Comparison of two measures of relative importance of predictors in logistic regression

This paper focuses on two frequently used methods for determining the relative importance of the predictors in explaining the response variable in the framework of logistic regression. The methods under consideration are relative weights (RWs) analysis and general dominance (GD) analysis, which are thought to correspond closely with each other. The unique contribution of this research lies in comparing the methods through an extensive simulation study, as they were only previously compared using an illustrative example. We employed the Bangladesh Demographic and Health Survey (BDHS) 2017–18 data set as a practical example, focusing on the dichotomous response variable of whether a mother in Bangladesh attends a sufficient number of antenatal care (ANC) visits. The real data example showed a higher degree of correspondence between GD and RWs. The absolute difference between the weights from two methods for all variables were negligible while wealth status and media exposure were the only variables where predictor rankings differ between the two methods. Furthermore, during the simulation phase, we obtained similar results by creating data 100 times under the same conditions. The average total R2 from GD analysis and RWs analysis across all iterations were 0.1330 and 0.1311, respectively. In the simulation, the weight ranges of the two methods were similar and overlap, reflecting what is observed in the real data set.