Geographically Weighted Poisson Regression (GWPR) for Analyzing The Malnutrition Data in Java-Indonesia

Many regression models are used to provide some recommendations in private sectors or government public policy. Data are usually obtained from several districts which may varies from one to the others. Assuming there is no significant variation among local data, a single global model may provide appropriate recommendations for all districts. Unfortunately this is not common in Indonesia where regional disparities are very large. Geographically weighted regression (GWR) is an alternative approach to provide local specific recommendations. The paper compares between global model and local specific models of Poisson regression. The secondary data set used in this study is obtained from Podes (Village Potential Data) of 2008 in Java. Malnutrition as the outcome variable is the number of malnourished patients in a district. The parameter estimation in the local models used a weighting matrix accommodating the proximity among locations. Iterative Fisher scoring is used to solve the parameter estimation process. The corrected AIC shows that geographically weighted Poisson model produces better performance than the global model. Variables indicating poverty are the most influencing factors to the number of malnourished patients in a region followed by variables related to health, education, and food. The local parameter estimates based on the geographically weighted Poisson models can be used for specific recommendations.