Model Selection and Model Averaging with Missing Data
Missing data can impair the reliability of statistical inference as they may affect the representativity of the sample. Nonetheless, under some conditions guaranteeing that the missing data mechanism is ignorable, reliable conclusions can be still drawn from the incomplete sample. Ignorability conditions are well-understood for parameter estimation but when the inference task involves the computation of the posterior probability of a data model, as required by Bayesian model selection and prediction through model averaging, these conditions are not sufficient. This paper defines new ignorability conditions for model selection and model averaging from incomplete data.
1. Knowledge Media Institute, The Open University.
2. Department of Actuarial Science and Statistics, City University.