McHahan & Porter (1997) How Much Does Industry Matter

Strategic Management Journal, Vol.
18 (Summer Special Issue), 15–30 (1997)
HOW MUCH DOES INDUSTRY MATTER, REALLY?

ANITA M. McGAHAN* and MICHAEL E. PORTER
Graduate School of Business Administration, Harvard University, Boston, Massa-
chusetts, U.S.A.
In this paper, we examine the importance of year, industry, corporate-parent, and business-
specific effects on the profitability of U.S. public corporations within specific 4-digit SIC
categories. Our results indicate that year, industry, corporate-parent, and business-specific
effects account for 2 percent, 19 percent, 4 percent, and 32 percent, respectively, of the
aggregate variance in profitability. We also find that the importance of the effects differs
substantially across broad economic sectors. Industry effects account for a smaller portion of
profit variance in manufacturing but a larger portion in lodging/entertainment, services,
wholesale/retail trade, and transportation. Across all sectors we find a negative covariance
between corporate-parent and industry effects. A detailed analysis suggests that industry,
corporate-parent, and business-specific effects are related in complex ways.  1997 by John
Wiley & Sons, Ltd.
Debate in strategy has long focused on the performance has received scant empirical study,
sources of performance differences among firms. reflecting both the unavailability of data and chal-
In the research growing out of the industrial- lenging statistical difficulties. Rumelt (1991) is
organization tradition, industry structure is a cen- perhaps the most influential study. Rumelt’s
tral determinant of firm performance, and firm research followed methods introduced by Schma-
differences are considered against an industry lensee (1985) for disaggregating business-unit
background.1 More recently, a line of thought profits into components associated with industry
sometimes called the resource-based view argues effects, corporate-parent effects, and market-share
that firm performance is most influenced by effects. Neither Rumelt (1991) nor Schmalensee
unique organizational processes.2 Under this view, (1985) made claims about the economic or
industry structure is less important than idiosyn- organizational processes underlying their results;
cratic historical factors giving rise to firm differ- both papers were descriptive rather than norma-
ences. tive. Nevertheless some have interpreted Rumelt’s
Despite the importance of these questions, the finding of low stable industry effects to support
relative influence of firm and industry effects on the resource-based view.3
In this paper, we revisit the influence of indus-
try, business-specific, and corporate-parent influ-
Key words: profit components, firm performance, ences on profitability using comprehensive data
industry effects and enhanced statistical methods. We examine
* Correspondence to: Anita M. McGahan, Graduate School of the relative effects of these influences on prof-
Business Administration, Harvard University, Soldiers Field,
Boston, MA 02163, U.S.A. itability for the economy as a whole as well as
1
Porter (1980) and Oster (1990) are in the industrial-organi- in broad economic sectors. Finally, we begin to
zation tradition.
2
For discussion of the resource-based view, see Conner
(1991), Dierickx and Cool (1989a, 1989b), and Barney
3
(1986, 1989). For example, see Levinthal (1995: 20).
CCC 0143–2095/97/S10015–16 $17.50

 1997 by John Wiley & Sons, Ltd.
10970266, 1997, S1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/(SICI)1097-0266(199707)18:1+<15::AID-SMJ916>3.0.CO;2-1 by The University Of Manchester, Wiley Online Library on [10/06/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
16 A. M. McGahan and M. E. Porter
explore how the effects interact. Industry proves Our analysis differs from prior work in several
to have a powerful direct and indirect influence ways. First, we use recently compiled data from
on profitability. the Compustat Business Segment Reports for
1981 through 1994. This dataset covers activity
in all sectors of the American economy (except
ANTECEDENTS the financial sector), whereas the prior studies
cover only manufacturing. The breadth of cover-
Schmalensee (1985) examined the accounting age provides not only a representative sample on
profits of American manufacturing firms that were the economy but also allows examination of profit
covered in the Federal Trade Commission’s Line influences across sectors. The average time series
of Business Report for a single year, 1975. He on each economic unit in our dataset is 5.7 years,
found that industry effects accounted for about which compares favorably with the 4-year series
20 percent of variation in business-unit profits on each business unit in Rumelt’s data. Because
(and nearly 100% of total variance explained), our dataset covers a 14-year period, our results
and that corporate-parent effects (or ‘firm effects’, reflect several phases of the business cycle.
in his terminology) had no impact on variation. Second, we show how the results are affected
Schmalensee’s only measure of heterogeneity by a more robust statistical approach to inter-
among participants in the same industry was mar- temporal persistence. Rumelt’s specification
ket share. He reported that share positively affec- allows for transient industry effects, but does not
ted business-unit profits, but only by a negli- similarly allow for transient year, corporate-par-
gible amount. ent, or business-unit effects. Our specification
Rumelt (1991) extended Schmalensee’s allows for transience in all effects, and we report
approach by including data from the FTC Reports the effect of the difference in method.
on manufacturing firms for all available years— Third, our unit of analysis differs. The Compu-
1974 through 1977. With data on more than one stat Reports contain information on firm profit by
year, Rumelt generalized Schmalensee’s measure SIC code (i.e., by business segment), not by
of intraindustry heterogeneity to all business-unit business unit. Schmalensee and Rumelt examined
effects rather than just market-share effects. He the business-unit returns given in the FTC data.
reported that business-unit effects explain 44–46 We believe that the average business segment
percent of variation (about 73% of the explained covers the activity of several business units. All
variation), stable and transient industry effects else equal, the diversity of business-unit activity
account for a total of 9–16 percent of variation, attributed to a single 4-digit SIC code may arti-
and corporate-parent effects explain 1–2 percent ficially reduce the measured influence of industry
of variation. It is these results—the relatively low relative to Schmalensee and Rumelt. Moreover,
proportion attributed to industry effects compared our need to rely on the SIC system for industry
with business-unit effects—that have been inter- classification further diminishes the measured esti-
preted to support the resource-based perspective.4 mates of industry influence because SIC industries
err primarily in being overly broad. In our dis-
4
Rumelt’s report of low corporate-parent influence is not cussion, we suggest that the influence of industry
consistent with a resource-based view of diversification, and might be even stronger if data of finer grain
has stimulated additional research. In a study of diversified were available.
firms from the Compustat Business Segment Reports, Roque-
bert, Andrisani, and Phillips (1996) challenge Rumelt’s result Like Rumelt, our specification includes a num-
on corporate-parent effects. The authors find that corporate- ber of potential sources of variation in accounting
parent effects are significantly more important than indicated returns: yearly macroeconomic fluctuations,
by Rumelt (1991). The Roquebert et al. study is not directly
comparable to Rumelt’s (1991) or to ours because Roquebert industry factors, corporate-parent effects, and seg-
et al. exclude single-business firms from their analysis. This ment-specific effects. This last category, segment-
exclusion means that estimates of industry effects are con- specific effects, encompasses all business-segment
structed from the performance of only diversified firms. In
our Compustat Business Segment data, single-business firms
account for half of all assets. When we exclude single-
business firms from our data set, the estimated influence of covariance between industry and corporate-parent effects. This
industry decreases substantially. Low estimates of industry tendency may compound an upward bias in the estimated
influence under the Roquebert et al. approach may distort corporate-parent influence that arises from the exclusion of
estimates of corporate-parent influence because of negative all single-business firms.
How Much Does Industry Matter, Really? 17
differences, including diversity in market share, ventions may influence all four types of effects
differentiation, heterogeneity in fixed assets, dif- on profitability (i.e., year, industry, corporate-
ferences in organizational processes, differences parent, and segment-specific).5 Because we have
in organizational effectiveness, heterogeneity in no a priori hypothesis about the nature and direc-
activity configurations, anomalies in accounting tion of these biases, and because the Compustat
practices, and differences in managerial com- Business Segment Reports are the best source of
petence. Our objective is to understand the rela- available data on profitability, we proceed with
tive significance of industry, corporate-parent, and the analysis but interpret the results with caution.6
segment-specific differences in explaining profit Our specification differs from Schmalensee’s
variation when industries are defined by the (1985) in several ways. Because Schmalensee
SIC system. had only one year of data, his analysis excluded
both the year effect (gt ) and the segment-specific
effect (fi,k ). The segment-specific effect can only
METHODS be identified when multiple years of data are
available on each segment because only multiple
Our analysis relies on the following model, which years identify when a segment’s performance dif-
draws on the models used by Schmalensee and fers systematically from the mean given the
Rumelt: simultaneity of year, industry, and corporate-
parent effects. Instead, Schmalensee included
ri,k,t = m + gt + ai + bk + fi,k + e i,k,t (1) measures of market share that had been developed
by David Ravenscraft for an earlier study on the
In this equation, ri,k,t is the accounting profit in FTC data (Ravenscraft, 1983).
year t of corporate-parent k’s business in industry Our model also differs from Rumelt’s (1991),
i. Profit is measured as the ratio of operating which is reproduced as Equation 2:
income to identifiable assets in percent. The first
right-hand-side term is m, which is the average ri,k,t = m + gt + ai + bk + dit + fi,k + ei,k,t (2)
profit over the entire period for all business seg-
ments. The second term is gt, which represents Rumelt’s model includes an additional term to
the difference between m and the average profit represent industry–year interactions, dit. By
of all business segments in year t. The next three including both ai and dit, he distinguishes ‘stable’
terms represent industry, corporate-parent, and industry effects from ‘transient’ industry effects.
segment-specific effects. The term ai is the Transient industry effects occur when all mem-
increment to profit associated with participation bers of an industry have high or low profits
in industry i; bk is the increment to profit con- in year t.
ferred by membership in a diversified corporate- Rumelt proceeds by assuming that the error in
parent k; and fi,k is the increment to profit Equation 2 is drawn independently. In making
associated with the specific situation of business this assumption, he suppresses the possibility that
segment i,k given the other effects. We assume a shock to the year, corporate-parent or business-
that a corporate-parent effect arises only if a specific effect at time t − 1 influences the year,
business segment is a member of a diversified corporate parent or business-specific effect at time
firm. The final term, ei,k,t, is the residual. Any of t. Suppose, for example, that a specific segment
the increments to profit may be positive or nega- has an unusually good year at time t − 1. Rumelt’s
tive. The model is estimated using dummy vari- specification does not account for the possibility
ables to represent industry, corporate-parent, and
segment-specific effects.
Our study is limited by shortcomings in 5
For example, consider the ethical-pharmaceutical industry,
accounting measures of profit. Because account- which is relatively research-intensive. The industry effect for
ethical pharmaceuticals reflects idiosyncratic conventions in
ing conventions exclude intangible assets from accounting for research that similarly affect all members of
the balance sheet, measured assets may be too the industry.
6
low for some segments. The use of operating Powell (1996) uses executives’ perceptions instead of
accounting profit to assess the influence of industry. He finds
income excludes the effects of differences in that industry accounts for about 20 percent of performance
financing. Measurement error and accounting con- variation among the 54 single-business firms in his survey.
of a spillover effect on the segment in year t. parent, or segment-specific. The error, vi,k,t, for
By including industry–year interaction dummies, which we assume independence, is the portion of
however, his model does capture spillovers that the transient shock that is not influenced by the
affect all the members of an industry. Now sup- shock in the prior year. This specification accom-
pose that the business cycle generates unusually modates both new shocks and spillovers from the
high year effects in successive years. Rumelt’s prior year, although it cannot capture differences
industry–year interaction term may partly capture in the rate at which shocks resound across years
the influence of the business cycle, which would (Rumelt’s model can capture differences in the
be attributed to persistence in the year effect in rate of persistence in industry shocks). We
a complete model. Rumelt justified his approach acknowledge this deficiency, but argue that
by reporting no autocorrelation in residuals from changes in the rate of persistence are important
his estimation. Nonetheless, this justification does in the second order whereas the simple presence
not address the possibility that industry–year of persistence in year, corporate-parent and seg-
interaction effects may proxy for persistence in ment-specific shocks is important in the first
year, corporate-parent, and business-specific order. As a result of this difference in speci-
effects in his specification. This possibility is fication, our principal estimates are comparable
salient given that Rumelt’s data cover the period only with the stable effects in Rumelt’s work.
immediately subsequent to the 1973 oil shock It is important to note that ai, bk, and fi,k
and to the removal of wage and price controls describe how a business segment is affected in
under the Nixon administration. all years by its industry, corporate parent, and
Although we appreciate the benefits of model- segment-specific situation. The rate of persistence,
ing transient industry effects, we exclude them r, reflects the influence of a shock in any single
because the model would be overspecified if we year on the performance in just the subsequent
equally represented transient year effects, transi- year. To isolate the portions of effects that are
ent corporate-parent effects, and transient busi- stable, we subtract from (1) the rate of persist-
ness-specific effects. This point is important to ence, r, multiplied by the lagged value of ri,k,t:
the differences in our econometric model com-
pared with Rumelt’s. In Rumelt’s view, an asym- ri,k,t = rri,k,t−1 + (1 − r)m + gt − rgt−1
metry in treatment of industry effects is justified
+ (1 − r)ai + (1 − r)bk + (1 − r)fi,k
when the data cover a relatively short period
because corporate-parent and business-specific + vi,k,t (4)
effects will not change much (i.e., when shocks
are small so that the persistence of shocks The left-hand side of this equation is the same
between years is not important). However, transi- as in (1): it is the profit to business segment i,k
ence may arise at any level, and it is at least at time t in percent. The first term on the right-
plausible that industry effects will change slower hand side is the rate of persistence multiplied by
than business-specific or corporate-parent effects. the profit to the same business segment at time
Indeed, in another paper (McGahan and Porter, t − 1. In calculating lagged variables, we lose data
1997) based on the same data set but somewhat for the first year for which we have information
different methods, we show that shocks to busi- on each segment. The other terms on the right-
ness-specific and corporate-parent effects may be hand side include the year, industry, corporate
larger than industry shocks. parent, and segment-specific effects.
To deal with the possibility that a shock in We analyze this model in two ways, following
year t − 1 might influence profits in year t, we Schmalensee and Rumelt. First, we conduct a
allow for serial correlation on the errors in components-of-variance (COV) estimation under
Equation 1 according to the following process: the assumption that random processes generate
each of the effects in Equation 1.7 Consider, for
ei,k,t = r e i,k,t−1 + vi,k,t (3)
7
This assumption is completely separate from our prior dis-
The parameter r captures the intertemporal per- cussion of stable and transient effects. Here, we are describing
the technical assumptions by which we estimate the model
sistence of effects regardless of source: in Equation 4 given that the model includes only stable effects,
macroeconomic fluctuations, industry, corporate- year effects, and the error. At this point, we are interested in
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, r; and
effect is drawn randomly at some early date from of the population variances in year (sg2 ), industry
an underlying population of possible industry (sa2 ), corporate-parent (sb2 ), and segment-specific
effects. Once the effect is established, it remains effects (sf2 ); 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 sv2 = [(1 − r2 )s2e ]. We use this
but is of primary interest because it determines expression to state our results in terms of s2r, the
the importance of industry. Because the observed total variance in ri,k,t. Rumelt decomposed s2r by
industry effects are randomly drawn from this estimating the following equation:
population, they may not perfectly represent the
population (we expand on this point below).8 s2r = s2g + s2a + s2b + s2d + s2f + 2Cab + s2e
We also assume that the random processes that (6)
generate one type of effect are not correlated
with the processes that generate other types of In Equation 6, the term s2d represents the popu-
effects, with one exception. Following Schma- lation variance of the distribution of industry–
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. 10 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
attractive 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.11 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-
s2R = (1 + r2 )s2g + (1 − r)2 (s2a + s2b + s2f ) try, segment-specific and corporate-parent effects.
For example, the sample variation among industry
+ 2(1 − r)2Cab + s2v (5)
effects is the sum of an unbiased estimate of
industry variation plus a small portion of the
The dependent variable in this equation, sR2, is
underlying variation in the year effects plus a
the variance of Ri,k,t, which is defined by Ri,k,t =
small portion of the underlying variation in the
ri,k,t − rri,k,t−1 as the portion of the return to busi-
corporate-parent effects plus a portion of the
ness segment i,k at time t that is not influenced
underlying variation in the segment-specific
by the shock in the prior year.9 Equation 4
effects. This equation may be expressed analyti-
the processes that make the stable effects differ across busi-
ness segments. Searle (1971), chapter 9, provides a detailed + (1 − r)2Var(ai + bk ) + (1 − r)2sf2 + sv2. Note that
discussion and several helpful examples of components-of- Var(ai + bk ) = sa2 + sb2 + 2Cov(ai ,bk ). We therefore have
variance analysis. Also see Chamberlain (1984: 1254–1270; sR = (1 + r )sg + (1 − r) (sa + sb + sf) + 2(1 − r) Cab + sv2.
2 2 2 2 2 2 2 2
10
Chow (1984), Griliches (1984), and Rumelt (1991). Searle (1971) provides a comprehensive discussion of the
8
Note that the assumption of random effects is somewhat approach. Rumelt (1991) provides an excellent discussion of
different than an assumption of random sampling. Random how the method applies to a simple equation akin to the one
samples arise when observations are independently drawn here. He also describes the random-effects assumption in
from a population. Random effects occur when observations intuitive terms (pp. 172–173) and develops an example of the
are generated through draws from an underlying and unobserv- COV approach (pp. 174 –176). Abowd et al. (1995) develop a
able probability distribution. model of covariance among persistence rates that accounts
9
We obtain Equation 5 from Ri,k,t = (1 − r)m + g t − rgt−1 for the endogeneity of exit decisions.
+ (1 − r)ai + (1 − r)bk + (1 − r)fi,k + vi,k,t. This expression 11
There has been some criticism of Schmalensee’s and
gives s2R = Var[gt − rgt−1 + (1 − r)ai + (1 − r)bk + (1 − r)fi,k Rumelt’s approach on the grounds that the FTC data do not
+ vi,k,t ]. Using our assumptions about the independence of represent a random sample of the population. We believe that
random effects, we reduce the equation to s2R = sg2 + r2s2g this criticism is based on a misconception.
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, r, 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.12 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 R2 implied by
estimates of population variances. By estimating r, 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 were 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 year effects 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 R2 in a
variables, 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 r. 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.13 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
by 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 r. 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
12
For example, the equation that represents the relationship COV approach does not generate any test of
between the expected variance in the observed industry effects, significance, however, whereas the ANOVA
Es2i , and the population variances is:
approach generates F-statistics. The COV
Es2i = s2a + ags2g + a bs 2b + ads2d + a fs 2f + aC(2Cab) + a es2e approach also incorporates a controversial
assumption of independence in the random proc-
where ag, a b, a d, af, aC and ae each represent complex ratios esses that generate the effects. The assumption
that account for number of draws on each type of effect. The
intuition is based on the idea that the observed industry 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
ment-specific effects, covariance effects, and errors. The
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 analysis.
these ratios is obtained.
13
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
test for significance is not a prerequisite to COV drop 2743 records that do not contain a primary
estimation. By Rumelt’s logic, a COV analysis SIC designation. A total of 22,041 segments are
is a simple statistical description of the data, and excluded because they are in SICs identified as
he offers it as his flagship approach. Rumelt ‘not elsewhere classified,’ ‘nonclassifiable estab-
includes an ANOVA estimation because it has lishments,’ or ‘government, excluding finance.’
independent merit as a method for estimating the We also drop segments designated as ‘depository
importance of effects. He suggests that further institutions’ (15,689 financial business segments
research on assumptions would be warranted if with SICs in the 6000s) because returns are not
the two methods generate results that are very dif- comparable with those in other industries. We
ferent. exclude 2529 segments that are the only organi-
We subscribe to Rumelt’s logic, but add two zations covered by Compustat in their primary
additional words of caution. First, the nested SIC classifications in specific years (analogous to
ANOVA analysis is inherently imprecise because monopolies) because we cannot distinguish their
it largely attributes covariance between types of industry effects from their segment-specific
effects to the first effect introduced. In contrast, effects. Another 1433 observations are excluded
the COV approach is based on the assumption because they are associated with segments that
that effects are independently generated; for are in Compustat for only one year. We then
example, an incidence of exit by a segment is exclude 29,077 very small segments with sales
assumed to be unrelated to the industry effect for less than $10 million and an additional 5675
the segment. Although the COV approach is lim- segments with assets less than $10 million.
ited for this reason, we have no a priori hypoth- Single-year appearances and small segments are
esis that the COV analysis is biased, and hence often anomalous because they are created for the
we also offer it as our flagship approach. We also disposition of assets prior to exit, for example.
concur with Rumelt’s suggestion that qualitatively The exclusion of small segments is comparable
important differences in results should motivate with Schmalensee’s (1985) exclusion of units that
further research on the appropriateness of the account for less than 1 percent market share in
assumptions in each model. his FTC Line of Business data set.
Our screened data set includes 72,742 obser-
vations, or an average of 5196 business segments
DATA per year. This figure is substantially larger than in
previous studies. Schmalensee’s dataset included
The Compustat Business Segment Reports include 1775 observations. Rumelt ran his analysis on
information on companies with equity that is two datasets, which he labeled Sample A and
publicly traded in American markets. For each Sample B. Sample A excluded the small business
corporate parent, the Compustat Reports identify units that Schmalensee had excluded, and
up to 10 lines of business because SEC guidelines Sample B did not. Rumelt had 6932 observations
require the reporting of information on segments in Sample A and 10,866 observations in
that comprise 10 percent or more of the parent’s Sample B. Because his results for the two data-
business. Each line is identified by a segment sets were similar, we focus our comparison with
number, which allows tracking of performance his Sample A to accommodate a simultaneous
between years even if the name or primary SIC comparison with Schmalensee.
of the segment changed. For each business seg- The raw Compustat Business Segment data
ment, the data set contains a primary 4-digit SIC (after screening only for missing observations and
code, operating income, sales, and identifiable for financial firms) account for about two-thirds
assets. We used Compustat’s conventions for of the corporate sales and 45 percent of the
dealing with the SIC revisions in 1981, 1987, corporate assets reported to the Internal Revenue
and 1992. Service for nonfinancial sectors from 1985 to
We screened the Compustat data base in sev- 1992, the last year for which data are available.
eral ways. Before screening, the data set con- After the application of our screens, the data
tained 151,929 records, each of which described cover slightly more than half of corporate sales
a single business segment in a particular year and slightly more than a quarter of corporate
between 1981 and 1994. From this dataset, we assets in nonfinancial sectors. Schmalensee
reports that the FTC data in his study accounted tries, corporate parents, and business units in
for about half of manufacturing sales and two- varying degrees. Our results are less vulnerable
thirds of manufacturing assets in 1975. Thus, to anomalies because the entire period of our
our analysis covers a comparable percentage of study is longer.
competitive activity but over a much broader A third advantage of the Compustat data is
range of economic sectors. that it captures a large portion of activity in
In our analyses, we require that each obser- all sectors of the American economy. Whereas
vation include information on lagged perform- Schmalensee’s and Rumelt’s studies focused on
ance. Thus, the first observation on each business the manufacturing sector, our study covers the
segment is excluded. Our COV and nested retail sector, wholesale services, mining and
ANOVA results are therefore based on 58,132 agriculture, food and textiles production, chemical
remaining observations. The screened dataset rep- businesses, transportation services, lodging and
resents the activities of 12,296 distinct business other services, entertainment, and all other indus-
segments in a total of 628 different industries, tries except those in the financial and government
which are represented by their 4-digit SIC codes. sectors. This difference also means that we have
The average business segment posts 5.7 years of many more observations than previous authors.
data (including lagged information for the first In our final report of results, we exploit the
observation). Each industry includes the activities variety in the Compustat data to show how results
of 7.7 business segments in the average year and differ by economic sector.
21.3 business segments on average over the entire There is evidence that the business segments in
period. Our analysis covers 7003 corporations, of our data set are considerably larger than operating
which 1791 participate in more than one industry business units. The average segment in our
in at least one year for which we have data. screened data base has assets of $903 million,
Slightly less than half our observations are and diversified corporate parents post information
associated with diversified corporate parents. on 2.6 segments on average. Montgomery
Table 1 describes the business segments in the (1994: 164) indicates that the Fortune 500 partici-
screened data base by year and by economic pated in 10.65, 10.85, and 10.90 different SICs
sector. The mean profit is 9.3 percent with a on average in 1985, 1989, and 1992, respectively.
variance of 248 percent. Thus, it is likely that a typical Compustat seg-
There are several potent advantages to the ment, which is characterized by a 4-digit SIC
Compustat data. First, the 5.7-year time series on code, actually reflects operating activity in several
each segment allows us to identify those effects related 4-digit SIC codes.15 As a consequence,
that are stable over a somewhat longer period the operations posted to each SIC in the Compu-
than the Rumelt study. Any measure of stability stat data are probably more diverse than the actual
is integrally related to the number of years of operations in each SIC.
data on each economic unit. A longer time series There are also some disadvantages to the Com-
inherently leads to lower estimates of stability in pustat data. The broadening of industry definition
effects. We believe that the 5.7-year average pro- beyond the actual 4-digit level probably tends to
vides a useful benchmark. dampen industry and corporate-parent effects, and
A second advantage of the Compustat data is potentially to distort segment-specific effects.
that it covers a longer period of time: 14 years These problems are exacerbated because the SIC
vs. 4 years for the FTC dataset. The longer period system does not map closely to strategically dis-
allows us to measure the influence of various tinct industries in some cases. As a result of
effects over several phases of the business cycle.14 these limitations, our results must be interpreted
Although Rumelt’s year effects surely capture with caution. A finding of high segment-specific
part of the impact of these macroeconomic con- effects with low industry and low corporate-par-
ditions, he does not have the latitude to examine ent effects may reflect aggregation in Compustat
whether the unusual conditions influenced indus-
15
Seven out of our 628 industries are 2-digit, and 50 are 3-
14
The FTC dataset covers an unusual period in economic digit. The 2-digit industries account for 1.9 percent of assets
history immediately subsequent to the first oil shock and in the dataset, and the 3-digit industries account for another
Nixon’s wage and price controls. 4.7 percent of assets.
Table 1. Screened Compustat Business-segment Data
(A) By year
Avg. Avg.
assets Avg. Median assets Avg. Median
No. ($mil) profita profit No. ($mil) profita profit
1982 5200 528 11.3 10.9 1989 5030 948 9.8 9.4
1983 5285 556 11.4 10.8 1990 5029 1028 9.5 9.0
1984 5205 598 11.9 11.6 1991 5114 1040 8.4 8.4
1985 5195 660 10.3 10.3 1992 5232 1051 9.0 8.7
1986 5249 699 9.2 10.0 1993 5396 1127 9.1 8.7
1987 5319 772 10.0 9.8 1994 5733 1161 9.1 9.0
1988 5112 865 10.4 9.9
(B) By sector
No. of No. of Avg. assets

First digit represented represented per segment Avg. profit
of SIC Brief description of sector SICs segments ($mil) per segment a
0,1,2 Agriculture, Mining 203 3661 966 11.2

3 Manufacturing 219 4068 677 8.1
4 Transportation 39 1651 2006 8.9
5 Wholesale & Retail Trade 91 1768 450 10.4
7 Lodging & Entertainment 53 987 315 9.2
8 Services 23 492 279 10.3
a
Average ratio in percent of operating income to identifiable assets.
rather than real economic differences. This prob-

lem is related to the general question of appropri-
ate industry definition.
Table 2. COV results developed from estimates of
Equation 5 a
EMPIRICAL RESULTS
Percent of
In this section, we present our estimates of total variance
Equation 5 through COV methods and of
Equation 4 through nested ANOVA methods. Year (s2g) 2.39
Industry (sa2 ) 18.68
Table 2 shows the COV estimation of Equation 5. Corporate parent (s2b) 4.33
Results are expressed as a percent of s2r, the total Segment specific (sf2 ) 31.71
variance in business-segment profits. To trans- Corp.-par.–industry covariance (2Cab) −5.51
form our COV decomposition from sR2 to s2r, we Model 51.60
use an estimate of r equal to 0.3777. This esti- Error 48.40
mate is obtained from the nested ANOVA analy- Total (s2r ) 100.00
sis and is described in detail below.
The results in Table 2 indicate that 51.60 per- a
Results are expressed as a percent of total variance, s2r ,
cent of the total variance in business-segment rather than of sR2 using an estimate of r equal to 0.3777.
profits is explained by the model. The error,
which equals 48.40 percent of the total variance,
arises because business-segment profits are sub-
ject to shocks, a portion of which may carry market-share effects are small but significant; and
from one year to the next. that corporate-parent effects are not significant.
About 2 percent of the variance in profits is Rumelt’s results for his Sample A are repro-
associated with year effects. By definition, these duced in the second column. They indicate that
effects are macroeconomic fluctuations that affect business-unit effects account for 46.37 percent of
all business segments to the same degree in a total variance, and that stable industry effects
particular year. account for only 8.32 percent of variance. An
Nearly 19 percent of variance is attributable to additional 7.84 percent is explained by industry–
stable industry effects. This result provides strong year interaction (also called ‘transient industry
support for the idea that industry membership effects’). Rumelt originally estimated the variance
has an important influence on profitability. The in year effects as a small negative number, and
estimate is much higher than Rumelt’s stable the industry–corporate-parent covariance as a
industry effect (8.32%) and is comparable with small positive number. As he notes, the negative
Schmalensee’s result (19.59%). estimate of the variance of year effects may
Stable effects of corporate-parent membership be evidence of specification error; however, the
account for nearly 4 percent of the variance in absolute value of the estimate is so close to zero
business segment profit. Although this estimate that it may be safely ignored. The impact of both
is absolutely low, it is higher than in prior the year variance and the industry–corporate-par-
research and suggests that the effects of corporate ent covariance are held to zero in the estimates
parents may have been greater during the 1980s reported in Rumelt’s Table 3 and reproduced here.
and early 1990s than in previous decades. The The third column of Table 3 is our estimate of
negative covariance between corporate-parent and Rumelt’s model, Equation 6, on the segments in
industry effects is consistent with the idea that our screened Compustat data classified as manu-
corporate parents have a greater positive influence facturing. Stable and transient industry effects in
when they participate in unattractive industries. the Compustat manufacturing data are modestly
Stable segment-specific effects account for lower than in Rumelt’s study, the influence of
nearly 32 percent of the variance in business- corporate-parent effects is higher, and a negative
segment profitability. Our estimate of stable seg- corporate-parent–industry covariance arises. The
ment-specific influence is lower as a percent of most striking difference between columns (2) and
total variance and of explained variance than (3) is the 33.79 percent of variance attributable
Rumelt’s business-unit effects, perhaps because to segment-specific effects in the Compustat
of differences in method, in the unit of analysis, manufacturing data vs. the 46.37 percent of vari-
and in the average length of the series on each ance attributed to business-unit effects in
segment. A comparison with Schmalensee’s Rumelt’s study.
model indicates that market share is a poor proxy Because Rumelt’s study covers only manufac-
for the segment-specific effect. turing firms, differences between columns (2) and
It is apparent that differences both in data and (3) arise from the unit of analysis, the average
in method account for the differences between length of the series, and the period from which
our results and those of previous authors. To the data are drawn. Problems in the SIC system
obtain some insight into the relative importance for manufacturing probably lower the influence
of these two causes, we analyzed Rumelt’s model of industry effects in column (3). The difference
on our manufacturing data and on our entire between the results in columns (2) and (3) also
screened Compustat dataset. Table 3 compares may be affected by anomalies in the FTC dataset
Schmalensee’s and Rumelt’s reported results with because of the 1973 oil shock and the removal
our estimates. of price and wage controls from the Nixon admin-
The first column of Table 3 is a reproduction istration. Conditions in the macroeconomic
of Schmalensee’s results. Because Schmalensee environment may have exacerbated differences in
analyzed one year of data, the question of transi- business-unit positions to a greater extent than
ent effects is not relevant, and his results are differences between industries. The unusual
directly comparable to ours. His principal findings macroeconomic conditions also may have gener-
are that industry effects account for about 20 ated an unusually large variance in the industry–
percent of variation in business-unit return; that year interaction effects.
Table 3. Comparison of COV results (percent of total variance attributed to various effects)
(1) (2) (3) (4) (5)

Rumelt model
on McGahan
and Porter Rumelt model McGahan
Schmalenseeb Rumelt c manufacturing on all McGahan and
(as reported) (as reported) data and Porter data Portera
No. of observations 1775 6932 18,298 58,132 58,132

Year (sg2 )
Industry (sa2 ) N/A N/A 0.40 0.37 2.39
Corporate parent (sb2 ) 19.59 8.32 7.20 17.32 18.68
Segment specific (sf2 ) N/A 0.80 2.05 6.96 4.33
Business unit (sf2 ) N/A N/A N/A N/A 31.71
Corp.-par.–ind. N/A 46.37 33.79* 29.57* N/A
covariance (2Cab ) −0.62 N/A −1.42 −5.37 −5.51
Market share 0.62 N/A N/A N/A N/A
Industry–year N/A 7.84 4.44 4.39 N/A
MODEL 19.59 63.33 46.46 54.23 51.60
Error 80.41 36.87d 53.54 46.77 48.40
TOTAL (s2r ) 100.00 100.00 100.00 100.00 100.00
*Based on business segments rather than business units.

a
Results in column (5) are reproduced from Table 2.
b
Results in this column are reproduced from Schmalensee’s Table 1 (1985: 348 ). Schmalensee’s model is a variant of
Equation 6 with an additional term for market share.
c
These results apply to Rumelt’s data set A and are based on his Table 3 (Rumelt, 1991: 178); Rumelt’s Sample A excludes
the same small business units that Schmalensee had excluded in his analysis of the FTC data.
d
We calculate that this number, which is reproduced directly from Rumelt’s Table 3, should be 36.67 given Rumelt’s reports
of the actual values of his estimates.
The fourth column of Table 3 shows our esti- disaggregate the Compustat data to the business
mate of Rumelt’s model, Equation 6, on our entire unit level, it is likely that the figures for stable
screened Compustat dataset. If we had used industry effects in columns (4) and (5) would be
Rumelt’s model, we would have found that stable even greater than 17.32 percent and 18.68 per-
industry effects account for 17.32 percent of vari- cent.
ance and that transient industry effects account Before conducting the estimation, we were con-
for 4.39 percent of variance. Although we would cerned that our unit of analysis would suppress
have attributed a greater portion of variance to corporate-parent effects because of aggregation
corporate-parent effects (6.96%) than in Rumelt’s across business units. The results in columns (3),
study (0.80%), our attribution of variance to seg- (4), and (5) indicate a greater influence of corpo-
ment-specific effects (29.57%) would have been rate parents than in previous studies. This out-
significantly lower than Rumelt’s attribution to come might reflect general changes in the climate
business-unit effects (46.37%). for diversification. During the 1970s, unrelated
A comparison of columns (3) and (4) provides conglomerates were probably more prevalent than
support for the idea that differences in sectoral during the 1980s and early 1990s.16
coverage influence the results. The Compustat A comparison of the results in columns (4)
dataset covers activity in all sectors of the econ- and (5), which differ only in the employed model,
omy, whereas the FTC data set used in Rumelt’s is also instructive. The percent of variance
study covers only manufacturers. The greater explained by stable industry effects in column
diversity of industries covered in the Compustat (5) is not much greater than the stable industry
data generates higher industry effects. This ten-
dency is probably offset in part by differences in 16
See Wernerfelt and Montgomery (1988) for a study of
the unit of analysis. If we had been able to corporate-parent effects that accounts for relatedness.
effects in column (4). Also note that the portion in the services sector (−2.66%). Negative vari-
of variance attributed to the corporate-parent ance estimates may be evidence of specification
effect is lower in column (5), but that the portion error, or they may reflect unimportant anomalies
attributed to segment-specific effects is larger. in the data. Concerns about data made us least
The suppression of persistence in segment-specific confident about COV results for the service sector
and corporate-parent effects in Rumelt’s model because the Compustat data excludes the many
also results in a smaller error in column (4). service-sector firms that are privately owned and
These comparisons suggest that the difference not traded on major financial exchanges. Follow-
in method affects the results in an important way. ing Rumelt, we constrain the negative estimates
When industry–year interaction is excluded from of variance to zero and report results in Table 4
the decomposition, the variance attributed to year under the assumption that the negative estimates
effects and to segment-specific effects increases. are too low to be significant.
This supports the possibility that the interaction There are several striking patterns in panel B
term in Rumelt’s model may indeed proxy for of Table 4. The first is a remarkable variation in
persistence in other types of effects. The the importance of industry effects. In
exclusion of the industry–year interaction is also wholesale/retail, lodging/entertainment, and ser-
associated with a lower attribution of variance to vices, industry accounts for over 40 percent of
corporate-parent effects. This change indicates the the variance in profitability. In agriculture/mining
possibility of strong covariance between the per- and transportation, industry accounts for 39.50
sistence of industry and corporate-parent effects, percent and 29.35 percent, respectively, of vari-
which is consistent with the strong covariance in ance. Manufacturing is the outlier, with industry
the stable effects reported in the table. We cannot accounting for just 10.81 percent of variance in
test for these relationships in this study, however, profitability (and 23.3% of explained variance).
without overspecifying the model. In addition, manufacturing is characterized by the
Table 4 shows the COV analysis by broad eco- lowest total explained variance of the sectors.
nomic sector. The first section of the table, panel Second, corporate-parent effects also vary
(A), provides information on the data. The greatest markedly in impact. For manufacturing and ser-
number of observations are present in the first two vice segments, corporate parents have no direct
sectors: agriculture and mining,17 and manufactur- influence on variance in profitability. In
ing. Over 200 different 4-digit industries fall within wholesale/retail, however, variance in corporate-
these two sectors. About half as many observations parent effects contributes more to the model than
are associated with the transportation and variance from any other source. This may reflect
wholesale/retail sectors, and somewhat fewer are more pervasive opportunities for cross-unit
associated with the lodging/entertainment and ser- relationships in wholesale/retail than in the more
vices sectors. Separate rates of persistence in shocks variegated manufacturing and service sectors.
are obtained for each sector through nested Corporate-parent effects are also large in transpor-
ANOVAs by type of effect for each sector. Persist- tation and agriculture/mining, perhaps for simi-
ence is greatest in the manufacturing and lar reasons.
lodging/entertainment sectors, with wholesale/retail Third, variance in segment-specific effects is
trade, agriculture/mining, and transportation show- more important in manufacturing than in any
ing moderate persistence rates. Persistence is lowest other sector. On average, manufacturing may
in the service sector. offer richer possibilities for sustainable posi-
Panel (B) of Table 4 shows the results. The tioning than other sectors, a possibility also sup-
analysis generated three negative estimates of ported by the results of the previous studies. The
variance: for the year effect in the only other sector in which segment-specific
lodging/entertainment sector (−0.43%); for the effects are comparable in importance is services.
corporate-parent effect in the manufacturing sec- In the agriculture/mining, wholesale/retail and
tor (−1.98%); and for the corporate-parent effect transportation sectors, variance in segment-speci-
fic effects contributes relatively little to total vari-
17
ance in profitability. In agriculture/mining and
We pooled the agricultural and mining sectors in the analy-
sis because there is apparent overlap in the SIC system transportation, commoditization and regulatory
between the sectors. influences may dampen interfirm differences.
Table 4. COV by economic sector
W’sale,
Agriculture, retail Lodging
All mining Mfrs. Transp. trade and ent. Services
(A) Characteristics of the data

SIC designation 0,1,2 3 4 5 7 8
No. observations 58,132 16,829 18,298 9221 7937 4016 1831
No. industries 628 203 219 39 91 53 23
No. segments 12,296 3,661 4,068 1651 1768 987 492
No. diversified corp.’s 1,791 920 836 428 432 225 122
Total corporations 7,003 2,131 2,432 1164 1383 827 400
No. years/segment 5.7 5.6 5.5 6.6 5.5 5.1 4.7
Average profita 9.3% 11.2 8.1 8.9 10.4 9.2 10.3
Est. persistenceb 37.77% 29.91 40.52 26.90 34.31 44.20 14.03
(B) COV analysis developed from estimates of Equation 5c
Year (sg2 )
Industry (sa2 ) 2.39 2.25 2.34 3.25 2.64 N/A 4.17
Corporate parent (sb2 ) 18.68 29.35 10.81 39.50 41.79 64.30 47.37
Segment specific (sf2 ) 4.33 22.35 N/A 28.33 44.06 14.71 N/A
Corp.-par.–industry 31.71 5.02 35.45 9.72 2.04 19.41 33.46
covariance (2Cab ) −5.51 −9.45 −2.27 −16.49 −20.24 −29.80 −23.98
MODEL 51.60 49.52 46.33 64.31 70.29 68.63 61.02
Error 48.40 50.48 53.67 35.69 29.71 31.37 38.98
TOTAL (s2r ) 100.00 100.00 100.00 100.00 100.00 100.00 100.00

a
Average ratio in percent of operating income to identifiable assets.
b
Obtained through a nested ANOVA on segments in the economic sector.
c
Results are expressed as a percent of total variance, s2r , in the profitability of segments in the sector.
Finally, the negative corporate-parent–industry with the null hypothesis that the class of effects
covariance is remarkably robust. Again, manufac- has no explanatory power. In all cases, F-values
turing is an outlier, with an influence of covari- are significant at the 0.001 level, indicating that
ance at −2.27%. In services, the negative corpo- all types of effects are significant. The F-tests
rate-parent–industry covariance is large even are flawed, however, because the estimates do
though corporate-parent effects do not contribute not account for covariance. The final two columns
directly to the total variance. This result suggests show the minimum and maximum values of the
that diversification in the service sector may be effects implied by the estimates. In each case,
more affected by industry selection than by the value is in the same terms as the expression in
relationships between business segments. parentheses in the first column (marked ‘source’).
Table 5 shows the sequential ANOVA analysis. The estimate of r indicates that shocks in year
Panel A shows the results when the industry t − 1 persist in year t at the rate of 37.77 per-
effects are introduced before corporate-parent cent.18 This persistence generates an R2 of about
effects. In panel B, corporate-parent effects are 13 percent in the null model, which accounts for
introduced first. In each panel, the column labeled lagged effects and the grand mean. It is in this
‘d.f.’ shows the number of degrees of freedom null model that we capture the intertemporal per-
associated with the introduction of each type of sistence of shocks between years, regardless of
effect. The next column shows the increment to
R2 when each class of effect is introduced. Recall,
18
however, that the ANOVA results are not reliable McGahan and Porter (1997) study this persistence in detail
and show that it differs for high and low performers. For all
measures of the sizes of effects. The column classes of effects, the rate of persistence is significantly greater
labeled ‘F-value’ shows the F-statistic associated than zero.
Table 5. Results of nested ANOVA on Equation 4
Source d.f. Incr. R2 F-value Min. est.a Max. est.b
(A) Ordering of effects: year, industry, corporate-parent, segment-specific
Null model (rri,k,t−1 + (1 − r)m)* 2 0.129 8613.64** −164.02 174.06

Year (gt − rgt−1) 12 0.003 18.26** −1.40 1.75
Industry ((1 − r)ai ) 616 0.094 10.14** −30.21 44.71
Corporate-parent ((1 − r)bk ) 1,791 0.091 3.65** −77.36 46.73
Segment-specific ((1 − r)fi,k ) 9,877 0.351 2.39** −113.34 171.39
MODEL 12,298 0.668
Error 45,834 0.332
TOTAL 58,132
(B) Ordering of effects: year, corporate-parent, industry, segment-specific
Null model (rri,k,t−1 + (1 − r)m)* 2 0.129 8613.64** −164.02 174.06

Year (gt − rgt−1) 12 0.003 18.26** −1.40 1.75
Corporate-parent ((1 − r)bk ) 1,791 0.119 4.32** −80.42 47.68
Industry ((1 − r)ai ) 616 0.068 8.21** −34.49 36.80
Segment-specific ((1 − r)fi,k ) 9,877 0.349 2.39** −112.84 171.91
MODEL 12,298 0.668
Error 45,834 0.332
TOTAL 58,132
*Estimate of r: 0.3777.
**Significant at the 0.0001 level.
a
Minimum value of expression in parentheses implied by estimated coefficients.
b
Maximum value of expression in parentheses implied by estimated coefficients.
source. Intertemporal persistence is introduced of effects.19 The negative covariance between

first in the sequence because in theory it is industry and corporate-parent effects in the COV
determined by activity in the prior year. The estimation has an impact of the same order as
percent of variation that is not explained by the the variance of corporate-parent effects.20 The
model, 33.2 percent, reflects transient shocks that sequential ANOVA tends to attribute this covari-
do not persist between years. ance to the first effect in the pair that is intro-
The marginal R2 for industry effects is 7–
9 percent, for corporate-parent effects is 10–13
percent, and for segment-specific effects is 35 19
Results from James (1996) suggest that corporate-parent
percent. The increment to R2 for all three types effects and perhaps segment-specific effects might partially
arise from managerial choice of generic strategy and from
of effects is smaller than estimated by Rumelt organizational learning. (James cannot identify segment-speci-
(as reported in his Table 2), perhaps because of fic effects through dummy-variable techniques because she
a greater diversity of sectoral coverage and the has a single observation on each business. Thus, her analysis
focuses on the disaggregation of corporate-parent effects.)
longer time series in the Compustat data. James’s results are based on survey data from only 99 Compu-
In the sequential ANOVA, the percentage of stat firms in 14 industries; each corporation participates in at
variation explained by industry effects is smaller least three different 4-digit SIC categories.
20
Brush and Bromiley (1998) argue against this style of
than in the COV estimation, and the percentage analysis because variance decomposition does not account
of variation explained by corporate-parent effects for the details of relationships between corporate-parent and
is larger than in the COV estimation. The sequen- business-specific effects. For example, they note that the
approach may attribute the effect of a corporate parent to one
tial ANOVA generates an incremental R2 for of its segments if the corporate parent improves only the
segment-specific effects that is comparable to the performance of the one segment. Although we share the
percent of variance for segment-specific effects concerns expressed by Brush and Bromiley, we argue that a
variance decomposition generates insights about aggregate
in the COV estimation. These differences may performance that are important to establish the context for
arise from relationships in the data between types detailed analysis of specific relationships.
duced to the model, and thus the industry influ- Through first-order differencing, we accommodate
ence may be significantly lower than in a the possibility of persistence in shocks to year,
simultaneous ANOVA.21 industry, corporate-parent and segment-specific
As Rumelt originally suggested, the differences effects. As a result, we report results for only
in the results from the COV analysis and the stable portions of effects. These differences
ANOVA warrant research on the relationships more than offset the reduction in the estimated
between the processes through which year, indus- influence of industry that results from the use of
try, corporate-parent, and segment-specific effects data aggregated to the Compustat business seg-
arise. A close examination of these relationships ment rather than to the FTC business unit.22
is critical for verifying and perhaps revising the We also estimate the results by broad economic
estimation approaches. sector, and find large and interesting differences
in the attribution of variance. In manufacturing,
industry and corporate-parent effects account for
CONCLUSIONS a relatively lower portion of variance, while seg-
ment-specific effects account for a relatively high
In this study, we revisit fundamental questions in portion of variance (although less of total vari-
strategy and economics about the relative impor- ance is explained). When we apply Rumelt’s
tance of year, industry, corporate-parent, and seg- model to our manufacturing data, we obtain
ment-specific effects on business-segment profits results quite similar to those reported in his 1991
using comprehensive data covering most broad study (which covered the manufacturing firms in
economic sectors drawn from Compustat’s Busi- the FTC Line-of-Business survey). Our analysis
ness Segment Reports for 1981 through 1994. indicates that manufacturing, which has been the
The results indicate that variation in year effects, focus of previous studies, is an outlier: generaliza-
stable industry effects, stable corporate-parent tions about the economy as a whole that are
effects, and stable segment-specific effects based on the results for manufacturing understate
account for 2 percent, 19 percent, 4 percent, and the importance of industry and corporate-parent
32 percent, respectively, of the aggregate variance effects, and overstate the importance of segment-
in business-segment profits. We find a negative specific effects.
covariance between stable industry and stable Our analyses provide strong support that indus-
corporate-parent effects that dampens the variance try really matters in three important ways. First,
in business-segment profits by about 6 percent. industry directly accounts for 19 percent of aggre-
These results support Schmalensee’s (1985) prin- gate variation in business-specific profits, and 36
cipal conclusion that industry effects contribute percent of explained variation. Second, industry
importantly to variation in business-specific prof- influences the effect of the corporate parent on
itability, and call into question Rumelt’s finding business-specific profitability. Third, the absolute
that stable industry effects have low influence. and relative influence of industry, corporate-par-
Our results differ from those of previous stud- ent, and business-specific effects differs substan-
ies for two reasons. First and most important, tially across broad economic sectors in ways
our data represent all economic sectors (except which suggest characteristic differences in their
finance), and cover a longer period encompassing industry structural context. To these three find-
several phases of the business cycle. Second, our ings, we add a fourth from a related study. We
method treats transient effects differently. find that industry effects are more persistent over
time than business-specific or corporate-parent
21
effects, which is consistent with the view that
A simultaneous ANOVA would generate estimates of coef-
ficients which could be examined for covariance. We are
prevented from this analysis by limitations in computing
22
capacity. Even among the coefficients as estimated by the The difference in our unit of analysis affects results in
nested ANOVA, we have evidence of complex relationships subtle ways. Although business segments nominally represent
among types of effects, however. The simple correlation a corporation’s activity by 4-digit SIC code, we have evidence
between estimates of industry and segment-specific coef- that aggregation occurs in reporting. Aggregation probably
ficients is −0.0203 in panel (A) and 0.0111 in panel (B) of dampens industry effects and may obscure interesting differ-
Table 5. The simple correlation between estimates of corpo- ences in business-unit performance. The longer average series
rate-parent and segment-specific coefficients is −0.0101 in in our data set—5.7 years vs. 4 years in Rumelt’s study—
panel (A ) and 0.0343 in panel (B) of Table 5. also tends to diminish the portion of effects that are stable.
industry structure changes relatively slowly Chamberlain, G. (1984). ‘Panel data’. In Z. Griliches
(McGahan and Porter, 1997). These results do and M. D. Intriligator (eds), Handbook of Econo-
metrics, Vol. 2. Elsevier, Amsterdam, pp. 1247–
not support the assertion that rapid change in the 1318.
economy has diminished the influence of industry. Chow, G. C. (1984). ‘Random and changing coefficient
While the organizational differences emphasized models’. In Z. Griliches and M. D. Intriligator (eds),
by the resource-based view are surely meaningful Handbook of Econometrics, Vol. 2. Elsevier,
(and would be included in our estimates of seg- Amsterdam, pp. 1213–1245.
Connor, K. R. (1991). ‘A historical comparison of
ment-specific differences), it would be misguided resource-based theory and five schools of thought
to disconnect the influence of organization from within industrial organization economics: Do we
the industry and competitive contexts in which have a new theory of the firm?’, Journal of Manage-
firms operate. ment, 17(1), pp. 121–154.
Dierickx, I. and K. Cool (1989a). ‘Asset stock accumu-
lation and sustainability of competitive advantage’,
Management Science, 35, pp. 1504–1511.
ACKNOWLEDGEMENTS Dierickx, I. and K. Cool (1989b). ‘Asset stock accumu-
lation and sustainability of competitive advantage:
We are grateful for comments and discussions to Reply’, Management Science, 35, p. 1514.
Richard E. Caves, Todd Eckler, Pankaj Ghema- Griliches, Z. (1984). ‘Economic data issues’. In Z.
Griliches and M. D. Intriligator (eds), Handbook of
wat, Ann L. McGill, Jan Rivkin, Richard P. Econometrics, Vol. 3. Elsevier, Amsterdam,
Rumelt, Al Silk, Harbir Singh, an anonymous pp. 1465–1514.
referee, the editors, and seminar participants at James, C. R. (1996). ‘In search of firm effects: The
MIT’s Sloan School, the NBER Productivity managerial and behavioral implications of competi-
Lunch, Northwestern’s Kellogg Graduate School tive advantage’, working paper, Anderson School of
Management, University of California, Los Angeles.
of Management, Stanford’s Graduate School of Levinthal, D. A. (1995). ‘Strategic management and
Business, the Academy of Management, and the the exploration of diversity’, In C. A. Montgomery
SMJ Special Issue Conference. Special thanks (ed.), Resource-based and Evolutionary Theories of
to Cynthia A. Montgomery for comments and the Firm. Kluwer, Boston, MA, pp. 19–42.
discussions, to Arthur Schleifer for extensive dis- McGahan, A. M. and M. E. Porter (1997). ‘The persis-
tence of profitability: Comparing the market-struc-
cussions on statistical methods, and to Jan Rivkin ture and Chicago Views’, manuscript, Harvard Busi-
for assistance in the components-of-variance ness School.
analysis. Thanks to Todd Eckler, Jan Rivkin, Montgomery, C. A. (1994). ‘Corporate diversification’,
and Sarah Woolverton for help in extracting and Journal of Economic Perspectives, 8(3), pp. 163–
assimilating data. The Division of Research at 178.
Oster, S. (1990). Modern Competitive Analysis. Oxford
the Harvard Graduate School of Business Admin- University Press, New York.
istration provided generous financial support for Porter, M. E. (1980). Competitive Strategy. Free Press,
this project. New York.
Powell, T. C. (1996). ‘How much does industry matter?
An alternative empirical test’, Strategic Management
Journal, 17(4), pp. 323–334.
REFERENCES Ravenscraft, D. J. (1983). ‘Structure–profit relation-
ships at the line of business and industry level’,
Abowd, J. M., B. Crepon, F. Kramarz and A. Trognon Review of Economics and Statistics, 65, pp. 22–31.
(1995). ‘A La Recherche Des Moments Perdus: Roquebert, J. A., P. J. Andrisani and R. L. Phillips
Covariance models for unbalanced panels with (1996). ‘The relative contribution to SBU profit-
endogenous death’, NBER Technical Working Paper ability of industry, corporate, and SBU effects’,
No. 180, Cambridge, MA. working paper, University of Texas.
Barney, J. B. (1986). ‘Strategic factor markets: Expec- Rumelt, R. P. (1991). ‘How much does industry mat-
tations, luck, and business strategy’, Management ter?’, Strategic Management Journal, 12(3),
Science, 32, pp. 1231–1241. pp. 167–185.
Barney, J. B. (1989). ‘Asset stocks and sustained com- Schmalensee, R. (1985). ‘Do markets differ much?’,
petitive advantage: A comment’, Management American Economic Review, 75(3), pp. 341–351.
Science, 35, pp. 1511–1513. Searle, S. (1971). Linear Models. Wiley, New York.
Brush, T. H. and P. Bromiley (1998). ‘What does a Searle, S. R. (1987). Linear Models for Unbalanced
small corporate effect mean? A variance components Data. Wiley, New York.
simulation of corporate and business effects’, Stra- Wernerfelt, B. and C. A. Montgomery (1988). ‘Tobin’s
tegic Management Journal, forthcoming. q and the importance of focus in firm performance’,
American Economic Review, 78(1), pp. 246–250.

McHahan & Porter (1997) How Much Does Industry Matter

Uploaded by

Copyright:

Available Formats

McHahan & Porter (1997) How Much Does Industry Matter

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

McHahan & Porter (1997) How Much Does Industry Matter

Uploaded by

Copyright:

Available Formats

Strategic Management Journal, Vol.

18 (Summer Special Issue), 15–30 (1997)

HOW MUCH DOES INDUSTRY MATTER, REALLY?

CCC 0143–2095/97/S10015–16 $17.50

No. of No. of Avg. assets

0,1,2 Agriculture, Mining 203 3661 966 11.2

rather than real economic differences. This prob-

(1) (2) (3) (4) (5)

No. of observations 1775 6932 18,298 58,132 58,132

*Based on business segments rather than business units.

(A) Characteristics of the data

TOTAL (s2r ) 100.00 100.00 100.00 100.00 100.00 100.00 100.00

Source d.f. Incr. R2 F-value Min. est.a Max. est.b

(A) Ordering of effects: year, industry, corporate-parent, segment-specific

Null model (rri,k,t−1 + (1 − r)m)* 2 0.129 8613.64** −164.02 174.06

(B) Ordering of effects: year, corporate-parent, industry, segment-specific

Null model (rri,k,t−1 + (1 − r)m)* 2 0.129 8613.64** −164.02 174.06

source. Intertemporal persistence is introduced of effects.19 The negative covariance between

You might also like