B0735
Title: Avoiding highly-informative ``nonparametric'' priors
Authors: Antonio Linero - University of Texas at Austin (United States) [presenting]
Abstract: The problem of specifying nonparametric priors when the goal is to estimate (conditional) average causal effects is considered. Ironically, the flexibility aimed for in specifying a nonparametric prior can sometimes accomplish the opposite of what we want. Rather than flexibly modeling causal effects, we may inadvertently (i) encode information that highly constrains them or (ii) encode unrealistically large amounts of heterogeneity. We illustrate the problem through simple examples using Gaussian processes and ridge regression and present solutions. The proposed corrections take the form of propensity score adjustments, giving a Bayesian nonparametric take on the wisdom of controlling for elements of the design in constructing estimates of causal effects.