, produced smaller values estimates, particularly in large samples(see Searle for the parameters and a much larger value for et al, 1992, for further discussion) the error term(industry =4.6%0; corporate=7.0%, Due to computational constraints with the SBU=32.6%; industry by year interaction maximum likelihood method, we employed a ran- 2.9%0, and error= 52.6%). We suggest that the dom sampling procedure, without replacement. relative size of the parameters of the MIVQUEO Ten samples of 100-150 corporations were ana- results tend to support the maximum likelihood lyzed. Each sample, on the average, contained parameter estimates 100-150 corporations in over 160 industries. The However, there is a significant divergence 10 samples resulted in 16,596 different sbu between the present results and both the schma ROAs (all that met the selection criteria in the lensee (1985) and Rumelt (1991) results data base) being analyzed. The average number Specifically, we estimate a nontrivial corporate of SBUs per corporation was 4.01. Table 2 pre- effect, i.e., 18 percent of variance for us vs. less sents the descriptive statistics for each sample. than 2 percent for Rumelt(1991)and near zero In order to further investigate the robustness of for Schmalensee (1985). See Table 4 for a com he procedure, we used the MIVQUEO estimates, parison of Schmalensee's (1985), Rumelt's which are more easily computed(but less efficient (1991), and current results. (Note: we compare than the maximum likelihood estimates), allowing our results to Rumelt's, 1991, Sample B as that us to analyze the entire data set with a single run. was the sample that was less restrictive with respect to the size of sBU market share and thus, more comparable to our data) Table 3 shows the results for each includ- ing the parameter estimates for the various effects, TOWARD A POSSIBLE the percentage of total variance parameter RECONCILIATION represents, the value of the dia asymptotic covariance matrix, and the estimated It should be acknowledged that Schmalensee significance (1985)recognized that 80 percent of the variance Table 2. Descriptive statistics for each sample Sample Average S.D. Minimum Maximum Levels ROA ROA ROA Industry Cor SBU Sample 1 1, 752 l134 22.53-138.2413239 437 Sample 2 1, 773 0521973-1376014938 266 45l Sample 3 1,711 11 20.3 131.87 13559 432 1096 21.79 -13847 128,35 Sample51,60310.36 -134.38 145.16 6909039 387 415 Sample 6 1, 512 20.90 31.22 1464 403 Sample 7 1, 699 1003 22.27-144.68121.74 129920.55-147.0614241 Sample 9 1,721 11.4 13868 143.10 Sample 10 1, 568 9.29 23.48-143.22144.6 403 16,59611.00991 -147.0614938 237.5 105.1 4214