EcoSta 2021: Start Registration
View Submission - EcoSta2021
A0438
Title: Adversarial Monte Carlo meta-learning in partially identified problems Authors:  Alex Luedtke - University of Washington (United States) [presenting]
Abstract: Traditionally, estimation in missing data and causal inference problems has been performed using the following three-step approach: (1) posit the existence of some unidentifiable full data distribution, (2) identify the quantity of interest as a feature of the observed data distribution, and (3) develop an estimator of this feature. We propose a new numerical approach for developing estimators in these problems that entirely circumvents the identification step (2). In our approach, missing data problems are framed as two-player games in which Nature adversarially selects a full data distribution that makes it difficult for the Statistician to answer the scientific question using a coarsening of data drawn from this distribution. The players' strategies are parameterized via neural networks, and optimal play is learned by modifying the network weights over many repetitions of the game. This approach performs favorably compared to standard practice in numerical experiments.