Abstract:
|
Causal inference in observational studies often relies on estimating an unobserved treatment assignment mechanism and using estimates of treatment assignment probabilities in the analysis of outcomes. This paper assesses the performance of a Bayesian nonparametric framework utilizing Gaussian process priors for the estimation of treatment assignment probabilities as a function of covariates. Gaussian process priors are ideal for several reasons. First, by using a Gaussian process prior the analyst does not have to specify the functional form of the assignment mechanism, avoiding the possibility of model specification bias. Additionally, they can provide insight into the role of the covariates and, in particular, the areas of the covariate space where causal inference is possible. This talk provides simulation results that compare the performance of two different methods of obtaining approximations to the posterior distribution of the treatment assignment mechanism: MCMC sampling of the posterior distribution of the treatment assignment function (and underlying parameters) and an approximation to the posterior distribution derived using expectation propagation.
|