Resampling Methods in Generalized Linear Models Useful in Environmetrics

Abstract

Generalized linear models are important tools for analysing relationships between binary, count or continuous response variables and predictors with fixed effects. In this paper we present a survey on bootstrap methods based on (extended) quasi-likelihood assumptions. We discuss two approaches: one-step residual resampling and score resampling to estimate the variability of functions in the linear parameters of the model, and an iterative procedure which allows us to define replicates of the dependent variate. With the latter we are able to estimate non-linear parameters in the variance function and to compare non-nested models. The power of these resampling schemes is illustrated by air sampler data concentrating on the number of bacteria colonies observed at outdoor sites in the area of Graz. © 1997 John Wiley & Sons, Ltd.