A0825
Title: A generalized global optimal design for GLMs
Authors: Yiou Li - DePaul University (United States) [presenting]
Abstract: Optimal experimental design for generalized linear models (GLMs) is often sensitive to model specifications, including the choice of link function, predictors, and unknown parameters such as regression coefficients. To deal with the uncertainty inherent in these specifications, it is essential to construct designs that maintain high efficiency across a range of model specifications. Existing methods include Bayesian designs that incorporate prior distributions to handle such uncertainty and maximin designs that focus on optimizing the worst-case performance. The aim is to propose a unified framework of generalized global-optimal designs that includes Bayesian and maximin designs as special cases. Based on the theoretical properties of the proposed design criterion, an efficient algorithm is developed with a sound convergence property to construct optimal designs. Through a series of numerical examples, the robustness and superior performance of the proposed approach are demonstrated under various model uncertainties.
