Superiority of Bayesian Model Averaging to Stepwise Model in Selection of Factors Related to the Incidence of Type II diabetes in Pre-diabetic Women

 The world prevalence of type 2 diabetes and its related increment mortality rate which needs high controls cost has attracted high scientific attention. Early detection of individuals who face this disease more than the others can prevent getting sick or at least reduce the disease consequences on public health. Regarding the costs and limitations of diagnostic tests, a statistical model is presented that helps predict the time of diabetes incidence and determines its risk factors. Furthermore, this model determines the significant predictor variables on response and considers them as model equation parameters. In this study, 803 pre-diabetic women in the age range of more than 20 years were selected from Tehran lipid and glucose study (TLGS) to examine the predictor variables on time of diabetes incidence. They were entered into the study in the phases 1 and 2 and were followed up to the phase 4. The predictor variables selection was performed using the Stepwise Model (SM) and the Bayesian Model Averaging (BMA). Then, the predictive discrimination was used to compare the results of both models. The Log-rank test was performed and the Kaplan-Meier Curve was plotted. The statistical analyses were performed using R software (version 3.1.3). The Backward Stepwise Model (BSM), the Forward Stepwise Model (FSM) and the BMA have used 9, 10 and 6 variables, respectively. Although the BMA selected predictor variables number is much lower than the SM, the prediction ability remains nearly constant. The BMA has averaged on the supported models using dataset. This model has shown nearly constant accuracy despite the selection of lower predictor variables number in comparison to the SM.