|
Abstract:
|
We present a Bayesian nonparametric generative model for predicting medical costs in the presence of zero-inflation. Our work is motivated by the need for estimating differences in costs that would have accrued under different treatments. Even if the usual casual assumptions hold, estimation critically relies on accurate cost predictions under each intervention. These are difficult to obtain due to the complexity of cost data – characterized by zero-inflation, multi-modality, and skewness. Our approach models costs using an infinite mixture of zero-inflated regressions. Unlike finite mixtures, new clusters are dynamically introduced to accommodate the complexity of the data – yielding high-quality predictions. We incorporate our prediction model into a fully Bayesian standardization procedure – yielding posterior point and interval estimates of various causal contrasts such as differences, quantiles, and ratios. We present an MCMC posterior sampling procedure, a Monte Carlo standardization procedure, and a method for assessing overlap a posteriori. We end with an application to inpatient cost estimation among endometrial cancer using the SEER Medicare database.
|
