Inferences in non-stationary longitudinal binary models

Abstract

Longitudinal binary data has been analyzed over the last three decades either by using odds ratio or 'working' correlations as a measure of association between the repeated binary responses. Recently, this type of data has been analyzed by modeling the correlations parametrically and estimating the parameters by a generalized quasi-likelihood (GQL) approach. In this thesis, we consider a specific correlation model, namely, the binary autoregressive order 1 (AR(1)) model to generate the data, and study the relative performance of the odds ratio and equi-correlations based estimation approaches with the GQL approach. This comparison is mainly done by simulations under both stationary and non-stationary AR(1) correlation models. A real life data set containing repeated asthma status of a group of children is also analyzed.