Comparing the logistic regression and different classification models

In this thesis we aimed to compare the performance of classification models by simulating data sets using 2 different models (with and without interaction terms) where sample size, prevalence, and coefficient of determination combinations changed. In addition, their performances were compared for 12 real data sets from the literature. In simulations without interaction terms, the performance of the NB method was higher than the other methods and comparable with the LR method. In simulations with interaction terms, the NB method performed better than the other methods at low sample size but SVM and RF methods performed better in medium and large data sets. We observed that DT and SVM methods were not able to make classifications (50%) in simulation settings with low prevalence and low coefficient of determination and small sample size. Real data set analysis showed that SVM and RF methods perform better than LR, DT and NB in some real data sets. Key words: Logistic Regression, Prevalence, Linearity, Decision Tree