Comparing Estimation Methods for Multilevel Spatial Models

• Activity Number: 167
• Type: Topic Contributed
• Date/Time: Tuesday, August 13, 2002 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Alastair Leyland*+ and Carolyn Davies
• Affiliation(s): Medical Research Council and Medical Research Council
• Keywords: multilevel models ; IGLS ; RIGLS ; maximum likelihood ; Poisson data

Abstract

This paper makes a comparison between three estimation methods for multilevel spatial models of Poisson data for disease mapping iterative generalized least squares (IGLS), restricted iterative generalized least squares (RIGLS), and maximum likelihood through numeric integration (ML)--all based in or derived from MLwiN. The data used are counts of all cancer deaths standardised for age and sex in 188 regions of 12 countries of the European Union in 1991. Covariates considered include socioeconomic variables (GDP), lifestyle variables (tobacco consumption), and diet (consumption of animal fats, fruit and vegetables). Some covariates are measured at the regional level and some at the national level. We have two specific areas of interest: Firstly, to what extent do the estimation methods differ in terms of the smoothed disease rates and rankings they produce? Secondly, we consider comparisons of the estimates of the degree of spatial smoothing and of the significance of covariates; in a multilevel model, this corresponds to testing the significance of random and fixed effects respectively.