658 J. A Roquebert, R. L Phillips and p A. Westfall since both Schmalensee's(1985)and Rumelt's The Frc data were designed to obtain data from (1991)corporate effects were trivial, at best only the corporations with the greatest market Data b cessful,corporations. On the other hand, there vantage in looking at a sample In the present effort, we selected the COMPU- not only better represents businesses in general, STAT@ Business Information Industry Segment but is much more recent. Despite the above Data as the population from which to sample. advantages, it must be acknowledged that the In addition to the financial information at the Ftc data were collected under regulatory force consolidated corporate level, COMPUSTAT com- whereas the COMPUSTAT data are taken from piles information required by the Securities and data that companies provide in their respective Exchange Commission (SEC) and the Financial statements. ccounting Standards Board (FASB) Statement Number 14. Corporations are required to provide nformation on their principal lines of business Sample selection anmd statistical procedure (industry segments). In the data base, each cor- In order to parallel Schmalensee (1985)and poration is required to report information on seg- Rumelt (1991) we restricted our sample to 4 ments which account for 10 percent or more of digit SIc codes connected with manufacturing consolidated sales, operating profits, or assets. (codes 2000-3999). Second, we only selected Before proceeding, it may be well to compare those corporations with two or more SBUs. Third, the comPustat data to the ftc data used by we eliminated all SBU ROAS greater than four both Schmalensee(1985) and Rumelt (1991). standard deviations from the mean in order that Table 1 presents key aspects of the two data extreme values did not have undue influence on ases. As can be seen from Table 1. there are the anal important differences between the two data bases. Of the four estimation procedures in the The compustat data base is SAS/STATO package for the variance comp nents analysis(PROC VARCOMP), we selected o more recent(1985-91)VS. FTC (1974-77); the maximum likelihood procedure because of o broader in scope(746 manufacturing SICs vs. certain advantages. One benefit is that the asymp 260 for FTC); totic covariance matrix of the parameter estimates for a greater time period(7 years of data Vs. is readily available as a by-product of the esti for FTC) o larger (almost 3000 corporations involve manufacturing vs. fewer than 500; over Y 6000 onal elements of the covariance matrix are stan- dard errors of the parameter estimates which can manufacturing lines of business vs. fewer than in turn, be used to create asymptotically valid 4000 for the Ftc data) confidence limits on the parameters. In other o less restrictive(includes smaller corporations- words, we can estimate the significance levels of average SBUs per corporation of 1.95 vs 9.17 the resulting z-values. We must, however, add a for the ftc data ). caveat. The validity of the asymptotic approxi- mation requires that the true underlying parameter To be sure, both data bases have their advantages. be nonzero That the use of these standard errors Table 1. Comparison of the COMPUSTAT and FTc zero is, strictly speaking, not valid for hypothes data bases testing. Nevertheless, it is useful to supply the standard errors along with the estimates to pro Average of vide a measure of accuracy. Further, the order FTC sample COMPUSTAT of magnitude of the parameter(expressed as a Industry categories 942 percentage of the total variance accounted for) Corporations 6873 an be used as an indicator of the likelihood that Lines of business (LOB) 4, 193 13.398 the underlying true value is nonzero--that is, the LOB rporation 917 195 greater the parameter, the more likely the basi is met