ANALYZING OVER-DISPERSED COUNT DATA IN TWO-WAY CROSS-CLASSIFICATION PROBLEMS

Abstract


In recent years the generalized linear model has received increasing attention in the analysis of designed experiments, in part because the distributional form of the data can be specified rather than relying on robustness or asymptotic results. In the case of two-way cross-classification studies, count data is often encountered, especially in biological applications. If the data are assumed to follow Poisson distributions, the generalized linear model can be applied in a straightforward manner. However, what if the data are over-dispersed, i.e., with variance larger than the mean? The generalized linear model requires some sort of adjustment in this case. Several possible adjustments will be mentioned. Alternatively, one can use the traditional general linear model on the raw counts or on a transformation of the data. Results of a simulation study involving these alternatives will be presented, with general observations and preliminary recommendations.

---