Comparison of Nelder Mead and BFGS Algorithms on Geographically Weighted Multivariate Negative Binomial

Abstract

Geographically Weighted Negative Binomial Regression (GWNBR) was proposed related to univariate spatial count data with overdispersion using MLE via Newton Raphson algorithm. However, the Newton Raphson algorithm has the weakness, it tends to depend on the initial value. Therefore, it can have false convergence if the initial value is mistaken. In this research, we derive estimating the mean of dependent variables of multivariate spatial count data with overdispersion, Geographically Weighted Multivariate Negative Binomial (GWMNB) and compare it to the global method, multivariate negative binomial (MNB). We use MLE via Nelder Mead and Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithms. We conduct the simulation study and application of mortality data to find out the characteristics of the methods. They show that GWMNB performs better than global method (MNB) in estimating the means of dependent variables of the spatial data. The Nelder Mead tends to be more successful in estimating the means for all locations than BFGS algorithm. Although BFGS is a stable algorithm in MNB related to the initial value, it tends to have false convergence in GWMNB. The mortality rate of infant is larger than it of toddler and preschool and also maternal. The highest deaths of infant, toddler, and preschool and also maternal tend to happen in east parts of East Java.