Abstract
Interval estimates for the mean response from a generalised linear model are used to indicate uncertainty in the predictions from the model. The commonly used asymptotic Wald confidence intervals for the mean response are shown to perform poorly, with respect to coverage and precision, for small sample sizes which frequently occur from designed experiments.
Several alternative methods for constructing interval estimates for the mean response are considered including a modification of the Wald interval, Bayesian methods and bias-reduced logistic regression for binomial responses.