Online Program

An all-in-one Bayesian nonparametric model for medical cost prediction, clustering, and causal estimation (306483)

Keywords: costs, Bayesian, causal inference, prediction, clustering, nonparametric, cancer

Health policy researchers are often interested in both medical cost prediction and identifying clusters of patients with different cost-covariate profiles. In terms of policy evaluation, they are also interested in causality: estimating the difference in accrued costs under competing interventions. These have historically been treated as separate tasks, each complicated by zero-inflation, skewness, and multi-modality. We present a Bayesian nonparametric generative model that accounts for these complications while accomplishing all three tasks simultaneously. Costs are modeled with an infinite mixture of zero-inflated regressions using a Dirichlet process prior. This prior also induces a partition of the cost data into more homogeneous clusters. Posterior cost predictions can be computed under hypothetical interventions to form nonparametric estimates of various causal contrasts. The model also produces posterior propensity scores which can be used to evaluate positivity. We present relevant posterior quantities, simulation results, and an application to endometrial cancer cost data in the SEER Medicare database.