A Bayesian semiparametric framework for causal inference in high-dimensional data
share
We introduce a Bayesian framework for estimating causal effects of binary and continuous treatments in high-dimensional data. The proposed framework extends to high-dimensional settings many of the existing semiparametric estimators introduced in the causal inference literature. Our approach has the following features: it 1) considers semiparametric estimators that reduce model dependence; 2) introduces flexible Bayesian priors for dimension reduction of the covariate space that accommodates non linearity; 3) provides posterior distributions of any causal estimator that can broadly be defined as a function of the treatment and outcome model (e.g. standard doubly robust estimator or the inverse probability weighted estimator); 4) provides posterior credible intervals with improved finite sample coverage compared to frequentist measures of uncertainty which rely on asymptotic properties. We show that the posterior contraction rate of the proposed doubly robust estimator is the product of the posterior contraction rates of the treatment and outcome models, allowing for faster posterior contraction. Via simulation we illustrate the ability of the proposed estimators to flexibly estimate causal effects in high-dimensions, and show that it performs well relative to existing approaches. Finally, we apply our proposed procedure to estimate the effect of continuous environmental exposures.
READ FULL TEXT