Bayesian binary regression using power and power reverse link functions: an application to premature birth data

This study aims to determine factors associated (and quantify) with prematurity through binary regression models, considering power and reverse power link functions, with asymmetric characteristics. As criteria for the model selection, the Bayesian Deviance Information Criterion (DIC), predictive evaluation, and residual analysis. All models analyzed presented similar predictive capacity, however, the model with a reverse power logit link function, with asymmetry parameter =0.336 was chosen, since it presented the lowest value of DIC=3,203, residues that indicated a good fitted of the model. There was an association of prematurity with the following variables: maternal - age over 35 years (OR=1.485), with a partner (OR=0.731), and primiparous (OR=1.307); of pregnancy and childbirth - multiple pregnancy (OR=36.360), cesarean childbirth (OR=1.337) and number of prenatal consultations less than seven (OR=3.305); and newborns of white race/skin (OR=0.731) and presence of congenital malformation (OR=2.663). Considering the proposed criteria, an asymmetric link function (reverse power logit) was the most parsimonious for the model. From this, there were high chances of factors associated with the occurrence of prematurity, indicating the need for actions to minimize them.