A Bayesian Nonparametric Approach to Estimate Causal Effects of Mediation in the Presence of Nonignorable Missingness

• Activity Number: 164 - SPEED: Causal Inference and Related Methodology
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Dandan Xu* and Michael Daniels
• Companies: US Food and Drug Administration and University of Florida
• Keywords: Causal Inference; Mediation Analysis; Nonignorable Missingness; Bayesian Nonparametric

Abstract

We extend a Bayesian nonparametric (BNP) framework for estimating causal effects of mediation (Kim et al. 2016) to the situation when both mediator and outcome have nonignorable missingness. We specify a pattern mixture model for the full data and for each pattern (defined by the missing data indicators), we specify a Dirichlet process mixture of multivariate normals model on the joint distribution of the outcome, mediator, and covariates with sensitivity parameters introduced for extrapolating the missing data. The posterior of model parameters can be obtained as if the missing data is ignorable under this model specification. By making sequential ignorability assumptions (Imai et al., 2010) and incorporating (informative) priors on sensitivity parameters, causal parameters can be identified and estimated as functions of the posterior of model parameters and sensitivity parameters. We apply this approach to assess the mediation effect of body weight change on HbA1c change in a weight management trial.