Modeling of Welfare Indicators in Java Island using Biresponses MARS

Abstract

Poverty and Human Development Index (HDI) are indicators that usually used by the government to measure the level of prosperity in Indonesia. Some literatures state that there are several predictor variables which could affect the both indicators, where the effect of each predictor can be modeled by statistics method. In addition, according to Statistics Indonesia, there was a significant correlation between the poverty rate and HDI, so the modeling of poverty and HDI can also be done by accomodating their correlation. This study aims to model welfare indicators (the poverty rate and HDI) in Java island in 2012 using biresponses MARS. In general, biresponses MARS is generalization of original MARS (uniresponse), which not only models the effect of predictors but also accomodates the effect of correlated responses. We use MARS in this study, because based on relationship pattern in variables used, indicates for using nonparametric regression. Beside that, MARS has many advantages compared to other methods in nonparametric regression, including it can solve high-dimensionality case and accomodates interaction in its estimation, so it can increase the accuracy of the model. The results show that poverty and HDI can be influenced by some variables such as the percentage per capita expenditure for food, economic growth and unemployment, however the influence from those variables are not the same for each area in Java island in 2012. Based on goodness of fit criterion, the result shows that biresponses MARS gives better model than uniresponse MARS on modeling welfare indicators in Java island in 2012. The reason could affect the result is biresponses MARS accomodates the effect of correlated responses whereas uniresponse MARS ignore it.