BAYESIAN LOGISTIC REGRESSION WITH ROBUST PREDICTORS FOR GESTATIONAL DIABETES PREDICTION

Abstract

The presence of outlying observations within the predictor space of the dataset for binary logistic regression can impact significantly the predictive performance of the developed classifier for public health problems. In this regard, this paper considers improving the predictive performance of the binary logistic classifier for Gestational Diabetes Mellitus prediction using alternative predictors termed robust predictors. These predictors are based on the computation of non-central moments of probability density functions of the original predictors. With this treatment, the relative importance of predictor-specific observations is easily assessed with outlying observations handled automatically and utilized for model fitting without being deleted. This way, the predictors become robust to extreme observations, and predictor-specific autocorrelations, allowing easy extension of binary logistic classifiers to public health problems for which outlying observations are inevitable. Appropriate Binary logistic regression models using the alternative predictors were developed within both the Classical and Bayesian paradigms with their utility illustrated with both simulated datasets and real gestational diabetes mellitus datasets, in comparison with existing current proposals.