structural analysis of competitive behavior: new empirical ... · the new empirical industrial...

Ž .Intern. J. of Research in Marketing 18 2001 161–186www.elsevier.comrlocaterijresmar

Structural analysis of competitive behavior: New EmpiricalIndustrial Organization methods in marketing

Vrinda Kadiyali a, K. Sudhir b,), Vithala R. Rao a

a Johnson Graduate School of Management, Cornell UniÕersity, Sage Hall, Ithaca, NY 14853-6201, USAb Stern School of Business, New York UniÕersity, 44 West Fourth St, MEC 8-80, New York, NY 10012, USA

Abstract

The impact of a firm’s strategic marketing mix choices on profitability can be evaluated by understanding the impact ofthose choices on consumer demand for the firm’s products and on the firm’s costs. Additionally, a firm’s strategic marketingmix choices, and its demand and costs can be affected by rival firms’ strategic choices. Therefore, to understand the effectsof choice of marketing mix on profitability, we have to understand its effects on demand, cost and competitor reactions. Theeffects of choices of marketing mix on consumer demand have been analyzed in great depth in marketing, but research onthe strategic reactions of competitors to such choices have been far more limited. The New Empirical Industrial OrganizationŽ .NEIO framework provides us with a source of methods that has potential to substantially add to our insights aboutcompetitive interactions among firms.

In this paper, we first discuss a simple NEIO model to illustrate the basic methodology. We then discuss the contributionsof this literature to our knowledge of competitive marketing strategy. In the process, we discuss methodological extensionsof the basic model that are needed to model the institutional realities of specific markets. We also summarize how theexisting literature has evolved, and provide our view of where the literature might profitably proceed from here. Inparticular, we discuss how future methodological innovations in the dynamics of competition, discrete strategy choice, andasymmetric information estimation will enable wider application of this methodology to competitive marketing strategyissues. The main advantage of NEIO studies is that they provide greater understanding of the competitive behavior inspecific markets or industries compared to cross-sectional studies across industries. Bountiful opportunities exist foradditional studies that focus on similar phenomena in different markets to draw generalizable conclusions from this line ofresearch. q 2001 Published by Elsevier Science B.V.

Keywords: Structural models; Competition; Strategic behavior; New empirical industrial organization

1. Introduction

The impact of a firm’s strategic marketing mixchoices on profitability can be evaluated by under-

) Corresponding author. Tel.: q1-212-998-0541; fax: q1-212-995-4006.

Ž .E-mail address: [email protected] K. Sudhir .

standing the impact of those choices on consumerdemand for the firm’s products and on the firm’scosts. Additionally, a firm’s strategic marketing mixchoices, and its demand and costs can be affected byrival firms’ strategic choices. Therefore, to under-stand the effects of choice of marketing mix onprofitability, we have to understand its effects ondemand, cost and competitor reactions. The effectsof choices of marketing mix on consumer demand

0167-8116r01r$ - see front matter q 2001 Published by Elsevier Science B.V.Ž .PII: S0167-8116 01 00031-3

( )V. Kadiyali et al.r Intern. J. of Research in Marketing 18 2001 161–186162

have been analyzed in great depth in marketing, butresearch on the strategic reactions of competitors tosuch choices have been far more limited.

There is a rich tradition of empirical research inmarketing strategy beginning in the 1950s that exam-ines the impact of cost and competitive character-istics of a market on the profitability of firms. Thisempirical tradition following the structure–

Ž .conduct–performance hereafter referred to as SCPparadigm of empirical industrial organization usescross-sectional data across industries to find empiri-cal regularities across industries. Many of these stud-ies in marketing have used the Profit Impact of

Ž . ŽMarketing Strategies PIMS data see Buzzell and.Gale, 1987 for a survey . These studies have pro-

vided valuable insights about the empirical regulari-ties of relationships between marketing mix choiceslike advertising, cost components including R&Detc., and profits of firms.

Beginning in the late seventies, advances in gametheory have led to a large amount of theoreticalresearch analyzing strategic issues in the context ofcompetition between firms, firms and channel mem-

Žbers, firms and their advertising agencies, etc. For a.review, see Moorthy, 1993. This research convinced

Žempirical researchers that market outcomes i.e.,firms’ strategic marketing mix choices and the result-

.ing sales, etc. and profitability are not merely afunction of the broad structural characteristics usedin SCP studies. Rather, these market outcomes andprofitability are affected by specific industry andfirm specific demand and cost characteristics that aredifficult to model within the SCP framework ofcross-industry analysis. A consequence of these in-sights has been the birth of the ANew Empirical

ŽIndustrial OrganizationB literature henceforth re-ferred to as NEIO; for a review see Bresnahan,

.1989 . This literature incorporates more industry-and firm-specific details in modeling demand, cost,and competition as steps in analyzing the relation-ship between marketing mix and profits. Therefore,this approach should be seen as the next step in thestream of empirical research in marketing strategyafter the SCP literature. The goal of this paper is toreview this literature and provide an agenda forfuture work.

The NEIO approach involves the developmentand estimation of structural econometric models of

strategic, competitive behavior by firms. By a struc-tural model, we mean a model where firms’ choicesare based on some kind of optimizing behaviorŽ .usually profit maximization . In this respect thesemodels are similar to structural models of consumerchoice, which are built on the assumption of utilitymaximization behavior of consumers. Where theydiffer is that NEIO structural models are strategicwhile structural models of consumer choice are non-strategic. Consumer models are non-strategic be-cause one consumer’s choice has no impact on an-other consumer’s choice and therefore, these choicescan be assumed to be independent. In contrast, NEIOmodels of firms need to account for the interdepen-dency of firm choices: a firm’s choice will cause areaction from its competitor. This modeling of strate-gic behavior is the key difference between structuralmodels of consumer choice and structural models offirm choice.

This difference between consumer choice and firmchoice has two econometric implications. The firstissue is simultaneity. Firms make their strategic mar-keting mix choices simultaneously. That is, any onefirm’s choice is a function of its rivals’ choice, andrivals’ choice a function of this firm’s choices. Fur-ther, firm choices affect demand and demand charac-teristics affect firm choices. Therefore, from an esti-mation viewpoint, the realized demand and the firm’sstrategic choices are simultaneous. The simultaneityis accounted for by estimating the demand equationsand the choice equations of firms as a system ofsimultaneous equations. The second issue is endo-geneity. In consumer choice models, it is assumedthat each individual’s choice by itself has no effecton the firm’s choices such as prices and promotionsin consumer choice studies; therefore, firm’s deci-sions are treated as exogenous. However, in a struc-tural model of firm choice where separate equationsfor firms’ choices are estimated, the choices have tobe treated as endogenous. We therefore have to useinstruments for these choice variables to account forthe endogeneity.

A structural model of competitive interaction pro-vides at least four benefits, which we will elaborateon in various parts of the paper.

Ž .1 Theory testing: The structural approach pro-vides an opportunity to empirically compare and testalternative theories of strategic behavior. A better

( )V. Kadiyali et al.r Intern. J. of Research in Marketing 18 2001 161–186 163

fitting model can be deemed to be more descriptiveof the phenomenon.

Ž .2 Ease of interpretation: Structural models areŽlinked to a behavioral theory for example, profit

.maximization of firms; hence, the estimated parame-ters have economic or behavioral meanings that canbe interpreted easily. This advantage is parallel touncovering the parameters of the consumer utilityfunction in consumer choice studies.

Ž .3 GWhat-if H analysis: Estimated parameters ofa structural model are invariant to policy changesŽi.e., they are not subject to the well-known prob-

.lems of the Lucas critique . Managers can thereforeuse these estimates to perform Awhat-ifB analysis tounderstand what effect their actions will have on themarket.

Ž .4 Decomposing the determinants of marketpower and profitability: Many studies in NEIO havefocussed on measuring market power. The Lernerindex1 is a good measure of market power andserves as a measure of the profitability of firms.Structural NEIO models decompose the sources of

Ž .market power profitability into components associ-ated with demand structure, cost structure and com-petitive interaction. Hence, differences in profitabil-ity among firms in an industry can be attributed to

Ž .consumer preferences demand structure , efficiencyŽ .cost advantages , and anti-competitive conductŽ .tacitly cooperative behavior . Armed with this de-scription about the structure of their market, eachfirm can figure out how they can use the marketingmix to most efficiently boost their market power. Asimilar analysis can be done for better understandingthe determinants of profitability in closely related

Žmarkets for e.g., different geographical markets.within the same industry .

The rest of this paper is organized as follows. InSection 2, we compare other modeling traditions inempirical industrial organization to the NEIO ap-proach. We use a simple example to illustrate theprinciples of NEIO modeling in Section 3. We re-view the literature and its contributions in Section 4.In the process, we discuss the current limitations inthe literature and suggest several methodological ex-

1 Ž .Lerner indexs priceymarginal cost rprice.

tensions to the basic model that are needed to repre-sent specific institutional realities. We conclude thepaper with a summary in Section 5.

2. Modeling traditions in Empirical IndustrialOrganization

We classify the modeling traditions in EmpiricalIndustrial Organization in Fig. 1 to provide anoverview of various methods and their differences.The new empirical industrial organization paradigmis primarily focused on the analysis of firm behaviorin a specific market or closely related markets. Thisis in contrast to the previously dominant empiricalparadigm in the field of industrial organization, thestructure–conduct–performance or SCP paradigm.We briefly review the SCP paradigm, the reduced-form hypothesis testing approach and the reactionfunction approaches, before discussing in detail the

ŽNew Empirical Industrial Organization paradigm or.structural analysis of strategic behavior .

2.1. The structure–conduct–performance paradigm

Until the early eighties, empirical analysis of firmbehavior in industrial organization was based on the

Ž .structure–conduct–performance SCP paradigm fol-Ž .lowing the seminal work by Bain 1951 . The SCP

paradigm is illustrated in Fig. 2.The industry competitive structure is measured by

such factors related to both demand- and cost-sideprimitives; common measures include concentration,growth, scale economies, buyer and supplier power,barriers to entry, and degree of product differentia-tion. Industry conduct is defined by the factors re-lated to competitive behavior as choices of market-

Fig. 1. A taxonomy of Empirical Industrial Organization modelingapproaches.


Ž .Fig. 2. Structure conduct performance SCP approach.

ing mix, decisions to enter a market, R&D invest-ment. Industry performance is measured by prof-itability, rate of innovation, etc. The SCP paradigmstates that industry structure drives industry conduct,which in turn drives industry performance. Much ofthe analysis of firm behavior in marketing and strat-egy using PIMS data was also based on this paradigmŽsee Buzzell and Gale, 1987 for a comprehensive

.review . To marketers and business strategy re-searchers, the structure- and conduct-based determi-nants of profitability were of great interest, becausethey could provide managers with guidelines onwhat structural characteristics make markets prof-itable and what kinds of conduct are conducive toprofitability. For example, whether managers shouldpursue market share for higher profits has been oneof the most controversial questions in this literatureŽ .for a meta analysis, see Szymansky et al., 1993 .Therefore, the literature using SCP paradigm hasprovided many important insights on empirical regu-larities across markets and within markets.

There are some conceptual, methodological anddata issues with this literature. One conceptual con-cern is the inability to effectively capture the hetero-geneity in the structural characteristics and the result-ing optimum marketing mix strategies of firms acrossindustries, and across firms within a given industry.Typically, studies in this literature pool cross-sec-tional data across disparate industries in estimatingthe SCP relationships. To address this criticism,researchers have used econometric methods to con-trol for heterogeneity across the population. Onetechnique employed in the literature is to run sepa-rate regressions for different homogeneous sub-

Ž .groups of firms Prescott et al., 1986 . Anotherapproach is to allow fixed effects for firms; thisapproach has exploited the panel nature of PIMS

Ž .data Boulding and Staelin, 1990, 1993 . Controllingfor heterogeneity in this fashion addresses the econo-metric problem.

This issue of heterogeneity is, however, deeperthan a debate about the appropriateness of economet-ric methodology. The relevant issue is that the het-erogeneity among industries is more than what canbe captured by the structural variables used in SCPstudies. Markets differ in more fundamental wayswithin the industry, such as differences in cost anddemand structures among firms and order of movesby different competitors. Game theory shows thatthese kinds of market-specific characteristics have asubstantial impact on market outcomes. This impliesthat the structural variables used in SCP studies arenot sufficient statistics to account for the vast hetero-geneity across industries and the resulting marketoutcomes. Such heterogeneity can best be accountedfor by carefully modeling the details of a specificmarket. Clearly there is a trade-off between depthand breadth. SCP studies by virtue of their breadth ofanalysis across industries provide generalizable in-sights across industries. In contrast, since NEIOstudies are case studies, they do not lend themselvesto easy generalizations, though there is greater depthin the observed relationships. Generalizability inNEIO should come through replication of results inmultiple case studies across similar markets. Wediscuss this tradeoff in more detail in Section 6.

A major methodological issue with the SCPparadigm is the non-exogeneity of structural vari-ables. That is, the structural characteristics that deter-mine conduct are not truly exogenous. While struc-ture has an impact on conduct, conduct in turn canaffect structure. Consider these relationships betweenconduct and structure: marketing strategies likechoice of product characteristics and advertising af-fect product differentiation; barriers to entry are afunction of the product positioning and pricing be-havior of incumbents; R&D investments affect thecost structure of firms in an industry. As Wind and

Ž .Lilien 1993 say, AThe PIMS results are norms;therefore, the equations do not have a causal inter-pretation. Although it is tempting to predict theconsequences on profitability of changes in the inde-

Ž .pendent variables of the PAR model, it is notreasonable to do so.B The estimated relationships inSCP models are therefore correlational, not causal.One solution to this endogeneity issue is to useinstrumental variables for the endogenous structuralvariables. However, good instruments are hard to


find. A more complete solution would be to modelthe process by which these endogenous structuralvariables are chosen by firms, or in other words,follow the structural estimation route of NEIO stud-ies.

A third issue with the SCP literature is the type ofdata used in several of these studies. Researchershave typically used three- or four-digit SIC industrycodes to define industries in which to study SCPrelationships. These code-based industry definitions

Žmay be too narrow sometimes e.g., aspartame shouldbe included in the competitive set when examining

. Žthe market for sugar , and too broad sometimes e.g.,the market for instant photography is quite distinctfrom the market for regular developing photographic

.film . Of course, industry definitions are critical inthe construction of structural variables such as con-centration and market share, and in estimating mar-keting mix–profitability relationships. PIMS studiesare more careful in this regard, because they analyzecompetition at the SBU level. Nevertheless, evenSBUs usually have product lines targeted to differentsegments. The appropriate structural measures wouldbe at the segment level, but SCP studies usuallywould find it hard to find such detailed data across a

Ž .broad cross-section of industries. Sudhir’s 2001aanalysis is an illustration of how a segment levelanalysis of the auto market provides more detailedinsights than would be possible with an aggregatelevel analysis.

A more vexing data problem is the measurementof costs. Costs are crucial in the construction ofmarket performance variables such as profitability.Many studies have used accounting costs for measur-ing profits. But accounting costs are usually average

Ž .costs and not correlated with marginal economiccosts. Similarly, accounting profits are not the sameas economic profits. Therefore, conclusions of SCPusing accounting data can be suspect.

Managerially, while cross-industry observationsno doubt provide some guidance in selecting opti-mum marketing mix, managers are often interestedin the details of their own industry because theseprovide invaluable inputs in to strategic decisions.

Ž .Consider the study of Suslow 1986 , where a de-mand-and-supply model of aluminum is estimated tounderstand where and why Alcoa, the leading alu-minum producer, makes profits. Suslow carefully

includes the recycled aluminum suppliers in definingthe industry. She finds that despite the presence ofthese additional suppliers in the industry, Alcoa hassignificant market power because recycled aluminumis viewed as inferior to primary aluminum. Neverthe-less, recycled aluminum does act as a brake on thepricing power of Alcoa for primary aluminum. Sincethe quantity of recycled aluminum in a market is a

Žlagged function it takes several years for primary.aluminum to become recycled aluminum of the

primary aluminum produced by Alcoa, Alcoa’s cur-rent pricing policies will have an impact on its futuremarket power. Managers can therefore use Suslow’sestimates of demand, supply and costs to determinean optimal long-term pricing policy. Such an exer-cise would be impossible to do using a standard SCPapproach.

Summarizing, SCP studies are unable to ade-quately capture heterogeneity across industries andfirms in the marketing mix–profitability relation-ships. Additionally, there are econometric issues ofendogeneity of some of the explanatory variables inthe SCP paradigm. Last, there also appear to beserious data issues with these studies. Therefore, theresults of the SCP studies should be seen as provid-ing us stylized facts of marketing mix–profitabilityrelationship, rather than giving us insights in to whyand how this relationship is structured.2 Further,since the estimates of a structural approach are spe-cific to a market, they are managerially useful indetermining the optimal marketing mix.

2.2. ComparatiÕe statics or hypothesis testing

A reduced-form method is to test comparativestatics generated from structural game-theoreticmodels with regressions between structural charac-teristics and industry conduct. The crucial differencebetween the SCP approach and the reduced-formhypothesis testing lies in the fact that hypothesistesting is done on a cross-section of closely relatedmarkets or using time series data on a single market.Therefore, while SCP regressions provide correla-

2 For a critique of the SCP approach and also a summary of themain insights gained from this stream of research, refer to

Ž .Schmalensee 1989 .


tional rather than causal relationships, these regres-sions give causal relationships resulting from strate-gic competitive optimization resulting from game-theoretic behavior models. But unlike structuralNEIO studies, these reduced-form approaches do notdecompose demand, cost and competitive effects onfirm strategic choices.

We illustrate the principles underlying reduced-form model-based hypothesis testing with a paper by

Ž .Agrawal and Lal 1995 . This paper tests severalhypotheses about the role of royalty and monitoringarrangements in a franchising system. Based on the

Ž .comparative statics in Lal 1990 and Agrawal andŽ .Lal 1995 , Agrawal and Lal regress brand name

investments, service levels and monitoring frequencyagainst royalty rate and monitoring costs. Their theo-retical models guide their expectations on the signsof these effects. For example, a higher monitoringcost should lead to lower monitoring frequency.Based on the evidence from these regressions, they

Ž .argue that the model of Lal 1990 and its extensionŽ .in Agrawal and Lal 1995 are consistent with the

data.An issue with the above hypothesis-testing ap-

proach is that reduced-form analyses provide nopolicy invariant descriptions of specific markets. Thatis, since we do not obtain demand, cost and competi-tion primitives from the estimation process, the esti-mates cannot be used in an optimizing model formanagers making decisions. The advantage is thatthey place very limited structure either on demand-or supply-side and hence are useful in exploratoryempirical analysis of competitive behavior.

2.3. Reaction function studies

The reaction function approach is another re-duced-form method that measures how firms react tocompetitors within a particular industry by regress-ing a firm’s marketing mix against other elements ofits own marketing mix and those of its competitorsŽ .including lagged values . The pioneering work in

Ž .this area was by Lambin et al. 1975 and HanssensŽ . Ž .1980 . Leeflang and Wittink 1996 use this tech-nique to analyze pricing and promotional competi-tive behavior using store level scanner data. Theyprovide a very useful link to structural models be-

cause they estimate both market share equationsŽ .demand-side of the market as well as reaction

Ž .functions supply- or competitive-side of the market .Unlike the hypothesis testing described in Section2.2 above, these regressions do not result from anydirect optimization behavior. But unlike the SCPliterature, reaction functions are always at the firmlevel within an industry and hence, account for theindustry- and firm-level heterogeneity in marketingmix–profitability relationships.

Though the reaction function technique is usefulas a description of competitive reactions, it does notprovide insight into the underlying reasons for theobserved reactions. This is because the reactions arenot based on primitives of demand and cost charac-teristics facing firms and the nature of the equilib-rium between firms. For example, if reaction func-tions show that firms’ optimal choices are veryresponsive to their rivals’ choices, is this becausefirms have very similar cost conditions and thesecosts are changing over the time period studied? Oris it because competition is very intensive and hence,there is price matching every time? Or is it becausefirms collude, but demand shifts cause price shifts?Such questions can be answered in a structural ap-proach because the reactions are linked to demand,cost and equilibrium characteristics underlying themarket.

Nevertheless, the reaction function approach doesprovide insights into off-equilibrium behavior offirms or to model dynamic interactions, both ofwhich, as we will discuss below, are difficult tomodel in structural models. Reaction functions canalso provide interesting starting points for furtherempirical work using the NEIO framework.

3. Implementing the NEIO framework: An illus-tration

In this paper, we will use a simple example toillustrate the workings of a NEIO model. We willalso address some issues that a modeler needs toresolve in using these models. Comparing the NEIOmodels to those discussed in the Section 2, the gains

Ž .come from two features: i these are estimated at theindustry level, and more specifically, at the firm


Ž .level within an industry, and ii these are derivedfrom explicit demand and cost, and profit-maximiz-ing primitives.

Because an NEIO model is specified at the levelof the industry, the first task is to define the industry.In many studies, the focus has been on the largesttwo or three firms in an industry or market. Re-cently, methods have been developed to deal withmarkets where there are a large number of products.

Ž .An NEIO model has three basic ingredients: 1 aŽ . Ž .demand specification, 2 a cost specification and 3

a specification for competitive interactions.3 Demandspecifications have ranged from simple models suchas linear or log-linear demand to flexible nonlinearmodels such as the logit model with random coeffi-cients to account for heterogeneity of customer pref-erences. Costs have been specified using a simpleconstant marginal cost model or as a linear or log-linear function of production cost factors. The speci-fication of competitive interactions needs a littlemore elaboration. In our illustrative example, weexplain three commonly used approaches to model

Ž . Ž .competitive interactions: 1 the menu approach, 2Ž .the conjectural variation approach and 3 the con-

duct parameter and the weighted profits approach.Consider a duopoly market with each firm having

a differentiated product. The demand for the twoŽ .products q and q is as follows.1 2

q sa qb p qb p ,1 1 11 1 12 2

q sa qb p qb p .2 2 21 1 22 2

Let the marginal costs of the two products be c and1

c . We will assume each firm chooses prices in such2

3 Ž .Studies such as those of Panzar and Rosse 1987 and Ashen-Ž .felter and Sullivan 1987 have many ingredients of NEIO studies

but are not completely structural. Specifically, Panzar and Rosseestimate a reduced-form equation of revenue as a function offactor prices. Although their specification can differentiate be-tween competitive and monopolistic markets, they make an as-sumption of identical firms in the marketplace. The NEIO frame-work has shown this assumption to be a restrictive one. Ashenfel-ter and Sullivan estimate a supply curve for firms in the cigaretteindustry as function of tax and other cost shifters. However, NEIOstudies have demonstrated that to accurately estimate price andquantity movements, demand-side variables have accounted for aswell.

a way as to maximize profits in each period. Theprofits for firm i is given by,

P s p yc q , is1, 2.Ž .i i i i

3.1. The menu approach

We derive the first-order conditions under differ-ent equilibrium interactions between the two firmsand then choose the equilibrium that best fits thedata. As an illustration, we derive the econometricestimation models under three equilibria here:Bertrand, Stackelberg leader–follower and collusiveequilibrium. In the estimation step, each of these setsof first-order conditions for the different games isestimated separately, and goodness of fit test per-formed to select the best-fitting one.

3.1.1. The Bertrand equilibriumIn the Bertrand equilibrium, each firm i chooses

its prices assuming that competitors do not react tochanges in their prices, i.e., they assume Ep rEp s0.j i

The first-order condition for firm i is:

EP Eqi is p yc qq s0.Ž .i i i

Ep Epi i

For the linear demand equation this is

EPis p yc b qq s0, is1, 2.Ž .i i i i i

Epi

These first-order conditions now serve as thesupply-side equations describing firm behavior. Notethat the equations link demand parameters and costs,i.e., there are cross-equation restrictions.

3.1.2. The Stackelberg leader–follower equilibrium4

Let firm 1 be the leader and firm 2 be thefollower. In the Stackelberg equilibrium, the fol-lower firm 2 chooses its prices, assuming that theleader does not react to changes in its prices, i.e., itassumes Ep rEp s0. In contrast, the leader firm1 2

anticipates the follower’s reaction and incorporates

4 An astructural approach to inferring leader–follower behaviorŽ .is Granger Causality. See Roy et al. 1994 for an illustration.


this into its pricing behavior. The first-order condi-tion for the follower firm 2 is similar to the Bertrandequilibrium.

EP2s p yc b qq s0.Ž .2 2 22 2

Ep2

Taking the derivative for the first-order conditionwith respect to p , we get1

Ep yb2 21s .

Ep 2b1 22

The leader’s first-order condition is,

E q E q E p1 1 2p yc q qq s0.Ž .1 1 1ž /E p E p E p1 2 1

For the linear demand model,

b21p yc b yb qq s0.Ž .1 1 11 12 1ž /2b22

3.1.3. The collusiÕe equilibriumIn the collusive equilibrium, firms choose their

prices as if they collusively maximize their totalprofits. Each firm chooses prices for its products asif it was a monopoly firm selling both differentiated

Ž .products. The firms’ objective is Ps p yc q q1 1 1Ž .p yc q . The first-order condition for firm i is2 2 2

given by:

EPs p yc b q p yc b qq s0,Ž . Ž .i i i i j j i j i

Epi

j/ i , is1, 2.

3.1.4. EstimationFor each of the models, there are four equations,

two describing the demand for each firm and twodescribing the supply for each firm. In performingthe estimation, all these models can be nested in onegeneral model with the following four equations.

q sa qb p qb p ,1 1 11 1 12 2

q sa qb p qb p ,2 2 21 1 22 2

q ql p ql p qd c qd c s0,1 11 1 12 2 11 1 12 2

q ql p ql p qd c qd c s0.2 21 1 22 2 21 1 22 2

The restrictions given in Table 1 apply to the param-eters when estimating the individual equilibriummodels. These models form a simultaneous equation

Table 1Parameter restrictions for different equilibrium models

Parameter Bertrand Leader–follower Cooperative

l b b yb b r2b b11 11 11 12 21 22 11

l 0 0 b12 12Ž .d yb y b yb b r2b yb11 11 11 12 21 22 11

d 0 0 yb12 12

l b b b22 22 22 22

l 0 0 b21 21

d yb yb yb22 22 22 22

d 0 0 yb21 21

system and a simultaneous equation instrumentalvariable estimation approach will yield consistentand efficient estimates of the parameters. Three-Stage

Ž .Least Squares 3SLS , Full Information MaximumŽ .Likelihood FIML , and Generalized Method of Mo-

Ž .ments GMM may be used for estimation.

3.1.5. Model selectionSince the equilibrium models discussed above are

nested within the general model, a test of restrictionscan be performed. This test enables us to rejectequilibrium models that are not consistent with thedata, i.e., enables us to Alet the data speakB. Sincethese are structural models, the parameters them-selves should be of expected signs. Any estimationthat produces unintuitive signs provides indication ofmodel mis-specification and therefore may be re-jected. However, even after these tests, one may notbe able to reject several models. Eventhough all ofthese models are nested within the general model,the specific models themselves are non-nested ineach other. Hence, non-nested tests are used to chooseamong the several models that are still not rejectedby testing restrictions. More generally, it is possiblethat the games being tested are not nested in anygeneral model nor in each other. Again, in suchcases non-nested model tests have to be performed.

With 3SLS estimates, usually the model that min-imizes the sum of squares is considered the ArightBmodel. With FIML as the estimation procedure,

Ž .Bresnahan 1987 uses Cox tests for non-nestedŽmodels which are based on the ratio of the likeli-

.hoods of the models to choose the right model.Ž . Ž .Gasmi et al. 1992 and Roy et al. 1994 also use

Žlikelihood-based non-nested hypotheses tests Vuong,


.1989 for model selection. When using GMM, wemay use the equivalents of Cox tests since in manycases, the maximum likelihood estimation has been

Žshown to be a special case of GMM Hamilton,.1994, p. 14.4 .

In discussing model selection among alternativemodels of competition, we are making the assump-tion here that the functional form for demand that weuse is appropriate. Given that firms’ pricing behavioris closely related to the demand functional form,recent studies have focussed on using highly flexible

Ždemand models to limit potential for bias random.coefficients logitrLA-AIDS, etc. or tested among

Žalternative models of demand Genesove and Mullin,.1998; Sudhir, 2001a .

3.2. The Conjectural Variations approach

Ž .The Conjectural Variation CV approach, pio-Ž .neered by Iwata 1974 , offers a method by which

different types of equilibria can be captured by oneparameter. In the CV models, firms are postulated tohave conjectures about how competitors will react tochanges in their marketing mix and incorporate theseconjectures into their decisions. The first-order con-ditions for firm i in the duopoly case we set upearlier is given by,

Eq Eq Epi i jp yc q qq s0,Ž .i i iž /Ep Ep Epi j i

j/ i , is1, 2.

For the linear model this is

Epjp yc b qb qq s0, j/ i , is1, 2.Ž .i i i i i j iž /Epi

The term Ep rEp may be interpreted as a conjec-j i

ture of firm i about how firm j will change its pricein response to a change by firm i. Hence, the termAconjectural variationB. The CVs are estimated fromthe data. As described in Table 2, different equilibriaimply different CV estimates for the demand andoptimization system discussed above. Note that theCV approach above can be extended to advertisingand other instruments as well. As in the menu ap-proach, the two demand equations and the supplyequations are estimated using a simultaneous equa-

Table 2CV parameters associated with different equilibria for this demandsystem

Equilibrium CV parameters

Bertrand ZeroLeader–follower Zero for follower,

Non-zero for leaderDominant firm–fringe Zero for dominant with respect tofirm fringe

Negative for fringe with respect todominant

Cooperative Positive for bothCompetitive Negative for both

CVs for different competitive instruments, e.g., advertising, andfor different demand systems can be different from the values inthe table.

tion instrumental variable estimation approach like3SLS, FIML or GMM.

Industrial Organization textbooks typically intro-duce the notion of conjectural variations with ahomogenous product market with n firms. In thismarket, the Bertrand model has a zero CV, collusionhas a CV equal to 1 and Cournot competition has aCV of 1rn. Hence, there is a common notion thatthe estimated CVs should be below 1, but this isapplicable only in the case of a market with homoge-neous products. In the case of price competition withdifferentiated products, there is no unambiguous up-per bound. In practice, it is useful to test whether theCV estimates make sense by computing the marginsresulting from the estimates. Very high CVs may bemeaningless because beyond an upper bound, discon-tinuities occur and margins become negative. This ismeaningless because positive CVs for price implycooperative behavior and higher margins. For anexample of how an upper bound on the CV isderived for a differentiated product model, see Nor-

Ž .man 1983 .There are some advantages and disadvantages of

the CV approach relative to the menu approach. Asthe number of players and the number of marketinginstruments increase, the number of potential com-petitive equilibria increases combinatorially. In themenu approach, since a separate estimation is re-quired for each of these equilibria, the estimation andmodel selection can become potentially cumbersome.The CV approach, on the other hand, nests thedifferent equilibria and therefore avoids the explo-


sion in the number of estimations. Usually, however,when the number of firms and the number of market-ing instruments increase, the CV models may not beidentified. This means that identifying restrictionsneed to be made to estimate the model. Sometimesthere may not be justifiable reasons for the choice of

Ž .these restrictions. Nevo 1998 discusses identifica-tion issues in CV models.

Some researchers have argued that the estimatedconjectures in the CV approach may not be consis-tent and hence, the CV approach is theoreticallyunsound. However, recent theoretical work hasshown that the conjectural variations approach shouldbe interpreted as a reduced-form way to capture

Ž .dynamic behavior in a static model. Dockner 1992derives a conjectural variations equilibrium as the

Ž .steady state of a subgame perfect closed loop Nashequilibrium in a non-cooperative differential gamewith adjustment costs. This model provides a theoret-ical foundation for negative conjectural variations.

Ž .Cabral 1995 formulates a repeated game and showsthat the collusive equilibrium is equivalent to a CVequilibrium with a strictly positive conjectural va-

Ž .riation. We take the perspective of Lindh 1992in saying A . . . conjectural variations should beviewed . . . as a measure of the deviation from somecommon standard like the Cournot behaviorB.

Ž .Corts 1999 has investigated conditions underwhich a CV estimation approach will fail to estimatethe right competitive interaction. He finds that in

Žmarkets with high seasonality negative correlation.in demand states , the potential to mismeasure com-

petitive interaction is severe. In recent empiricalwork, armed with cost data for the sugar and the spot

Ž .electricity markets, Genesove and Mullin 1998 andŽ .Wolfram 1999 , respectively, find that the differ-

ence in their CV estimates are not very differentfrom direct measures obtained using cost data. Whilethese studies give us some faith in the empiricalvalidity of the CV approach, more research is neededto reassure us that CV approaches are effective ininferring the right competitive interaction or to high-light conditions under which the technique fails.

3.3. The conduct parameter and the weighted profitapproaches

Ž .Bresnahan 1989 explains the equivalence of themenu and the CV approach from the point of view of

empirical work. In particular, the use of the AconductŽ .parameterB Porter, 1983 , a close cousin of the CV

parameter, is proposed. Porter’s model is based onquantity competition. To be consistent with our pre-vious discussion, we develop an equivalent model interms of price competition.

Eq Eq Epi i jp yc q qq s0,Ž .i i iž /Ep Ep Epi j i

j/ i , is1, 2,

is rewritten as:

Eqip yc 1yu qq s0, is1, 2,Ž . Ž .i i i

Epi

where

Eq rEp Epi j ju syi ž /Eq rEp Epi i i

or

qip yc sy , is1, 2.i i Eqi

1yuŽ .Epi

Therefore, us0 gives the Bertrand–Nash equi-librium. Given that own price-effect is negative,u) 0 gives a price–cost margin larger thanBertrand–Nash, and therefore, competition isAsofterB than Bertrand competition. Similarly, u-0gives a price–cost margin smaller than Bertrand–Nash, and therefore, competition is AfiercerB thanBertrand competition. Therefore, the least structuralinterpretation of the CV parameter is to interpret it asconduct parameter, which benchmarks competitionagainst Bertrand–Nash competition.

Another continuous parameter approach thatavoids the criticisms of the CV approach, but enablesus to measure the level of cooperation in a market, isthe weighted profits approach advocated by Gasmi et

Ž .al. 1992 . In this approach the model allows anunequal but positive weight on the competitor’s prof-its to allow for a more flexible model than the


Table 3Summary of steps in implementing NEIO models

Step Alternatives

v Ž .Identification of industry Industry level analysis with homogeneous product Porter, 1983v Ž .Few major productsrfirms most studies in the literature, e.g., Roy et al., 1994v ŽCompressing several players into a composite firm Suslow, 1986, compresses all small firms

into a fringe firm; Putsis and Dhar, 1999, combine all national brands into a composite national.player, competing against a private label

v Ž .More complete sample of productsrfirms Sudhir, 2001a

v Ž .Choice of data Aggregate market level data most studiesv Ž .Individual level data Horsky and Nelson, 1992; Goldberg, 1995

v Ž .Choice of the objective function Profits most studies in literature, e.g., Kadiyali et al., 1996, 2000v Revenuesv Ž .Target sales Roy et al., 1994v Ž .Category profitsr brand profits Sudhir, 2001b

v Ž .Specification of demand functions Linear most studies in literature, e.g., Kadiyali et al., 1996v Ž .Log-linear Kadiyali et al., 2000v Ž .Log-log functions Bresnahan, 1987v Ž .LA-AIDS Putsis and Dhar, 1999v Ž .Logit functions Besanko et al., 1998; Sudhir, 2001b; Chintagunta et al., 1999v Ž . Ž .Logit functions with random normal distribution coefficients Berry et al., 1995; Sudhir, 2001av Ž . Ž .Logit functions with random finite mixture coefficients Berry et al., 1996; Besanko et al., 2000

v Ž .Specification of cost functions Constant marginal cost many studies in literature, e.g., Kadiyali et al., 1996v Ž .Linear function of factor costs Besanko et al., 1998v Ž .Log-linear function of factor costs Sudhir, 2001a

v Ž .Competitive interaction specification Menu approach Roy et al., 1994; Kadiyali et al., 1996v ŽConjectural variation approach many studies, e.g., Vilcassim et al., 1999; Putsis

. Ž .and Dhar, 1999 ; conduct parameter or weighted profits approach Sudhir, 2001aŽ .These techniques have different ways of inferring the type of competitive equilibrium

vHypotheses on competitive equilibria Bertrandv Leader–followerv Dominant firm–fringe firmv Collusivev Competitive

v Ž .Method of estimation 3SLS many studiesv Ž .FIML Bresnahan, 1987; Gasmi et al., 1992; Roy et al., 1994; Sudhir, 2001bv Ž .GMM Berry et al., 1995; Sudhir, 2001a

v Ž .Model selection methods Cox’s likelihood ratio tests for non-nested models Bresnahan, 1987v Ž .Vuong’s test Roy et al., 1994; Sudhir, 2001bv Ž .Minimum sum of squares many studies

collusive equilibrium discussed in Section 3.1. Thisgeneral notion of collusion would allow firms to set

Žprices anywhere on the firms’ Pareto frontier i.e.,the set of possible prices that firms might reach ifthey were able to negotiate, so that they can allachieve higher profits relative to the single period

. 5 Ž .Nash profits . Sudhir 2001a extends this argumentto allow for firms to place negative weights oncompetitor profits in his analysis of the auto market.

5 We thank a reviewer for suggesting this interpretation.


The intuition is that firms may price aggressively soas to gain market share in the small car segments togain the loyalty of first time buyers, so that they canreap the rewards of greater loyalty from these buyersover their lifetimes. However, the same firms willprice cooperatively in larger car segments wherethere is greater loyalty and therefore, aggression doesnot pay off in additional sales. In general, the nega-tive weights can be used to detect any form ofpredatory pricing.

For a market with two firms, the objective func-tion for the two firms are assumed to be:

P s p yc q qf p yc q ,Ž . Ž .i i i i i j j j

i , js1, 2, i/ j.

Firm i is thus assumed to place a weight f on firmi

j’s profits. f sf s1 indicates perfect collusion;i j

f sf s0 indicates Bertrand competition. Wheni j

0-f -1, is1, 2, we have intermediate levels ofi

cooperation. When f , f -0, we have highly com-i j

petitive behavior, relative to Bertrand competition.The first-order conditions for this model are

EPs p yc b qf p yc b qq s0,Ž . Ž .i i i i i j j i j i

Epi

j/ i , is1, 2.

3.4. NEIO modeling and estimation: A summary ofsteps

The preceding example provided a quick reviewof the basic methodology. We now summarize thevarious steps and the alternative approaches thatresearchers have used in building and estimatingNEIO models in Table 3. These steps begin with anidentification of an industry context, specification ofobjectives of players, specification of demand andcost functions, specification of competitive interac-

Ž .tion among players game employed , choice of rele-vant data, estimation and testing models and hy-potheses. The specific choice would depend on theproblem of interest and the industryrcontext that isbeing studied.

4. Applications of NEIO framework to issues ofinterest to marketers

As the example above shows, the essence of anNEIO model is modeling demand and first-order

optimization conditions of firms, and the economet-ric estimation of the same. In this section, we elabo-rate on some issues in applying the above frameworkto marketing problems. We will discuss issues in thefollowing three dimensions: modeling demand-sideissues of interest to marketers, modeling cost, andmodeling issues in competitive behavior. We nextsummarize the pattern of, and reasons for, the evolu-tion of the literature. We end with a discussion aboutpossible directions for future research in this area.

4.1. Modeling richer demand structures

4.1.1. Other functional forms of demandIn empirical applications, it is important to ensure

that the demand function used describes accuratelythe industry and firm being studied. Recall that thefunctional form of demand has implications for opti-mal profit margins and thus, the supply equations. Ina simple linear demand model, the own and crossprice effects are constant at all levels of prices; in alog–log model, the own and cross price elasticitiesare constant at all levels of prices. While thesefunctional forms are useful in estimating pure de-mand models, greater flexibility in the demand mod-els are needed when also estimating the supply-sidemodels. Marketing researchers have been using moreflexible demand models in their research of late:

Ž .Putsis and Dhar 1999 use an LA-AIDS demandŽ . Ž .model, Besanko et al. 1998 , Sudhir 2001b and

Ž .Chintagunta et al. 1999 use a logit model andŽ .Sudhir 2001a uses a random parameters logit model.

Researchers also test the appropriateness of func-tional form used for their markets; for example, see

Ž .Genesove and Mullin 1998 .In recent theoretical work on marketing channels,

Ž . Ž .Lee and Staelin 1997 and Tyagi 1999 have shownthat the functional form crucially determines

Ž .a retailer’s strategic behavior retail passthrough .Demand models characterized by Vertical Strategic

ŽSubstitutability VSS hereafter; e.g., linear, homoge-.neous logit models always have retail passthrough

of less than 100%, while models characterized byŽvertical strategic complementarity VSC hereafter;

.e.g., multiplicative model have passthrough greaterŽ .than 100%. Sudhir 2001b explicitly tests whether


the VSS or VSC implications best fit the data byŽ .using both a homogeneous logit VSS or the multi-

Ž .plicative model VSC . He finds the VSS implica-tions of the logit are more consistent with the data.Flexible models such as heterogeneous logit andLA-AIDS allow passthrough to be greater than orbelow 100% and should be potentially useful infuture research modeling channel behavior.

The use of discrete choice models has vastlyincreased the usability of the NEIO approach toproblems of relevance to marketers. Most studies inthe NEIO approach have focussed on multiple firmsselling fairly homogeneous products or a limitednumber of differentiated products. In some cases, asmall number of products truly reflect the competi-

Žtion in the market for example, rivalry betweenCoke and Pepsi modeled by Gasmi et al., 1992;rivalry between Kodak and Fuji in the film market

.by Kadiyali, 1996 . In other cases, the number ofcompeting products is restricted to a few majorplayers to enable ease of estimation. However, mostmarkets of interest to marketers have multiple play-ers, and multiple products, and also horizontal andvertical differentiation. The simple linear model isincapable of handling such markets for the followingreason.

In the example of Section 3 above, there are twofirms, each manufacturing one product, leading tofour price effects to be estimated in the demandfunction. When the number of products increases as

Žin the auto market or breakfast cereal market or the.number of firms , the number of parameters to be

estimated explodes. For example, in a market with30 products there will be 900 price effects in a lineardemand model.6 A solution to this problem is toimpose greater structure on the demand specification,so that fewer parameters need to be estimated. A

Ž .Lancasterian approach Lancaster, 1971 , where util-ity is based on the product characteristics, can serveto impose a suitable structure on demand. The de-mand equations are derived by aggregating individ-ual choices in a utility maximizing framework. Boththe deterministic and random utility approaches havebeen used.

6 This problem occurs even if we use a log-linear or log-logmodel.

An example of a deterministic utility approach isŽ .that of Bresnahan 1987 . He assumes that all prod-

uct characteristics can be summarized to a singlenumber called product quality by an appropriatefunction of product characteristics. Consumer valua-tions of quality are assumed to be uniformly dis-

Ž .tributed. Feenstra and Levinsohn 1995 allow prod-uct characteristics and consumer preferences for thesecharacteristics to be multidimensional. With theseassumptions, the demand equations for the productsare constructed. Products with similar characteristicswill compete against each other than products withdivergent characteristics under these specifications.

In many situations, several unobserved character-istics affect consumer choice. Accordingly, the ran-dom utility framework has been adopted in recentwork. In this framework, there are two componentsof utility: a certain deterministic component ob-served by the researcher, and a random componentunobserved by the researcher. Assumptions about therandom component of the utility are made explicit,taking into account the market characteristics. Whenthe random component of utility is assumed to beindependently and identically extreme value dis-tributed across products, we get the logit model. Byassuming appropriate correlations across products forthe random component, we can generate the nestedlogit model.

Ž .Berry 1994 develops a general contraction map-ping approach to compute the mean level of utilityfor a product across consumers for a general class ofmodels that include the logit, the nested logit andrandom parameters logit models, using observedmarket shares. This mean utility can then be used asa dependent variable for the demand equation inestimating the parameters for the utility associatedwith the characteristics of the product. VerbovenŽ .1996 estimates a nested logit model in analyzinginter-country price discrimination by firms in the

ŽEuropean car market. Berry et al. 1995, henceforth.BLP implement this method in a random parameters

logit model of the US auto market, where the coeffi-cients of utility for characteristics are assumed to benormally distributed. In their paper, they assume thatthe firms are involved in Bertrand competition. Sud-

Ž .hir 2001a generalizes the BLP analysis by allowingfor deviations from Bertrand competition in differentsegments of the US auto market. He finds that firms


are more competitive in the smaller car segments andmore cooperative in the larger car segments.

4.1.2. Aggregate Õersus indiÕidual-leÕel dataIn addition to the product differentiation issue

discussed above, another important issue in demandspecification is that of the level at which demand isspecified. Marketing researchers have debated theuse of individual choice data versus market-levelchoice data. Most studies in the NEIO tradition inmarketing have used aggregate firm-level data onsales and prices to specify both the demand and themarketing mix optimization rules. Marketing re-search in the consumer choice area has used bothaggregate and individual level, and in particular, theadvances made in estimating various dimensions ofchoices using individual-level data are very impres-sive. Using mainly aggregate data in the NEIO tradi-tion means that the degree of sophistication in de-mand modeling is much less than the one in theconsumer choice literature using individual-leveldata.

There are some notable exceptions, however, inthe NEIO literature, which use individual-level data.

Ž . Ž .Horsky and Nelson 1992 and Goldberg 1995estimate models of the automobile market usingsurvey data from individuals. This enables them tomodel demand heterogeneity at a great level ofdetail. Horsky and Nelson use their model to dooptimal product positioning given the estimated de-mand and cost characteristics under the assumptionof Bertrand competition among the auto manufactur-ers. Goldberg attempts to infer whether the voluntaryexport restraint imposed by Japanese auto manufac-turers had any impact on the auto market in the USduring the 1980, assuming that firms were in Bertrandcompetition with each other.

When dealing with individual data, researchersuse a two-stage estimation procedure. In the firststage, they estimate the demand parameters usingstandard choice modeling techniques. In the secondstage, they aggregate market shares from individualchoice data and then estimate supply-side parametersŽ .costs and competitive interactions , subject to thedemand estimates from the first stage. The gap thatremains in using individual-level data in NEIO stud-ies is that, techniques to estimate general models ofcompetition have not yet been developed. Both

Horsky and Nelson and Goldberg assumed theBertrand–Nash equilibrium, rather than infer thecompetitive interactions from the data.

4.2. Modeling richer cost structure

There has been much less sophistication in cost-side specification relative to the demand-side devel-opments discussed above. Typically, constant

Žmarginal cost specification has been used see Jarmin,1994, who allows costs to fall over time due to

.learning, for an exception . While many studies haveestimated costs as a single parameter, Besanko et al.Ž .1998 estimate a linear function of production cost

Ž .factors while Sudhir 2001a estimates a log-linearfunction of production cost factors.

Functional forms of demand and cost need to becarefully considered in studies that use conduct pa-

Ž .rameter estimation procedure. Bresnahan 1989 dis-cusses the issue of identification of the conductparameter under a variety of functional forms fordemand and cost. A major conclusion is that aminimum amount of nonlinearity is needed to iden-tify conduct parameters. For example, Parker and

Ž .Roller 1997 use a semi-log form of demand toenable identification of this parameter. Researchersare advised to verify that the conduct parameter isindeed identified in their models by using appropri-ate functional forms of demand and cost.

4.3. Richer specification of competition

4.3.1. AlternatiÕe objectiÕesThe model in Section 3 assumes that firms maxi-

mize profits. This is the NEIO counterpart of con-sumer choice model assumption that consumers max-imize utility. Researchers have speculated that firmsoften maintain market share and may have other

Ž .goals as well. Roy et al. 1994 assume that theobjective of firms is to be close to a preset target,where the targets are treated as exogenous to the

Ž .analysis. Sudhir 2001b allows for alternative objec-Žtives category profit maximization, brand profit

.maximization, charging a constant markup and findsthat the best fitting model is one where the retailermaximizes category profits. Other possible objec-tives include maximizing profits subject to marketshare, or subject to a certain level of consumer


Ž .surplus to avoid antitrust activities , or minimizingtime to market. One difficulty to keep in mind is thatclosed-form, unique and econometrically identifiableoptimization rules cannot be found for all games forthese alternative objectives.

A related issue is that most NEIO studies haveŽassumed that firms choose prices, or advertising or

.capacity to maximize profits. The theoretical gametheory literature has shown us that changing thecompetitive game from price to quantity competitionsignificantly alters results. Therefore, an importantquestion is whether the assumption of price competi-tion rather than quantity competition is correct. Raju

Ž .and Roy 1999 find that for the US microprocesserindustry, price competition fits the data better thanquantity competition. They argue that a reasonablecase can be made that quantity competition is likelyin this industry, and therefore, finding that pricecompetition better describes the data is reassuring,given that many marketing studies use price competi-tion. It must however be remembered that the notionof quantity competition is widely used when model-ing homogeneous or monopoly markets, because in-dustry quantities essentially determine industryprices. In contrast, in differentiated product markets,

Žwith excess capacity most marketing studies fall in.this category , price competition is theoretically more

appealing and is therefore widely used in theoreticalstudies.

4.3.2. Continuous Õersus discrete choices of market-ing mix

In the example discussed in Section 3, the pricedecision of firms is a continuous choice variable.However, in many classes of games, strategic deci-sions may be only qualitative, e.g., entry and exitgames are 0–1 variables. Modeling of the retailer’sdecision to sell a product offered by a manufactureris similar. In these kinds of situations, we need tomake inferences about player’s incentives from qual-itative data.

Ž .Bresnahan and Reiss 1991 extend the approachŽof single person discrete choice models McFadden,

.1982; Hausman and Wise, 1978 that has been widelyused in marketing to analyze consumer choice to themulti-person discrete choice framework. In singeperson choice models, parameters of consumer pref-erences are estimated using threshold models of con-

sumer behavior. In multi-person choice models pa-Ž .rameters of agents’ firms’ payoffs are estimated

using threshold models of games.As an illustration of the methodology, consider

the example of an entry game used by Bresnahanand Reiss involving two firms. Each firm has twoactions: to enter or not to enter. Let us denote theaction of player i as a where a s0 indicates thati i

firm i does not enter and a s1 indicates that firm ii

enters. The following tables represent the two firms’payoff matrices. P i indicates the payoff to firm ijk

where firm 1 takes actions j and firm 2 takes actionk. D i indicates the increase in profits to firm i when0

firm i enters a market with zero competitors. D i1

indicates the decrease in profits to firm i when firmj enters a market in which it had a monopoly. Byeconomic theory, we expect D i (0.1

Player 1’s payoffs

a s0 a s12 2

1 1a s0 P P1 00 011 1 1 1 1a s1 P qD P qD qD1 00 0 01 0 1

Player 2’s payoffs

a s1 a s12 2

2 2a s0 P P1 00 012 2 2 2 2a s1 P qD P qD qD1 00 0 01 0 1

In a Nash equilibrium,

a1s0mD1qa2 D1(0,0 1

a2s0mD2qa1 D2(0.0 1

For an econometric model, we need a stochasticspecification for unobserved profits. For their pur-pose, Bresnahan and Reiss treat the D i as constants1

and the D i as random variables that vary across0

firms and markets. Assuming that D i , the incremen-0

tal profit firm i receives as a monopolist in anyŽ .given game, is a linear function of observables X

Ž . i 0 iand unobservables ´ , i.e., D sXb y´ , then the0 i

structural equations determining the Nash equilibriaof this game are

y)sX b 1qa2 D1y´ 1 ,1 1 0 1

y)sX b 2qa1 D2y´ 2 ,2 2 0 1


where

0 if y)-0iia s for is1, 2.)½ 1 if y -0i

These two equations form a discrete dummy variableendogenous system for the unobserved y). Thisi

system is a special case of the systems considered byŽ .Heckman 1978 . From here, the joint probability

distribution of D i , is1, 2, can be estimated and the1

thresholds of entry recovered. Their model is used toexplain the number of entrants in the market, and thethreshold of market size that is needed to sustainthese entrants. They also illustrate how to develop

Ž .models for sequential leader–follower models andcollusive games.

Summarizing, the NEIO framework can be ex-tended to estimate discrete choice marketing strategyvariables, but the general conduct parameter estima-tion approach is yet to be implemented.

4.3.3. The dynamics of competitionThe model discussed in Section 3 is a one-period

or static model of competition. The reaction functionliterature allows for dynamic competitive effects, butis not structural in nature. Modeling dynamics in afully structural context turns out to be a challengingtask. We discuss three approaches used in the litera-ture. Fully structural dynamic models of competitionhave not yet been estimated in the NEIO literature;in Section 5.3, we discuss the shortfalls of currentdevelopments in the area.

An attempt at modeling dynamics is done byŽ .Roberts and Samuelson 1988 . They develop a

Žmodel of advertising as a goodwill stock see also.Vilcassim et al., 1999 . All competitive effects of

advertising greater than two periods are summarizedby a constant in this model as these effects cannot beseparately identified econometrically. In other words,only two-period dynamic effects are estimable in thismodel.

Another attempt at modeling dynamics is made byŽ .Slade 1995 . Her model allows inter-temporal dy-

namics in price and advertising to last several peri-ods and chooses the appropriate lags based on the

Ž .data. That is, unlike Roberts and Samuelson 1988where effects greater than two periods cannot beestimated, she allows the data to determine how

Žmany lags should be included see also Chintagunta.et al., 1999, for a similar approach . Slade finds that

for the saltine cracker market, the appropriate num-ber of lags to consider was two. She uses a Kalmanfilter approach to estimate the model. Interestingly,she finds that what may be usually considered as acollusive equilibrium in a static context is an out-come of a non-cooperative game when inter-tem-poral dynamics are considered. This shows that mod-eling dynamics can give us greater insights into thebehavioral rationale underlying the observed equilib-rium. One should add that the descriptive value ofthe static analysis, however, continues to be relevant.

Ž .A third approach by Slade 1992 is to focus onoff-equilibrium behavior. She estimates a dynamicmodel of demand and supply of the Vancouvergasoline market during a price war. While this paperis close in spirit to the reaction function estimationdiscussed earlier, it differs from that stream of re-search in that, reactions are explicitly linked to de-mand and cost characteristics in the market. Thus,this paper provides a more structural approach toreaction functions. She classifies the firms in themarket into two types: majors representing major oilcompanies and independents. She attempts to under-stand the out of equilibrium strategies used by firmsduring price wars, by estimating the reaction func-tions using a Kalman filter. She allows for asymmet-ric reactions to price increasesrdecreases and also

Žallows for nonlinearity of reactions by allowing.reactions to depend on price levels , a feature similar

to reaction function studies. She finds that reactionparameters are not constant over time. However, thestochastic process that generates the variation is sta-tionary, indicating that there are long run stablevalues for these parameters.

4.3.4. The issues of endogeneity, simultaneity andperiodicity

As discussed in Section 1, the question of endo-geneity and simultaneity need to be addressed inapplying the model of Section 3 to marketing con-texts, as well as in extending the model to othermarketing strategy instruments. If the model of Sec-tion 3 were a model of competition between tworetailers, are the prices truly endogenous to retailersor are they determined by manufacturers? Since de-


mand for the product is also a function of nationalmanufacturer advertising, should we treat pricingdecisions as endogenous to the retailer but treat theadvertising as exogenous? NEIO studies have typi-cally examined manufacturer pricing and advertisingdecisions, and these decisions are both endogenousto manufacturers. The issue of simultaneity of deci-sion-making is also relevant here. For example, ifcapacity decisions or product line decisions or entrydecisions are made prior to the decision on pricingand advertising, then these decisions can be treatedas exogenous.

NEIO studies in marketing have dealt with theissues at various levels. In many papers in marketing

Žthat use scanner panel data for example, Kadiyali et.al., 1996 , features in newspapers are treated as

exogenous decisions to manufacturers. In Besanko etŽ . Ž . Ž .al. 1998 , Kadiyali et al. 2000 and Sudhir 2001b ,

both manufacturer and retail price are modeled asendogenous, with optimization rules for both.

Ž .Kadiyali et al. 2000 treat features and displays asendogenous, but no optimization rules for these are

Ž .specified. Sudhir 2001b , however, could not rejectexogeneity of features and therefore treats it as anexogenous variable. Researchers are therefore ad-vised to perform econometric tests to determinewhether or not to treat variables as endogenous, andthen determine if the endogeneity needs to be mod-eled as a separate process or whether it is enough touse instruments to account for endogeneity.

An econometric issue here is the choice of appro-priate instruments for variables that are assumed tobe endogenous. Typically, demand and cost shiftershave been used. In certain situations, lagged valuesof endogenous variables may be appropriate. BLPŽ . Ž .1995 and Sudhir 2001a use averages of ownfirm’s other products and other firm’s products asinstruments by making a supply-side argument inanalyzing the auto market. In an analysis of multiple

Ž .markets for breakfast cereals, Nevo 2001a usesprices in other markets as instruments for prices.

The issue of periodicity of decision-making istrickier. For example, in modeling manufacturercompetition using scanner data, do we believe thatmanufacturers set price weekly? An econometric testof endogeneity may be needed to determine whetherthese variables are indeed exogenous to weekly scan-ner data. A variation of the Hausman test to test

Ž .endogeneity was introduced by Besanko et al. 1998Ž .and has been applied by Kadiyali et al. 2000 and

Ž .Sudhir 2001b . A potentially difficult situationwould be if different decisions have different period-icity, e.g., pricing decisions made weekly, but adver-tising decisions made monthly. That is, pricing deci-sions are endogenous in weekly data, but advertisingis endogenous only once in four weeks. Therefore,periodicity of decision making and time aggrega-tionrdisaggregation are important issues to bear inmind in implementing NEIO models.

4.3.5. Incomplete informationThe model in Section 3 above assumes that firms

have full information about demand functions aswell as each other’s cost functions. In many competi-tive situations, however, there may be incompleteinformation about demand and cost conditions. Mod-eling this in a fully structural framework is limitedby the lack of detailed data on firm decisions. Hence,reduced-form hypothesis testing approaches are morewidely used for such testing. Game theory papers onagency or principal-agent problems describe the ef-fects of lack of information on equilibrium out-comes, and these effects can be tested with data. For

Ž .example, Shepard 1993 tests the role of monitoringissues on which gas stations are franchised out andwhich ones are company owned. She finds thatfull-serve stations are more likely to be franchisedout because of the difficulty in monitoring a man-ager’s effort in providing this service.

However, as discussed earlier, hypothesis testingdoes not provide structural estimates of demand andcost which are policy invariant. The task of buildingand estimating a structural model is significantlyharder under incomplete information. MiravateŽ .1997 estimates consumer preferences for local tele-phone calling deals, given private information be-tween consumers and local telecommunication com-panies. In a market, consumers are faced with anonlinear price schedule that consists of a fixed feeand a marginal rate that are both functions of the

Žminutes called. He assumes that consumer types in.terms of their calling preferences in the population

may be known by their demographic characteristicswith any asymmetry of information being capturedby a beta distribution. Based on this assumption, hederives the optimal demand for the various types and


the optimal fixed fee and marginal rates for eachtype. He then computes the expected total paymentsfor consumers and finds that total expected tariffs area linear function of demographic characteristics whenall information about consumer types is captured bythe demographic characteristics. However, if theasymmetry of information is relevant, i.e., the betadistribution is not degenerate, then the expected tar-iffs include interaction terms of demographic charac-teristics. By testing the significance of the interactionterms he concludes that there exists significant asym-metry of information. He also finds that failure toaccount for the asymmetry of information in hismodel leads to systematic overestimation of demand.

Therefore, accounting for the effects of incom-plete information, not just between consumers andfirms, but also between firms, is critical to estimatingcorrect demand, cost, and competitive conduct pa-rameters.

4.3.6. Modeling intra-channel strategic behaÕiorManufacturers in many product categories sell

their products through a retailer to the consumer. Amodel of market response for a manufacturer there-fore needs to take into account not only how con-sumers and competing manufacturers react to themarketing mix, but also how the members of thechannel react to the marketing mix. Measuring chan-nel response, simultaneously with consumer andcompetitive response, is therefore managerially im-portant.

With scanner data, in general, the wholesale pricesare unobserved. Researchers have in the absence ofwholesale price information made the assumptionthat retailers charge a constant margin. With thisassumption, wholesale prices can be known up to aconstant margin and they then infer competitive be-havior among manufacturers. Prior studies have usu-ally found cooperative behavior among manufactur-ers with this assumption. Given that retail prices areused for the analysis, it raises the question whetherwe may be misinterpreting retailer category manage-ment as manufacturer cooperation. To test this issuewhen wholesale price data is unavailable, SudhirŽ .2001b uses a game theoretic modeling approach toinfer wholesale prices. He tests for alternative mod-els of manufacturer–retailer interaction such as man-ufacturer Stackelberg Leader–Follower behavior and

Ž .Vertical Nash behavior. Cotterill and Putsis 2001also tests for alternative models of manufacturer–re-tailer interaction in the absence of wholesale price

Ž .data. Alternatively, Kadiyali et al. 2000 use actualwholesale price data to tackle the issue of incorporat-ing channel response and infer the balance of power

Žbetween manufacturers and retailers see also Parker.and Kim, 1999 .

5. The evolution of literature: Retrospective andprospective

We now discuss how the literature has evolvedand how we expect it to evolve in the future. As inthe previous section, we organize this discussion intothose relating to demand, cost and competition is-sues. Clearly, theoretical developments in game the-ory, and marketers’ empirical observations from SCPstudies have given us enough puzzles to empiricallyexplore further. We argue that the evolution of thisliterature has been mainly data- and methodology-driven, and these two constraints will also shape thefuture direction of this literature.

5.1. Demand: looking back, looking ahead

The most active area of development has been insophisticated demand modeling using widely avail-able aggregate data. Advances in the economics

Ž .literature e.g., BLP have fueled these develop-ments. Its quick diffusion in marketing is not surpris-ing given the rich tradition of demand modeling inmarketing. In fact, this is an area where more bridgesbetween marketing and economics literature can bebuilt. There are several directions this literature can

Ž .profitably pursue. BLP 1995 illustrated how we canmodel heterogeneity among consumers using the dis-

Ž .tribution of consumer demographics income thatare usually available with consumer data. This tech-nique of using aggregate distribution information toeffectively account for consumer heterogeneity is

Ž .further enhanced in the work of Nevo 2001b andŽ .Petrin 1999 . These methods to infer demand pa-

rameters with purely aggregate data should be ofgreat interest to marketers, because in practice man-agers usually have much greater access to aggregatestore-level or market-level data rather than individuallevel data.


Secondly, as mentioned previously, marketershave advanced techniques for estimation of demandusing individual data as well as analyzing the biasesin using aggregate versus individual-level demandŽ .Gupta et al., 1996 . This would be a promising areafor future work, because with individual level data,dynamics in consumer preferences can be better

Ž .captured Erdem, 1996 . Studying the role of con-sumer dynamics on firm pricing dynamics is a diffi-cult but important issue. Bayesian estimation toolsfor modeling demand offers great promise in estimat-ing finer levels of heterogeneity across firm demand,cost and competition parameters in the NEIO frame-work. In the short- to medium-term, we thereforeexpect the literature to focus on further sophistica-tion in modeling heterogeneity in the aggregate de-mand-discrete choice arena and the use of Bayesianmethods to capture the heterogeneity on both thedemand and the firm side.

5.2. Cost and supply: looking back, looking ahead

On the cost dimension, specifications used inmarketing have been relatively simple. An importantreason for this is that given weekly scanner data usedin several marketing studies, it is not always possibleto find weekly cost-side instruments. Second, NEIOstudies in economics have been more varied in theirapplication across industries, whereas marketingstudies have been mainly confined to packaged goodindustries. Given issues of retailer versus manufac-turer margins when using local market data, the lackof data on various manufacturer–retailer transfer

Ž .payments for e.g., slotting allowances and the diffi-culty of estimating hedonics or characteristic-basedcosts for these industries, a richer cost specificationis not easy. That said, the hope is that marketingstudies in the future will explore non-packaged goodsindustries and get past some of the cost specificationissues associated with these.

Another area of importance is the modeling ofinstitutional details of supply in a particular market.An example of detailed supply-side modeling ineconomics is the previously cited Suslow’s model,where first-time and recycled aluminum are treateddifferently in modeling and estimation. Berry et al.Ž .1996 model in great detail the hub and spokestructure of the airline industry. Researchers in mar-

keting have paid great attention to institutional de-Žtails in modeling channel behavior see Section

.4.3.5 . Marketers are now making strides in this areaas theoretical and data constraints are relaxed. As wediscussed in Section 4.3.5, a unique innovation ofmarketing to the NEIO literature has been the studyof manufacturer–retailer relations. These studies havebeen mainly confined to the study of pricing. Itwould be instructive to expand the scope of thisresearch to other channel decisions like advertisingor promotions. Additionally, models of multiple re-tailers interacting with multiple manufacturers in acomplete intra- and inter-channel model of competi-tion are yet to be estimated. Relevant data for thistype of analysis may be hard to get. While therelationship between manufacturers and retailers hasbeen explored in some detail because of data avail-ability, the relationships between a firm and other

Žintermediaries including suppliers rather than retail-.ers remains an area yet to explored in the marketing

NEIO literature, mainly because of data constraints.ŽNon-packaged goods industries e.g., high-tech prod-

ucts with complex supply chain partnerships, or.pharmaceuticals with R&D processes are especially

well-suited to expand the literature in this area.

5.3. Competition: looking back, looking ahead

On the competition dimension, several marketingstudies have explored a variety of issues in pricingcompetition. Pricing is clearly a very important mar-keting mix element, and therefore, it is not surprisingthat it has been the focus of the majority of attentionas the literature has evolved. Examples include Roy

Ž .et al.’s 1994 study of price competition betweenŽ .Ford and Chrysler, Kadiyali et al.’s 1996 study of

product line pricing decisions of two firms, Sudhir’sŽ . Ž2001a study of price competition across car mar-

. Ž .ket segments, Vilcassim et al.’s 1999 study ofhierarchy in with three firms in the market, Putsis

Ž .and Dhar’s 1998 study of national brand-privatelabel price competition. There have also been at leasttwo studies of manufacturer–retailer pricing compe-

Ž .tition Sudhir, 2001b; Kadiyali et al., 2000 .All these studies have found that simple

Bertrand–Nash pricing is not a good assumption tomake, and firms in markets have varying degrees ofmarket power because of the efficacy of brand posi-


Žtioning measured by own- and cross-price elastici-.ties , and their low costs. All these applications have

typically been driven by researchers’ interest in pric-ing policies under different marketing strategy situa-tions. We expect the future direction of this literatureto be comprised of other applications of the NEIOframework to other marketing strategy situations. Inparticular, we expect a more complete analysis ofmanufacturer–retailer pricing interactions and the ef-fect of new channels of distribution on these rela-tionships in the immediate future. Data availability islikely to be the crucial limiting factor.

An issue with discrete choice demand model NEIOŽ .papers using aggregate data e.g., BLP is that they

impose the assumption of Bertrand pricing amongŽfirms. As we mentioned in Section 4.1, this is also

true for models that use discrete choice models using.individual-level data . For these functions, there is

no solution yet to estimate separate competitive in-Ž .teractions among several firms. Sudhir 2001a has

overcome this issue by estimating segment-levelcompetitive interaction parameters among firms, ef-fectively reducing the number of parameters thatneed to be estimated. However, future work willneed to decompose the competitive interactions be-tween products in terms of simpler basic primitivesto estimate a full-blown model of competitive inter-

Žactions among all products whether the applicationis to pricing competition, or other types of competi-

.tion .Applications of the NEIO framework to advertis-

Žing competition have been more limited e.g., Vil-.cassim et al., 1999 and therefore, several issues

remain to be explored. This is possibly because it isharder to obtain advertising GRP data for markets,especially at the weekly or other high frequencylevel at which pricing data are available. Also, adver-tising competition is likely to be more dynamic thanpricing competition given its long-lived demand ef-fects. This poses a significant challenge to re-searchers, and further developments in the dynamicsof competition will enable researchers to study theseissues in greater detail.

An example of a future application in the advertis-ing competition arena is promotion scheduling. The-oretical research has identified demand and advertis-ing cost conditions for when it makes sense to pulseadvertising or to have to steady stream of advertis-

Ž .ing. Roberts and Samuelson’s 1988 study of adver-tising accounting for dynamic advertising effects andadvertising costs provides a building block for anempirical analysis of these issues. It would be in-structive to generalize their demand, cost, as well asgoodwill specifications, to test optimal market andcompetitive conduct scenarios for alternative adver-tising schedules.

Competitive analysis of promotions like features,displays, etc., has not yet received attention in themarketing literature. This could be because it is noteasy to formulate the variable cost structures of thesemarketing mix variables. Additionally, the tech-niques of estimating discrete choice models de-scribed above in Section 4.3 could be used to studyfeature and display competition, but to combine themwith continuous choice of pricing is a challengingtask. Future applications of these techniques espe-cially to retailer competition can be very instructive.This could also answer the long-standing debateabout whether there are mixed-strategy equilibria in

Žpromotions although we await the finessing of dy-.namic equilibria testing to better test this .

Another marketing mix variable that has been lessstudied is choice of location. The limited attention tothis area is possibly again because of data con-straints, as well as the discrete nature of this choicevariable. Data of firms in an area, with locationcoordinates, as well as other dimensions of competi-tion like price, quantity, etc., are needed to estimatesuch models. An example of such work is DavisŽ .1997 , who analyzes spatial competition amongmovie theaters. However, the insights offered bymodeling competition in an NEIO framework are yetto be explored fully.

A marketing mixrproduct design variable thathas received some attention in the marketing litera-ture is network externalities. Shankar and BayusŽ .1999 analyze the impact of network externalitieson competition in the home video game player indus-try. This literature points the way for studyingswitching costs in general in the context of eitherproduct life-cycle or entry. Given the widespreadimportance of network externalities in the emergingarea of e-commerce, insights on this topic should beof great value to managers.

A general issue with measuring marketing mixcompetition is to test if competition varies over time.


There are two ways to do this. The first is to developfully structural dynamic models of competition, andthe second is to estimate a static model of competi-tion with a time-varying conduct parameter. Thetechniques for fully structural dynamic models ofcompetition have not yet been developed in the

Ž .NEIO literature. Pakes and McGuire 1994 havedeveloped techniques using Markov-perfect Nashequilibrium concepts to model the dynamics of com-

Žpetition where last period’s firm choices are suffi-.cient statistics for predicting future behavior . The

estimation is based on a computationally intensivedynamic programming technique. However, there aretwo roadblocks in widespread application of thesemethods to NEIO framework. The first is that theMarkov-perfect Nash equilibrium assumption needsto be relaxed because it is restrictive, and second, theestimation needs to be extended to forms on compe-

Žtition other than the currently possibly set i.e.,.Bertrand, fully collusive, etc. .

A second solution to estimating whether competi-tion varies over time is to measure time-varyingconduct parameters. This has so far not been at-tempted in marketing. An example in economics is

Ž .Parker and Roller 1997 , who, in a study of compe-tition in the cellular phone market, estimate theconduct parameter as a function of multimarket con-tact and cross-ownership. As the multimarket contactand cross-ownership changes over time, they analyzethe impact of competitive behavior. For future appli-cations in marketing, one simple alternative is tospecify the conduct parameter as a hazard functionor as a logistic function and to examine whetherthere are any cycles of marketing mix competition.

Ž .The current NEIO solution is Porter’s 1983 en-Ž .dogenous switching regression model of pricing

conduct parameters that alternate between periods ofcollusion and breakdown of collusion. A generalspecification of the time-varying conduct parameteris needed to provide more insights into the dynamicsof competition. This is a simpler solution to captur-ing the dynamics of competition than awaiting morerealistic models of Markov-perfect competition.

A direct and especially interesting future applica-tion of time-varying competition is the study ofmarketing mix efficacy in different stages of theproduct life cycle. In the NEIO literature, Greenstein

Ž .and Wade 1998 have addressed this issue. They

analyze how competitive conduct changes over theproduct life cycle in the mainframe computer market.The constraint to expanding this study to marketing

Žmight be the data requirements we would needprice, sales and other marketing strategy instrumentsby brand from the start of the market to its matura-

.tion . Another future application might be the studyof early and late mover advantages, which can alsobe addressed by estimating time-varying demand,cost and competition parameters.

Another application of time-varying competitionis in the area of new product introduction. Twopapers in marketing have addressed some aspects of

Ž .this. Kadiyali et al. 1996 analyze how a newproduct introduction changes competition in the mar-

Ž .ketplace. Shankar 1997 analyzes what the bestentry strategies are when accounting for competition.However, many aspects remain to be explored inmuch greater detail. For example, how the marketmoves from the competitive disequilibrium causedby market entry to a new competitive equilibrium isof interest to marketers, and can be examined byusing a time-varying conduct parameter. Addition-ally, models of how firms learn of each other’s coststructures or of consumer demand for a new productand the resulting competitive structure remain to beestimated.

In contrast to pricing and advertising, many mar-keting decisions are discrete and dynamic in nature.The dynamics of these are also slowly being devel-

Ž .oped. Ericson and Pakes 1995 develop estimationtechniques for dynamic models for R&D investmentand entry-exit decision of firms using the Markov-perfect equilibrium concept discussed above. WhileR&D investment decisions are continuous, entry-exitdecisions are discrete. They use their model to ana-lyze the entry and exit of differentiated products in amarket. This approach is used to test alternativetheories of firm dynamics and the evolution of mar-

Ž .kets in Pakes and Ericson 1998 . However, the sameissues discussed earlier, i.e., modeling dynamics moregenerally than Markov-perfect behavior and moregeneral models of competition in the conduct param-eter framework, need to be developed before thesemethods can be applied to gain insights.

In the age of the Internet, firms can very easilytrack pricing behavior of their competitors and reactvirtually instantaneously by appropriately changing


their prices. Markov-perfect equilibrium models aretherefore appropriate to model price competition onthe Internet. We therefore believe these types ofmodels can be useful to model the ease with whichcompetitors can react in many Internet markets.

A final frontier in analyzing more realistic modelsin competition is accounting for incomplete informa-tion in structural models, as we discussed in Section4.3.4. Apart from the methodological difficulties, amajor bottleneck in estimating structural models ofincomplete information is that it is very difficult toobtain detailed data on decisions at a very disaggre-gate level. Fortunately, in the area of auctions, thereis detailed data on the bidding process across multi-ple auctions. Here is where the most progress onestimating structural models of asymmetric informa-

Ž .tion has been made Porter, 1995 . With the easyavailability of auction bid data on internet sites suchas eBay , this could potentially be a very promisingarea for research; for e.g., see Bajari and HortacsuŽ .2000 . Given the importance of consumer-to-con-sumer auctions in the internet age, this should be ofsubstantive interest to researchers in marketing. An-other potentially attractive area would be the analysisof sales force compensation. Sales force compensa-tion models are principal-agent models in an asym-metric information framework. Since detailed dataon sales force performance under different compen-sation schemes should be available in many firms,one can estimate a structural model of asymmetricinformation to infer distributions of sales force riskpreferences and utilities. In the absence of suchdetailed data, researcher could collect experimentaldata to refine estimation techniques for this class ofproblems. In short, we believe there are a number ofopportunities waiting to be explored in this area.

6. Conclusion

Structural models in the NEIO framework provideestimates of, and insights about, the underlying com-petitive game, demand and cost structures for aspecific industry. We have argued in this paper thatthey are therefore extremely useful for managerialdecision-making, and specifically, for marketing mixdecisions. As we stated in the introduction, there areseveral benefits to estimating NEIO structural mod-els.

A major benefit of structural NEIO estimation isthat it enables theory testing. Game theory offerscompeting theories under different assumptions lead-ing to different predictions. For example, Green and

Ž .Porter 1984 suggest a theory of dynamic competi-tive behavior that suggests procyclical behavior, i.e.,prices rise during periods of high demand of amarket and fall during periods of low demand for a

Ž .market. Rotemberg and Saloner 1986 in contrastdevelop a theory that predicts countercyclical behav-ior. Ratomberg and Saloner use estimates of compet-

Ž .itive behavior from Porter 1983 to show that therailroad cartels of the 1880s behaved in a mannerconsistent with procyclical pricing implications.

Ž .Bresnahan 1987 shows that the auto market be-haves consistently with the procyclical prediction ofRotemberg and Saloner. NEIO research can thusserve to judge the appropriateness of theories.

An issue with such theory testing is that while thetests are based on structurally rich and industry-specific studies, it is not clear whether and how thesefindings can be generalized. We have argued previ-ously that managers care both about strategic gener-alizations across industries and strategic specifics oftheir industry. There are several studies that combinethe methodology of NEIO and yet study severalclosely related markets. For example, Parker and

Ž . Ž .Roller 1997 and Nevo 2001a analyze firm behav-ior in the cellular telephone and breakfast cerealmarkets, respectively, across multiple geographical

Ž .markets within the United States. Verboven 1996analyzes price discrimination behavior in different

Ž .countries of the European car market. Sudhir 2001aanalyzes multiple segments of the US auto market toinfer differences in competitive behavior across dif-ferent segments. The studies across closely relatedmarkets analyze how similar firms adapt their behav-ior in different markets to account for the variationsin structural characteristics of these markets.

Ž .Putsis and Dhar 1999 use a hybrid of NEIO andSCP approaches. They infer competitive behavior forabout 42 different frequently purchased product cate-gories across 59 local markets using NEIO tech-niques that account for the endogeneity and simul-

Žtaneity of firm choices. In a second step in the SCP.tradition , they meta-analyze their estimates of com-

petitive behavior against structural characteristics ofthe different product markets to draw conclusions


about the relationship between competition andstructural characteristics of a market. The criticismabout the appropriateness of pooling competitionestimates across disparate markets is somewhat tem-pered by the fact that all of these product categoriesare frequently purchased consumer goods.

To obtain generalizable results, researchers need anumber of such industry specific studies in similarcompetitive contexts. At this stage, the volume ofempirical research in this area is still too limited toperform a meaningful meta-analysis; hence, thelament of researchers in industrial organization andcompetitive strategy researchers in the areas of busi-ness strategy and marketing about the paucity ofempirical studies in this area. In a survey of applica-tions of game theory to competitive marketing strat-

Ž .egy, Moorthy 1993 says A . . . the ratio of theoreticalto empirical work is . . . unpleasantly largeB. Ghe-

Ž .mawat 1997 argues that the inability of game the-ory-based new industrial organization to influencebusiness strategy is primarily due to the lack ofempirical work. We look forward to more studiesthat would enable robust generalizations.

In addition to theory testing, the structural esti-mates of NEIO models are managerially useful. First,the estimates have behavioral meaning and hence,managers can easily interpret them. Second, sincethe estimates are policy invariant, they can be usedto perform Awhat-ifB analyses. That is, by estimatingdetailed structural parameters of demand, cost andcompetitive interactions, a manager can evaluate theimpact of decisions such as the effect of a introduc-ing a new product or line extension, merger with acompetitor, impact of a price change, etc. In one ofthe early papers in marketing in the NEIO tradition,

Ž .Horsky and Nelson 1992 obtain detailed estimatesof demand and cost, and then investigate the impactof product repositioning and price changes for carsassuming Bertrand competition. These results arevery informative for strategic marketing mix choicesand are possible because of the firm-level modelingof this industry.

However, in order to be more broadly applicableto the auto market, the estimation should allow forgeneral forms of competition. For example, in a

Ž .recent study, Sudhir 2001a finds that firm behaviorin the auto market indicate more aggressive behaviorthan Bertrand in some segments, but more coopera-

tive behavior in other segments. Additionally, thenature of competition can vary in different periods.

Ž .For example, research by Bresnahan 1987 suggeststhat competitive behavior changes depending onwhether markets are in an expansionary or recession-ary state. Such information about competition, whenaccounted for in Awhat-ifB simulation models, willgive managers a more accurate picture of the impacttheir choices of marketing mix and positioningstrategies on profits compared to the original Bertrandassumption in Horsky and Nelson. A manager couldalso use these estimates to evaluate profitability im-plications of a merger between firms. For example,when Chrysler bought American Motors in 1987, itwould have been possible to study the profit impactof such a merger using such estimates. Antitrustofficials could have used such a model to evaluatethe social welfare implications of the merger. In fact,

Ž .Nevo 2001b uses his estimates of a structural modelof the cereal market to evaluate the profitability andsocial welfare implications of the merger betweenGeneral Mills and Ralston Purina’s branded cereal

Ž .line that included the Chex line .In general the structural models enable us to

better understand the sources of market power andŽprofitability. We had argued in the Section 1 that

.this is the fourth benefit of using NEIO models .This is useful for managers and anti-trust officials. Isthe superior profitability of a firm due to its better

Žability to provide customers what they want demand. Ž .advantages , superior efficiency cost advantages or

Žanti-competitive conduct tacitly cooperative behav-. Ž .ior ? Nevo 2001a finds that the market power of

firms in the cereal market can be best explained bythe firms’ demand advantages through differentiatedproducts targeted to different segments of the marketthan by anti-competitive conduct.

Given these benefits of the NEIO approach, webelieve that there are several applications of this tothe empirical study of competitive marketing strat-egy. Thus far, much of the applied work in this areahas been on market power, because these methodshave been of great interest to researchers exploringantitrust issues. Some of the most pervasive findingshave been that even in the most homogeneous ofindustries, there is significant market power if thereis a sufficient degree of concentration. BresnahanŽ .1989 presents the Lerner indices for a number of


markets based on prior NEIO studies. Firms alsoseem to be able to tacitly achieve cooperative con-duct. As these methods diffuse through applied areassuch as business strategy and marketing, there shouldbe a rapid increase in empirical applications and weshould be able to address issues of broader interest tobusiness researchers. Researchers in marketing havealready begun to explore many such issues. Forexample:

1. testing hypothesis about the nature of strategicinteraction and relative power between manu-

Žfacturers and retailers Kadiyali et al., 2000;.Sudhir, 2001b; Cotterill and Putsis, 2001 ;

2. analyzing change in competitive behavioramong firms due to an entry into the productline by one of the firms; the impact of a privatelabel on competitive interactions between man-

Žufacturers and retailer Kadiyali et al., 2000;.Putsis and Dhar, 1999; Cotterill et al., 2000 ;

Ž3. retailer pricing objectives and behavior Chin-.tagunta, 2000; Sudhir, 2001b ;

4. examining whether firms’ cooperative or ag-gressive strategies are consistent with their

Ž .long-term profit objectives Sudhir, 2001a

.Methodological innovations in estimating models

of dynamics of competition, discrete strategy choice,and asymmetric information are expected in the nearfuture, given the high degree of interest amongempirical industrial organization economists. Perhapsa measure of this interest is the large number of highquality doctoral dissertations in this area from manyprestigious economics departments. As the numberof researchers in marketing interested in this areaalso increases, we hope the Agap between theoreticaland empirical research in competitive marketingstrategyB will become smaller. With enough newpapers on issues of interest to marketers, we hope ameta-analysis of NEIO based industry and market-specific findings will also lead to generalizable con-clusions from this literature.

References

Agrawal, D., Lal, R., 1995. Contractual arrangements in franchis-ing: An empirical investigation. Journal of Marketing Re-

Ž .search 32 2 , 213–221.

Ashenfelter, O., Sullivan, D., 1987. Nonparametric tests of marketstructure: An application to the cigarette industry. Journal ofIndustrial Economics.

Bain, J.S., 1951. Relation of profit rate to concentration: Ameri-can manufacturing, 1936–1940. Quarterly Journal of Eco-nomics 65, 293–324.

Bajari, P., Hortacsu, A., 2000. Winner’s curse, reserve prices andendogenous entry: Empirical insights from eBay auctions.Working Paper, Department of Economics, Stanford Univer-sity, Stanford, CA.

Berry, S., 1994. Estimating discrete-choice models of productŽ .differentiation. Rand Journal of Economics 25 2 , 242–262.

Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices inŽ .market equilibrium. Econometrica 63 4 , 841–890.

Berry, S., Carnall, M., Spiller, P.T., 1996. Airline hubs: Costs,markups and the implications of customer heterogeneity.Working Paper No. 5561, NBER Inc.

Besanko, D., Gupta, S., Jain, D., 1998. Logit demand estimationunder competitive pricing behavior: An equilibrium frame-

Ž .work. Management Science 44 11 , 1533–1547.Besanko, D., Dube, J.P., Gupta, S., 2000. Heterogeneity and

target marketing using aggregate retail data: A structural ap-proach. Working Paper, Northwestern University, Evanston,IL.

Boulding, W., Staelin, R., 1990. Environment, market share, andŽ .market power. Management Science 36 10 , 1160–1177.

Boulding, W., Staelin, R., 1993. A look on the cost side: Marketshare and the competitive environment. Marketing Science 12Ž .2 , 144–166.

Bresnahan, T., 1987. Competition and collusion in the Americanautomobile industry: The 1955 price war. Journal of Industrial

Ž .Economics 35 4 , 457–482.Bresnahan, T., 1989. Industries and market power. In:

Ž .Schmalensee, R., Willig, R. Eds. , Handbook of IndustrialOrganization. North Holland, Amsterdam.

Bresnahan, T., Reiss, P.C., 1991. Empirical models of discretegames. Journal of Econometrics 48, 57–81.

Buzzell, R.D., Gale, B.T., 1987. The PIMS Principles—LinkingStrategy to Performance. The Free Press, New York.

Cabral, L., 1995. Conjectural variations as a reduced form. Eco-nomics Letters 49, 397–402.

Chintagunta, 2000. Investigating category pricing behavior at aretail chain. Working Paper, Graduate School of Business,University of Chicago.

Ž .Chintagunta, P., Kadiyali, V., Vilcassim, N. 1999 , Endogeneityand simultaneity in competitive pricing and advertising: Alogit demand analysis. Working Paper, Johnson GraduateSchool of Management, Cornell University, Ithaca, NY.

Corts, K., 1999. Conduct parameters and the measurement ofmarket power. Journal of Econometrics 88, 227–250.

Cotterill, R.W., Putsis, W.P., 2001. Testing theory: Do assump-tions of models of vertical strategic interaction of national andstore brands meet the market test? Journal of Retailing.

Cotterill, R.W., Putsis Jr., W.P., Dhar, R., 2000. Assessing thecompetitive interaction between private labels and national

Ž .brands. Journal of Business 73 1 , 109–137.Davis, P., 1997. Spatial competition in retail markets: Movie


theaters. Working Paper, Department of Economics, YaleUniversity, New Haven, CT.

Dockner, E.J., 1992. A dynamic theory of conjectural variations.Journal of Industrial Economics 40, 377–395.

Erdem, T., 1996. A Dynamic analysis of market structure basedŽ .on panel data. Marketing Science 15 4 , 359–378.

Ericson, R., Pakes, A., 1995. Markov-perfect industry dynamics:A framework for empirical work. Review of Economic Studies62, 53–82.

Feenstra, R., Levinsohn, J., 1995. Estimating markets and marketconduct with multidimensional product attributes. Review ofEconomic Studies 62, 19–52.

Gasmi, F., Laffont, J.J., Vuong, Q., 1992. Econometric analysis ofcollusive behavior in a soft drink market. Journal of Eco-

Ž .nomics and Management Strategy 1 2 , 277–311.Genesove, D., Mullin, W.P., 1998. Testing static oligopoly mod-

els: conduct and cost in the sugar industry. Rand Journal ofŽ .Economics 29 2 , 355–377.

Ghemawat, P., 1997. Games Businesses Play: Cases and Models.MIT Press, Cambridge, MA.

Goldberg, P., 1995. Imperfect competition in international mar-kets: The case of the US automobile industry. Econometrica

Ž .63 4 , 891–951.Green, E., Porter, R., 1984. Noncooperative collusion under im-

perfect price information. Econometrica 57, 87–100.Greenstein, S., Wade, J.B., 1998. The product life cycle in the

commercial mainframe computer market. Rand Journal ofŽ .Economics 29 4 , 772–789.

Gupta, S., Chintagunta, P., Kaul, A., Wittink, D.R., 1996. Dohousehold scanner data provide representative inferences frombrand choices: A comparison with store data. Journal of

Ž .Marketing Research 33 4 , 383–398.Hamilton, J.D., 1994. Time Series Analysis. Princeton Univ.

Press, Princeton, NJ.Hanssens, D.M., 1980. Market response, competitive behavior and

time series analysis. Journal of Marketing Research 17, 470–485.

Hausman, J.A., Wise, D.A., 1978. A conditional probit model forqualitative choice: Discrete decisions recognizing interdepen-dence and heterogeneous preferences. Econometrica 46, 403–426.

Heckman, J.J., 1978. Dummy endogenous variables in a simulta-neous equation system. Econometrica 46, 931–959.

Horsky, D., Nelson, P., 1992. New brand positioning and pricingŽ .in an oligopolistic market. Marketing Science 11 2 , 133–153.

Iwata, G., 1974. Measurement of conjectural variations inŽ .oligopoly. Econometrica 42 5 , 947–966.

Jarmin, S.R., 1994. Learning by doing and competition in theŽ .early rayon industry. Rand Journal of Economics 25 3 ,

441–454.Kadiyali, V., 1996. Entry, its deterrence and its accommodation:

A study of the US photographic film industry. Rand Journal ofŽ .Economics 27 3 , 452–478.

Kadiyali, V., Vilcassim, N., Chintagunta, P., 1996. Empiricalanalysis of intertemporal competitive product line pricing deci-sions: Lead, follow or move together? Journal of Business 69Ž .4 , 459–487.

Kadiyali, V., Chintagunta, P., Vilcassim, N., 2000. Power inmanufacturer–retailer interactions: An empirical investigation

Ž .of pricing in a local market. Marketing Science 19 2 , 127–148.

Lal, R., 1990. Improving channel coordination through franchis-Ž .ing. Marketing Science 9 4 , 299–318.

Lambin, J.J., Naert, P.A., Bultez, A., 1975. Optimal marketingbehavior in oligopoly. European Economic Review 6, 105–128.

Lancaster, K.J., 1971. Consumer Demand: A New Approach.Columbia Univ. Press, New York.

Lee, E., Staelin, R., 1997. Vertical strategic interaction: Implica-Ž .tions for channel pricing strategy. Marketing Science 16 3 ,

185–207.Leeflang, P.S.H., Wittink, D.R., 1996. Competitive reaction ver-

sus consumer response: Do managers overreact? InternationalJournal of Research In Marketing 13, 103–119.

Lindh, T., 1992. The inconsistency of consistent conjectures.Journal of Economic Behavior and Organization 18, 69–90.

McFadden, D., 1982. Qualitative response models. In: Hilden-Ž .brand, W. Ed. , Advances In Econometrics. Cambridge Univ.

Press, Cambridge.Miravate, E.J., 1997. Estimating demand for local telephone ser-

vice with asymmetric information and optional calling plans.Working Paper, Department of Economics, University ofPennsylvania, Philadelphia PA.

Moorthy, S.K., 1993. Competitive marketing strategies. In: Eliash-Ž .berg, J., Lilien, G.L. Eds. , Handbooks In Operations Re-

search and Management Science, Vol. 5: Marketing. NorthHolland, Amsterdam.

Nevo, A., 1998. Identification of the oligopoly solution concept ina differentiated products industry. Economics Letters 59, 391–395.

Nevo, A., 2001a. Measuring market power in the ready-to-eatcereal industry. Econometrica.

Nevo, A., 2001b. Mergers with differentiated products: The caseof the ready-to-eat cereal industry. Rand Journal of Eco-nomics.

Norman, G., 1983. Spatial pricing with differentiated products.Ž . Ž .Quarterly Journal of Economics 98 2 , 291–310, May .

Pakes, A., Ericson, R., 1998. Empirical implications of alternativemodels of firm dynamics. Journal of Economic Theory 79,1–45.

Pakes, A., McGuire, P., 1994. Computing Markov-perfect Nashequilibria: Numerical implications of a dynamic differentiatedproduct model. Rand Journal of Economics 25, 555–589.

Panzar, J., Rosse, J., 1987. Testing for AmonopolyB equilibrium.Journal of Industrial Economics.

Parker, P.M., Kim, N., 1999. Collusive conduct in private labelmarkets. International Journal of Research In Marketing 16Ž .2 , 143–145.

Parker, P.M., Roller, L., 1997. Collusive conduct in dupolies:Multimarket contact and cross-ownership in the mobile tele-phone industry. Rand Journal of Economics 28, 304–322.

Petrin, A., 1999. Quantifying the benefits of new products: Thecase of the minivan. Working Paper, Graduate School ofBusiness, University of Chicago.


Porter, R.H., 1983. A study of cartel stability: The joint executivecommittee. The Bell Journal of Economics 14, 301–314.

Porter, R., 1995. Role of information in US offshore oil and gaslease auctions. Econometrica 63, 1–28.

Prescott, J.E., Kohli, A.K., Venkatraman, N., 1986. The market-share profitability relationship: An empirical assessment ofmajor assertions and contradictions. Strategic ManagementJournal 7, 377–394.

Putsis Jr., W.P., Dhar, R., 1998. The many faces of competition.Ž .Marketing Letters 9 3 , 269–284.

Putsis, W.P., Jr., Dhar, R., 1999. Category expenditure, promotionand competitive market interactions: Can private labels expandthe pie? Working Paper, School of Management, Yale Univer-sity, New Haven, CT.

Raju, J.S., Roy, A., 1999. Competitive conduct in the US micro-processer industry. Working Paper, Wharton School, Univer-sity of Pennsylvania, Philadelphia, PA.

Roberts, M.J., Samuelson, L., 1988. An empirical analysis ofdynamic nonprice competition in an oligopolistic industry.Rand Journal of Economics 19, 200–220.

Rotemberg, J., Saloner, G., 1986. A supergame-theoretic model ofbusiness cycles and price wars during booms. American Eco-nomic Review 76, 390–407.

Roy, A., Hanssens, D.M., Raju, J.S., 1994. Competitive pricing byŽ .a price leader. Management Science 40 7 , 809–823.

Schmalensee, R., 1989. Inter-industry studies of structure andŽ .performance. In: Schmalensee, R., Willig, R. Eds. , Hand-

book of Industrial Organization. North Holland, Amsterdam.Shankar, V., 1997. Pioneers’ marketing mix reactions to entry in

different competitive game structures: Theoretical analysis andŽ .empirical illustration. Marketing Science 16 3 , 271–293.

Shankar, V., Bayus, B., 1999. Network effects and competition:An empirical analysis of the home video game industry.Working Paper, University of Maryland, College Park, MD.

Shepard, 1993. Contractual form, retail price, and asset character-Ž .istics in gasoline retailing. Rand Journal of Economics 24 1 ,

58–77.

Slade, M.E., 1992. Vancouver’s gasoline-price wars: An empiricalexercise in uncovering supergame strategies. Review of Eco-nomic Studies 59, 257–276.

Slade, M.E., 1995. Product rivalry and multiple strategic weapons:An analysis of price and advertising competition. Journal of

Ž .Economics and Management Strategy 4 3 , 445–476.Sudhir, K., 2001a. Competitive pricing behavior in the auto

market: a structural analysis. Marketing Science.Sudhir, K., 2001b. Structural analysis of manufacturer pricing in

the presence of a strategic retailer. Marketing Science.Suslow, V., 1986. Estimating monopoly behavior with competi-

tive recycling: An application to Alcoa. Rand Journal ofŽ .Economics 17 3 , 381–403.

Szymansky, D.M., Bharadwaj, S.G., Varadarajan, P.R., 1993. Ananalysis of the market share–profitability relationship. Journalof Marketing 57, 1–18.

Tyagi, R.K., 1999. A characterization of retailer response tomanufacturer trade deals. Journal of Marketing Research 36Ž .10 , 510–516.

Verboven, M., 1996. International price discrimination in theŽ .European car market. Rand Journal of Economics 27 2 ,

242–262.Vilcassim, N., Kadiyali, V., Chintagunta, P., 1999. Investigating

dynamic multifirm market interactions in price and advertis-ing: A conjectural variations approach. Management Science

Ž .45 4 , 499–518.Vuong, Q.H., 1989. Likelihood ratio tests for model selection and

non-nested hypothesis. Econometrica 57, 307–333.Wind, J., Lilien, G.L., 1993. Marketing strategy models. In:

Ž .Eliashberg, J., Lilien, G.L. Eds. , Handbook of OperationsResearch. Management Science, Vol. 5: Marketing. NorthHolland, Amsterdam.

Wolfram, C.D., 1999. Measuring duopoly power in the Britishelectricity spot market. American Economic Review 89, 805–

Ž .826, September .

structural analysis of competitive behavior: new empirical ... · the new empirical industrial...

Documents