A Parametric Bayesian Approach to Estimating Causal Treatment Effect on Medical Costs

• Activity Number: 69 - Modern Statistical Methods for Multi-Scale and Time Series Data
• Type: Contributed
• Date/Time: Sunday, July 30, 2017 : 4:00 PM to 5:50 PM
• Sponsor: International Indian Statistical Association
• Author(s): Arman Oganisian* and Andrew Spieker and Jason A Roy and Nandita Mitra
• Companies: University of Pennsylvania and University of Pennsylvania and University of Pennsylvania and University of Pennslyvania
• Keywords: healthcare costs ; causal inference ; health policy ; g-computation ; time-varying confounding

Abstract

Estimates of cumulative medical costs can be useful in many settings, including cost-effectiveness analyses and informing public policy. However, time-varying treatment and informative censoring of total costs can impede accurate estimation. Semi-parametric frequentist methods have been developed to account for informative censoring. To date, however, there is a lack of available approaches to account for time-varying treatment and confounding, which are often encountered in practice. We address these challenges by developing a parametric Bayesian approach. Our method partitions the total time range of interest into discrete intervals and utilizes a Bayesian g-formula to estimate differences in cumulative costs between two treatment regimes. Under the assumption of conditional independence of censoring and cost within each partition, this approach can account for time-varying treatment while also addressing informative censoring. Based on simulation results with weakly-informative priors, we obtain low-bias estimates of model parameters under correct model specification. Performance was assessed under a variety of scenarios and compared with existing methods.