Model uncertainty quantification in the presence of missing data

S. Cabras, M. E. Castellanos Nueda, A. Forte Deltell, G. Garcia-Donato, A. Quirós Carretero

Statistical procedures for dealing with model selection problems are always challenging but they can even be invalidated when the data are affected by missingness. Missing data is a pervasive problem in statistical analysis. Current methodologies such as list-wise deletion or single imputation can distort statistical inference and predictive accuracy. We present a comprehensive Bayesian approach for evaluating model uncertainty when dealing with missing data in general linear models. In particular, this approach includes the computation of marginal distributions that accommodate missing data, in order to obtain Bayes Factors and subsequent posterior probabilities for models. Our work delves into the issue of variable selection in linear models, with the innovation of g-priors extensions. Through theoretical arguments and simulation studies, we demonstrate that our proposal outperforms prevalent techniques like the list-wise deletion method, effectively handling missing data.

Keywords: Bayes Factors, Bayesian Statistics, g-priors, General Linear Models, Missing data, Model Uncertainty, Variable Selection

GT11.BAYES1 Variable Selection
November 7, 2023  4:50 PM
CC1: Audience

D. Rossell, P. Rognon, P. Zwiernick

V. Peña

G. Garcia-Donato, M. E. Castellanos Nueda