Seleção de variáveis em modelos de regressão: uma avaliação do uso de redes de probabilidades condicionais

Abstract

The Bayesian network is a method presented by Judea Pearl in 1985 that describes a probabi- listic graphical model, which represents a set of variables and their conditional dependencies

with a directed acyclic graph. The vertices (or nodes) represent propositions (or variables), and directed edges (or arcs) signify the probabilistic dependencies between these variables. The objective of this study is to evaluate the use of Bayesian networks for the selection of variables

in regression models. This technique is compared with stepwise methods in simulation sce- narios that consider different sample sizes, correlations between the variables (responses and

variables) and different numbers of variables. In addition to the simulation study, we present a practical application of Bayesian networks in this context. For this purpose, data from a study

conducted between 2018 and 2019 involving veterinarians in Minas Gerais were used to iden- tify the most important risk factors associated with accidental exposure to antiviral vaccines,

specifically, vaccines for Brucella abortus (Brucellosis). One of the results of interest in this study was the prevalence of brucellosis among these professionals, which was estimated using a logistic regression model. According to the Bayesian network, the most important covariates

associated with accidental exposure to vaccines were knowledge about the symptoms of brucel- losis, whether the professional had performed premature childbirth procedures or abortions in

the previous six months and the frequency with which the professional used personal protective

equipment. All analyses were performed in R software using the bnlearn package. We recom- mend combining stepwise methods with Bayesian Networks, as stepwise methods are effective

for automatic variable selection, while Bayesian Networks excel at visualizing and understan- ding indirect associations between variables. This combined approach enriches the analysis,

providing a more comprehensive and detailed view of the results.