Regression models for binary response applied to data on neonatal deaths in newborns

Abstract

The present study presents binary data modeling regarding 1.6% of neonatal deaths in 3,448 newborns from an epidemiological and observational study with a cross-sectional design, involving the retrospective analysis of 4,293 medical records of high-risk pregnant women followed in a gestational outpatient clinic from September 2012 to September 2017. Different symmetric and asymmetric link functions were considered by means of Bayesian inference. The support of more accurate inferences regarding the parameters of the model will provide biological interpretations that are more reliable and consistent with the reality. The model that presented, significantly, the lowest value for the deviance information criterion (DIC = 398.8), was the binomial with power logit (PL) link function, whose median posterior value estimated and significant for the parameter asymmetry was l = 0.25 (0.14;1.17). This significance is observed in all other models of the power family, however with very different values ​​and significantly higher DIC values, indicating less parsimonious models. The Bayesian methodology proved to be flexible. Additionally, the results show that such model shows an accuracy = 97.4% and area under the ROC curve AUC = 89.4% in the prediction of neonatal deaths based on the weight of children at birth. Specifically, for 2.500g, a value predicted in the medical literature for low weight, the model predicts a probability of 1.43%.