Comparative examination of outlier detection methods in binary logistics regression analysis

In this thesis, statistical based, proximity based, cluster based, deviation based and isolation based outlier detection methods were compared. The data set was simulated to have the standard normal distribution in R. The sample size of the data set was selected as 3000. The data set suitable for binary logistic regression with 3 independent and 1 dependent variable were produced. In order to compare the methods, the data was modified by adding 30 outliers to the data set. As a result of the study, the iForest algorithm from isolation based outlier detection methods has found all the added outliers and has the highest performance. In a real data set, outliers were determined using the iForest algorithm. In addition, the model was estimated in both artificial and real data sets for all observations and without outliers. The goodness of fit and adequacy of the model were examined. Thus, the effect of outliers on the model was determined.