Perbandingan Performa Bandwidth CV, AICc, dan BIC pada Model Geographically Weighted Regression (Aplikasi pada Data Pengangguran di Pulau Jawa)

Carisa Putri Salsabila Purnamasari (Matematika, Universitas Indonesia, Depok, Indonesia)
Yekti Widyaningsih (Matematika, Universitas Indonesia, Depok, Indonesia)

Inferensi
Special Issue: Seminar Nasional Statistika XI 2022

Unemployment is a social phenomenon, a problem faced by every region in Indonesia. One way that can be carried out to reduce the unemployment rate is analyzing the factors that affect the open unemployment rate. Rather than using linear regression analysis, Geographically Weighted Regression (GWR) was preferable since it gave a better representative model by effectively resolve spatial heterogeneity problem which is generally exist in spatial data of social phenomenon. Spatial heterogeneity show that linear regression analysis will give a misleading interpretation results in some locations. GWR solve this problem by generating a single model in each observation location so the regression parameters can be different at each observation location. Parameter estimation in the GWR model uses weights based on the location of each observation so that the estimate model applies only to this location. The weighting determination depends on the bandwidth value. Bandwidth is a circle with radius ℎ from the center point of the observation location which is used as the basis for determining the weight of each observation location. Smaller bandwidth value will result a large variance. It can happen because when the bandwidth is very small, there will be a small number observations in the radius h, which can makes the estimate model is very rough (undersmoothing) because it uses few observations, and vice versa. Therefore, choosing the optimum bandwidth is very important in determining the weights where it can affect the accuracy of the model formed. This study aims to compare the performance of the GWR model using the Cross Validation (CV), Akaike Information Criterion Corrected (AICc), and Bayesian Information Criterion (BIC) bandwidth methods in the formation of Fixed Gaussian Kernel weighted function which is applied to unemployment data in districts/cities in Java. The results show that the GWR model with CV bandwidth is better at explaining district/city unemployment data on Java Island in 2020 which it has the smallest RMSE value, 1.0904, and the largest R2 and Adjusted-R2 values, namely 0.8539011 and 0.7937159, respectively.