You are here

Causal Effect Estimation in The Presence of Unmeasured Confounders

Event Category:
Statistics and Data Science Seminar Series
Rohit Bhattacharya
Williams College

In this talk, I discuss recent developments in (i) semiparametric estimation of causal effects in graphical models with unmeasured confounders, and (ii) the design of semiparametric tests for verifying the key identifying assumptions in such models. In particular, I discuss doubly robust estimation strategies for a class of causal graphical models defined by a simple graphical criterion on the treatment variable (this class includes the popular conditionally ignorable model and front-door model as special cases.) I then discuss two newly proposed goodness-of-fit tests, which under mild assumptions, can be used to verify the key identifying assumptions in this class of models. These tests rely on variationally independent pieces of a natural parameterization of the observed data likelihood, and have the appealing property that they require no additional modeling than what is used in the downstream semiparametric estimators. That is, the same models used to perform the pre-test can be re-used for downstream causal effect estimation. I end with a short discussion on theoretical and empirical comparisons of this approach to instrumental variable approaches to handling unmeasured confounding.

Friday, September 30, 2022 - 11:00am
LGRT 171