A Bayesian Nonparametric Approach for Causal Inference with Missing Covariates

• Activity Number: 247 - Causal Inference and Statistical Learning of Intervention and Policy Effects
• Type: Contributed
• Date/Time: Wednesday, August 11, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Huaiyu Zang* and Hang Kim and Bin Huang
• Companies: Cincinnati Children's Hospital Medical Center and University of Cincinnati and Cincinnati Children's Hospital Medical Center
• Keywords: Bayesian nonparametrics; Causal effect; Missing covariates; Multiple Imputation; Observational studies; Stick-breaking process

Abstract

Missing covariates in observational studies are common. Inappropriately handling missing covariates could result in inefficiency as well as potentially biased causal estimation. Besides missing data problems, the complexity of data structure makes the causal inference more difficult. In real data, the data distribution could be very complex; a standard parametric model lacks its flexibility. To address these problems, we introduce a Bayesian nonparametric causal model to estimate causal effect with missing covariates, that simultaneously imputes missing values and estimates causal effect under a potential outcome framework. We compare the performance of our method to the existing causal models applied to complete-case only and completed data by the sequential-chain imputation via repeated sampling simulations. Our simulations show that our method produces a more accurate average treatment effect estimate. The proposed method is also applied to Juvenile Idiopathic Arthritis (JIA) data, extracted from electronic medical records. The case study compares the effectiveness of early aggressive use of biological medication with a more conservative approach in treating children with JIA.