Predictive distributions in binary models with missing data

Abstract

We assess the performance of several methods to deal with an incomplete data set, when interest lies in the evaluation of the predictive probability for a future individual in a binary response model.

Results are presented from simulation studies for both the sampling and diagnostic approaches within the Bayesian paradigm (Dawid, 1976), with normal and logistic models being used. The process that causes missing values in the continuous covariate available to predict the binary response is assumed to be the missing-completely-at-random mechanism (Rubin, 1976).

Keywords:

• binary response
• diagnostic approach
• imputation
• incomplete data
• predictive probability
• sampling approach