The World’s Largest Open Access Agricultural & Applied Economics Digital Library This document is discoverable and free to researchers across the globe due to the work of AgEcon Search. Help ensure our sustainability. Give to AgE con Search AgEcon Search http://ageconsearch.umn.edu [email protected] Papers downloaded from AgEcon Search may be used for non-commercial purposes and personal study only. No other use, including posting to another Internet site, is permitted without permission from the copyright owner (not AgEcon Search), or as allowed under the provisions of Fair Use, U.S. Copyright Act, Title 17 U.S.C. DEPARTMENT OF AGRICULTURAL AND RESOURCE ECONOMICS DIVISION OF AGRICULTURE AND NATURAL RESOURCES UNIVERSITY OF CALIFORNIA AT BERKELEY WORKING PAPER NO. 780 ESTIMATING A MIXED STRATEGY EMPLOYING MAXIMUM ENTROPY by Amos Golan Larry S. Karp and Jeffrey M. Perloff Copyright © 1996 by Amos Golan, Larry S. Karp, Jeffrey M. Perloff. All rights reserved. Readers may make verbatim copies of this document for non-commercial purposes by any means, provided that this copyright notice appears on all such copies. California Agricultural Experiment Station Giannini Foundation of Agricultural Economics February, 1996 Estimating a Mixed Strategy Employing Maximum Entropy Amos Golan Larry S. Karp Jeffrey M. Perloff Abstract Generalized maximum entropy may be used to estimate mixed strategies subject to restrictions from game theory. This method avoids distributional assumptions and is consistent and efficient. We use this method to estimate the mixed strategies of duopolistic airlines. KEYWORDS: Mixed strategies, noncooperative games, oligopoly, maximum entropy, airlines JEL: C13, C35, C72, L13, L93 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 [email protected] George Judge was involved in every stage of this paper and was a major con- tributor, but is too modest to agree to be a coauthor. He should be. We are very grateful to Jim Brander and Anming Zhang for generously providing us with the data used in this study. Table of Contents 1. INTRODUCTION 1 2. OLIGOPOLY GAME 3 2.1 Strategies 4 2.2 Econometric Implications 5 3. GENERALIZED-MAXIMUM-ENTROPY ESTIMATION APPROACH 8 3.1 Background: Classical Maximum Entropy Formulation 8 3.2 The Basic Generalized Maximum Entropy Formulation 10 3.4 Generalized Maximum Entropy Formulation of the Nash Model 13 3.5 Properties of the Estimators and Normalized Entropy 15 4. AIRLINES 16 4.1 The Airline Model Specification 17 4.2 Airline Estimates 18 4.3 Comparing Estimators 20 4.4 Sample Size Sensitivity Experiments 21 5. CONCLUSIONS 22 References 24 Appendix 1: GME-Nash with Unknown Demand Coefficients 27 Appendix 2: Consistency 30 Abstract Generalized maximum entropy may be used to estimate mixed strategies subject to restrictions from game theory. This method avoids distributional assumptions and is consistent and efficient. We use this method to estimate the mixed strategies of duopolistic airlines. KEYWORDS: Mixed strategies, noncooperative games, oligopoly, maximum entropy, airlines JEL: C13, C35, C72, L13, L93 1. INTRODUCTION We develop a method for estimating oligopoly strategies subject to restrictions implied by a game-theoretic model. Using this method, we estimate the pricing strategies of American and United Airlines. Unlike most previous empirical applications, 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 Our method allows firms to use either pure or mixed strategies consistent with game theory. First, we approximate a firm’s continuous action space (such as price, quantity, or advertising) with a discrete grid. Then, we estimate the vector of probabilities — the mixed or pure strategies — that a firm chooses an action within each possible interval in the grid. We use these estimated strategies to calculate the Lerner index of market structure. The main advantage of our method is that it can flexibly estimate firms’ strategies subject to restrictions implied by game theory. The restrictions we impose are consistent with a variety of assumptions regarding the information that firms have when making their decisions. Firms may use different pure or mixed strategies in each state of nature. Firms may have private or common knowledge about the state of nature, which is unobserved by the econometrician. For example, a firm may observe a random variable that affects its marginal profit and know the distribution (but not the realization) of the random variable that affects its rival’s marginal profit. Each firm may choose a pure strategy in every state of nature and regard its rival’s action as a 1 Breshnahan (1989) and Perloff (1992) survey conjectural variations and other structural and reduced-form "new empirical industrial organization" studies. 1 random variable. Alternatively, there may be no exogenous randomness, but the firm uses a mixed strategy. To the econometrician, who does not observe the firm’s information or state of nature, the distribution of actions looks like the outcome of a mixed strategy in either case. The econometrician is not ableto determine the true information structure of the game. Nevertheless, the equilibrium conditions for a variety of games have the same form, and by imposing these conditions we can estimate strategies that are consistent with theory. There have been few previous studies that estimated strategies based on a game-theoretic model. All of the studies of which we are aware (Bjorn and Vuong 1985, Bresnahan and Reiss 1991, and Kooreman 1994) involve discrete games. For example, Kooreman estimates mixed strategies in a game involving spouses’ joint labor market participation decisions using a maxi- mum likelihood (ML) technique. Our approach differs from his in three important ways. First, Kooreman assumes that there is no exogenous uncertainty. Second, he allows each agent a choice of only two possible actions. Third, because he uses a ML approach, Kooreman assumes a specific error distribution and likelihood function. Despite the limited number of actions, his ML estimation problem is complex. Our problem requires that we include a large number of possible actions so as to analyze oligopoly behavior and allow for mixed strategies. To do so using a ML approach would be ex- tremely difficult. Instead, we use a generalized-maximum-entropy (GME) estimator. An impor- tant advantage of our GME estimator is its computational simplicity. With it, we can estimate a model with a large number of possible actions while imposing inequality and equality restrictions implied by the equilibrium conditions of the game. In addition to this practical advantage,theGMEestimatordoesnotrequirestrong,arbitrarydistributionalassumptions,unlike 2 ML estimators. However, a special case of the GME estimator is identical to an ML estimator. In the next section, we present a game-theoretic model of firms’ behavior. In the third section, we describe a GME approach to estimating this game. The fourth section contains esti- mates of the strategies of United and American Airlines, and sampling experiments that illustrate the small sample properties of our GME estimator. In the final section, we discuss our results and possible extensions. 2. OLIGOPOLY GAME Our objective is to determine the strategies of oligopolistic firms using time-series data on prices, quantities, and, when available, variables that condition the cost or demand relations. We assume that two firms, i and j, play a static game in each period of the sample. (The generalization to several firms is straightforward.) Firm i (and possibly Firm j), but not the econometrician, observes the random variable e i(t) in period t. For notational simplicity, we suppress the time variable t. The set of K possible realizations, {e , e , ..., e }, is the same every period and for both firms. This assumption does 1 2 K not lead to a loss of generality because the distribution may be different for the two firms. The firms, but not the econometrician, know the distributions of e . We consider three possible k stochastic structures: (1) Firms face no exogenous randomness (K = 1); (2) e is private informa- k tion for Firm i; (3) e is common-knowledge for the firms. Because the econometrician does not k observe e , even if the firms use a pure strategy in each period, it appears to the econometrician k that they are using a mixed strategy whenever their actions vary over time. 3 2.1 Strategies The set of n possible actions for either firm is {x , x , ..., x }. The assumption that the 1 2 n action space is the same for both firms entails no loss of generality because the profit functions can be specified so that certain actions are never chosen. The notation xi means that Firm i s chooses action x. We now describe the problem where the random state of nature is private s information and then discuss alternative assumptions of a single state of nature or common information. In determining its own strategy, Firm i forms a prior, b i , about the probability that Firm sk j will pick action xj when i observes e i. If the firms’ private information is correlated, it is s k reasonable for Firm i to base its beliefs about j’s actions on e i. If the private information is k uncorrelated, Firm i form priors that are independent of e i. We do not, however, assume k independence. In state k, Firm i’s strategy is a = (a i , a i , ..., a i ), where a i is the probability k k1 k2 kn ks that Firm i chooses action xi. If Firm i uses a pure strategy, a i is one for a particular s and zero s ks otherwise. The profit of Firm i is p i = p i(xj, xi, e i), where r indexes the strategies of Firm j and s rsk r s k indexes the actions of Firm i. In state k, Firm i chooses a to maximize expected profits, S k r b i p , where the expectation is taken over the rival’s actions. If Yi is Firm i’s maximum rk rsk k expected profits when e i occurs, then Li ” S b i p - Yi is Firm i’s expected loss of using action k sk r rk rsk k xi in k. Because Yi is the maximum possible expected profit, the expected loss when Firm i uses s k action s must be nonpositive, (2.1) Li £ 0. sk 4 For a to be optimal, the product of the expected loss and the corresponding probability must k equal zero: (2.2) Li a i 0. sk sk Equation 2.2 says that there is a positive probability that Firm i will use action s only if the expected profits when action s is used are equal to the maximum expected profit. This problem may have more than one pure or mixed strategy. Our estimation method selects a particular pure or mixed strategy consistent with these restrictions and the data. 2.2 Econometric Implications Our objective is to estimate the firms’ strategies subject to the constraints implied by optimization, Equations 2.1 and 2.2. We cannot use these constraints directly, however, because they involve the unobserved random variables e i. By taking expectations, we eliminate these k unobserved variables and obtain usable restrictions. Using the expectations operator E , we define b i ” E b , Yi ” E Yi, a i ” E a i , p i ” E k r k rk k k s k sk rs k p i , and E Li ” Li. If we define q i ” Li - (S b ip i - Yi) and take expectations, then E q i = S rsk k sk s sk sk r r rs k sk r cov(b i , p i ) ” q i. Thus, Li ” E Li = S b ip i - Yi + q i. Taking expectations with respect to k rk rsk s s k sk r r rs s of Equation 2.1, we obtain (2.3) b i p i Yi q i £ 0. r rs s r Taking expectations with respect to k of Equation 2.2, we find that (cid:230) (cid:246) (2.4) (cid:231) b i p i Yi(cid:247) a i d i 0, r rs s s Ł ł r 5