B0596
Title: The impact of positivity assumption on causal inference using Bayesian nonparametric methods
Authors: Jason Roy - Rutgers University (United States) [presenting]
Nandita Mitra - University of Pennsylvania (United States)
Yaqian Zhu - University of Pennslyvania (United States)
Abstract: In observational studies, differences between the treatment and control groups may be due to confounding variables. To assess the causal effects of a treatment in a population, an important identifiability condition is the positivity assumption (or `overlap'), which requires the probability of treatment to be bounded away from 0 and 1. That is, for every covariate combination, we should be able to observe both treatment and control subjects if the sample size is large enough. We discuss how different causal inference methods (parametric and Bayesian non-parametric) implicitly deal with non-overlap. We assess the performance of these approaches with respect to bias and efficiency in simulations.
