Improvement of estimation based on small number of events per variable (EPV) using bootstrap logistics regression model

Abstract

In this research, a bootstrap approach model is proposed, namely as Bootstrap Logistics Regression Model (BLRM) that is specifically used to solve the small events per variable (EPV) problem. Considering a sample data from study case of endemic dengue at several localities in Kelantan, Malaysia, a simulation study is conducted.  We generated 5, 10, 20 and 25 mean samples with 500 times replacement, 1500 times bootstrap for each small EPV value (EPV= 2, 3, 4 and 5) according to the basic reproduction number, R0 for endemic dengue.  The performance of the propose BLRM revealed that the frequency distribution of estimated regression coefficient became less peaked and possessed thinner tails; the average percent relative bias consistently decreased and was closed to true parameter; the sample variance (MSE and RMSE) of the estimated regression coefficients of were smaller than original model.