G. Garcia-Donato, M. E. Castellanos, A. Quirós Carretero, S. Cabras, A. Forte
We consider the problem of model uncertainty when the working database contains missing (NA) entries. This is a very common situation in applied statistics and is normally handled casually either removing all rows with unobserved cells or resorting to imputation mechanisms. The literature on the topic is very scarce and has not been properly documented which are the implications of these standard practices and if there is any better alternative. We try to shed some light on these questions.
We approach the ensuing problem revisiting the fundaments of Bayes factors, focusing on the situation with missing values in explanatory variables in regression models. Handling missing values from a formal perspective compels us to use an imputation model that determines how the prior marginals have to be computed. The standard priors (g-priors) are no longer valid due to its dependence on the data and novel possibilities are explored.
Keywords: Bayes factor; Incomplete Information; g-priors
Scheduled
GT21 Bayesian Inference II
June 7, 2022 3:30 PM
Audiovisual room