Yannick Guyonvarch (PSAE) joint with Alexis Derumigny (TU Delft) and Lucas Girard (CREST-ENSAE)
Parameters identified through moment conditions are ubiquitous in Econometrics (OLS, linear IV, quantile regression to name a few). The construction of confidence sets (CSs) for such parameters typically relies on the asymptotic normality of the Generalized Method of Moments (GMM) estimator. These CSs are asymptotically exact but they have two major drawbacks: i) they are not guaranteed to contain the true parameter with the right probability in finite samples, no matter how large the sample, ii) they can behave poorly even asymptotically when the parameter is weakly identified (e.g: when instruments are weakly correlated with endogenous variables in the linear IV model). We propose a generic approach to construct CSs that addresses points i) and ii). We discuss in details the theoretical properties of our method, show that it is especially computationally tractable in linear models and illustrate its practical relevance through simulations. This presentation relies on findings from and extensions to recent works co-authored with Alexis Derumigny and Lucas Girard.