Data Panel Regression Implementation for Predicting Divorce in Regencies and Cities of West Java Province

Abstract

Family is an important component in a community. Divorce, as a major indicator of a family, is used as a metric to gauge family health in a community. A Prediction of the total number of divorces for regency/city in West Java Province was developed. Three models are used for prediction and compared to each other through MAPE, MAE, and RMSE tests, with an additional classic assumption test and model selection test for additional guidance. The models are the common effects model, fixed effects model, and random effects model. The first iteration of prediction has a best MAPE of 33.25%, achieved by the common effects model, and it violated nearly all classic assumption tests. A log transformation for the features was conducted. In this iteration of prediction, the random effects model has the best MAPE with 19.95% and no major assumption test violations. It has an RMSE of 1550.66 and MAE of 1002.35. In addition, during this study, an entry has a completely missing divorce value for the entire training dataset. An attempt at zero-shot was conducted; the first iteration had 73.5% MAPE, and the second iteration had a MAPE of 5.3%, a significant improvement from log transformation.  For further study, an analysis could be made to determine what the common denominator is between city/regency that has low MAPE. Through this analysis, an adjustment could be made for the model to improve predictive ability. Additionally, new data could be introduced to the model.