Adversarial Inference Is Efficient

Abstract

We study properties of the adversarial framework, introduced in Kaji, Manresa and Pouliot (2020). We show that the adversarial inference with an oracle classifier is statistically efficient. In addition, we study the finite sample properties of the adversarial estimation framework for the autoregressive parameter of a linear dynamic fixed effects panel data model with Gaussian errors. Unlike maximum likelihood, but similarly as other minimum distance estimators, the adversarial estimators do not suffer from the incidental parameter bias. In our simulations, using a one-hidden-layer neural network as discriminator delivers the estimates with smallest root mean squared error.