Bayesian sensitivity analysis of incomplete data: bridging pattern-mixture and selection models.

Abstract

Pattern-mixture models (PMM) and selection models (SM) are alternative approaches for statistical analysis when faced with incomplete data and a nonignorable missing-data mechanism. Both models make empirically unverifiable assumptions and need additional constraints to identify the parameters. Here, we first introduce intuitive parameterizations to identify PMM for different types of outcome with distribution in the exponential family; then we translate these to their equivalent SM approach. This provides a unified framework for performing sensitivity analysis under either setting. These new parameterizations are transparent, easy-to-use, and provide dual interpretation from both the PMM and SM perspectives. A Bayesian approach is used to perform sensitivity analysis, deriving inferences using informative prior distributions on the sensitivity parameters. These models can be fitted using software that implements Gibbs sampling.