The Comparisons of Four Splitting Rules for Fitting a Classification Tree with Simulation and an Application Related to Albuminuria Data in Type 2 Diabetes Mellitus

Abstract

The objective of this study was to compare the performances of splitting rules for predicting an ordinal response with simulation and a real data set. In the case of simulations, we compared across the methods using different sample sizes and the number of independent variables by employing the Monte Carlo simulation method. In the real data application, an analysis was performed with 265 cases. The results showed that the performances of the generalized Gini with the linear and quadratic costs of misclassification were better suited for analysis based on the gamma ordinal association measure and misclassification error rate than the other approaches. According to the gamma ordinal association measure, the generalized Gini (linear and quadratic) to the major risk factors determined for albuminuria in type 2 diabetes mellitus patients showed a slightly better performance than the other approaches. The predictive capability of splitting rules based on generalized Gini for predicting an ordinal response can be used for different sample sizes, number of independent variables and potential future suitable classification data problems. Consequently, our study will move towards choosing the generalized Gini (linear or quadratic) as the splitting rule and evaluate the data by using the Classification Trees (CT) in future studies, focusing on predicting an ordinal response.