Improved resampling technique for the choice of stopping criterion and model selection in stepwise logistic regression

In recent years,logistic regression models are widely used in many research fields for establishing relation between a discrete outcome variable and predictor variable(s). various automated selection procedures are used for the selection of predictor variables that might influence the outcome variable. the significance level (x2 (α)) is the standard stopping criterion in these automated model selection methods in logistic regression. the problem with these automated model selection methods is the choice of appropriate stopping criterion for entry and removal of predictor variables. most of the statistical packages typically used the default significance value (α = 0.05) for an entry that may be unreasonable and sometimes even dangerous because it results either too many variables in the model for a reliable interpretation or too few variables for best prediction. besides different recommendations concerning the entry and removal criteria,there is still a problem for the true best choice of these values in automated selection methods in logistic regression models. we propose to resolve this problem by using cross-validation resampling technique that will optimize the stopping criterion each time for different data sets and different number of significant predictor variables. we further proposed the bootstrap resampling screening test for validating the final parsimonious logistic regression model. moreover,for moderate correlated predictor variables,our strategy had shown better results as compared to other model selection methods. © 2016 pakistan journal of statistics.