An exponential tilt approach to generalized linear models

This thesis consists of two largely self-contained papers concerning the joint estimation of parameters in the mean model and the error distribution in the generalized linear model framework from an exponential tilt regression approach. We show how consistent estimators for both parameters can be constructed via an empirical likelihood approach and establish a joint central limit theorem for the estimators. We also show how an empirical profile likelihood can be used for hypothesis testing and confidence intervals for parameters in the mean model, much in the same way as analysis-of-deviance techniques in classical generalized linear models. Small sample performance are examined via simulations and the method will be applied to some real data analysis examples.