A0394
Title: Optimal discrimination designs for a constrained type of random effects models
Authors: Victor Casero-Alonso - University of Castilla-La Mancha (Spain)
Jesus Lopez-Fidalgo - University of Navarra (Spain)
Sergio Pozuelo Campos - University of Castilla-La Mancha (Spain) [presenting]
Chiara Tommasi - University of Milan (Italy)
WengKee Wong - UCLA (United States)
Abstract: It is common in the optimal design literature to assume that the model has no random effects and it is known, apart from model parameters. Random effects models are widely applied across all disciplines, particularly in the life sciences and clinical studies. The assumption is that there are several plausible models, and the aim is to find a design that optimally discriminates among models with random effects. A common design criterion for discriminating between models with fixed effects is the Kullback-Leibler divergence criterion. It is a maximin-type of criterion, not differentiable and is a 2 or 3-layer nested optimization problem over very distinct domains. Consequently, optimal designs for discriminating models are notoriously difficult to determine, not only analytically but also computationally. Theoretical results are provided for the KL-optimality criterion value for discriminating among random effects models, and a nature-inspired metaheuristic algorithm is implemented to facilitate the search for an optimal discrimination design. The methodology is quite general and applies to discriminating random effects models with multiple interacting experimental conditions, which may be continuous or discrete. Two applications are provided; the first finds a design that optimally discriminates among fractional polynomials with a single continuous variable, and the second identifies the best design to discriminate among several multi-factor random effects models.
