Identification, Estimation and Inference in Empirical Games

摘要

This thesis collects three papers studying topics related to the econometrics of empirical games. In Chapter 1, I investigate the identification and the estimation of empirical games of incomplete information with common-knowledge unobservable heterogeneity and potentially multiple equilibria realized in the data. I introduce pre-determined outcome variables to recover such unobserved heterogeneity. The recovered unobservables provide an extra source of exogenous variation that helps to identify the primitives of the model. I apply this method to study mobile telecommunications in Canada. I estimate a game in which national incumbents and new entrants choose the number of transceivers they install in different markets. The results highlight sizeable economies of density in transceivers location decisions. Counterfactual experiments shed light on the government's attempt to increase competition in this industry.;In Chapter 2, I propose a test of an assumption commonly maintained when estimating discrete games of incomplete information, i.e. the assumption of equilibrium uniqueness in the data generating process. The test I propose is robust to player-specific common-knowledge unobservables. The main identifying assumption is the existence of an observable variable interpreted as a proxy for these unobservables. It must (i) have sufficient variation; (ii) be correlated with the common-knowledge unobservables; and (iii) provide only redundant information regarding the players' decisions and the equilibrium selection, were these unobservables actually observed.;In Chapter 3, I study bias reduction when estimating dynamic discrete games. An iterative approach (the K-step estimator) is known to reduce finite sample bias, provided that some equilibrium stability conditions are satisfied. Modified versions of the K-step estimator have been proposed to deal with this stability issue. Alternatively, there exist other bias reduction techniques which do not rely on equilibrium's stability, but have not received much attention in this class of models. Using a dynamic game of market entry and exit, I compare the finite sample properties of the K-step approach with alternative methods. The results show that, even when the K-step estimator does not converge to a single point after a large number of iterations, it still considerably reduces finite sample bias for small values of K..