Optimal block designs for experiments with responses drawn from a Poisson distribution

TL;DR

This paper develops a new method for designing experiments with responses following a Poisson distribution, improving efficiency over traditional additive model-based designs by using Poisson GLMMs and simulated annealing.

Contribution

It introduces a novel approach for creating optimal block designs for Poisson-distributed responses using D_A- and C-optimality criteria within a Poisson GLMM framework.

Findings

Poisson GLMM-based designs outperform classical designs for significant treatment effects.

Treatment replication is inversely related to expected counts in optimal designs.

Efficient, matrix-inversion-free computation of objective functions enables practical design optimization.

Abstract

Optimal block designs for additive models achieve their efficiency by dividing experimental units among relatively homogenous blocks and allocating treatments equally to blocks. Responses in many modern experiments, however, are drawn from distributions such as the one- and two-parameter exponential families, e.g., RNA sequence counts from a negative binomial distribution. These violate additivity. Yet, designs generated by assuming additivity continue to be used, because better approaches are not available, and because the issues are not widely recognised. We solve this problem for single-factor experiments in which treatments, taking categorical values only, are arranged in blocks and responses drawn from a Poisson distribution. We derive expressions for two objective functions, based on D_A- and C-optimality, with efficient estimation of linear contrasts of the fixed effects parameters in a Poisson generalised linear mixed model (GLMM) being the objective. These objective functions are shown to be computational efficient, requiring no matrix inversion. Using simulated annealing to generate Poisson GLMM-based locally optimal designs, we show that the replication numbers of treatments in these designs are inversely proportional to the relative magnitudes of the treatments' expected counts. Importantly, for non-negligible treatment effect sizes, Poisson GLMM-based optimal designs may be substantially more efficient than their classically optimal counterparts.