《Innovation, Diversity and Diffusion：A Self-Organisation Model》5 Schmalensee-1985-do markets differ much

The American Economic Review is currently published by American Economic Association. Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/about/terms.htmlJstOr'sTermsandConditionsofUseprovidesinpartthatunlessyou

团购合买资源类别：文库，文档格式：PDF，文档页数：12，文件大小：1.53MB

Do Markets Differ Much? By RIChARD SChmALenSEe This essay reports the results of a cross- easily support full-blown structural estima section study of differences in accounting tion One can view the sort of search for profitability that sheds light on some basic stylized facts conducted here as either a re- controversies in industrial economics. Most placement for or an input to interindustry cross-section studies in this field structural estimation, depending have been concerned with testing hypotheses feeling about the long- run potential of that about structural coefficients in models meant research approach. This study also departs to apply to essentially all markets. As we from much of the cross-section literature by have learned more about the difficulties of being fundamentally concerned with the im- nstructing such general models and of per- portance of various effects, not just with forming tests on their structural parameters coefficient signs and t-statistics erly, structural cross-section analysis has In particular, this essay provides estimates out of fashion. In contrast to most of of the relative importance of firm, market, the cross-section literature, the analysis re- and market share differences in the de ported here is fundamentally descriptive; it termination of business unit ( divisional) does not attempt directly to estimate or profitability in U.S. manufacturing. Using test hypotheses about structural parameter 1975 data from the line of business pro- hope to show by example that one can gram of the U.s. Federal Trade Commission perform illuminating analysis of cross-sec- (FTC), I find support neither for the ex tion data without a host of controversial istence of firm effects nor for the importance maintained hypotheses. Cross-section data of market share effects. Moreover, while in an yield interesting stylized facts to guide dustry effects apparently exist and are im- both general theorizing and empirical anal portant, they appear to be negatively corre sis of specific industries, even if they cannot lated with seller concentration in these data Section I relates firm, market, and share ffects to current issues and controversies in industrial economics and thus supplies the Sloan School of Management, Massachusetts In stitute of Technology. motivation for our empirical analysis. The indebted to Stephen Postrel for excellent research assis remainder of the essay treats the data and ance to the FtCs Line of Bus taff, particularly statistical methods employed(Section II), the David Lean, William Long, and David Ravenscraft, for empirical results obtained (Section Im), and the main implications of those results (Sec chard Caves, Jerry Hausman, Paul Jos tion I luable advice. Seminar audiences at Stanford, Berke Chicago, Northwestern, and British Columbia pro I. Sources of Profitability Differences ided useful comments on earlier versions of this essa am grateful for financial support from the In the classical tradition, following Joe National Science Foundation, the U.S. Federal Trade Bain(1951, 1956), industrial economists grant to MIT). The representations and conclusior treated the industry or market as the unit of presented herein are my own and have not been adopted whole or in part by the FTC or its Bureau of onomics.The Manager of the Line of Business Pro- gram has certified that he has reviewed and approved ms(1980)has expressed a similar the disclosure avoidance procedures used by the staff of the Line of busines to ensure that the data ach taken by Michael gort and ra included in this paper do not identify individual com- Singamsetti (1976)is close in some espects to th Line of Business data. I alone can be held respon- taken here. They use firm-level data, howeve ible for this papers contents ain very different result 341

THEAMERICAN ECONOMIC REVIEW JUNE 1985 study. Differences among firms were as- thus predicts a positive correlation between sumed transitory or unimportant unless concentration and profitability in cross based on scale economies, which were gener- tion at the industry level even though, by ally found to be insubstantial. Equilibrium assumption, concentration does not facilitate ndustry profitability was generally assumed the exercise of market power to be primarily determined by the ability of At the firm or(for multiproduct firms) established firms to restrict rivalry among business unit level, the revisionist view themselves and the protection afforded them plies that market share should appear as the by barriers to entry. A central hypothesis in primary determinant of profitability in cross virtually all the classical work was that in- section regressions, while market concentra creases in seller concentration tend to raise tion should have no impact. David Ravens- industrywide profits by facilitating collusion. craft (1983)checked these predictions with Most classical studies thus included con- FTC Line of Business data. 5 He found the centration among the independent variables impact of share on business unit profitability in regression analysis of industry average to be positive and highly significant, while rates of return, and most published studies the coefficient of concentration in the same reported the coefficient of concentration to regression was negative and significant. be positive and significant Ravenscraft interpreted his results as provid An anticlassical. revisionist view of in Ing strong support for the revisionist argu- dustrial economics has emerged in the last ment that the significance of concentration decade. In the simplest model consistent with in traditional industry-level cross-section re- this view, all markets are(at least approxi gressions arises because concentration mately) competitive, and scale economies are related with share(and thus efficiency) absent(or negligible). The key assumption is differences, not because it facilitates collr that within at least some industries there are sion. Stephen Martin has recently obtained persistent efficiency differences among sell- similar results in a simultaneous equations ers. Because more efficient enterprises tend analysis of the FTC data. The strong relation both to grow at the expense of their rivals between market share and profitability found and to be more profitable, these differences by these and other authors is difficult tend to induce a positive intra-industry cor- interpret within the classical tradition, given relation between share and profitability even the apparent absence of important scale in the absence of scale economies. moreover. economies in most industries the more important are efficiency differences A third tradition, which I will call man- in any industry, the less equal are market aerial, has yet another set of implications shares (and thus the higher is market con- for business unit profitability. Business centration) and the higher are the profits of the leading firms (and thus the higher is ndustry average profitability). This model also Sam Peltzman (1977). Interesting formal models nsistent with this view have recently been 2Leonard Weiss( 1974)provides a survey of cross- by Boyan Jovanovic (1982), S. A. Lippman and R.P. section studies in the classical tradition; see also F, M. Rumelt(1982), and others. It is important to note that rer(1980,ch.9) something like the classical notion of entry or mobility Efficiency should not be interpreted barriers( richard C and Michael Porter, 1977)must cess terms here be invoked to explain why imitation does not the production of astrian climinate efficiency differences among firms characteristics it supplies to an existing lier studies of the eff reating something approaching a new market)canr of market share. Most obtained results n. It seem stent with those of Ravenscraft and ste propriate to think of nondramatic product innovations in efficiency terms for purposes of positive analysis of See Scherer(ch. 4) for an excellent survey of the rofitabili available evidence on economies of scale

VOL. 75 NO. 3 SCHMALENSEE: DO MARKETS DIFFER MUCH? schools and management consultants exist Conventional, classical industry-level vari- because it is widely believed that some firms ables may thus perform poorly at least in are better managed than others and that part because they are poor, incomplete mea one can learn important management skills sures of the(classical and other) market that are not industry specific. In a widely effects present in available data. Since many cclaimed best seller, Thomas Peters and of the usual classical industry-level variables Robert Waterman, Jr.(1982) stress the are endogeneous in the long run, and it mportance of firm-level efficiency differences difficult to formulate enough noncontrol- asure on differences in versial exclusion restrictions to identify organizational cultures. Dennis Mueller 1977, 1983)has recently reported economet- that problems of measurement and disequ ric results implying the existence of substan- librium can be successfully attacked by tial, long-lived differences measured firm structural modeling using available cre profitability. When profit rates in 1950 are section data taken into account, Mueller (1983)finds that concentration has a significant negative co- II. Methods and Data efficient in an equation explaining projected firm profit rates in 1972, and industry effects Instead of attempting structural analysis, in general are relatively unimportant. this study employs a simple analysis of vari- Both the revisionist and managerial alter- ance framework that allows us to focus di natives to the classical tradition are based on rectly on the existence and importance of plausible arguments and suggestive evidence. firm, market, and market share effects with But I do not think that it has been shown out having to deal simultaneously with that the classical attention to the industry specific hypotheses and measurement issues was in any sense a mistake: case studies related to their determinants. Specifically, I f real markets clearly reveal important deal in all that follows with the following differences. Why, then, do conventional mar- basic descriptive model ket-level variables perform poorly or per- versely when firm or share effects are in- (1) r,=u+a; +B+YS, +ei cluded in cross-section regressions One probable reason comes readily to where ri is the(accounting)rate of return of have very imperfect measures of the classic market share, the a's are firm effec,, is its dimensions of market structure and basic are industry effects, u and y are constants, onditions. Conditions of entry have proven and the e's are disturbances. The assump particularly difficult to measure in a satisfac- tions that market share enters linearly in( tory fashion. moreover, the link between the and that y is the same for all industries are eal, economic profitability dealt with in made mainly for comparability to the litera heoretical discussions and the accounting ture, though both also simplify computation returns used in empirical work is weakened and interpretation. The 1975 FTC Line of by inflation (Ge eoffrey Whittington, 1983), Business data set, which I use, contains in- depreciation policy(Thomas Stauffer, 1971; formation on large multidivisional firms Franklin Fisher and John McGowan, 1983), Such information is clearly required to sep- risk (myself, 1981), and both cyclical arate firm and industry effects in (1) (Leonard Weiss) and secular (Ralph Brad While none of the coefficients in (1)can be burd and Richard Caves, 1982)disequilibria. given a defensible structural interpretation analysis of that model as a whole can shed An additional individual lines of this, spurious industry effects added to bus trary. If firms fol

344 THE AMERICAN ECONOMIC REVIEW JUNE1985 light on the relative merits of at least the decomposed as follows extreme versions of the classical, revisionist and managerial positions. An extreme classi- (2 )02(r)=02(a)+o2(B)+r202(S) differ substantially with a =y=0 for al, o cist, for instance, would expect the B's a2()+2p(a,B)o(a)o(B) Estimates consistent with these expectations would of course not exclude the possibility +2yp(a, S)o(a)o(s) that industry effects simply reflect industry- wide differences between accounting and +2Yp(B,S)o(β)o(S) economic rates of return or industry-level disequilibria, with variations in monopoly where the p's are correlation coefficients and power of little or no importance. But a find- the as are standard deviations. Depending ing that the as and y did not differ signifi on which effects are revealed to exist by the cantly from zero would cast doubt on ex- analysis of (1),I estimate either (2)or a treme managerial or revisionist positions. special case thereof to provide information The implications of these last two posi- on the importance of the determinants of ons for the parameters of equation(1)are observed profitability. Estimates of (2)relate slightly less clear cut. An extreme revisionist directly to the predictions of the alternative would presumably expect a large y with all traditions discussed above. The particular the as and B's near zero if the r were (random effects) estimation techniques used observations on equilibrium rates of return. in this phase of the analysis are presented in But, since our data are in fact for a single Section Ill year and thus reflect the effects of cyclical In most of the statistical literature con- and other short-run industry-level disequi- cerned with variance decomposition, orthog libria, an extreme revisionist would not likely onality of effects is assumed, so that covari be surprised to find significant differences ance terms like the last three on the right of among the B's estimated here. Similarly, an (2) are set to zero. But that assumption is extreme managerial position might be that not plausible here. If an important attribute ariations in the a's should be much more of efficient firms is their ability to pick prof important in equilibrium than those in the itable industries in which to operate, for Bs or in the yS,, terms. But an extreme instance, we would expect this feature of the managerialist would also not likely be sur- data generation process on which I must rised to find differences in the Bs in a condition the estimates to produce a positiv single year's data. Moreover, firm-level p(a, B). Similarly, one expects efficient firm efficiency differences might affect business to have low costs and high shares, so that unit profitability through the revisionist p(a, S)should be positive. Finally, if one mechanism, so that firm-level and share knows that some particular Sii is above aver- effects might be hard to distinguish. (I in- age, one's conditional expectation must be vestigate this possibility below.) that concentration in market j is above aver Using firm and industry dummy variables, age. If one expects industry concentration to I first use ordinary least squares(fixed effects be positively related to industry profitability, estimation)and the usual F-statistics to test it then follows that one expects p(B, S)to be for the existence of market effects(noniden- positive. On the other hand, since e captures tical B's), firm effects(nonidentical as), and all profitability differences unrelated to firm, share effects (nonzero y )in(1)and the natu- industry, or market share diffe erences ral special cases thereof. To analyze the im- assumption that it is orthogonal to those portance of these effects, I treat the actual effects seems natural and reasonable ' s, B's, S's, and es in any particular sam- The strength of this descriptive approach ple as(unobservable)realizations of random is that my conclusions about the three rele- variables with some joint population distri- vant types of effects will not be conditioned bution. Under the usual assumption that e is distributed independently of the other vari- ables, the population variance of r can be See. for instance, S.R. Searle(1971, chs 9-11)

VOL, 75 NO. 3 SCHMALENSEE: DO MARKETS DIFFER MUCH? by maintained hypotheses regarding the de- seemed unlikely to correspond even ap terminants of those effects. I can focus proximately to meaningful markets. This rectly on the general implications of extreme removed 340 observations. In order to miti classical, revisionist, and managerial posi- gate scale-related heteroscedascity problems tions without having to deal with issues of and to focus on the revisionist mechanism endogeneity or identification. In addition, if (as distinguished from scale economies), the one doubts a priori that any of these extreme 1,070 remaining observations with market positions is tenable, one can look to quanti- shares of less than 1.0 percent were excluded tative evidence on the importance of firm, (Note that none of these involve small firms market, and market share effects and the all are small divisions of the 471 large firms correlations among them to suggest tenable sampled by the FTC. )Finally, one outlier compromise positions as well as questions (with operating losses exceeding sales and and strategies for future research assets/sales several times larger than other One important issue of research strategy business units in its industry) was excl an be very easily addressed within this before analysis began. Our final data set framework: is it defensible to work with contained 1.775 observations on business industry-level data? Given the central role of units operated by 456 firms in 242 of the 261 profits in industrial economics, the answer FTC manufacturing industries must depend critically on how important In equation (1),ri was measured as the industry effects are in determining industry ratio of operating income to total assets, rates of return. Only if industry profitability expressed as a percentage. This quantity pro- mainly reflects industry-level effects can one vided an estimate of the total pre-tax rate of ope that hypotheses about the(classical, return(profits plus interest) on total capital counting, disequilibrium, and other) de- employed; it seemed superior on theoretical terminants of those effects can be produc- grounds to the frequently employed price- tively tested with industry-level data. If R, is cost margin as a measure of profitability the(appropriately weighted) average rate of Its mean was 13.66, and its variance, s(r), return of business units operating in industry was 348.97. For each industry in the sample, j, equation (1)implies also computed the asset-weighted average ate of return, r. The mean and variance (3)R,=u+B, +terms in a's, Ss, and e's. these 242 numbers were 13.08 and 86.91 Industry-level analysis would seem to be mates computed and kindly supplied by sensible if and only if (estimates of)o2(B) Ravenscraft. The mean percentage market are large relative to the cross-section vari- share in this sample was 6.14, with a variance ance of the R, so that industry-level dif- of 59. 23(=S-(S)) ferences are important determinants of in Ill. Empirical Findings All empirical results reported below are based on a subset of the 1975 data on indi- Figure 1 summarizes the results of least vidual business units gathered and compiled squares estimation of equation (1) and re by the FTCs line of Business Program These business units account for about one- half of manufacturing sales and about 9 The industrie opped were the two-thirds of manufacturing assets.(See 22. 12. 23.06, 23.07 24.05, 25.06, 28 Ravenscraft and the sources he cites for de. 32. 18.33.13,34.21,35.37.36.28.37.14 tailed discussions of the FTC data. In order rates of return on investment not on sales. The case for to minimize the influence of newly born and ng rate of return on sales as a measure of the lerner nearly dead operations, only the 3, 816 busi- index rests a be lief that accounting aver ness units present in the ftc data in both good proxy for marginal cost, which I doubt, and the 1975 and 1976 were considered. Sixteen in- undeniable proposition that sales are measured more dustries that appeared to be primarily resid accurately than assets I This variable is 100 times the variable MS used by ual classifications were excluded because they Ravenscraft

VOL 75 NO. 3 SCHMALENSEE: DO MARKETS DIFFER MUCH? any noticeable extent Firm dumm smaller; it is between 0.53 percent (GLS) serve as proxies for market share and 0.82 percent(OLS). 4 While avenscraft is no difficulty disentangling firm and also uses 1975 Line of business data. he uses effects the ratio of operating income to sales to Oe In sharp contrast, all tests for the existence measure profitability, does not delete"mis- cant results. All four tests of the null hy- small shares, uses classical variables like con- thesis of no share effects(arrows pointing centration in place of industry dummies, and to the left in Figure 1)signal rejection at attempts(in his GLS estimates)to correct for P-levels below 4.5 percent, while the null complex pattern of heteroscedascity. The hypothesis of no industry effects is always statistical significance but quantitative unim- rejected at below the 0.01 percent level ance of market share effects thus seems a (vertical arrows in Figure 1). 3 robust feature of these data Let us now consider the importance of One final pattern in the statistics presented share and industry effects, postponing until above deserves mention. Market share add Section IV a discussion of the implications of more to adjusted R in the presence of in the absence of firm effects in these data. It is dustry effects(0.62 vS. 0. 17 percent), and most instructive first to present an informal industry effects add more in the presence of treatment based on information in Figure 1 hare effects (19.29 vs. 18. 84 percent). This and then to employ the relevant special case sort of complementarity is suggestive of a of(2) negative correlation between market share Comparing adjusted R's of models not and industry effects. Pointing in the same nvolving firm effects, market effects seem to direction are the drops in the p-levels account for between 18.84 and 19.29 percent ated with share effects when industry of the sample variance of r. Following the are added and the corresponding changes these percentages correspond to 75.65 and ciated with industry effects. Finally, the fact 77.46 percent of s(R), the sample variance that the estimate of y25(S)discussed above of industry average rates of return. Industry exceeds the contribution of share effects to effects thus seem to be quite important, ap- adjusted R is also suggestive of a negative parently accounting for the bulk of inter- correlation between share and industry ef- industry differences in accounting rates of fects. (See equation (5)below return. The industry seems an easily defense- Let us now provide a more systemat ble unit of analysis analysis of the issues raised in the preceding On the other hand, the adjusted Rs in paragraphs. With no firm effects pres- Figure 1 indicate that market share effects ent, the relevant special cases of (1)and(2) add only between 0. 17 and 0.62 percent to are the following varlance ex plained. Similarly, using Y 0.2304 from Figure 1,ys2(S)amounts to (4) μ+月+yS+ only 0.90 percent of s(r). It is interesting to note that in Ravenscraft's paper, which (5)02(r)=02(B)+r02(S)+0(e) 2Yp(B,S)σ(B)o(S) Readers uninterested in estimation technique of the evidence for the presence of industry effects. The and persuaded by the evidence presented F-statistics and corresponding restricted models are the above bearing on (5) may wish to glance following: F(241, 1533)=2.709, null model; F(241, briefly at Table 1, which summarizes the 1532)=2.762, share effects only:F(241,1078)=2007 effects. I calculate the probability of obtaining Fs above ny one of these values under the null hypothesis to be I4 The necessary statistics are in Tables 1 and A I of Ravenscraft

THEAMERICAN ECONOMIC REVIEW JUNE 1985 TABLE1— ESTIMATEDⅤ ARIANCE DECOMPOSITIONS Population Name Estimate Estimate Percentage Market 68466 1959 02(B)H y2a2(S)(1-(1-H)p2 Covariance 2 ypa(B)a(s) 2Hypa(B)o(s) -2.159 Error 281.049 81.049 a2(r) 100.00 10000 Note:See text for sources and definitions, Totals may not add because of rounding results developed below, and then skip and the defensibility of industry-level analy Section Iv re again clear Ordinary least squares estimation of (4), In order to estimate the two remaining which appears in Figure 1 as the""Industry terms on the right of (5), it is necessary to be and Share Effects"model, yields a consistent more specific about what is meant by a non and unbiased estimate of 281.05 for a2(e). zero population correlation between market Following Searle's(chs. 9-11)treatment of share and market effects. Imagine the data variance components estimation in unbal- generation process first fixing the N, then nced models, I next compute consistent drawing the ps independently from their analysis-of-variance"estimates of the re- unconditional distribution, and finally draw maining quantities on the right of (5) ing the S's for each industry from the cond et the operator can"expected sum tional distribution determined by of squares about the sample mean, "let n be of B previously drawn. Assume without loss he total number of observations, let N, be of generality that the unconditional mean of the number of observations in industry j, the B's is zero and of the S's is u, I then and let M be the total number of industries. impose the following assumptions a bit of algebra yields (6)ESS(, -YS,) i=k =(N-1)2(a)+(N-G)02(B,E(S,S) (k2)2+o2(S) ()+(s)2(B,s)1≠k where (8b)E(BS)=p(B, S)o(B)o(S) The first part of(8a)and(8b)are not restric (7)G=∑(N)2/N tive,the second part of (8a) is consi tent with but does not impose normality These expectations are taken with respe If all industries had only one firm, G would to the unconditional population distribution equal one. If there were only one industry, G would equal N, since industry effects would not contribute to overall variance. In these data,G=15.55. Using y= 2304 and o(e) 15As a final check on the robustness of 81.05 from above, setting the expectation sion, I computed MIVQUEQO estimates of on the left of (6)equal to its sample value, firm, market, and error variance compone and solving yields an estimate of 68.47 for (r,.-ySiyj). See H. 0. Hartley et al., 1978 0(B). This is equal to 19.62 percent of the estimates of a(B)of 62.03 and 64.88,respectively.This of ple variance of the r. and 78.78 percent close to those in the text, further strengthening the case the sample variance of the R;. The or the qua e importance of in quantitative importance of industry effects these data

VOL, 75 NO. 3 SCHMALENSEE: DO MARKETS DIFFER MUCH? 349 ut they are conditional on the assignment I. Conclusions and Implications of firms to markets. Similarly, for h*j, E(BhB,)=(nS)=0, and E(SinSk) The analysis of Section Ill indicates that the 1975 FTC line of Business data provide Let r, be the unweighted mean of the rates strong support for the following four em- of return of business units in industry j. pirical propositions Then if (4)is the true model, (8a) and some PROPOSITION 1: Firm effects do not exist (9)ESS(;-r)=(N-M)yo2(S) PROPOSITION 2: Industry effects exi are important, accounting for at least 7 x[1-P2(B, S)+(N-M)02(e). on assets cent of the variance of industry rates of The quantity on the left is the expected sum PROPOSITiON 3: Market share effects ex of squared residuals from a regression of ible fraction of the rii on M industry dummy variables variance of business unit rates of return regression appears as the "Industry Effects Only" model in Figure 1. Use of( 8)and a PROPOSITION 4: Industry and market bit more algebra yields share effects are negatively correlate (10)ESS)(N-1)=E[o() The apparent nonexistence of firm effects is somewhat surprising. This finding is per =Ho2(B)+y2(S)[1-(1-H)p2(B,s ctly consistent with substantial intra-in dustry profitability differences, which Table 1 shows to be present in these data. The +o2(e)+2HYp(B, s)o(B)o(S), absence of firm effects in(1) merely means that knowing a firms profitability in market here A tells nothing about its likely profitability in randomly selected market B. This is con H=(N-G)/(N-1) sistent with the conglomerate bust of the past decade and with a central prescriptive Equation(10) provides a decomposition of thrust of Peters and Waterman(ch. 10): wise ability corresponding to the d s unit profit firms do not diversify beyond their demon- composition strated spheres of competence. The nonex of the population variance given by (5) istence of firm effects suggests that Mueller's Setting expectations equal to sample val-( 1983)persistent firm-level profitability dif ues, solving (9) for yo(S)and substituting ferences are traceable to persistent differ- into(10), an equation is obtained involving ences at the business unit or industry level p(B, S), sample statistics, and estimates combined with relatively stable patterns of derived above. A search of the interval activity at the firm level. 6 eveals a unique root; the esti The finding that industry effects are im- mated value of p(B, S)is-0.089. This con- portant supports the classical focus on in- firms the negative correlation between in- dustry-level analysis as against the revisionist dustry and share effects. Equation (9)then tendancy to downplay industry differences. ields an estimate of 2. 182 for y202(S).As this is only about 3.2 percent of the esti- share effects is also confirmed. Table 1 re- FTC Line of Business data, Johny variables to analyze mated value of a(B), the unimportance of 16 Using firm and industry dumn (5)and(10), respectively, implied by these reflects policy rather than performance, it is not surpris- ing that firm effects show up there but not here

点击进入文档下载页（PDF格式）

共12页，试读已结束，阅读完整版请下载

点击下载（PDF格式）

浏览记录