West Coast Ballroom
Comparative Studies of Bayesian Causal Inference with Gaussian Process Prior (307797)
Chen Chen, Cincinnati Children's Hospital Medical Center*Bin Huang, Cincinnati Children's Hospital Medical Center
Keywords: Bayesian Causal Inference, real-world data (RWD), Gaussian Process (GP) prior
The large real-world data (RWD), such as patient registry or the electronic health records, are increasingly used for comparative effectiveness research. However, they post great challenges, and demand for robust and accurate causal inference methods. Bayesian Gaussian Process (GP) regression provides a nonparametric approach. Various modeling strategies have been used.
We discuss and compare different strategies of constructing Bayesian causal inference with GP prior: 1) use estimated propensity score (PS) in the GP prior; 2) use estimated PS in the mean function; 3) using covariates in the GP prior with single length scale parameters; 4) using covariates in the GP prior with different length scale parameters; and 5) additive-interactive GP regression.
Extensive simulation studies are conducted for evaluating the performances of Bayesian GP for estimating averaged treatment effect, conditional treatment effect, with and without correct specification of the propensity score, under RWD like setting. For benchmark, PS matching, PS weighting, PS adjustment method, and oracle model are also compared. The case study considers a RWD from a large patient registry.