Comparison performance between rare event weighted logistic regression and truncated regularized prior correction on modelling imbalanced welfare classification in Bali

Poverty is a problem that remains to be resolved and requires appropriate poverty reduction programs. Error classification in the implementation of poverty reduction programs such as inclusion error and exclusion error should be minimized as little as possible. Therefore, good methodology is needed in determining the proper classification of household welfare. Logistic regression is an analytical method that can be used to classify data. However, when used in cases of welfare in Indonesia where the data are imbalanced, the usual logistic regression method is not appropriately used. This is due to logistic regression tends to nullify the opportunities of minority groups because predicted values will tend to the category of data that the majority so that the resulting accuracy becomes less good. Rare Events Weighted Logistic Regression (RE-WLR) and Truncated Regularized Prior Correction (TR-PC) are the development of logistic regression used to overcome weaknesses in the case of imbalanced data. This research applies RE-WLR and TR-PC method using 14 research variables and the result is found 10 significant variables in RE-WLR and TR-PC model and 4 other variables are not significant in both models. Indicators used to measure the accuracy of model performance in predicting positive (poor household) and negative class (non-poor households in this study were AUC and G-mean values. This study showed that RE-WLR model was able to give better accuracy result than TR-PC model and logistic regression with G-mean value of 83.46 percent and AUC value is 0.835 (good category). This value is above the G-mean value of TR-PC model is 73.65 percent and the AUC value of model TR-PC is 0.761 (fair category).