RIDGE AND LASSO PERFORMANCE IN SPATIAL DATA WITH HETEROGENEITY AND MULTICOLLINEARITY

Abstract

Spatial heterogeneity becomes a separate issue on the analysis of spatial data. GWR (Geographically Weighted Regression) is a statistical technique to explore spatial nonstationarity by form the differrent regression models at different point in observation space. Multicollinearity is a condition that the independent variables in model have linear relationship. It would be a problem for estimation parameters process, because that condition produces unstable model. This problem may be found in GWR models, which allow the linear relationship between independent variables at each location called local multicollinearity. GWRR (Geographically Weighted Ridge Regression) and GWL (Geographically Weighted Lasso) which use the concept of ridge and lasso is shrink the regression coefficient in GWR model. GWRR and GWL techniques are consider to be capable of overcoming local multicollinearity to produce more stable models with lower variance. In this study, GWRR and GWL is used to model Gross Regional Domestic Product (GRDP) in Java using kernel exponential weighted function. The results showed that GWL has better performance to predict GRDP with lower RMSE and higher value than GWRR.
Keyword : Spatial Heterogeneity, GWR, Local Multicollinearity, Ridge, Lasso