This action might not be possible to undo. Are you sure you want to continue?

# Estimating Coke and Pepsi’s Price and Advertising Strategies

Amos Golan* Larry S. Karp** Jeffrey M. Perloff** March 1999

* **

American University University of California, Berkeley, and Giannini Foundation We benefitted greatly from George Judge’s comments about econometrics and Leo Simon’s comments on game theory. We are very grateful to Jean Jacques Laffont, Quang Vuong, and especially Farid Gasmi for generously providing us with the data used in this study and for advice. We received useful suggestions from participants at the "Industrial Organization and Food-Processing Industry" conference at the Institute D’Economie Industrielle at the Université des Sciences Sociales de Toulouse and anonymous referees and the associate editor. We thank Gary Casterline and Dana Keil for substantial help with computer programs.

Contact: Jeffrey M. Perloff (510/642-9574; 510/643-8911 fax) Department of Agricultural and Resource Economics 207 Giannini Hall University of California Berkeley, California 94720 perloff@are.Berkeley.Edu

2 Table of Contents

1. INTRODUCTION 2. OLIGOPOLY GAME 2.1 Strategies 2.2 Econometric Implications 3. GENERALIZED-MAXIMUM-ENTROPY ESTIMATION APPROACH 3.1 Background: Classical Maximum Entropy Formulation 3.2 Incorporating Sample Information 3.3 Incorporating the Non-Sample (Game-Theoretic) Information 3.4 Properties of the Estimators and Normalized Entropy 4. COLAS 4.1 Cola Estimates 4.2 Tests 4.3 Lerner Measures 4.4 Effects of the Exogenous Variables 5. CONCLUSIONS REFERENCES Appendix 1: The GME-Nash Estimator

1 3 4 5 7 7 9 11 14 15 17 20 22 23 24 26 30

Abstract A semi-parametric, information-based estimator is used to estimate strategies in prices and advertising for Coca-Cola and Pepsi-Cola. Separate strategies for each firm are estimated with and without restrictions from game theory. These information/entropy estimators are consistent and efficient. These estimates are used to test theories about the strategies of firms and to see how changes in incomes or factor prices affect these strategies.

KEYWORDS: strategies, noncooperative games, oligopoly, generalized maximum entropy, beverages JEL: C13, C35, C72, L13, L66

1. INTRODUCTION This paper presents two methods for estimating oligopoly strategies. The first method allows strategies to depend on variables that affect demand and cost. The second method adds restrictions from game theory. We use these methods to estimate the pricing and advertising strategies of Coca-Cola and Pepsi-Cola. Unlike most previous empirical studies of oligopoly behavior, we do not assume that firms use a single pure strategy nor do we make the sort of ad hoc assumptions used in conjectural variations models.1 Both our approaches recognize that firms may use either pure or mixed (perhaps more accurately, distributional) strategies. In our application to Coca-Cola and Pepsi-Cola, we assume that the firms’ decision variables are prices and advertising. We divide each firm’s continuous price-advertising action space into a grid over prices and advertising. Then we estimate the vector of probabilities — the mixed or pure strategies — that a firm chooses an action (a rectangle in the priceadvertising grid). We use our estimates to calculate the Lerner index of market structure and examine how changes in exogenous variables affect strategies. The main advantages of using our method are that we can flexibly estimate firms’ strategies subject to restrictions implied by game theory and test hypotheses based on these estimated strategies. The restrictions we impose are consistent with a variety of assumptions regarding the information that firms have when making their decisions and with either pure or mixed strategies.

Bresnahan (1989) and Perloff (1992) survey conjectural variations and other structural and reduced-form "new empirical industrial organization" studies.

1

These studies (Bjorn and Vuong 1985. or a firm only knows that its private information is correlated or uncorrelated with its rival’s — lead to restrictions of the same form. The econometrician observes only the firm’s action and not the marginal cost. Given the realization of marginal cost. it appears to both the rival and the econometrician that a firm is using a mixed strategy. firms observe each other’s private information but the econometrician does not. Due to the randomness of the marginal cost. Firms choose either pure or mixed strategies. firms might use pure strategies and know the distribution but not the realization of their rival’s cost. we concentrate on the situation where firms have private. the econometrician cannot distinguish between pure or mixed strategies. The equilibrium depends on whether firms’ private information is correlated. suppose that a firm’s marginal cost in a period is a random variable observed by the firm but not by the econometrician. As a consequence.marginal costs (or some other payoff-relevant variable).2 For example. If both firms in a market use pure strategies and each observes its rival’s marginal cost. Bjorn and Vuong and Kooreman estimate mixed strategies in a game involving spouses’ joint labor market participation decisions using a maximum likelihood (ML) technique. Bresnahan and Reiss 1991. and Kooreman 1994) involve discrete action spaces. each firm can anticipate its rival’s action in each period. uncorrelated information about their own . For expositional simplicity. All of these possibilities — firms have only public information. the firm chooses either a pure or a mixed strategy. Alteratively. which results in an action: a priceadvertising pair.but not their rival’s . For example. There have been few previous studies that estimated mixed or pure strategies based on a game-theoretic model. Our approach differs .

advertising. the GME estimator does not require the same strong. which indicates that those restrictions to the GME model is identical to the distributional assumption of the standard approach. their ML estimation problems are complex. An important advantage of our GME estimator is its computational simplicity. they assume that there is no exogenous uncertainty. we present a game-theoretic model of firms’ behavior. Despite the limited number of actions. we discuss our results and possible extensions. we can estimate a model with a large number of possible actions and impose inequality and equality restrictions implied by the equilibrium conditions of the game. 2. and variables that affect cost or demand. The fourth section contains estimates of the strategies of Coke and Pepsi. we use a generalized-maximum-entropy (GME) estimator. Our problem requires that we include a large number of possible actions in order to analyze oligopoly behavior and allow for mixed strategies. explicit distributional assumptions used in standard ML approaches. Doing so using a ML approach would be difficult if not impossible. First. Instead. Third. quantities. In the third section.3 from these studies in three important ways. in order to use a ML approach. OLIGOPOLY GAME Our objective is to determine the strategies of oligopolistic firms using time-series data on prices. they assume a specific error distribution and likelihood function. we describe a GME approach to estimating this game. In the next section. such as input . A special case of our GME estimator is identical to the ML multinomial logit estimator (when the ML multinomial logit has a unique solution). they allow each agent a choice of only two possible actions. In the final section. Second. In addition to this practical advantage. Using GME.

we assume that εi and εj are private. The econometrician observes payoff-relevant public information. In state k. To simplify the description of the problem. . Firm i’s strategy i i i i i is αk(z) = (αk1(z). We assume that two firms.. the distribution of Firm i’s private informa- . however. where αkr(z) is the probability that Firm i chooses i action xr given private information εk and public information z.. Firm i (and possibly Firm j). The firms. εk.. xn}..1 Strategies i i i The set of n possible actions (price-advertising pairs) for Firm i is {x1. but does not observe private information known only to the firms. αkr(z) is one for a particular r and zero otherwise. ... uncorrelated information.. where r is the action chosen by Firm i and s is the action chosen by Firm j. The distributions are constant over time but may differ across firms. . but not the econometrician. xs.. αk2(z). 2. so it does not know the condii tional probability αkr(z)... such as demand and cost shifters..4 prices or seasonal dummies. T. Firm j does not observe Firm i’s private information. We now describe the problem where the random state of nature is private information and uncorrelated across firms. but not the econometrician. i and j. The set of K possible realizations. x2. {ε1. εK}. play a static game in each period of the sample. know these distributions. αkn(z)). is the same every period for both firms. observes Firm i’s marginal cost or some other payoff-relevant random variable εi(t) in period t = 1. z). . we suppress the time variable t for notational simplicity. If Firm i uses a pure i strategy. Firm j knows. ε2. Where possible. z. i i j i The profit of Firm i in a particular time period is πrsk(z) = πi(xr..

Firm i chooses αk(z) to maximize expected profits.2) L r k ( z ) αr k (z ) i i 0. We cannot use these constraints directly.5 tion. where Ek is the expectations operator. Equations 2. . By taking expectations. If it is optimal for Firm i to use action r with positive probability. 2.1) L r k (z ) ≡ s i αs (z ) πr s k (z ) j i Y k (z ) ≤ 0 . Similarly Firm i knows the unconditional j probability αs(z) of Firm j. then Firm i’s expected loss from using action xr is (2. however. the expected loss of using that action must be 0. Σs αs(z)πrsk(z). If Yk(z) is Firm i’s maximum expected profits i given εk and z. we eliminate these unobserved variables and obtain usable restrictions. i which is non-positive.1 and 2. optimality requires that (2. Our estimation method selects the pure or mixed strategy equilibrium that is most consistent with the data.2 Econometric Implications Our objective is to estimate the firms’ strategies subject to the constraints implied by optimization. This probability is the expectation over Firm i’s private information: αr(z) = Ek i αkr(z). The Nash assumption is that Firm j knows the unconditional probability of Firm i using i action r. Hence. The equilibrium to this game may not be unique. i because they involve private information εk. j In state k.2. where i the expectation is taken over its rival’s actions.

First. or there may be measurement error. Taking expectations with respect to k of Equations 2. If information is correlated.3 .3) s αs ( z ) πr s ( z ) αs ( z ) πr s ( z ) s j i j i Y i (z ) ≤ 0 . however. Firm i’s beliefs would also be conditioned on the realization of εj.4 as a stochastic restriction and include additive errors in estimation. If we tried to estimate this model (Equations 2.3 and 2. δ(z). we can take expectations with respect to the private information and obtain equations analogous to 2. we can estimate the unobservable strategies αi(z) subject to the conditions implied by Firm i’s optimization problem. i i i where δr ≡ cov(Lrk.4) using traditional techniques. In our empirical application to the cola market. Equations 2. For each Firm i = 1. If information is observed by both firms. it would be reasonable to suppose that Firm i’s j j beliefs about j’s actions depend on the realization of εi.3. and the definition of δ would be changed. already has an additive function. Thus. The signs of both θ and δ would be indeterminate.4.2 Firms may use approximately optimal decisions due to bounded rationality. In both cases. we treat Equation 2. Equation 2. all the estimated δ are positive. we would have an additional additive term in 2.3 and 2. so that αs is replaced by αks.2. we obtain (2. the restrictions are slightly more complicated. we would run into two problems.6 i i i We define Yi(z) ≡ Ek Yk(z) and πrs(z) ≡ Ek πrsk(z). αrk) ≥ 0. 2 . However. imposing the various equality and inequality restrictions If εi and εj are correlated or observed by both firms.4. which is consistent with the model in the text where εi and εj are uncorrelated.4) i Y i ( z ) αr ( z ) 0. which we cannot distinguish from the additive error in 2. δr ( z ) i (2.2 and using the previous definitions. say θ.1 and 2. with either generalization. Therefore. δ(z) is the only "error term" we include in this equation.4.4. 2.

performs well with small samples. 3. 2. Shannon’s (1948) entropy is used to measure the uncertainty (state of knowledge) we have about the occurrence of a collection of events. and can incorporate inequality restrictions. as the problem is ill posed in small samples (there may be more parameters than observations). Unlike ML estimators. Our GME method is closely related to the GME multinomial choice approach in Golan. and Miller (1996). s = 1. Csiszár (1991). Jaynes (1984). Jaynes (1957a. Judge. Soofi (1992. Gokhale and Kullback (1978). we propose an alternative approach. Kullback (1959).1 Background: Classical Maximum Entropy Formulation The traditional entropy formulation is described in Shannon (1948). we would have to impose additional assumptions to make the problem well posed. In this section. and Perloff (1996 — henceforth GJP). 3. the GME approach does not require explicit distributional assumptions. Gibbons. Second. In this approach. To avoid these and other estimation and inference problems. GENERALIZED-MAXIMUM-ENTROPY ESTIMATION APPROACH We use generalized maximum entropy (GME) to estimate the firms’ strategies. we start by briefly describing the traditional maximum entropy (ME) estimation procedure. Letting x be a random variable with possible outcomes xs. Judge. we present the GME formulation as a method of recovering information from the data consistent with our game. and Greenberg (1989). Skilling (1989). . Shore and Johnson (1980). Then.7 from our game-theoretic model would be very difficult if not impossible with standard techniques. Denzau. 1957b). Levine (1980). 1994) and Golan.

which are their probability distributions over their actions. the sample information for cola manufacturers consists of time series of price-advertising pairs for each firm. and time series of exogenous variables that affect demand (a seasonal . such as restrictions from economic theory. The next two subsections explain how we incorporate sample and non-sample (theory) information. n. 1957b) proposed maximizing entropy. which Shannon interprets as a measure of the uncertainty in the mind of someone about to receive a message. subject to the sample and non-sample information... quantities sold. or non-sample information about the random variable. with probabilities αs such that Σs αs = 1. such as the sample moments. The function H. α2. The frequency that maximizes entropy is an intuitively reasonable estimate of the true distribution when we lack any other information. the firms’ price-advertising decisions are the random variables that correspond to x in the previous example. we want to alter our "intuitively reasonable" estimate. reaches a maximum when α1 = α2 = … = αn = 1⁄n. To recover the unknown probabilities α. subject to available sample-moment information and adding up constraints on the probabilities. as (3. We want to estimate the firms’ strategies. . Shannon (1948) defined the entropy of the distribution α = (α1.1) H ≡ s αs ln αs . The method of Maximum Entropy proceeds by choosing the distribution that maximizes entropy. In our game. Jaynes (1957a. If we have information from the experiment. In our application.8 …. αn)’.. where 0 ln 0 ≡ 0.

we can estimate this model using maximum likelihood multinomial logit or probit. 3. The variable yi tr equals one in period t if action r is observed and zero otherwise. If we view the moment conditions as stochastic restrictions. we follow GJP and relate the set of covariates zt to the data yi and the unknown αi and ei . and ζr are unknown parameters to be tr tr tr estimated. which is a generalization of the multinomial logit. 2. and the price of sugar). ei .4. which are price-advertising pairs.2 Incorporating Sample Information We incorporate the sample information into our GME estimator of the strategies. we derive a ME estimator. αi.3 . we obtain estimates of the probabilities αi as a function of the public information.2) yt r i G(z t ζ i ) r et r i αt r i et r .2 and assuming that G(·) is a known cdf such as the logistic or the normal. We multiply Equation (3. The game-theoretic restrictions.9 dummy and income) and cost (an interest rate. by maximizing the entropy of αi subject to the moment or consistency conditions that contain sample information.2. z. With either the ME or GME approaches. i i for i = 1. To avoid having to assume a specific cdf. there are n actions.2) by each tr tr tr . which is identical to the ML multinomial logit estimator (when the ML estimate is unique). By eliminating ei from Equation 3. We can use either of two approaches (as GJP shows). contain all the non-sample information. a wage rate. Equations 2. we obtain a GME estimator. In our problem. where yi and zt are observed and αi . If we require that the moment restrictions hold exactly. The variable yi is a tr function of the public information: (3.

ei. ztl. and span the interval [ 1 / T . i i i We rewrite each error element as er ≡ Σm vmwrm. symmetric around zero. L. 1].. We obtain the basic GME estimator by maximizing the sum of the entropy corresponding to the strategy probabilities. wrM]’ that have the i i properties of probabilities: 0 ≤ wrm ≤ 1 and Σm wrm = 1.. As we discussed in Section 3.. .10 covariate variable.. i for l = 1.. that are at uniform intervals.. . w2. For example. .3) t yt r z t l t i αt r z t l t i et r z t l . we represent the GME consistency conditions.. 0. vi = i i i [v1. 1 / T ]. and there exists w1. where T is the number of observations in the sample. as .. Using this parameterization. v2.. 1 / T )’. To determine the entropy of ei. if M = 3.. αi.3).1. i i i i i Each error term er has corresponding unknown weights wr = [wr1. we reparameterize its elements using probabilities. and r = 1. vM]’ of dimension M ≥ 2.3. . n.. and w3 such that each noise component can be i written as er = wr 1 / T i i wr 3 / T . and sum over observations to obtain the stochastic sample-moment (data consistency) restrictions: (3. but the elements of ei range over the interval [-1. wr2. called the support space. We start by choosing a set of discrete points. subject to that data consistency condition (3. the arguments of the Shannon’s entropy measures must be probabilities. Equation 3.. then vi = i i i ( 1 / T . and the entropy from the noise. which is the number of covariates in z. The elements of αi are probabilities.

4) t t i αt r z t l t t m i αt r z t l t i et r z t l wt r m v m z t l . the GME problem for each firm is (3. i i i i For notational simplicity. we may estimate αi and αj separately.3 and 2. are independent and define w as the vector which contains the elements wtrm.4. e. x. GJP shows how to estimate this model and demonstrates that the GME problem can be viewed as a generalized logit likelihood function. 1.4. which includes the traditional logit as a special case. 2. The GME-Nash estimator. to the GME estimator. …. 2.7) 1 αt 1 wts 1. w α ′ ln α w ′ ln w . When using the GME estimator. Equations 2. Equation 3.11 yt r z t l (3. Henceforth we refer to the GME estimator that uses only sample information as "the GME" estimator. for s = 1. and the errors. …. If we assume that the actions. subject to the GME consistency conditions.6) (3. T.3 Incorporating the Non-Sample (Game-Theoretic) Information We obtain the "GME-Nash" estimator by adding the game-theoretic restrictions. and the normalization constraints (3. and its application to . w α. we now drop the firm superscript. n and t = 1.5) max H α . 3.

δ2. Miller 1994. as we noted above. w = (wi´. [In contrast. the stategies can be estimated separately when using the GME estimator.3 Let y = (yi´. so we use the "three-sigma rule" (Pukelsheim. To explain how to implement this i estimator. Let va be a vector of dimension J a ≥ 2 with corresponding unknown weights ωa such that r (3. we also have to estimate δi(z). 1994. i j We do not have natural boundaries for δi.12 data. 2. from Equation 2. where sigma is the larger of the empirical standard deviation of the discrete action space of prices or advertising. πrs. Golan.4. and Miller 1996) to choose the limits of these support spaces. Judge. α = (αi´. we need to estimate αi and αj jointly because both strategy vectors appear in Equation 2. comprise the contributions of the current paper. The support spaces va are defined to be symmetric around zero for all ar. We then consider the more realistic case where we need to estimate the parameters of the profit function. yj´)´.9) ar .2. Our first step is to reparameterize δi(z) using probabilities. As we discussed in Section 2. αj´)´. This assumption allows us to concentrate on the game-theoretic restrictions. 3 . (3. To use the GME-Nash estimator.] Further. for a = δ1. wj´)´.8) j ωr j v a ωa r a 1. δi(z) is nonzero if the econometrician does not observe firms’ private information or if firms make mistakes in optimization. we first proceed as if we knew the parameters of the profit function. ωδ ) ′. and ω (ωδ . i = 1.4.

w . we need to determine how to impose the game-theoretic restrictions.3 and 2. hold at some or all of the values of z in our sample. we would require that the game-theoretic restrictions hold for all possible values of z. These restrictions. The GME-Nash problem is (3. qi. price of sugar. Instead. w. however. Solving the problem (3. we assume independence between the actions and the errors.8 and 3. but we observe Firm i’s output. The terms δr in Equation 2. which depends on the demand and cost functions.4 for each firm. w.4 are defined by Equations 3. w. ω α′ ln α w ′ ln w ω ln ω α.4 for each firm. We simultaneously estimate the strategies and the parameters of the profit function.3 and 2. and the adding-up conditions for α. We do not have observations on cost.10) Max H α. the necessary economic condii tions 2.2). Ideally. We now turn to the more realistic case where we do not know the parameters of the profit functions. Equations 2. ˜ ˜ ˜ In order to write the Lagrangean for this problem.10) yields the estimates α.13 As above. and ω. we impose the weaker condition that the game-theoretic restrictions.4. we impose the restrictions for only a subset of the observations. affect the α for all observations through the Lagrangean multipliers corresponding to the sample information (the moment conditions Equation 3. ω subject to the data consistency conditions 3. and . We cannot impose these restrictions for all values because we cannot write α(z) in closed form independent of the unknown Lagrangean multipliers. and some factor-cost (wages. Because of the large number of restrictions for each value of z in our application.9. and ω.

we choose a support for each such parameter and estimate the probability distribution over that support. . GJP shows that the GME estimator is consistent given an appropriate choice of the bounds of the error term in the data consistency constraint 3. zd. but they differ in efficiency and information content. We substitute these functions in the Constraints 2. zd. where zd are demand shifters (income and a seasonal dummy) and φi are parameters.3 and 2. Appendix 1 shows that the GME-Nash estimator is consistent. We assume that Firm i’s cost is given by a cost function Ci(qi.4 That is. qi = qi(xi. ηi).4 Properties of the Estimators and Normalized Entropy Both the GME and GME-Nash estimators are consistent. Under the assumption that a solution to the GME-Nash estimation problem exists for all samples. xj. We estimate the parameters φi and ηi using the same method described in the previous subsection for estimating parameters that are not probabilities. zc. where ui is an error term. Similarly. φi) + ui. zc. 3. 4 A formal derivation is available from the authors. where ηi are parameters.14 interest rate) variables. We perform this estimation by maximizing the sum of all the entropy measure in equation 3. we assume that the demand for Firm i’s product is qi(xi.10 plus the entropy associated with the unknown demand and cost parameters.4 and estimate ηi and φi jointly with the other parameters.4. φi). Because we observe demand (but not cost) we have an additional set of data consistency (sample) restrictions in the form of demand equations for each Firm i. xj.

15 GJP show that the GME estimates of α have smaller variances than the ME-ML multinomial logit estimates. The magnitude of the change in normalized entropy from imposing the gametheoretic restrictions provides a measure of the information they contain. Laffont. and Gasmi. The corresponding shares for Pepsi were 34.4% and its share of the national carbonated soft-drink market was 27. The possible solution space for the GME-Nash estimate of α is a subset of the solution space of the GME estimate of α.7 Monte Carlo sampling experiment that support this conjecture are available from the authors. We especially thank Farid Gasmi for patiently describing the data and making suggestions about the specification of our model.8% (according to the Beverage Industry Annual for 1986). 4. The normalized entropy measure is S(α) = 1 if all outcomes are equally likely. we estimate the price and advertising strategies for Coca-Cola and Pepsi-Cola using the GME and GME-Nash approaches. The data were generously provided by these authors. See Soofi (1992) and Appendix 1 for a derivation of the properties and inferences results for this estimator. In 1981 for example. The normalized entropy measure is S(α) = -(Σr αr ln αr)/(ln n).6%. dominate the cola and soft-drink markets. and Vuong (1992). Gasmi and Vuong (1991). Thus. Coca-Cola’s share of colas was 44. COLAS Using quarterly data for 1968-1986. Inc. 7 6 5 .5 We can quantify the added information contained in the game-theoretic restrictions by comparing the normalized entropy of α with and without the restrictions. The Coca Cola Company and Pepsico. and is S(α) = 0 if we know which action will be taken with certainty. we conjecture that the GMENash estimator has a smaller variance than the GME.6 We use quarterly data for 1968-86.6% and 21. which were obtained from a variety of secondary sources and are described in Gasmi (1988).

8 .1 so that it can be estimated along with the other parameters in the GME-Nash model. It would be possible to alter our model to allow for lagged advertising by assuming that the firms’ only decision variable is price and that advertising is only a demand shifter The earlier studies did not include a constant term. we use only one. We reparameterize 4. where η0. some of them used separate interest rates for the two companies. We assume that the marginal and average cost of Firm i i i i is ci = η0 + η1 × real price of sugar + η2 × real unit cost of labor in the nondurable i manufacturing sector + η3 × real yield on a Moody’s AAA corporate bond.16 We assume that firms set prices and advertising and use the demand specification from these earlier studies: (4. Were the demand curve to depend on lagged values of decision variables. qi is the quantity sold. and Vuong and we chose a demand curve that does not include lagged values of decision variables (prices and advertising) or lagged functions of those variables (lagged quantity). firms would realize that choices today influence profits in the future and hence would not act as though they were playing a repeated static game. Because the correlation between these two interest rate measures is 0. 2. pi is the real price charged.99.8 Both Gasmi. Laffont. and Ai is the real i advertising by Firm i. φ1 is negative. η3 ≥ 0. i i φ2 and φ3 are positive. I is income. i ≠ j.1) qt i φ0 i φ 1 pt i i φ 2 pt i j φ 3 (A t )½ i i φ 4 (A t )½ i j φ5 d i φ6 I i u i. This restrictive functional form is necessary given our economic model in which firms play a repeated (static) game and both price and advertising are decision variables. where i = 1. ui is an error term. η1. η2. d is a seasonal dummy. Moreover.

726 and 71. although the problem becomes difficult due to data and computational limitations. the prices range between $10. In principle our estimation methods can be applied to dynamic games.521 for Pepsi-Cola. our estimates did not vary greatly as we experimented with other action spaces. For example.646 and $9. Advertising expenditures range between 5. This difference in price levels is apparently due to the greater use of Coke syrup at fountains. 4. firms have 35 possible actions in each period. we do not believe that it is reasonable to view advertising as anything but a decision variable.814 for Pepsi-Cola.17 (like income).9 We divide the range of possible prices into seven intervals and the range of possible advertising levels into five intervals. 10 9 .1 Cola Estimates For both the GME and GME-Nash models. including Erickson (1992). Smoothing aside. There have been a number of efforts to estimate dynamic models of this sort.966 for Coca-Cola and from 7. we estimate strategies distributions that are always single peaked (unlike the results reported below where Pepsi has double-peaked mixed strategies).058 to 50. the strategy mode still occurs at the same area of price-advertising space. Karp and Perloff (1989) and Roberts and Samuelson (1988). This smoothing effect is the standard result from reducing the number of categories in a histogram. Though this approach would lead to a much simpler estimation problem than the one we examine. Erickson.10 If we reduce the number of possible actions. Another approach we could have taken would be to include lagged sales (arguing that consumers develop brand loyalties) and then estimated a dynamic oligopoly model.886 and $17. so that a stock effect exists. Within the sample.790 for Coca-Cola and between $6. for example. estimates a game in which Coke and Pepsi’s current market share depends on the lagged market share in addition to current advertising.

η1 = η2 = 0 (due to the theoretical restriction that the coefficient be non-negative). To save space. 0. For the GME-Nash. 0.079. φ5 = 7. The corresponding demand coefficients for PepsiCola are -20. z.730. 0.94 for Pepsi.93 for Coke and 0.056.208 for CocaCola. φ4 = -0.482. 5.737. We examined the sensitivity of our results to this assumption. 12 11 .592. -0. φ6 = 0.211. The corresponding cost coefficients for Pepsi-Cola are 7. These tables are available from the authors.741. We compared the estimate here where we imposed the theoretical restrictions on every fifth period (the frequency) starting with the third quarter to one with the same frequency where we started with the second or fourth quarter and found that the results were virtually identical. -1.12 From the estimated coefficients. we can calculate the strategy probabilities. φ1= -1.11 The Coca-Cola demand coefficients are φ0 = 4. φ2 = 2. the estimated cost parameters are η0 = 13.232. 0.021. φ3 = 0. we greatly reduce the size of the estimation problem. the GME (and GME-Nash) estimates have the opposite sign of the coefficients that would be produced by the roughly comparable ML multinomial logit approach.596. By only imposing the restrictions in about one-fifth of the periods. Due to an arbitary normalization convention. We also found that our estimates were not very sensitive to reducing the frequency. and η3 = 0.137.808. and 0. we do not report the coefficients for Pepsi or the two tables for the GME-Nash. we impose sign restrictions from economic theory on both the cost (all cost coefficients are non-negative) and demand parameters (demand falls with a firm’s own price and rises with the other firm’s price and its own advertising) and the game-theoretic restrictions in 15 periods (every fifth quarter starting with the third quarter). Table 1 shows the GME estimates of Coca-Cola’s coefficients on the exogenous variables.582. and 2.549.18 To estimate the GME-Nash model.251. The correlation coefficients between observed quantities and those predicted by the demand equation are 0.

The normalized entropy measures for the GME. If this theoretical information is true.31 for the GME model and 0.64 for the GME-Nash model. These numbers show the extent to which the game theoretic restrictions bind: They measure the amount of additional information contained in the restrictions. . The corresponding marginal distributions for price and advertising strategies for both the GME and GME-Nash models are shown in Figure 3 for Coke and in Figure 4 for Pepsi.41.73 (Pepsi). which is the expected value of 1 . near the midpoint of the sample.31 and 0. for each period. the pseudo-R2. for Coca-Cola in Figure 1 and Pepsi-Cola in Figure 2. is 0.S(·) for both firms (see Appendix 1). The GME-Nash model is flexible enough to allow for both pure and mixed strategies. 0.66 (Coke) and 0. the GME probability estimates are more uniform (reflect greater entropy) than the GME-Nash estimates. Similarly. Out of the 76 periods of the sample. This pattern is repeated in virtually all periods. 0. it improves our estimates. are closer to one (the upper bound of entropy) than are the corresponding GME-Nash measures. We show the estimates for the first quarter of 1977.19 α. panel a shows the GME estimates and panel b shows the GME-Nash estimates. The GME-Nash marginal distributions put more weight on the category with the largest probability than do the GME marginal distributions. In both figures. For both companies. We can compare the different estimators empirically using the normalized entropy (information) measure S(α). These figures illustrate that the game-theoretic conditions contain additional information beyond that in the data alone. there are three periods for each firm where it uses a pure strategy.

using a variety of tests (see Appendix 1.1) = 24 degrees of freedom. These test results are for a 0. We reject the null hypothesis that the two sets estimated strategies are identical in 74 out of the 76 periods. we reject the null hypothesis that the GME and GMENash estimated strategies are identical in 34 periods (out of the 76 total periods) for Coke and in 25 periods for Pepsi. so that imposing these restrictions affects our estimates of the strategies. Kullback. by comparing Figures 1b and 2b. we conclude that the profit-maximizing. Thus. Next. 0. we see that Coke and Pepsi had very different strategy distributions in the middle of the sample. Nash restrictions are consistent with the data and contain useful information. Coke had a single-modal strategy distribution with most weight on a moderate price-moderately intense advertising strategy.05 significance level.20 4.4) is consistent with the data. and Gokhale and Kullback.2 Tests We now test whether our theory is consistent with the firms’ behavior (data). we conclude that the economic theory represented by the set of conditions (2. 1978).2H(GME-Nash)] = 359. 1959. 13 .01.38 < χ2 1050. That is. Thus.1) × (5 . For example. As there are seven support points for the price strategy and five for the advertising strategy. we compare the strategies of the two firms for the GME-NASH model. the firms use different strategies. We now compare the strategies (estimated α) of the GME and the GME-Nash models using first cross-entropy χ2 tests (Appendix 1) and then Kolmogorov-Smirnov (KS) tests.3) and (2.13 On the basis of a cross-entropy χ2 test. there are (7 . where H(·) is the optimal value of the objective function. The entropy-ratio test statistic (which is analogous to the likelihood-ratio test from classical statistics) is [2H(GME) .

wage. The χ2 test-statistic values are 54. 34.41. and 56. are identical. and bond rate. price of sugar. 78. That is. we investigate the significance of the individual covariates.07.51 for income. and bond rate respectively. wage. the GME-Nash estimator correctly predicted which of the seven price categories Coke chose in 55% of the periods.14 If. These estimators fit the data reasonably well. however. We obtain similar results using the Kolmogorov-Smirnov (KS) test to examine whether the estimated GME and GME-Nash distributions differ systematically. For example. we test whether the estimated coefficient is zero (H0: zl = 0) or nonzero (H1: zl ≠ 0). though the factor prices do not greatly affect the marginal costs of the firms. We cannot reject the hypothesis that that estimate is the same as the average of the GME and GME-Nash estimates based on KS tests. averaged over all periods. we cannot reject the hypothesis that the GME and GME-Nash distributions and marginal distributions. we examine the hypothesis that the distributions are the same period by period. The ML multinomial logit (or ME model) for the entire sample ignoring z is the same as the empirical distribution (the frequencies) of the actions.21 whereas Pepsi had a bimodal distribution with the most weight on a high price-intensive advertising strategy. we reject the null hypothesis for all the covariates at the α = 0. price of sugar.38. we can reject the hypothesis for Coke in 66 (out of 76) periods and for Pepsi in 62 periods. Moreover.01 level.4). 14 . l = income. Thus. Next. As a result. as Table 2 shows for both the GME model and for the GME-Nash model. On the basis of a KS test at a 5% significance level. they do affect the strategies the firms use (see Section 4.

For comparison.45.16 We suppress the dependence of all functions on the public information. Usually. Laffont.15 4. i i i E[(pi . If the GME-Nash restrictions are correct.27 for Pepsi. the GME predictions are in each case more accurate than those of the GME-Nash: For example Coke’s price category is correctly predicted in 70% of the periods by the GME model and only 55% by the GME-Nash. we expect the GME-Nash to have lower mean squared errors than the GME. where ci is our estimate of Firm i’s marginal cost.24 for Coke and 0. the average adjusted Lerner index is 0. Averaged over the sample. 16 15 A formal derivation of the relevant equations is available from the authors. In this study.42 and for Pepsi is 0. In our study. the GME-Nash can predict better than the GME. and Vuong (1992). We use our estimates of probabilities to calculate the expected Lerner index. and hold z constant for purposes of this discussion. which is the percentage by which price is set above marginal cost. the index for Coke is 0. the GME-Nash estimates Based on our Monte Carlo simulations (available from the authors).3 Lerner Measures A standard measure of market power is the Lerner index. Thus. the Lerner index ranges between zero (competition) and one. we also calculated the Lerner index for the Bertrand-Nash model using the coefficients from Model 1 of Gasmi. .ci)/pi] = Σr αr[(pr .ci)/pr]. z.22 the model missed by more than one category in only 11% of the periods (this fact is not shown in the table).

but for Coke more probability These differences in the Lerner indexes are largely due to differences in the estimates of costs.154 and the corresponding elasticity for Pepsi is 0.17 4. By using the periods where we do not impose the constraints.18 Some of these elasticities are large in absolute value because the corresponding probabilities are close to zero. increases the probability that Coke.4 Effects of the Exogenous Variables Using our estimated models. Table 3 shows the average strategy elasticities using the GMENash estimates (the percentage change in expected action divided by the percentage change in a z variable). which shifts out demand. we summed over categories to compute the marginals of the averages. we average the probabilities for these periods. and to a lesser extent Pepsi. we can calculate the effect of a change in each of the exogenous z variables on the strategy probabilities. we are able to use an explicit formula for probabilities when calculating the derivatives. we calculate the derivatives of probabilities and average them over the 59 periods where we do not impose the game-theoretic constraints. First. α. By inspection of Table 3. The elasticity of Coke’s expected price with respect to income is 0. using the same approach as is used with logit and probit models. charge higher prices. Our GME-Nash cost estimates are substantially higher than their ML-Bertrand estimates. Pepsi’s advertising strategy does shift to the right. and used the results to calculate the elasticities of the marginals. we would have had to calculate the derivatives of a system of 70 implicit equations. Had we used the periods where the game theoretic constraints are imposed. An increase in income (and demand) spreads a unit cost of advertising over a greater volume of sales. Then. Using these averages.0013. we see that an increase in income. 18 17 .23 indicate that firms have less market power than do ML estimates of a Bertrand-Nash equilibrium. so we expect higher income to shift the distribution for advertising to the right.

This method is more flexible and . the actions are price-advertising pairs.03 for Pepsi.24 weight is shifted to both tails. Both methods are free of parametric assumptions about distributions and ad hoc specifications such as those used in conjectural-variations models. The elasticities with respect to the corporate bond rate are 0. we allow for both pure and mixed strategies. the corporate bond rate does not directly affect Pepsi’s costs.004 for Coke’s expected price. possibly because Pepsi thinks that it alters Coke’s strategy. the bond rate might indirectly affect Pepsi’s strategy. 0. an increase in the interest rate changes the mix of probabilities of charging a high price. Coke’s costs increase with the interest rate. Unlike previous studies of oligopoly behavior that only allowed for pure strategies.097 for Coke and 0. In our application to the cola market. An increase in the interest rate shifts more weight to the tails of Coke’s advertising strategy. CONCLUSIONS We developed two methods of estimating the strategies of firms. 0. We can calculate similar elasticities with respect to the other exogenous variables. Coke responds more than does Pepsi. For Coke. 5. Despite the absence of a direct effect (via costs). for changes in either income or bond rates. and 0.0005 for Pepsi’s expected advertising. which are the probabilities of taking particular actions. Thus. The elasticity of expected advertising with respect to income is -0. According to our estimates.032 for Coke’s expected advertising.004 for Pepsi’s expected price. and it has a negligible effect on Pepsi’s strategy. Our simplest approach is to use generalized maximum entropy (GME) to estimate the strategies for each firm using only sample information.

Tests show that the profit-maximizing. To do so only requires replacing profits with an appropriate alternative criterion. Both the traditional ML and the GME estimators ignore restrictions imposed by economic theory and some information about demand and costs. such as wars and joint decisions by husbands and wives. .25 efficient than the standard maximum likelihood multinomial logit (ML) estimator. We are able to use our estimates to show how changes in exogenous variables such as income or factor prices affect the firms’ strategies. because they contain information. The application to the cola market demonstrates that both the GME and GME-Nash models can be used practically. Our GME and GME-Nash approaches to estimating games can be applied to many problems in addition to oligopoly. alter our estimates of firms’ strategies. Nash restrictions are consistent with the data but that. Our generalized-maximum-entropy-Nash (GME-Nash) approach estimates firms’ strategies consistent with the underlying data generation process and the restrictions implied by game theory.

Timothy F." California Institute of Technology working paper 537. R. G.T..S. (1991): "Why Least Squares and Maximum Entropy? An Axiomatic Approach to Inference for Linear Inverse Problems. Gibbons. Greenberg (1989):: "Bayesian Estimation of Proportions with a Cross-Entropy Prior. 38. and J. Econometrics of Duopolistic Games in Prices and Advertising: The Case of the U.1748. New York: North Holland. MacKinnon (1993): Estimation and Inference in Econometrics. Reiss (1991): "Empirical Models of Discrete Games. Davidson. 19. and Quang H. Handbook of Industrial Organization. eds.M.D. 57-81. Vuong (1985): "Simultaneous Equations Models for Dummy Endogenous Variables: A Game Theoretic Formulation with an Application to Labor Force participation. Bresnahan. 1843-1861. 48. Erickson. P. . New York: Oxford University Press. 1732." The Annals of Statistics. and Peter C. (1992): "Closed Loop Duopoly Advertising Strategies" Management Science. 1989. G. I.26 REFERENCES Bjorn. Farid. (1989): "Studies of Industries with Market Power.. P." Journal of Econometrics. Csiszár. 1988. Bresnahan. and E. 18. Timothy F. Denzau.. 2032-2066. Soft Drink Industry..C. Gasmi. California Institute of Technology Ph." Communications in Statistics-Theory and Methods. A." in Richard Schmalensee and Robert Willig.. A.

Jean Jacques Laffont. (1957b): "Information Theory and Statistical Mechanics. T. Vuong (1992): "Econometric Analysis of Collusive Behavior in a Soft-Drink Market. and George Judge. Gasmi. 620-630. and Quang H. E.." Physics Review. ed." Journal of the American Statistical Association. (1957a): "Information Theory and Statistical Mechanics. Perloff (1996): "Recovering Information from Multinomial Response Data. II. 84153. and S. 171-190. Jaynes. Amos. New York. Advances in Econometrics. Berkeley.C. New York: John Wiley & Sons. Farid. Farid. Golan. 277-311." Physics Review.: Marcel Dekker. T. and Douglas J. Gokhale. 9. Jaynes. Golan." George F. (Presented at the 1997 Summer Meetings of the North America Econometric Society). Jr. Connecticut: JAI Press Inc. 91. Golan." Journal of Economics & Management Strategy. Amos. E. Rhodes. George Judge. Amos. Kullback (1978): The Information in Contingency Tables. D. "An Econometric Analysis of Some Duopolistic Games in Prices and Advertising.27 Gasmi. U.Y. 1. N. Greenwich. V. 1996. and Quang H. Miller (1996): Maximum Entropy Econometrics: Robust Estimation with Limited Data. George Judge. Vuong (1991)." Working Paper. . and Jeffrey M. 106. "A Maximum Entropy Approach to Empirical Likelihood: Estimation and Inference. 108.

Sheldon and Dennis R. thesis. University of California. (1980): "An Information Theoretical Approach to Inversion Problems. 255 . "Empirical Likelihood Ration Confidence Regions. 91-108." The Annals of Statistics." American Statistician. Perloff. L. (1992): "Econometric Analysis of Imperfect Competition and Implications for Trade Research. "Empirical Likelihood and General Estimating Equations. Kullback. Miller. Berkeley. Douglas J.D.28 Karp. C.. J. 22 (1994). Lawless. Jeffrey M." Working Paper. and J Perloff (1989) "Dynamic Oligopoly in the Rice Export Market" Review of Economics and Statistics. 1725-1747.470. New York: John Wiley & Sons. 48. Henderson. (1959): Information Theory and Statistics. ." Journal of Physics. Pukelsheim.68. (1994): "The Three Sigma Rule. 13. J. (1994): Entropy and Information Recovery in Linear Economic Models. LXXI 462 ." in Ian M.. A. Industrial Organization and International Trade: Methodological Foundations for International Food and Agricultural Market Research. A. Owen. Cardell. Kooreman. S. 88-91.. Peter (1994): "Estimation of Econometric Models of Some Discrete Games. 19 (1990). 9. Mittelhammer. 300-325. and N. and J. 1997. Levine. F." The Annals of Statistics. R. Rafi D. Qin. NC-194 Research Monograph Number 1.." Journal of Applied Econometrics. Ph. "On the Consistency and Asymptotic Normality of the Data-Constrained GME Estimator in the GLM. eds. Washington State University.

200 . E. 87. IT-26. ed. Soofi. and L Samuelson (1988): "An Empirical Analysis of Dynamic Nonprice Competition in an Oligopolistic Industry". (1992): "A Generalizable Formulation of Conditional Logit with Diagnostics. Rand Journal of Economics 19.220. Skilling. 812-816. Dordrecht: Kluwer Academic.." IEEE Transactions on Information Theory. (1994): "Capturing the Intangible Concept of Information. C. Shannon. E. 27.29 Roberts. E. Maximum Entropy and Bayesian Methods in Science and Engineering. (1948): "A Mathematical Theory of Communication. Soofi. 17387." in J. Johnson (1980): "Axiomatic Derivation of the Principle of Maximum Entropy and the Principle of Minimum Cross-Entropy. J." Journal of the American Statistical Association..J. E. S. 379-423. and R. Shore. 26-37. . S." Bell System Technical Journal. J. M." Journal of the American Statistical Association. 1243-1254. Skilling. W. (1989): "The Axioms of Maximum Entropy. 89.

w.30 Appendix 1: The GME-Nash Estimator A2. ˜ ˘ are equal to each other and to the true strategies in the limit as the sample size becomes infinite. . 135). all the estimators are consistent. its variance is finite. p. α. We want to prove Proposition: Given assumptions 1-3. the GME estimates α. plim(α) = plim(α) = α. We make the following assumptions: ˆ Assumption 1: A solution of the GME-Nash estimator (α. α. That is. and letting all the end points of the error support spaces v (for each firm) be normed by ˜ ˘ T . and the error distribution satisfies the Lindberg condition (Davidson and MacKinnon. ω) exists for any ˜ ˜ ˜ sample size. Assumption 2: The expected value of each error term is zero. According to this proposition. Assumption 3: The true value of each unknown parameter is in the interior of its support. T → ∞. This result holds even when the profit parameters are unknown. and the ME˜ ˘ ML estimates α. and the GME basic estimates. the GME-Nash estimates.1 Consistency Call the GME-Nash estimates of the strategies α. 1993.

Given Assumption 3. be .3 and 2.31 Proof: i) The consistency of the GME estimator is proved in GJP. or equivalently all the parameters (strategies as well as demand coefficients) are constrained to be zero (or at the center of their supports). ˜ 1996): plim φT = φ. 1990. Gokhale and Kullback. ˜ iii) The GME-Nash with unknown profit parameters is consistent. and Soofi. As T → ∞. ˘ ii) The GME-Nash with known profit parameters is consistent: By Assumption 1. v1 and vm. Golan and Judge. Let the end points of the error supports of v. plim αT = α. Doing so yields the total entropy value of the three sets of . the GME is a consistent estimator of φ in Equation A1. By the argument in (ii). The argument in (i) together with Assumption 2 implies that plim αT = α. ψs → 1 for all s in the dual-GME. Σs ln ψs(λ) → 0 and plim αT = α. Qin and Lawless. we can define an "entropy ratio statistic" which has a limiting χ2 distribution (see Kullback. A2. Let HM(λ*) be the entropy 0 value of the constrained problem where λ* = 0. Equation 3. These asymptotic properties can ˜ also be established via the empirical likelihood approach (Owen. Thus.1 / T and 1 / T respectively. 1992).12. 1959. we still have a solution.2 Hypothesis testing On the basis of the consistency of the estimators.1 (Mittelhammer and Cardell.4. after we have added the restrictions 2. In general. It can be 0 obtained by maximizing Equation (3. 1994. 1978. Thus. let λ* be the vector of Lagrange multiplies for all the model’s constraints. HM(λ*) is the maximum value of the joint entropies (objective function). We use this statistic to test hypotheses.11)subject to no constraints (except for the requirement that all distributions are proper). 1996).

Now. or any subset. uniform distributions α. we can test any other hypothesis of the form H0: α = α0 for all. represents additional potential information that may lower the value of the objective function but can never increase it). one can derive a "goodness of fit" measure for our estimator: .14) — where λ* is the set of estimated values (that is. The entropy-ratio statistic for testing the null hypothesis H0 that all parameters are zero is (parameters 0) 2H M (parameters 0) 2 H u (α . w . w. of the parameters.32 ˜ discrete. ˜ ˜ ˜ Under the mild assumptions we made above (or the assumptions of Owen. (parameters 2 0) → χK as T → ∞ when H0 is true and K is the number of restrictions. and ω. ω ) . or data point. they are not forced to equal zero). The approximate α-level confidence intervals for the estimates are obtained by setting 2 ( ) ≤ Cα . where Cα is chosen so that Pr(χK < Cα) = α. let Hu(λ*) be the objective (total entropy) value for the full GME-Nash model — the optimal value of Equation (3. 1990 and Qin and Lawless. Using the same line of reasoning as above (each constraint. We use these entropy-ratio statistics to test whether the economic and Nash restrictions are consistent with the data. 1994). Similarly.

taking into account all the data and all the restrictions (Equations 2. for the error terms of the entire system are estimated in the traditional way. 2. .6. 2. R*. Because our model is a system of a large number of equations.5. the relevant σ2 is estimated. where R* = 0 implies no informational value of the data set.3.6). This measure. the elements of the asymptotic variance-covariance matrix. for each set of equations. we calculate σi ˆ 2 1 T uit .33 ˜ H u (λ ) ˜ H M (λ 0) R 1 . The small-sample approximated variances can be computed in a number of ways. is the same as the information index in Soofi. We discuss two simplest approaches here. ˜ t 2 where ui t ≡ ˜ j u u i 2 ωt j vj and vˆ r (φk ) ≅ σi (X X ) ˜ a ˜ 1 i for each parameter φk. Similarly. and R* = 1 implies perfect certainty or perfect in-sample prediction. Ω. 1992. and A. for each equation (say the two sets of demand equations). First.

o where αo is a proper prior probability distribution. (k-1) . This relationship is used for comparison of various estimated strategies and various estimated distribution.34 Finally. o A second-order approximation of (A2. α o ) k αk ln (αk/αk ) .1) is I (α . we note the relationship between the entropy objective and the χ2 statistic. We conclude by noting that two times the entropy-ratio statistic corresponds (at the limit) to χ2 . Now. α o ) ≅ 1 2 1 k αk o (αk αk )2 . The cross-entropy measures is defined as (2.1) I (α . let {αk} be a set of K observed frequencies (strategies) over a set of K observed prices. Let the null hypothesis be H0: α = αo. o ˜ ˜ which is the entropy-ratio statistic (for evaluating α versus αo) that we previous discussed. then χ(k 2 1 1) k αk o (αk αk )2 .

202 25.428 77.354 -0.940 -90.190 -0.802 59.141 0.187 0.115 0.200 -2.275 -0.017 0.368 -0.040 -20.109 0.428 9.013 -0.909 -2.052 -0.168 -0.993 -19.526 1.788 11.420 0.196 0.851 -1.769 -0.534 -0.588 1.201 Price of Sugar 1.102 -61.002 -0.695 -0.203 1.860 -0.357 -0.500 -0.785 -0.441 -0.697 -1.207 0.110 -1.384 -0.393 49.332 65.646 2.955 -0.738 -43.266 -0.291 0.817 -5.465 -60.823 1.441 1.396 -0.098 -0.021 0.525 10.887 -13.339 Income 0.586 -0.090 -23.002 0.401 -0.512 -30.068 -59.096 Wage 49.840 1.671 0.070 -0.188 0.002 -0.263 11.846 7.262 0.274 0.834 0.107 -11.896 -1.756 2.295 -0.034 0.688 Dummy 1.542 -0.903 -2.841 1.077 -0.098 -0.287 -1.013 -0.041 0.832 0.128 -1.013 -1.098 -0.273 5.017 1.277 -0.699 33.023 Bond Rate -0.024 0.35 Table 1 GME Estimates of Coefficients for Coca-Cola Seasonal Price 1 1 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4 Advertising 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 Constant -46.281 .543 -0.017 0.822 -0.309 51.940 -12.719 21.

098 -0.013 -0.968 -0.428 1.36 4 5 5 5 5 5 6 6 6 6 6 7 7 7 7 7 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1.311 0.002 0.418 -23.002 0.017 -1.955 -0.977 1.002 0.017 0.013 -0.017 0.479 -0.013 -0.098 -0.002 0.002 0.452 1.098 -0.002 -0.002 0.017 0.013 -0.797 1.099 -0.453 1.013 -0.972 0. The coefficients in the missing first row are normalized to zero.017 0.098 -0.002 Notes: The first two columns show the price and advertising categories.017 0.938 23.013 -0.344 -0.098 0.013 -0.546 -0.099 -0.017 0.457 0.941 -0.017 1.099 -0.607 1.002 0.203 -0.977 -0.013 0.098 1.462 1.420 -29.952 -0.453 1.188 0.968 -0.098 -0.930 24.098 -0.017 0.013 -0.438 1.002 0.017 0.002 0.433 1.901 -0.930 -0.002 0.013 0.331 0.017 -0.017 -1.425 -22.543 -0.438 -0.418 1.037 0.017 0.932 29.099 -0.013 -0. .026 0.947 -0.449 -0.013 0.966 -0.441 1.014 -0.110 0.099 -0.379 -0.002 0.

37 Table 2: Percent of Categories Correctly Predicted GME Price Coke Pepsi 70 61 Advertising 68 33 Price 55 54 GME-Nash Advertising 63 32 .

000 .825 -.005 .262 -.000 .000 .204 -32.155 64.004 .077 .266 .000 -.000 75.003 -.767 .003 2.000 3.000 .000 .015 .737 .920 .635 .192 .000 .000 .000 .001 .38 Table 3: Strategy Elasticities Categories 1 Interest Rate Coke Price Pepsi Price Coke Advertising Pepsi Advertising Income Coke Price Pepsi Price Coke Advertising Pepsi Advertising .265 .000 -.000 .180 .000 -2.000 .000 .000 -.165 .000 .998 .001 15.000 -87.003 -1.000 2 3 4 5 6 7 .000 .000 -.000 10.756 .448 .

39 Figure 1a: GME Estimates of Coke’s Strategies (First Quarter 1977) .

40 Figure 1b: GME-Nash Estimates of Coke’s Strategies (First Quarter 1977) .

41 Figure 2a: GME Estimates of Pepsi’s Strategies (First Quarter 1977) .

42 Figure 2b: GME-Nash Estimates of Pepsi’s Strategies (First Quarter 1977) .

43 Figure 3: GME and GME-Nash Marginal Strategy Distributions for Coke (First Quarter 1977) a) Pricing Strategies b) Advertising Strategies .

44 Figure 4: GME and GME-Nash Marginal Strategy Distributions for Pepsi (First Quarter 1977) a) Pricing Strategies b) Advertising Strategies .