On comparisson between ordinary linear regression and geographically weighted regression: with application to indonesian poverty data

Abstract

Ordinary linear regression (OLR) is one of popular techniques in analyzing relationship between response variable and its predictors. It is an analysis that produces global models applied to all observations assuming no correlation among responses. In social studies, such as poverty analysis, response variable might be spatial nonstationarity, i.e. depends on the region or neighborhood. Therefore, of course, OLR model will not comply the assumption of independence. In dealing with the problem, an OLR model has to be calibrated by accommodating spatial variation. Alternatively, geographically weighted regression (GWR) involves geographical weights in estimating the parameters. GWR yields models in each region uniquely, i.e. local model, by setting the weights as a function of distance. The weights are greater as the distance is closer, and then continuously decrease to zero as the distance is farther. This paper shows an application of GWR in poverty analysis in Java, Indonesia. Performance of GWR and OLR model in describing poverty is compared. The results show that GWR has better performance than OLR does based on residuals, R2, AIC statistics and some formal tests.