How Much Does Industry Matter, really example, the industry effect. The random-effects expresses the variance of this portion of return assumption stipulates that each observed industry as a function of the rate of persistence, p; and effect is drawn randomly at some early date from of the population variances in year(ox), industry an underlying population of possible industry (oa), corporate-parent(oB), and segment-specific effects. Once the effect is established, it remains effects (o2); and of the population covariance fixed for the period under study. The population between industry and corporate-parent effects of possible industry effects cannot be observed, (2CaB). Note that oz=[(1-p2)o2]. We use this but is of hes expression to state our results in terms of o industry effects are randomly drawn from this estimating the following equation mposed o2 by ce in r rumelt de population(we expand on this point below ).8 σ2+aa+ob+G++2Ca+o2 (6) generate one type ect are not correlate with the processes that generate other types of In Equation 6, the term o represents the popu- effects, with one exception. Following Schma- lation variance lensee and Rumelt, we allow for covariance year interaction effects between corporate-parent and industry effects. a The Cov method is sufficiently unusual to positive covariance would arise if attractive indus- merit further discussion. o The main idea is that tries generated more opportunities for positive each effect is treated as though it were generated influence by corporate parents, or if corporate by an independent, random draw from an underly- parents skilled at exploiting relationships between ing population of the class of effects. Once drawn business units were also effective at selecting each effect is considered as fixed. The assumption tractive industries in which to compete. A nega- of random effects does not stipulate that the tive covariance would arise if the opportunities Compustat data represent a random sample of for positive influence by corporate parents were business segments in the economy. The assump particularly great in unattractive industries tion merely means that the represented effects are We then decompose the variance of business- generated by random processes segment returns using Equation 5 To estimate Equation 5, we exploit relation ships in the sample variation among year, indus oR=(1+p2)ox+(1-p)2(o2+o2+o3) try, segment-specific and corporate-parent effects +2(1-p)2Cg+02 (5) For example, the sample variation among industry effects is the sum of an unbiased estimate of industry variation plus a small portion of the The dependent variable in this equation, OR, Is underlying variation in the year effects plus a the variance of Ri k, which is defined by ri k. ri. -pri k -I as the portion of the return to busi- small portion of the underlying variation in the by the shock in the prior year. Equation 4 underlying. variation in the segment-specific ness segments. Searle(1971), chapter 9, provides a detailed +(1-p)Var(a;+Bx)+(1-p)os+ discussion and several helpful examples of components-of- Var(a;+B)=0a+oB+2Cov(a, B.). We therefore hav ariance analysis. Also see Chamberlain (1984 approach. Rumelt(1991)provides an excellent discussion of les te ation, Random effects occur when obs intuitive terms(pp. 172-173)and develops an example of the raws from an underlying v- COV approach(pp 174-176). Abowd et al.(1995)develop a ble probability distribution variance among persistence rates that accounts We obtain Equation 5 from Ri,=(1-p)u+ for the endogeneity of exit decisions Rumelt's approach on the grounds that the Ftc data do not +w,x 1. Using our assumptions about the independ represent a random sample of the population. We believe that random effects, we reduce the equation to oR=0, this criticism is based on a misconception
20 A.M. McGahan and M. E. Porter cally along with corresponding equations in which zero. The null model stipulates that profits are the dependent variable is the sample variation in entirely determined by shocks that persist at the each of year, corporate-parent, and segment-speci- rate, p, and the economic mean. The next step is fic effects, and the right-hand side is a linear to obtain year effects by regressing the residuals combination of the underlying population vari- from the null model on the year dummies. An ances of the effects. 2 The result is a system of F-test provides an assessment of the importance equations in unknowns that represent the unbiased of year effects by comparing the R implied by estimates of population variances, By estimating p, the economic mean, and the year effects with the parameters in each equation (i. e, the 'por- the R2 from the null model and by accounting tions'described above) and solving the system of for the number of year dummies that Intro equations, we estimate the portion of the variance duced to achieve the additional explanatory attributable to each type of effect power In our second approach to estimation, we ana- After we show that year effects are important lyze the variance of profits under the standard we obtain industry effects by regressing the assumptions of ordinary least squares. Dummy residuals from the model (of ffects with variables represent year, industry, corporate-par- the economic mean and accounting for persistence ent and segment-specific effects. Instead of exam- at the rate p) on the industry dummies. The ining each of the coefficients on the dummy industry effects are used to obtain an R in a ariables, we examine the percent of variance model that also includes the year effects, the explained by the models(R2 and adjusted R2) economic mean, and the persistence in shocks at and evaluate F-tests to assess the importance of rate p. We conduct an F-test to evaluate whether groups of effects. In theory, Equation 4 is esti- the industry dummies add significant explanatory mable through simultaneous analysis of variance power. The procedure is then repeated for corpo- (ANOVA)methods. In practice, however, compu- rate-parent and segment-specific effects. After we tational complexity prevented us from obtaining complete the procedure, we repeat it under a a simultaneous estimate of all types of effects. different ordering to verify that the significance Following Rumelt, we therefore estimate the of each group of effects is not sensitive to the model using nested ANOVA techniques. 3 The order of introduction nested ANOVA allows us to evaluate whether There is controversy in the literature about the each group of effects (i.e, year, industry, seg- suitability of the COv and ANOVA approaches ment-specific or corporate-parent) is significant for this type of model. Schmalensee suggests that y introducing them in order. Under the approach, ANOVA results establish whether each set of we first evaluate the full model in Equation 4 to effects is significant, and that COV results are obtain an estimate of p. We then obtain a null preferable for evaluating the relative importance model by restricting all of the year, industry, of each type of effect. He seems to prefer the corporate-parent, and segment-specific effects to COV over the ANOVA results(perhaps because he finds the random-effects assumption more natural than the fixed-effects assumption ). The For example, the equation that represents the rclationship Cov approach does not generate any test of between the expected variance in the observed industry effects, significance, however, whereas the AnovA approach generates F-statistics. The COV Es=+a,0+aG+a+a3+a(2Cma)+a,appoachalsoincorporatesacontroversial assumption of independence in the random proc- tuition is based on the idea that the observed indust of independence does not allow for the endogen effects are calculated from data that include noise associated eity of relationships between the levels of effects with the draws of year effects, corporate-parent effects, seg- and subsequent entry or exit, for example. These amount of the noise' from each source is related to the issues apparently motivated Schmalensee to number of draws from each distribution Rumelt(1991: 174- include the ANOVA estimates along with his 176)provides a detailed example and shows how each of Cov analys See Searle (1987: chapter 3), for a detailed discussion of Rumelt(1991)makes a different assessment nested ANOVA of the two approaches. He argues that an anova