What Have We Learned From Generic Competitive Strategy

7/22/2019 What Have We Learned From Generic Competitive Strategy

1/28

Strategic Management JournalStrat. Mgmt. J., 21: 127154 (2000)

WHAT HAVE WE LEARNED ABOUT GENERIC

COMPETITIVE STRATEGY? A META-ANALYSIS

COLIN CAMPBELL-HUNT*School of Business and Public Management, Victoria University of Wellington,Wellington, New Zealand

The dominant paradigm of competitive strategy is now nearly two decades old, but it hasproved difficult to assess its adequacy as a descriptive system, or progress its propositionsabout the performance consequences of different strategic designs. It is argued that this is dueto an inability to compare and cumulate empirical work in the field. A meta-analytic procedureis proposed by which the empirical record can be aggregated. Results suggest that, although

cost and differentiation do act as high-level discriminators of competitive strategy designs, theparadigms descriptions of competitive strategy should be enhanced, and that its theoreticalproposition on the performance of designs has yet to be supported. A considerable agenda forfurther work suggests that competitive strategy research should recover something of its formersalience. Copyright 2000 John Wiley & Sons, Ltd.

INTRODUCTION

Michael Porters theory of generic competitivestrategy is unquestionably among the most sub-stantial and influential contributions that have

been made to the study of strategic behavior inorganizations (Porter, 1980, 1985). In essence,the theory contains two elements: first, a schemefor describing firms competitive strategiesaccording to their market scope (focused orbroad), and their source of competitive advantage(cost or differentiation); and, second, a theoreticalproposition about the performance outcomes ofthese strategic designs: that failure to choosebetween one of cost- or differentiation-leadershipwill result in inferior performance, the so-calledstuck-in-the-middle hypothesis.

Within a few years of publication, the theory

Key words: generic strategy; competitive strategy;meta-analysis*Correspondence to: Colin Campbell-Hunt, School of Busi-

ness and Public Management, Victoria University of Welling-ton, PO Box 600, Wellington, New Zealand

CCC 01432095/2000/02012728 $17.50 Received 5 August 1996Copyright 2000 John Wiley & Sons, Ltd. Final revision received 30 July 1999

was recognized as the dominant paradigm ofcompetitive strategy ( Hill, 1988; Murray, 1988).But, despite widespread interest and application,it has proved difficult to progress its represen-tation of competitive behavior. In Kuhns account,

a paradigm gives a common platform and focusto subsequent empirical and theoretical investi-gation; it defines the scope of phenomena thatare deemed to be important, and the methodsused for investigation; and it becomes thereceived wisdom that is taught in the subjectstextbooks (Kuhn, 1962). In the following para-graphs it will be shown that Porters theory hasplayed all these roles.

But it is the thesis of this paper that theparadigm has so far failed to open up a periodof Kuhnian normal science, in which a detailedand immensely productive dialogue is establishedbetween fact and theory. Failure to establish thisdialogue threatens to leave the study of competi-tive strategy in a preparadigm state, as no morethan a series of brave beginnings, none of whichattract sufficient empirical or social support tomake the phase transition to normal science. The


2/28

128 C. Campbell-Hunt

impediment has been that there is no knownway to compare or cumulate individual empiricalstudies of the type suggested by the paradigm. Itis the objective of this paper to remove thisimpediment.

The dominant paradigm

The widespread acceptance of Porters descriptivescheme by researchers can be seen in the widerange of its application. These include industriesas diverse as shipping (Brooks, 1993), banking(Meidan and Chin, 1995), and hospital services(Kropf and Szafran, 1988); and countries asdiverse as Ireland (McNamee and McHugh,1989), Portugal (Green, Lisboa, and Yasin, 1993),Korea (Kim and Lim, 1988), and the PeoplesRepublic of China (Liff, He, and Steward, 1993).

The scheme has also been widely used byresearchers studying relationships between firmscompetitive strategy and other aspects of man-agement: i.e., their human relations strategy(Schuler and Jackson, 1989); information tech-nology (Huff, 1988); industrial engineering(Petersen, 1992); manufacturing strategy (Kothaand Orne, 1989); logistics (McGinnis and Kohn,1988); environmental scanning (Jennings andLumpkin, 1992); planning processes (Powell,1994); management selection (Govindarajan,1989; Sheibar, 1986); and managerial biases in

perceptions of competitive strategy (Nystrom,1994). The framework has also been used exten-sively in practice to structure managers percep-tions about their firms strategy. With few excep-tions (Bowman and Johnson, 1992), suchapplications are rarely reported.

The paradigms theoretical propositions havealso attracted intense debate. Early challengesto the stuck-in-the-middle hypothesis (Karnani,1984; Murray, 1988; Hill, 1988) argued that con-ditions which might favor cost-leadership (such asthe reduction of transaction costs through verticalintegration, process innovation and learning, andscale effects) were independent of conditions thatmight favor differentiation (such as consumerpreferences, product innovation, and quality dif-ferentiation based on a firms superiority in aparticularly complex value system). Hence, exter-nal conditions provide no a priori reason todiscriminate against mixed cost- and differen-tiation-strategic designs (Murray, 1988). More-over, in conditions where differentiation strategies

Copyright 2000 John Wiley & Sons, Ltd. Strat. Mgmt. J., 21: 127154 (2000)

can be used to expand market share, and this inturn permits greater capture of economies of scaleand scope, external conditions might activelyfavor mixed strategies (Hill, 1988; Phillips,Chang, and Buzzell, 1983). Conditions that havebeen considered in this way include the particular

nature of retailing as against manufacturing indus-tries (Cappel et al., 1994); and the distinctivecharacteristics of an industrys technology(Oskarsson and Sjoberg, 1994).

Beginning with Hambrick ( 1983), a series ofstudies has also begun the task of exploringthe paradigms empirical validity. These havefollowed the paradigms guidance to describe ge-neric strategies as polythetic gestalts or designs(Miller, 1981; Hambrick, 1984; Rich, 1992), atask best undertaken using principal componentsanalysis and cluster analysis (Hambrick, 1984;

Harrigan, 1985; McGee and Thomas, 1986).However, these techniques result in classificationsthat are specific to the sample of participatingfirms and cannot be cumulated with other find-ings. Thus, it has not been possible to assess theaccumulated weight of evidence on what genericcompetitive strategies look like in practice, norhow closely they accord with the paradigmsdescriptive and theoretical elements. The study ofcompetitive strategy is thus currently stuck insomething of a dead-end of its own design.

Compounding these difficulties, there have

evolved a number of different interpretations ofthe dominant paradigms descriptive system, sothat the paradigms descriptive and theoreticalpropositions may take a number of forms. Todate, these have not been systematically com-pared.

As a result of this impeded dialogue betweenparadigm and empirical investigation, the para-digms scheme for describing competitive strategyhas barely progressed in the two decades sinceit was first proposed. Attempts by Miller (1986)and Mintzberg (1988) to widen the set of stra-tegic competitive behaviors that are held to begeneric have met with little success, despiterecent empirical evidence which suggests thatthey offer a superior description of competitivebehavior ( Kotha and Vadlamani, 1995). Portersscheme remains unaltered as the typology set outin most contemporary textbooks (Thompson andStrickland, 1995; Pearce and Robinson, 1994;Bourgeois, 1996).

The study reported in this paper was accord-


3/28

Generic Competitive Strategy 129

ingly motivated to develop meta-analytic pro-cedures with which to aggregate empiricallyderived descriptions of generic competitive strat-egy. Study One reports a meta-analysis of theprincipal component solutions in the empiricalrecord; Study Two reports a meta-analysis of

clustered categories of competitive strategydesign. The resulting aggregates are comparedto alternative interpretations of the classificationsystem of the dominant paradigm. Study Threeuses these aggregate descriptions to assess theparadigms theoretical propositions on the per-formance of generic competitive strategies. Tobegin, alternative interpretations of the dominantparadigm and its propositions are discussed andformalized.

INTERPRETATIONS OF THEDOMINANT PARADIGM

Describing competitive strategy

All theory building requires a parsimonious wayto describe the intractable variety of nature. Thissection examines the four approaches that havebeen used to interpret the dominant paradigmsdescriptive system.

The taxonomic interpretation

The first approach is to interpret the system as ataxonomy, that is, a hierarchically ordered set ofclassifications, within which all designs can beallocated to a unique position, depending on theparticular set of strategic elements involved(Chrisman, Hofer, and Boulton, 1988). In thisapproach, the bewildering variety of strategicdesigns is reduced to a parsimonious set of allo-cation rules (Doty and Glick, 1994) by whicha specific design for competitive strategy is classi-fied within the hierarchy. This interpretation,clearly inspired by biological taxonomy, requiresthat allocation rules have a hierarchical structure,and that classifications be internally homo-geneous, mutually exclusive, and collectivelyexhaustive (Chrisman et al., 1988; Rich, 1992).

The order in which allocation rules enter intothe hierarchy is shown in Figure 1. At the top isthe paradigms distinction between designs thatplace distinctive emphasis, relative to competitors,on pursuing some source of advantage, anddesigns that spread their efforts more evenly and


become stuck-in-the-middle. The paradigmstheory of performance is based on this highest-level distinction. Within the class of distinctiveemphasis designs, Porters emphasis on thecost/differentiation dichotomy as two basic typesof competitive advantage, and as fundamentally

different route(s) to competitive advantage(Porter, 1985: 11), suggests that this allocationrule be placed above market scope in the rulehierarchy.

The life-science-inspired, taxonomic inter-pretation also places particular emphasis on themutual exclusion of class memberships. An essen-tialist rationale for the sharp distinction betweencost- and differentiation-emphasis would be thatthere are elements in the design of each thatnaturally repel the other. Each design has its ownfundamental essence, and attempts to mix them

will be quickly terminated by the unnatural natureof the experiment (Miller, 1981; Hannan andFreeman, 1989). Mixed-emphasis designs are notcompletely ruled out, but will be rare.

These key features of the taxonomic inter-pretation are set out in Table 1, and stressed inthe following proposition:

Proposition 1a: All competitive-strategydesigns can be precisely allocated to a numberof hierarchically ordered classes on the basisof (i) whether or not a design has some dis-

tinctive emphasis relative to competitors; (ii)whether that emphasis is towards cost- or

differentiation-advantage; and (iii) the market

scope adopted. Only a very small number ofmixed-emphasis designs will exist.

The empiricist interpretation

This second interpretation relaxes the restrictionsof taxonomy. The approach is best typified inan extensive series of studies by Danny Miller,including studies of competitive strategy (Miller,1992b; Miller and Friesen, 1986a, 1986b). Theapproach retains the assumption that the verylarge number of firm-level competitive-strategydesigns can be reduced to a smaller number ofclasses (Miller, 1981), but it differs from a taxo-nomic interpretation in four ways (Table 1).

First, it is no longer asserted that all designscan be so classified, just that a large proportioncan (Miller, 1986: 236). Room is left for idiosyn-


4/28


Figure 1. The taxonomic description of generic competitive strategy

cratic designs to flourish around the more com-monly observed classes. Secondly, the allocationof each individual design to a class is no longer

determined exactly by a precise set of allocationrules, but is in part stochastic. This uncertaintycan be reduced with more refined classification,so that a balance must be struck between a largernumber of more homogeneous classes, and amore parsimonious, but possibly less meaningful,classification ( Doty and Glick, 1994; Hambrick,1984; Miller, 1981). Thirdly, although all empiri-cally derived clusters are associated together inhierarchies of similarity, an empiricist inter-pretation does not impose an ex ante requirementthat cost and differentiation be high-level discrim-

inators in that hierarchy. Finally, the empiricistinterpretation does not anticipate a near-prohibition on mixed-emphasis designs, ex ante,but rather allows whatever common designs existto emerge from the data.

This less restrictive interpretation of the domi-nant paradigm can be summarized as follows:

Proposition 1b: Most competitive-strategy

Table 1. Interpretations of the dominant paradigm

Interpretation: Taxonomic Empiricist Nominalist Dimensional

Hierarchically ordered Yes No Yes Nodescriptions?Homogeneity of class Identical Approximate Variable n/amembersMutually exclusive Yes Approximate Approximate Noclassification?Mixed designs? Very few Yes Very few YesCollectively exhaustive? Yes Large proportion No Large proportion


designs can be meaningfully allocated to anumber of classes on criteria that include

whether or not a design has some distinctive

emphasis relative to competitors; whether thatemphasis is towards cost- or differentiation-

advantage; and the market scope adopted.

The nominalist interpretation

In this view, generic competitive strategies aretaken to represent ideal types, and the 2 2classification system of the dominant paradigmis interpreted as a general typology (Doty andGlick, 1994).

Correspondence between real designs and idealtypes will be both imperfect and variable (Mayr,1969; Rich, 1992), so that classifications will beneither fully homogeneous nor mutually exclusive(Table 1). Also, the nominalist interpretation doesnot require the four ideal types to be collectivelyexhaustive. To the contrary, and unlike all otherinterpretations of the dominant paradigm, theapproach seeks only to describe a limited number


5/28


of ideal types based on a few aspects of competi-tive-strategy design, selected for their importanceto the paradigms theory of performance.

The nominalist approach is hierarchical in thatthe limited number of characteristics chosen todescribe ideal types are held to be fundamentally

important to the design and performance of com-petitive strategies, and to be the basis on which todistinguish more richly described designs (Bakke,1959; Rich, 1992; Porter, 1980: 4041). All dif-ferentiation designs share the characteristic ofpursuing a price premium; cost designs are ori-ented to economy as the path to profit. Thisessentialist distinction between ideal types is com-mon to both the nominalist and taxonomicinterpretations and means that both expect thenumber of mixed designs to be small (Doty andGlick, 1994).

The nominalist interpretation of the dominantparadigm is accordingly formalized as follows:

Proposition 1c: Competitive-strategy designs

can be likened to a greater or lesser extent toone of two fundamentally different archetypes:

one emphasizing advantage from costs, the other

from differentiation, each with broad and

focused market scope variants. Only a very smallnumber of mixed-emphasis designs will exist.

Generics interpreted as dimensions of

competitive-strategy design

The fourth approach interprets the characteristicsof market scope, cost-, and differentiation-emphasis as independent dimensions of a multi-variate space encompassing most of the variationin competitive-strategy designs (Karnani, 1984;Miller and Dess, 1993). Distinctive features ofthis interpretation are summarized in Table 1.

Unlike all other interpretations of the dominantparadigm, the dimensional approach does notdefine classes of competitive-strategy designs, sothat the question of class homogeneity does notarise. Rather, the approach is restricted to describ-ing the space in which classes may be defined.The distinction is essentially that drawn betweentwo of Peppers world hypotheses: formism,which describes the world in categories; andmechanism, which describes the world inelements and the relationships between them(Pepper, 1942).


Because all designs are positioned relative toboth cost- and differentiation-dimensions, thepresence of one emphasis does not exclude theother, and unrestricted scope is allowed to mixed-emphasis designs (Miller and Dess, 1993; Parkerand Helms, 1992). Even the extreme archetypal

designs of cost- and differentiation-emphasis can-not be adequately described in their own termsalone, but must be positioned relative to bothparameters: cost leaders must not lose touch withthe competitive standards of differentiation, andvice versa. The descriptive parameters areexpected to be independent of each other andwithout hierarchical rank.

This fourth interpretation of the paradigmsdescriptive system can be stated as follows:

Proposition 1d: Most competitive-strategy

designs can be meaningfully positioned in thethree-dimensional space described by (i) rela-

tive emphasis on cost advantage; (ii) relative

emphasis on differentiation advantage; and

(iii) the market scope adopted.

The paradigms theory of performance

The fundamental theorem of the dominant para-digm is that above-average performance can onlybe achieved by adopting one of the four genericdesigns. Performance is defined as above-average

rate of return (Porter, 1980: 35), sustained overa period of years (Porter, 1985: 11). This theoremis formalized in different ways, depending onthe interpretation of the paradigms descriptivesystem. The dimensional interpretation is pri-marily concerned with defining the space in whichcompetitive strategy designs may be described.To support the paradigms theoretical proposi-tions, some classification of designs within thisspace is required, using one of the otherapproaches.

Taxonomic and empiricist approaches thatattempt a comprehensive classification of alldesigns specify those classes with high-performance attributes:

Proposition 2a: Classes of competitive-

strategy design will show above-average

performance that are characterized by a

distinctive emphasis, relative to competitors,

on one of cost advantage, or differentiation

advantage; and are either broad or focused in


6/28


market scope. Only a small number of

mixed-emphasis designs will show above-

average performance. The class of designs that

fail to achieve distinctive emphasis relative to

competitors will record average or

below-average performance.

The nominalist approach does not attempt com-prehensive classification, but rather posits a smallnumber of ideal types. Performance will improveas actual designs approximate these ideals:

Proposition 2b: The incidence of above-

average performance will increase as competi-

tive-strategy designs approach one of two fun-

damentally different archetypes: one emphasiz-

ing advantage from costs, the other from

differentiation, each with broad and focused

market-scope variants. Only a small numberof mixed-emphasis designs will show above-

average performance. As designs depart from

these ideals and fail to achieve distinctive

emphasis relative to competitors, they will re-

cord average or below-average performance.

As discussed above, measuring the distancebetween actual and ideal involves not only iden-tifying distinctive emphasis in terms of one ideal,but also measuring proximity to competitorsstandard in the other.

Both versions of the theory stem, as we haveseen, from interpretations that emphasize theessentialist differences between strategiesdesigned to support cost advantage and differen-tiation advantage. Failure to choose between themis theorized to violate their distinctive require-ments and to lead, in turn, to lower performance.

In a similar way, failure to choose either astrategy adapted to a broad market scope, span-ning many segments, or one that focuses on oneor a few segments, is theorized to produce lowerperformance. An important aspect of this choiceis that it defines the scope of competitors againstwhich the firm seeks to be distinctive. Failure todefine competitive scope results in poorly targeteddesigns and middling performance. The para-digms theory of performance is thus U-shapedwith respect to market scope, positing higherperformance when designs are well adapted toeither broad or focused target markets, and aver-age or below-average performance for inter-mediate designs. The authors of the PIMS study


pointed out that this U-shaped relationship withrespect to market scope was not necessarilyinconsistent with their clear result that perform-ance improves with market share, because PIMSdefines share relative to the firms served mar-ket, and this can be either broad or focused in

Porters terms (Buzzell and Gale, 1987: 8586).For both Porter and PIMS, successful competitivestrategies are likely to produce strong marketshare in the served market.

STUDY ONE: META-DIMENSIONS OFCOMPETITIVE STRATEGY

This section describes the meta-analytic methoddeveloped for this study, and applies it to summa-rizing the dimensions of competitive strategy, as

described in the empirical record.

Meta-analysis method

Meta-analysis is the term used to describe astructured, quantified analysis of a body ofempirical literature on a theorized relationship.Relative to literatures in applied psychology andorganization behavior from which meta-analysisemerged, use of these techniques has been slowto spread to management disciplines. Marketinghas been an early adopter (Farley, Lehmann, and

Sawyer, 1995), and there are a handful of meta-analyses on relationships of interest to strategicmanagement, i.e., the effect of formal planningon performance (Schwenk and Shrader, 1993);the association between industry concentrationand performance (Datta and Narayanan, 1989);the effect of mergers and acquisitions on share-holder wealth (Datta, Pinches, and Narayanan,1992); and the influence of a number of proposeddrivers on innovation (Damanpour, 1991).

Methods of meta-analysis

Several meta-analytic methodologies have beendeveloped (see Raju, Pappas, and Williams, 1989,and Hunter and Schmidt, 1990: 468489, forintroductions to the main methods). A distinctioncan be drawn between those methods that seek toproduce a consistent aggregation of the empiricalevidence on a relationship, and those whichfurther seek to draw inferences from these aggre-gations on the size and variance of relationship


7/28


effects in a population. Among the most widelyused methods, the meta-analysis introduced byGlass and colleagues (Glass, McGaw, and Smith,1981) is of the first, descriptive, type (Hunterand Schmidt, 1990: 479); and that developed bySchmidt and Hunter (Hunter and Schmidt, 1990)

is of the inferential type.Inferential meta-analyses have become apowerful tool for reducing estimated variance ina parameter (Hunter and Schmidt, 1990: 485)and hence uncovering nonzero effect sizes whichhad formerly been hidden by type II errors inindividual studies (Schmidt, 1992). However, thebenefits of inferential meta-analysis are gained atthe cost of stringent requirements for the consis-tency of data (see Hunter and Schmidt, 1990:480481), several of which are not met in theempirical literature on generic competitive strat-

egy. First, it must be possible to interpret eachstudy as a random sample from a population.Where one study reports more than one analysison the same data (as in Hambrick, 1983; Gal-braith and Schendel, 1983; and Douglas andRhee, 1989), use of both analyses violates theindependence assumption. Second, the studiesmust use the same variables in their specificationof the relationship. Violation of this requirementis empirically important: failure to use identicalmodel specification across studies has been foundto represent the largest source of effect-size vari-

ance in meta-analyses in marketing (Farley et al.,1995). Third, where regression coefficients (orfactor coefficients) are to be used, cumulationinto a meta-analysis requires that these be meas-ured using exactly the same scales (Hunter andSchmidt, 1990: 203204).

Noncomparability of scales and model speci-fication across studies is an inevitable feature ofthe comparative novelty of studies into competi-tive strategy, and its research designs. As shownabove, Porters paradigm of generic competitivestrategy has been cast in a number of differentinterpretations, and researchers have had goodreason to expand the list of elements of competi-tive strategy they wish to include in their analysis,and have often devised their own scales forthese constructs.

Furthermore, the polythetic nature of the con-cept of generic competitive strategy suggestsresearch designs involving principal componentanalysis and cluster analysis. As with regressioncoefficients, scale noncomparability across studies


makes the use of factor coefficients in a SchmidtHunter type meta-analysis problematic. More fun-damentally, what is of interest in a meta-analysisof this literature is the cumulation of multivariate

patterns of association between many elementsof competitive strategy, and not one single effect

size in a relationship. There are no establishedmeta-analytics of the inferential type to deal withthis situation.

A descriptive meta-analytic procedure for factor

and cluster analysis

Although the barriers to an inferential meta-analysis appear insuperable at present, themethods developed for this study permit adescriptive meta-analysis of the empirical litera-ture on competitive strategy. By constructing a

consistent aggregation of the patterns of competi-tive strategy design, the full weight of the empiri-cal record can be applied to assess the validityof the paradigms descriptive and theoreticalpropositions. Hence a descriptive meta-analysis issufficient for the purpose of establishing a dia-logue between the dominant paradigm and theempirical record, and the further development ofthe paradigm. Also, the research questions posedby the paradigm attach greater importance to theexistence or otherwise of multivariate patternsthan to the degree of closeness in those patterns.

The additional precision in estimation of effectsizes, which is an important advantage of inferen-tial meta-analyses, is of secondary importancehere.

The first step in building the required meta-analysis is to produce consistent aggregates ofthe principal component solutions that are usedto summarize and describe competitive strategy.Studies use principal component factor analysisto represent many elements of competitive strat-egy with a smaller number of factors, each ofwhich represents an orthogonally independentdimension of competitive-strategy design (Kimand Mueller, 1978). The estimated factor coef-ficients also identify those elements which aremost closely associated with each dimension.

The primary aim of a meta-analysis over sev-eral such studies should be to identify dimensionswhich best describe the totality of orthogonalfactor solutions in the empirical literature. It isnatural to refer to these as meta-dimensions ofcompetitive strategy. The procedure assumes the


8/28


presence of an unknown number of these orthog-onal dimensions in the population of all competi-tive strategy designs. Each studys factor solutionis taken as a sample estimate of these populationdimensions, and each studys estimate of theelements most closely associated with each factor

is taken as a sample estimate of the elementsmost closely associated with a meta-dimension.Because of the above-noted variability in con-

structs and measures, cumulation of factor scoresacross studies is not meaningful. What amountsto a voting procedure is used instead. Each vectorof factor coefficients reported in a study istransformed to a vector of votes: elements whichshow significant nonzero coefficients on the factorare coded 1; others coded 0. Each vote vector istaken as a sample record that identifies thoseelements that are significantly associated with a

meta-dimension of competitive strategy. Clusteranalysis is then used to aggregate these multi-element vote vectors across studies into com-monly occurring patterns. Each cluster of similarvote patterns, indicating which elements of com-petitive strategy are most often associatedtogether with an orthogonal factor, are taken asthe best aggregate description of a meta-dimension that can be derived from the empiricalliterature. Taken together, the set of clustersdescribe the number of orthogonal meta-dimensions of competitive strategy that have

been isolated.Finally, the incidence of votes for each

element clustered together in a meta-dimensionis compared to its overall frequency, using as ametric the standard test statistic for differencesin proportions. As is well known, the use ofcluster analysis to create categories violates theassumptions required to use these statistics todraw inferences from the sample of studies to apopulation. As discussed below, other methodsmust be used to assess the validity of the clustersolution (Ketchen and Shook, 1996). The statisticis used here for the simpler purpose of focusingthe description of each meta-dimension on thoseelements of competitive strategy that are mostdistinctive of that dimension in the availableempirical record.

The method follows the same logic, in a multi-variate context, as the statistically correct bivari-ate procedure of vote counting, in which theproportion of studies with significant effect sizesis compared to that expected under a null hypoth-


esis of no relationship between the variables(Hunter and Schmidt, 1990: 473). As Hunter andSchmidt note, a majority of nonzero effects istypically not needed to reject the null hypothesis,and it is the use of the majority criterion that isresponsible for the errors associated with vote

counting as a meta-analytic procedure. The focusof vote counting on the existence, rather thanthe effect size, of relationships is a recognizedlimitation of the method in univariate meta-analyses, but is appropriate for the purpose ofisolating patterns of relationships, as in this casebetween elements of competitive strategy.

One step in this procedure prohibits its use asan inferential meta-analysis, that is, the use ofcluster analysis to aggregate vote vectors, and theconsequent violation of the assumptions requiredto draw inferences to the population of all com-

petitive strategies. Instead, the procedure producesa descriptive aggregation of the accumulated evi-dence on the independent dimensions of competi-tive strategy, as they have emerged in the empiri-cal record to date. More universal claims mustawait more powerful procedures.

When assembling multiple studies into a meta-analysis, the question arises whether or not toweight each study by sample size. Monte Carlosimulations suggest that large samples are to bepreferred for their lower exposure to artifactualvariation (Koslowsky and Sagie, 1994). On the

other hand, when a meta-analysis includes studiesthat follow a skewed distribution of sample sizes,with outliers of very large or small samples,Osburn and Callender (1992) recommend the useof unweighted results, and conclude that there islittle to be gained from sample-size weighting inmost meta-analyses. The empirical literature oncompetitive strategy is highly skewed towardssample sizes of less than 100, with a long tailreaching out to n = 2578 (see Table 2). Accord-ingly, this meta-analysis uses vote vectorsunweighted for sample size.

Methods of clustering appropriate to this meta-

analysis

Use of cluster analysis in strategic managementresearch has been critically reviewed by Ketchenand Shook (1996). They conclude that the designof these analyses must be careful to match theanalysis to the type of data involved, and to assessthe reliability of results by triangulating the results


9/28


of several distinct clustering methodologies together.Following their advice, the distinctive demandsplaced on cluster analysis within the present meta-analytic methodology are now examined.

Since vote vectors involve binary data, themore familiar measures of similarity such as

Euclidian and Mahalanobis distances are inappro-priate. Of the binary measures of similarity, theJaccard coefficient was selected because itincludes only common occurrences of a pair andignores common absences (Aldenderfer andBlashfield, 1984: 29). This is appropriate to thedata because not all studies of competitive strat-egy include all elements, so that the absence ofan element can be due to differences in studydesign rather than strategic behavior.

The ultimate choice of clustering algorithmshould be guided by the clusters shape in the

n-dimensional space used to describe the data,and variation in cluster sizes (Aldenderfer andBlashfield, 1984: 5962). It is therefore advisableto explore several methods as a means ofenlightening the structure of the data (Aldenderferand Blashfield, 1984: 45; Miller, 1981; Ketchenand Shook, 1996).

Judgments on the shape of clusters and choiceof algorithm should also be guided by theory.The algorithm most often applied in the empiricalliterature on generic competitive strategy is hier-archical agglomeration, using Wards method to

link successive cases to the closest cluster. It isknown that this algorithm produces, and is mostappropriate to, clusters which are of approxi-mately equal size and are uniformly spread overthe various dimensions of analysis as hyper-spheres. Such shapes cannot be ruled out ex ante,and hierarchical agglomeration is used as oneapproach to representing the data of this study.

However, this algorithm is inconsistent withthe dominant paradigms assertion that eachdesign must emphasize a subset of strategies inwhich it will seek a distinctive advantage.Designs consistent with this assertion will formellipsoids aligned to those dimensions of competi-tive strategy that are emphasized within the clus-ter. Attempts to represent such data using hier-archical agglomeration have led to some famousfailures of cluster analysis in fields such as astron-omy (Wishart, 1969). Also, there is no reason toexpect designs to be equally popular, or clustersto be of equal size. There are thus reasons toprefer the density analysis algorithm (Wishart,


1969, 1987: 64), which is better suited to theidentification of ellipsoidal clusters, and whichalso protects small, dimensionally focused, clus-ters from being merged with larger neigh-boring hyperspheres.

The study accordingly applied both these algor-

ithms to the data, together with an iterative relo-cation algorithm which optimizes within-clustersimilarities and between-cluster dissimilarities.Like hierarchical agglomeration, iterative relo-cation is most suited to representing clusters thatform hyperspheres (Aldenderfer and Blashfield,1989: 48). The CLUSTAN procedures HIER-ARCHY, DENSITY, and RELOCATION wereused (Wishart, 1987). Hierarchical agglomerationused the average linking rule in place of themore familiar Wards method on the grounds thatit best preserves the structure of the data space

(Aldenderfer and Blashfield, 1984: 45; Milligan,1980). Density analysis estimated the density sur-rounding each case as the average of distancesto its seven nearest neighbors, as recommendedby Wishart (1987: 66) for this size of data set.

Data

This section discusses the measures of strategyused in the empirical record, assessing their com-parability, and their ability to support the meta-analysis.

Selection of studies

A search of the ABI-Inform data base was con-ducted in late 1995, yielding 126 entries forcompetitive strategy and 84 for generic strat-egy. From these records, and subsequent refer-ences, 17 studies were identified that describedcompetitive-strategy designs using principalcomponent factor analysis and cluster analysis.With two exceptions (Miller and Friesen, 1986a;Wright et al., 1991), elements of competitivestrategy are first aggregated into a smaller number(n) of dimensions of strategic design using princi-pal component factor analysis. Thirteen of thestudies use cluster analysis to isolate commondesigns within that n-dimensional space. Tenstudies also examined the performance of designs.Table 2 identifies which study applied each analy-sis.

It has been found that the selection of studiesis the most important source of difference


10/28


Table 2. Empirical studies of the dominant paradigm

Sample Factors Clusters Performance measures

Basis for No. of Report No. of PeriodSample rating factors factor clusters Financial of

Study size strategies identified S.D. identified return Growth years Scale

Carter et al.(1994) New ventures 2578 Internal 6 6Davis andSchul (1993) Pulp & paper 180 Internal 6 Yes 3 v v 5 ContinuousDess andDavis (1984) Paint industry 78 Internal 3 Yes 4 v v n/a IntervalDouglas and PIMS U.S.Rhee (1989) domestic 250 Competitors 7 6 v n/a Continuous

PIMSEuropean 187 Competitors 7 6 v n/a Continuous

Galbraith and PIMSSchendel consumer(1983) products ?600 Competitors 6 6 v v 5 Continuous

PMS industrialproducts ?600 Competitors 6 4 v v 5 ContinuousGreen et al. Portugese(1993) manufacturers 68 Internal 4Hambrick PIMS( 1983 ) industrial

products 400 Competitors 17 Yes 10 v 4 CategoryKim and Lim Korea(1988) electronics 54 Competitors 4 Yes 4 v v 3 ContinuousKotha et al. SIC 3439 177 Internal 6(1995)Miller (1992b) Single-

industry SME 45 Competitors 4 Yes 5Miller and PIMS

Friesen consumer(1986a, durables 102 Competitors Yes 10 v v n/a Continuous1986b)Morrison and GlobalRoth (1992) competition 306 Internal 5 4 v v 3 IntervalNayyar, 1993 Product

managers 496 Internal 3Parker and DecliningHelms (1992) industry 87 Internal 3 YesPrince (1992) Canadian

construction 240 Competitors 7 4Robinson and VariousPearce (1988) industries 97 Internal 4 5 v v 5 IntervalWright et al. Screw machine

(1991) products 56 Competitors Yes 3 v v 5 Continuous

between pairs of meta-analysis (Wanous, Sulli-van, and Malinak, 1989). Here, both computer-assisted and manual methods were used to iden-tify relevant studies, so that the only restrictionwas that a study be published before or during1995. Three limitations for the meta-analysis fol-low. First, the analysis is open to any publication


bias in favor of positive associations betweenelements of competitive strategy.

Second, the configurations of competitive strat-egy that emerge will be descriptive only of thekind of firms that have been studied to date. Itcan be seen from Table 2 that certain types offirm are likely to be overrepresented in the


11/28


empirical record: manufacturing as against servicesectors; and large enterprises. With the exceptionof Kim and Lim (1988), all studies are alsoof Western management practice, with imperfectvalidity in other cultures (Kotha, Dunbar, andBird, 1995). There is a clear need to widen the

range of contexts in which competitive strategyis studied.The frequent use of the PIMS data base is

evident, particularly for those studies that evaluateperformance. Repeated sampling from the PIMSdata base raises concerns that the meta-analysiswill be subject to a common methods bias, pro-ducing greater convergent validity in associationsbetween strategy elements than independentsamples would display. The analysis is not asexposed to this bias as Table 2 suggests, wherenearly half of the factors in the empirical record

stem from PIMS data. For reasons discussedbelow, only 54 of the factors listed in Table 2are used in the meta-analysis, and of these 17(31%) derive from PIMS-based studies. These inturn have sampled different sections of the database (consumer or industrial corporations; U.S.or European), and at different times, reducing thedanger of spurious convergence. In only six cases(11% of the factors used) are factors based onsamples drawn from the same data (Galbraithand Schendels and Hambricks 1983 studies ofU.S. industrials), and here a consolidation of the

results of these two studies is suggested. How-ever, removal of either study made no differenceto the resulting meta-analysis, and both areretained in the following report.

The period covered also limits the range ofcompetitive behaviors to those which have beeninvestigated so far. The consequences of thisconstraint are explored in the next section.

Measures of the elements of competitive

strategy

The elements of competitive strategy that areused in the meta-analysis must be comparableacross studies, but not necessarily identical. Tomaximize the use of the information in theempirical record, elements were included provid-ing only that they appeared in a minimum of twostudiesa bare minimum for clustering purposes.Aspects of the selection of variables are of con-cern to both meta-analysis and cluster analysismethodologies, and are discussed in turn.


Given the ability of cluster analysis to extractgroups of similar attributes regardless of anyrationale for the grouping (Ketchen and Shook,1996), it is advisable to begin with variables thathave a theoretical association with competitivestrategy. Table 3 shows the variables that have

been isolated from the empirical record as thebasis for this meta-analysis. All of these elementshave been cited as contributing to competitivestrategy, whether in Porters work, or in thePIMS program (Buzzell and Gale, 1987). Currentinterpretations of several of these elements wouldnow distinguish them as elements of the firmsresource portfolio, rather than its strategies.Skilled workforce, modern plant, reputation, andchannel influence are variables of this kind. Theiruse here may be justified as proxies for thestrategies used to create each resource: strategies

of acquisition and investment in human, physical,reputational, and relational assets.

Another concern is the use of only thoseelements associated with the small number offactors which each study has isolated from itsdata. Factors with low eigenvalues that have beendiscarded take with them information which istherefore lost to this analysisinformation whichmight influence the grouping of strategy elementsinto clusters. To assess the magnitude of thisloss, the proportion of variance in each elementthat is explained by factors was averaged over

the studies in which the element appears. Thisranges from 0.4 for skilled workforce andrefining existing products to 0.66 for lowprices, with a mean over all elements of 0.53.The common dimensions of strategic design rep-resented by these factor solutions thus leavenearly one-half of the variance in strategicelements unexplained. Interpretations of the domi-nant paradigm that seek powerful commondescriptors of strategic design are weakened bythis result: particularly the taxonomic approach(Proposition 1a); but also the empiricist anddimensional approaches (Propositions 1b and 1d).

Meta-analysis is similarly concerned with theselection of variables (Wanous et al., 1989; Bull-ock and Svyantek, 1985). As already noted, dif-fering model specifications have been found tomake the largest contribution to interstudy vari-ance in effect sizes (Farley et al., 1995). As anew field of investigation, it is likely that animportant limitation of this meta-analysis is itsrestriction to variables used in the published


12/28


record as of 1995. It can be seen from Table 3that the elements available offer a good coverageof strategies in the domain of marketing, oper-ations, products, and market scope. However, themeta-analysis must omit many important strate-gies which are explored, and which achieve sig-

nificance, in one study only: capacity utilization(Morrison and Roth, 1992); economies of scale(Kim and Lim, 1988); receivables management(Hambrick, 1983); backward integration(Galbraith and Schendel, 1983); and patents(Miller, 1992b). These omissions are particularlyevident in the domains of technology strategyand cost-emphasis strategies. The latter play acentral part in the dominant paradigm, and haveevidently not been studied to the degree appropri-ate, or possible (Amit, 1986; Fisher, Westney,and Gupta, 1994).

It has also been suggested that the scope ofcompetitive behaviors explored in the empiricalrecord is too narrow (Miller, 1992a). Strategiesthat have received little attention to date include

Table 3. Proportional frequency of strategy elements within meta-dimensions of competitive strategy

Elements of competitive strategy: Meta-dimensions of competitive strategy(relative emphasis on)

Quality Product MarketMarketing Sales reputation innovation Operations scope

1: advertising 0.70** 0.83** 0.092: brand identification 0.70**3: channel influence 0.80**4: marketing innovation 0.60** 0.115: promotion 1.00** 0.09 0.116: sales force 1.00**7: reputation 0.10 0.42* 0.09 0.228: high prices 0.10 0.25 0.82**9: new products 0.20 0.45**

10: refine products 0.18 0.2211: specialty products 0.08 0.73**12: product quality 0.10 0.83** 0.1113: quality control 0.08 0.09 0.33*14: service quality 0.20 1.00**15: procurement 0.08 0.09 0.44**16: skilled workforce 0.10 0.25 0.56*17: manufacturing innovation 0.08 0.56**18: operating efficiency 0.78**19: unit cost reduction 0.09 0.22*20: modern plant 0.33**21: product breadth 0.10 1.00**22: customer breadth 1.00**8a: low prices 0.10 0.11

**Significant at 0.01 level*Significant at 0.05 level


information management (Doll, 1989; Pyburn,1991); logistics (Christopher, 1993); and speedto market (Dess and Rasheed, 1992; Blackburn,1990; Stalk, 1989; Stonich, 1990). Of particularimportance to the paradigms descriptive systemis the very limited attention given to organi-

zational aspects of competitive-strategy design.As discussed above, organizational constraints arethe principal reason for expecting cost- and differ-entiation-emphasis designs to be mutually exclu-sive. The omission is particularly important giventhe considerable evidence that organizationalarrangements mediate and enhance the relation-ship between strategy and performance (Davisand Schul, 1993; Govindarajan and Fisher, 1990;Powell, 1992, 1993; Gupta, 1987; Miller, 1986,1987, 1988; White, 1986; Treacy and Wier-sema, 1995).

Another judgment call of a meta-analysis isthe extent to which variables that are given differ-ent names in studies are treated as identical(Wanous et al., 1989). In the empirical literature


13/28


on competitive strategy, there is a large degreeof consistency in variables across studies becausemany studies have used the PIMS data base (seeTable 2), and many others have used the set ofstrategy elements developed by Dess and Davis(1984), which are themselves closely comparable

to the PIMS measures. Beyond these identicalelements, some closely related keywords werealso grouped together: image with reputation;customized products with specialty products;customer service with service quality; rawmaterials and relations with suppliers with pro-curement. The element skilled workforce bringstogether a number of keywords associated withquality human resources: high caliber, trained,experienced, skilled. Also included here areelements used in one study to describe man-agement quality (Prince, 1992).

The 17 studies use two distinct bases of ratingin measuring the importance of each element toa respondents competitive strategy, as shown inTable 2. One rates emphasis on each elementof strategy relative to competitors (the PIMSapproach); the second relative to other elementsof the firms strategy. Only the first is consistentwith the descriptive and performance-theoreticpropositions of the dominant paradigm. Thesecond could be quite a poor proxy in that anemphasis on some element of competitive strat-egy that is distinctive for a firm may be well

short of distinguishing the firm from its competi-tors. Use of this proxy might therefore beexpected to lead to an overstatement of distinc-tive-emphasis designs, with a consequentialunderstatement of no-distinctive-emphasisdesigns, relative to classifications using the moreappropriate comparison with competitors. Asreported below in Study Two, this bias was notevident in the data.

Significance criterion for vote vectors

In contrast to other meta-analysis procedures wherecoding has been found to be a source of error(Wanouset al., 1989; Bullock and Svyantek, 1985),the coding of variables into vote vectors wasstraightforward. Studies use different criteria forsignificance, however, the most conservative requir-ing a minimum factor matrix coefficient of 0.5, asused by Carter et al. (1994), Dess and Davis(1984), and Green et al. (1993). For consistency,the most conservative criterion is applied uniformly


to all studies. The result of this conservatism isthat the meta-analysis acknowledges a relativelysparse degree of interconnectedness in meta-dimensions of competitive strategy.

The empirical record contains 54 factors witha minimum of two loadings of 0.5 or higher on

the 23 common elements shown in Table 3. Thusthere are 54 samples from which to estimate anunknown number of population dimensions ofcompetitive strategy design.

Results

The three clustering algorithms produced nearlyidentical results. Of the 54 cases classified in asix-cluster solution, 53 (98%) were given thesame classification by each algorithm. This highagreement indicated high levels of reliability in

the meta-analytic descriptions of dimensions ofcompetitive strategy (Ketchen and Shook, 1996).The iterative relocation solution was chosenbecause it optimizes the closeness of associationwithin clusters. Further, the close agreementbetween methods strongly suggests that this opti-mum is a global one.

The six-cluster solution was preferred becauseit provided a clean separation of elements: withone exception, elements are associated with onecluster only. This is desirable since the clustersare to represent meta-dimensions that are orthog-

onal to each other. The exception was advertisingstrategy, which is associated with both marketingand sales dimensions, and leads to the merger ofthese dimensions in the five-cluster solution. Thesix-cluster solution is nevertheless preferredbecause of the clear distinction between market-ing and sales on all other elements.

Table 3 shows the proportional frequency ofelements within each cluster in the iterative relo-cation six-cluster solution. The significance levelsof differences between this and the elementsoverall frequency is used to focus on thoseelements of strategy which are most distinctiveof each meta-dimension in the empirical record.Two elements which do not display distinctiveassociations are discarded at this stage: refineproducts (10), and low prices (8a).

The marketing dimension is distinguished fromsales by the long-term nature of the relationalassets targeted by these strategies: influence inthe channel, and consumer brand loyalty. Theassociation of marketing innovation with this


14/28


dimension suggests that such assets require sus-tained attention to retain their value.

Quality is identified as a distinct dimension ofcompetitive strategy, independent of operationsstrategy (and quality control), and involves atten-tion to both product and service quality. The

association of quality control with operationsstrategy and not the quality meta-dimension wasunexpected. It suggests that quality control ismostly perceived as a component of internaloperating processes, and that these can vary inde-pendently of a customers perception of the qual-ity of product or service.

The association of a firms reputation with thequality dimension replicates an association foundin the PIMS study (Buzzell and Gale, 1987). Thelink indicates the importance of quality for cre-ating and sustaining one of the firms most dis-

tinctive competitive assets (Kay, 1994) and asource of long-term competitive advantage.

The strategy of high pricing is shown to beassociated most strongly with product innovation,although there is some association with qualityalso, consistent with the findings of the PIMSstudy. Market scope, which embraces breadth inboth product lines and consumer segments, isclearly independent of other dimensions of com-petitive strategy. The strategic decision to beactive in a broad range of the market is shownto be invariably associated with a broad product

range. Products within the range do not, however,have to be distinguished from those of competi-tors. Specialist products are associated insteadwith the meta-dimension of product innovation.

The meta-dimensions are broadly similar to thedescriptive system proposed by Mintzberg, albeitwith some differences (Mintzberg, 1988). Mintz-bergs generics of quality, product design, market-ing image, and low price/low cost, correspond tothe meta-dimensions of quality reputation, productinnovation, marketing, and operations. The differ-ences are, first, that the meta-analysis isolates adistinct dimension of sales strategy, independentof marketing. The meta-analysis also suggeststhat marketing leadership involves not only themanagement of image (as suggested byMintzberg) through branding and advertising, butalso the building of channel power. Thirdly, themeta-analysis does not find a dimension anal-ogous to Mintzbergs support generic. Instead,the broad product scope that is one characteristicof support is found to be closely associated


with a broad range of customer segments in thesingle dimension of market scope. Acceptingthese differences, the meta-analysis suggests thatthe empirical record has more in common withMintzbergs more elaborate descriptive systemthan that of the dominant paradigm.

The accumulated evidence of the empirical rec-ord suggests that the use of Porters three dimen-sions of cost emphasis, differentiation emphasisand market scope to define the space of competi-tive strategy is insufficient to meaningfullydescribe the range of competitive strategy designs.The dimensional interpretation of the dominantparadigm, formalized in Proposition 1d, is notsupported. Instead, the record suggests thatadequate descriptions of competitive strategyrequire attention to each of six independent meta-dimensions. Within this descriptive system, the

distinction between types of differentiation is asimportant as the distinction between any one ofthem and operations-cost leadership.

Nominalist and taxonomic interpretations canaccommodate this variety while continuing toassert that cost and differentiation act as high-level discriminators of competitive strategydesigns. These claims are assessed in Study Two.

STUDY TWO: META-DESIGNS OFCOMPETITIVE STRATEGY

Description and classification of competitive-

strategy designs: The nominalist interpretation

The meta-dimensions isolated in Study One pro-vide a common language with which to describecompetitive strategy designs. Study Two usesthese constructs to produce consistent descriptionsof the designs found in the empirical record.Because of differences in their interpretation ofthe dominant paradigm, descriptions sought bythe nominalist approach are developed first; thoseof the taxonomic and empiricist approaches, inthe next section.

Thirteen studies use cluster analysis to producedescriptions of competitive strategy designs.Eleven of these use factor scores as the input toclustering, and two form clusters directly fromelements of competitive strategy (Miller andFriesen, 1986a; Wright et al., 1991). A total of80 clusters is reported (Table 2).

Studies report the differences between clustersin a table of J rows by K columns, showing


15/28


for each of the factors isolated in the study fj(j = 1, . . ., J), and for each cluster ck(k= 1, . . ., K), the average factor score forrespondents included in that cluster, Cjk. Whenclusters are formed directly from I elements ofcompetitive strategy, ei (i = 1, . . ., I), the table

reports the average element score for respon-dents, Cik.Neither the units of measurement of Cjk or

Cik, nor the factors, nor the clusters are directlycomparable across studies. To proceed, the meta-analysis must define variables and units ofmeasurement that are common to all.

Variables

As for Study One, the closest comparabilitybetween studies is in the individual elements of

competitive strategy, 21 of which can now beused to describe each design in terms of one ormore of the six meta-dimensions of competitivestrategy.

Two of the 13 studies report the requiredelement scores, Cik, directly. For the remainder,cluster descriptions must be restated from factorscores, Cjk, to elements. This was done by takingthe relative score of a factor over Kclusters, Cjk(k= 1, . . ., K), to index the relative score of eachof its constituent elements. In the few cases ofconflict, e.g., where a factor includes elements

that are associated with two different meta-dimensions, the conflict was resolved in favor ofthe element with the highest factor loading.

These consistent descriptions are gained, how-ever, at the cost of descriptive coverage: of the91 factors used by studies to describe clusters,only 59 (65%) are associated with common meta-dimensions. The taxonomic interpretation of thedominant paradigm, which seeks comprehensivedescription and classification of competitive strat-egy designs, is again weakened by this limitedcoverage.

Units of measurement

For the purposes of the dominant paradigm, therelevant characteristic of a competitive strategydesign is its success in achieving distinctiveemphasis on some aspects of strategy relative tocompetitors designs. This is operationalized asthe maximum over K clusters in a study ofelement scores Cik, or their associated factor


scores Cjk. Distinctive emphasis on an element iscoded 1, otherwise 0. Elements 112 and 14(Table 3) define the differentiation archetype, andelements 13 and 1520 define cost emphasis fromoperations strategies.

But Hambrick and others widen the definition

of cost emphasis to include clusters with distinc-tively low emphasis on strategies leading to dif-ferentiation advantage (Hambrick, 1983),reflecting Porters view that differentiation isusually costly (Porter, 1985: 18). To oper-ationalize this definition, elements associated withdifferentiation were represented with two vari-ables, one each for distinctively high and lowemphasis. These are listed in Table 5. Distinc-tively low emphasis on an element of competitivestrategy is measured as the minimum over Kclusters of element scores Cik, or equivalent factor

scores Cjk.These additional dimensions of cost advantage

are an attempt to capture the full scope of thatconcept in Porters account. However, it is opento question whether distinctively low emphasison, for example, marketing strategies can alwaysbe equated with a focused effort to specify anderadicate unnecessary marketing cost. Hence,these measures seem likely to overstate the extentof cost-emphasis strategies among respondingfirms, and consequently understate the number ofno-distinctive-emphasis, stuck-in-the-middle, designs.

Finally, each of the two elements of strategydesign that describe market scope are also oper-ationalized as two variables: one to describe dis-tinctive breadth (elements 21 and 22), the otherto describe distinctive market focus (elements 21aand 22a). A complete description of the distinc-tive features of a cluster, relative to others in astudy, is then produced as a vector of binaryvalues over all 35 variables shown in Table 5,showing 1 for those elements in which the clusterdisplays distinctively high or low emphasis, other-wise 0.

Classification

The nominalist interpretation of the dominantparadigm requires that the 80 clusters of theempirical record be allocated to one of four cate-gories: the two ideal types of distinctive cost-and differentiation-emphasis; a category of mixed-emphasis designs, which is expected to be smallin number; and a category of nonideal designs


16/28


that do not achieve distinctive emphasis of eithertype (stuck-in-the-middle). The classificationsresulting from the above descriptions are summa-rized in Table 4. Of the 80 clusters, 34 show oneform or other of cost emphasis, and 26 showsome form of differentiation emphasis. As noted,

the number of designs classified as cost emphasisseems likely to be an overstatement, and thenumber of no-emphasis designs understated.

While it is thus possible to aggregate the vari-ous forms of cost- and differentiation-emphasistogether, consistent with Proposition 1c, the valid-ity of this aggregation across independent dimen-sions of competitive-strategy design is now opento question. The nominalist interpretation of thedominant paradigm nevertheless insists that thetwo archetypes are the fundamental basis onwhich all distinctive designs are constructed.

Results of the following analysis shed furtherlight on this view. Also, relative to other interpre-tations, the nominalist view is less concernedwith descriptive accuracy and most concernedwith the performance consequences of departuresfrom the normative ideal. The nominalist inter-pretation must therefore be assessed primarily

Table 4. Summary descriptions of clusters of competi-tive strategy

Costemphasis Differen-(economy tiation

Meta-dimensions* in /from ) emphasis

Sales 10 8Marketing 8 7Product innovation 20 19Quality reputation 15 15Operations 13

Market scope Broad Narrow

9 9

Archetypes Totals

Cost emphasis 34Differentiation emphasis 26Mixed emphasis 9No emphasis 11

80

*Entries in this part of the table do not sum to totals belowdue to multiple entries.


on its predictive validity. This is reported inStudy Three.

As predicted by Proposition 1c, the number ofmixed-emphasis designs is small, representing 11percent of the total. Distinctive emphasis on mar-ket scope is also rare, and most designs (62 of

80) are neither clearly broad nor narrow in scope.Concern that the different bases of rating stra-tegic emphasis in these studies (internal or rela-tive to competitors) might lead to biases in thedescription of competitive-strategy designs provedto be unfounded. A chi-squared test of the inci-dence of single-, mixed- and no-emphasis designsin the two types of measure showed no significantdifference between them.

Meta-designs of competitive strategy:

Taxonomic and empiricist interpretations

The empiricist and taxonomic interpretations ofthe dominant paradigm seek to identify commonlyoccurring, or generic, classes of competitive-strategy design. Consistent description of all clus-ters in the empirical literature permits a searchfor these meta-designs.

Method

Cluster analysis was used to identify the mostcommon patterns in the binary vectors that

describe the distinctive features of each cluster,using the same procedures as those used in StudyOne. These meta-analytic estimates, which bestsummarize the cluster solutions of the empiricalrecord, are naturally referred to as meta-designsof competitive strategy.

Of the 80 clusters, only 63 show distinctiveemphasis on two or more of the 35 variablesused to describe them, a bare minimum for thepolythetic gestalts that are expected to describegeneric designs. In the density analysis clustering,a range of k-means was explored because thenumber of modal clusters formed proved to bevery sensitive to this parameter. Values greaterthan 4 produced a single-cluster solution; k= 2produced 16 modal clusters. The preferred so-lution of 6 modal clusters was produced at k= 3.This number of clusters suited the summativeobjectives of a meta-analysis, but still producedclusters that were distinct from each other andinterpretable in terms of the meta-dimensions ofcompetitive strategy.


17/28


Results

The six modal clusters produced by density analy-sis are reported in Table 5. The table shows theproportional frequency of elements within each

Table 5. Proportional frequency of strategy elements within meta-designs of competitive strategy

Meta-designs of competitive strategy

D1 D5 BroadInnovation D3 Focused quality D6 Focused

and ops D2 Cost quality D4 Sales and sales qualityleadership economy economy leadership leadership leadership

Elements of competitive strategy: n = 10 n = 26 n = 7 n = 5 n= 10 n= 5

Emphasis on:

1: advertising 0.10 0.12 0.80** 0.302: brand identification 0.08 0.203: channel influence 0.194: marketing innovation 0.12 0.205: promotion 0.80** 0.40**6: sales force 0.80** 0.40**7: reputation 0.20 0.12 0.30 0.408: high prices 0.80** 0.04 0.30 0.80**9: new products 0.80** 0.04 0.60* 0.10

11: specialty products 0.60** 0.04 0.1012: product quality 0.10 0.08 0.40* 1.00**13: quality control 0.10 0.0414: service quality 0.10 0.12 0.50** 1.00**15: procurement 0.10 0.0416: skilled workforce 0.20 0.08 0.2017: manufacturing innovation 0.20* 0.0818: operating efficiency 0.20**19: unit cost reduction 0.10 0.0420: modern plant 0.10 0.08 0.1421: product breadth 0.10 0.14 0.60**22: customer breadth 0.04 0.14 0.70**

Economies in/from:

1a: advertising 0.31* 0.14 0.202a: brand identification 0.083a: channel influence 0.10 0.194a: marketing innovation 0.10 0.085a: promotion 0.23 0.14 0.20 0.206a: sales force 0.19 0.14 0.20 0.207a: reputation 0.23 0.14 0.108a: low prices 0.10 0.42* 0.57*9a: new products 0.35** 0.10

11a: specialty products 0.19 0.2012a: product quality 0.15 0.72** 0.2014a: service quality 0.23 0.71** 0.1021a: product focus 0.71** 0.20* 0.10 0.20*22a: customer focus 0.04 0.71** 0.20 0.10 0.40**

**Significant at 0.01*Significant at 0.05


cluster, and the result of significance tests ofdifferences between this and the elements overallfrequency. As before, these are used as an indi-cator of those features of each design which mostdistinguish it from others in the empirical record.


18/28


In this analysis, the three clustering algorithmsproduced substantially different solutions. Across-tabulation of the meta-designs derived bydensity analysis and hierarchical agglomerationshowed only 25 of the 63 designs (40%) wereplaced in equivalent clusters by the two methods.

A choice between these two clusterings is notclear cut. Hierarchical agglomeration isolates amarketing- and operations-leadership clusterwhich the density algorithm subsumes within thecluster of cost-economy designs (D2), due to asingle shared characteristic of distinctively lowprices. The failure of the density algorithm todistinguish this from archetypal cost-economydesigns is an important shortcoming.

In other respects, however, the density solutionis to be preferred. As discussed in Study One, itis better suited to the ellipsoidal, dimensionally

focused shapes which theory predicts that com-petitive strategy designs will take; and it is ca-pable of isolating locally dense swarms of verysimilar designs from surrounding and more dif-fuse clusters. Hence a large agglomeration ofquality and innovation leadership designs pro-duced by hierarchical agglomeration is partitionedin the density solution into distinct innovation(D1) and quality-leadership designs, the latter inturn differentiated into broad and focused variants(D5 and D6) which have no parallel in theaverage-linkage solution. The density solution

was accordingly chosen as the best representationof the empirical record.1 Its major limitation ofconfusing a marketing- and operations-leadershipdesign with cost economy is compensated for inStudy Three.

Two cost-economy meta-designs are isolated,both associated with low prices, as Mintzberg(1988) suggests. A very large cluster, cost econ-omy (D2) is most distinguished by its lowemphasis on advertising and product innovation,but also with economies in a broad range ofother sources of differentiation (elements 3a, 5a14a). The focused-economy meta-design (D3)is distinctive for low emphasis on the qualitymeta-dimension. In addition to the low prices thatit shares with the more broadly defined costeconomy, it appears that the design also involvesthe further competitive protection of a limitedmarket scope.

1A tabulation of the average-linkage solution is availablefrom the author.


Neither of these cost-economy meta-designsis associated with operations leadership, as bothnominalist and taxonomic interpretations mightlead us to expect. Operations leadership is insteadassociated most strongly with leadership in prod-uct innovation (D1). This meta-design is the only

example of a mixed-emphasis design producedby the density algorithm; although it should berecalled that a marketing- and operations-leadership cluster (involving six cases) was alsoisolated by hierarchical agglomeration. The 10cases included in D1 represent 16 percent of the63 designs included in the analysis (rising to25% if the other six-case mixed design is added),consistent with the expectations of both nominal-ist and taxonomic interpretations that the fre-quency of mixed designs will be low.

The remaining three modal clusters involve

distinctive emphasis on a source of differentiationadvantage. Broad quality and sales leadership(D5) displays distinctive emphasis on the qualityand sales dimensions of competitive strategy, andis the only design to apply its strengths to adistinctively broad market scope. Distinguishedfrom this design are two small clusters, eachdisplaying distinctive emphasis on just one ofthese two dimensions: sales leadership (D4),and focused quality (D6). Both also show someevidence of being associated with focused-marketstrategies, in contrast to the broad scope of the

larger cluster.Focused quality (D6), and innovation- and

operations-leadership (D1), are the only twodesigns in the empirical record to support distinc-tively high prices. The ability of quality to sup-port a price premium has been found in the PIMSstudy (Buzzell and Gale, 1987: 110), and theimportance of commanding a premium for inno-vative products in order to maintain high levelsof internal reinvestment has been a feature of thestrategies of firms such as Hewlett-Packard, Intel,and others.

Figure 2 shows the dendogram of the densitysolutions final five cluster fusions. These exposethe hierarchy of associations between the sixmeta-designs. With the exception of focus qualityleadership, it is evident that the hierarchy ofassociations assembled in this meta-analysis isconsistent with that predicted by the taxonomicinterpretation and Proposition 1as hierarchicalclassification rules (compare Figure 2 withFigure 1). The behavior also supports the nomi-


19/28


Figure 2. Hierarchical classification of metadesigns

nalist view that cost and differentiation are funda-mental, high-level discriminators between com-petitive strategy designs. Cost-economy designscluster together, as do differentiation designs,before either type is combined with the other,or with the mixed design of innovation- and

operations-leadership.The reason why focused quality leadership is

excepted from this hierarchy is that it is a veryhomogeneous and dense cluster. Density analysisprotects the distinct identity of this cluster untilall of the other, more diffuse, clusters have been

joined. In other respects, as Table 5 shows, thecluster is most similar to the two differentiation-based designs, D4 and D5.

The empirical record is thus consistent withboth of the essential features of the nominalistinterpretation: the frequency of mixed designs isrelatively low; and cost and differentiation do actas high-level discriminators of competitive strat-egy designs. Proposition 1c is supported.

The same two features also characterize thetaxonomic interpretation. In other respects, how-ever, the requirements of a taxonomic inter-pretation are not met. As Table 5 shows, nodesign displays features which are universallydistinctive of their members and uniquely dis-tinguishing of the cluster (i.e., elements common


to 100% of a cluster but with no incidenceelsewhere). The allocation of individual designsto meta-design classes is thus stochastic ratherthan deterministic. The limited coverage of thepopulation of competitive strategies that is pos-sible with these meta-designs, as discussed above,

also fails to support the universal ambitions ofthe taxonomic approach. Proposition 1a is notsupported.

This studys success in deriving meta-designsthat display distinctive strategic emphasis, andembrace all of the clusters isolated in the empiri-cal literature to date, is, however, consistent withthe less restrictive requirements of the empiricistinterpretation. Proposition 1b is supported.

STUDY THREE: PERFORMANCE OFGENERIC COMPETITIVESTRATEGIES

The dominant paradigms central theorem pro-poses an association between competitive-strategydesigns and financial performance, the latterclassified into above average and average-or-less.The empirical literature has also explored theeffect on growth performance. The empirical testshave invariably taken the form of a simple model


20/28


linking competitive strategy designs to perform-ance, with no mediating variables or contin-gencies built into the model specification. In partthis has been due to the limited degrees of free-dom available to each study, and in part tothe limited articulation of these effects into the

dominant paradigms theory of performance. Theempirical record thus offers only a first steptowards testing the paradigms theory, in a partialmodel where the direct effects of interest totheory may be swamped by indirect effectsoperating through associations between strategyand other variables affecting performance.

Method

The effect of competitive strategy design on therelative odds of a grouping of designs being in

the above-average category, as against average-or-less, was estimated by logistic regression(Agresti, 1990) using various definitions of designas regressors. The choice of regressors was variedto represent the different interpretations of theparadigm. The use of these more powerful tech-niques of categorical data analysis, permittingmore precise modeling of the propositions of thedominant paradigm, is one of the advances overindividual studies made possible by meta-analysis.

The nominalist interpretation, which seeks todescribe individual competitive strategy designs

in terms of proximity to an ideal, was oper-ationalized by the fourfold archetype classificationdescribed in Study Two (Table 4). The nominalistinterpretation further requires a measure of howclose to cost standard are differentiation-emphasisdesigns, and vice versa. Considerable effort wasspent in devising a measure of competitive prox-imity from the empirical record; but this variableproved to be highly collinear with archetypeclassification, and had to be omitted. The logisticregression was designed to evaluate the log-oddsof performance of single- and mixed-emphasiscategories relative to the no-emphasis category.Proposition 2b requires both coefficients to bepositive.

Performance models for the taxonomic andempiricist interpretations (Proposition 2a) areframed in terms of generic classifications.Regressors for these models were the six meta-designs of competitive strategy that emerged fromStudy 2 (Table 5). Because the cost-economymeta-design has been found to confuse within it


a cluster of marketing-and-operations-leadershipdesigns, these were isolated as a seventh meta-design in a separate estimation of the model. Thelogistic regression was again designed to evaluatethe log-odds of performance of these meta-designs relative to clusters with no distinctive

strategic emphasis. Proposition 2a requires eachof these coefficients to be positive.

Data

Ten of the 17 studies investigate performancedifferences between the clusters they identify. Themeasures used are the average over a period ofyears in financial return, and in sales growth.Both measures are used in this meta-analysis.Table 2 identifies which studies assess the per-formance of generics, and which measure is used.

Of the 80 clusters isolated in the empirical record,65 include a measure of financial performance,and 43 a measure of growth performance. Forreasons explained in Study Two, not all of thesecould be included in the analysis of meta-designs,and these regressions were based on 61 and 33cases respectively. Although a variety of scalesare used, all are capable of identifying above-average performance, which is all that is requiredfor the performance propositions of the domi-nant paradigm.

The measures are subject to the well-known

limitations of accounting measures of perform-ance. Also, performance is typically self-assessed.Several studies attempt to assess the accuracy ofthese assessments from public sources and findcorrelations ranging up from 0.4 (Morrison andRoth, 1992; Robinson and Pearce, 1988). At thelower end, therefore, these measures are of mar-ginal value, explaining less than 20 percent of thevariance in true performance. Third, the period ofyears covered by performance data, which istypically 5 years, is towards the low end ofwhat would be regarded an acceptable period formeasurement of long-run sustained performance.Studies using a 3-year period offer a less reliableestimate. All of these limitations mean that theresults of this analysis must be treated with somecaution. The models nevertheless assemble all ofthe evidence that has been accumulated so far onthe explanatory power of the dominant paradigm.They can claim, on this basis, to offer the besttest of this theory to date.

The measure of sales growth is subject to


21/28


further reservations, on grounds of both measure-ment consistency and theoretical validity. Somestudies standardize for industry growth and recordgrowth in market share; others use unstandardizedsales growth. Growth performance is thereforesubject to an important inconsistency between

studies. Equivalent standardizations for industryprofitability are not made, however.Also open to question is the validity of using

growth measures in place of the financial returnsspecified by the theory. Profits, and the competi-tive balance between cost and price premiums,are central to the paradigm. Hence there aretheoretical as well as measurement reasons forthis meta-analysis to prefer financial return overgrowth as a measure of performance.

Results

Effects on performance of meta-designs

Table 6 shows the results of logistic regressionsof meta-designs on financial and growth perform-ance. Two specifications are reported for eachperformance measure: one using the no-emphasisdesign as the basis of comparison; the other usingaverage performance over all meta-designs. Inmodels of growth performance, some categorieswere omitted due to small membership size.

By Proposition 2a, both empiricist and taxo-

nomic interpretations require a positive coefficientin the logistic regressions reported in the left-hand column of Table 6. As can be seen, onlytwo metadesigns show this effect on financialperformance: innovation and operations leadership(D1), and broad quality and sales leadership(D5). For growth performance, only the lattershows the expected effect. In no case are theodds of above-average performance significantlypositive.

Several meta-designs record negative coef-ficients, exactly contrary to theory; although inonly one case is the coefficient significant, at the0.1 level. In the case of cost economy, some partof this result may be attributed to inadequatemeasurement in the empirical record, as discussedabove. But this concern does not apply to theother metadesigns which also produce lower oddsof above-average performance.

In Study Two it was found that the metadesignsproduced by density analysis confused a subsetof marketing- and operations-leadership designs


with the cost economy metadesign. The logisticregressions reported in Table 6 were accordinglyrerun with this subset identified as a separatecategory.2 The category showed positive, but notsignificant, coefficients for both financial andgrowth performance, relative to the no-distinctive-

emphasis design. The residual cost-economy cate-gory showed larger negative coefficients thanreported in Table 9, for both measures of per-formance, both of which were significant at the0.05 level.

Overall, the models have very limited explana-tory power: chi-square tests show no significantimprovement in log-likelihood due to the model,and only two-thirds of each models classi-fications are made correctly. Proposition 2a is notsupported. An alternative interpretation would bethat the poor predictive performance of these

models is evidence of poor criterion validity inthese meta-designs (Ketchen and Shook, 1996),particularly in view of doubts about the reliabilityof classifications derived using different clusteringmethods. However, the next section of the paperreports essentially the same result using classi-fications of competitive strategy design whichcircumvent the process of clustering designs. Theinability of the empirical record to display thebehavior predicted by the dominant paradigmdoes not appear to be an artifact of this particularclustering solution.

Instead the empirical record suggests that theodds of the no-emphasis design producing above-average financial and growth performance areclose to the average for all designs (as reportedin the right-hand columns of Table 6). Only onemeta-design, innovation and operations leader-ship, shows significantly higher-than-average oddsof superior financial performance, at the 0.1 level.With that exception, the analysis suggests thatany competitive strategy design is as capable asany other of producing above-average perform-ance. Importantly, the conclusion applies equallyto no-distinctive-emphasis designs.

Effects on performance: Archetypes

Table 7 reports a model of the essential featureof the nominalist interpretation of competitiveperformance: that single-emphasis designs, and a

2Full results are available from the author.


22/28


Table 6. Performance differences among meta-designs

Dependent variable: Financial return (n = 61) Financial return (n= 61)Basis category: No distinctive emphasis Average of all designs

Regression Regressioncoefficient Significance coefficient Significance

D1 Innovation and ops leadership 1.25 0.118 1.23 0.099D2 Cost economy 0.76 0.096 0.77 0.086D3 Focused quality economy 0.41 0.657 0.44 0.602D4 Sales leadership 1.10 0.341 1.15 0.250D5 Broad quality and sales leadership 0.69 0.327 0.67 0.314D6 Focused quality leadership 0.00 1.000 0.04 0.966No distinctive emphasis 0.05

2 log-likelihood 76.38 0.030 75.88 0.033Model chi-square 8.19 0.225 8.68 0.192% correct classifications 67.21% 67.21%

Dependent variable: Growth (n = 33) Growth (n = 33 )

Basis category: No distinctive emphasis Average of all desi

Documents

What Have We Learned From Generic Competitive Strategy