Robust Designs for Binary Data

Abstract

This article is concerned with experimental design of statistical models for binary data. Since in most commonly used models for binary data, the information matrix depends on the unknown parameters, the standard optimality criteria used in regression cannot be used. What is typically done to avoid this problem is to assume good initial parameter estimates. The main purpose of this article is to introduce a minimax procedure for obtaining designs that are robust to poor initial parameter estimates. The procedure yields designs with more design points, and larger spread, if precise knowledge of the parameters is unavailable. D-optimality, and Fieller and confidence intervals for the median response dose are used to construct optimality functions for the procedure.