Estimating Discrete Games of Complete Information: Bringing Logit Back in the Game

Abstract

Discrete games are central tools for empirical analysis of strategic interaction, but equilibrium multiplicity and partial identification often make them computationally difficult to estimate. This paper develops tractable methods for estimation and inference in complete-information discrete games. The key idea is to construct an outer set by comparing observed frequencies of action profiles with singleton-class generalized likelihoods: model-implied probabilities that those profiles can arise as equilibria. The resulting conditional moment inequalities avoid computationally expensive equilibrium enumeration, numerical simulation, and grid search. Under standard empirical assumptions used in discrete-game models, including logit payoff shocks, these restrictions have closed-form expressions and are convex in a subvector of structural parameters. I develop the approach for both unordered and ordered action spaces. Monte Carlo experiments and empirical applications show that the methods deliver informative outer sets and can reduce computation time by several orders of magnitude relative to existing approaches.