Semiparametric Bayesian causal inference (Q2215769)

The inference concerning the causal effect of a treatment or condition is conducted in many applications. Hence, the notion of ``causal'' needs to be carefully defined and the statistical analysis must take into account possible explanations for the observed outcomes. In the paper under review the missing data model, which is mathematically equivalent to observing one arm of the causal setup (and also is a template for a number of other models), is considered. The authors follow the approach of nonparametrically modeling the propensity score function and posing the estimation of the treatment effect as a problem of estimation of a functional on a semiparametric model. They introduce a semiparametric Bayesian approach for estimating the mean response in a missing data model with binary outcomes and a nonparametrically modeled propensity score. Equivalently, the causal effect of a treatment (correcting nonparametrically for confounding) is estimated. Next, it is shown that standard Gaussian process priors satisfy a semiparametric Bernstein-von Mises theorem under smoothness conditions and a propensity score-dependent prior (that provides efficient inference under strictly weaker conditions) is proposed. Finally (in the Supplementary Material) the authors present a result when the covariate density (rather than the distribution) is modeled.