Partitioning Variation In Multilevel Models For Count Data

A first step when fitting multilevel models to continuous responses is to explore the degree of clustering in the data. Researchers fit variance-component models and then report the proportion of variation in the response that is due to systematic differences between clusters. Equally they report the response correlation between units within a cluster. These statistics are popularly referred to as variance partition coefficients (VPCs) and intraclass correlation coefficients (ICCs). When fitting multilevel models to categorical (binary, ordinal, or nominal) and count responses, these statistics prove more challenging to calculate. For categorical response models, researchers appeal to their latent response formulations and report VPCs/ICCs in terms of latent continuous responses envisaged to underly the observed categorical responses. For standard count response models, however, there are no corresponding latent response formulations. More generally, there is a paucity of guidance on how to partition the variation. As a result, applied researchers are likely to avoid or inadequately report and discuss the substantive importance of clustering and cluster effects in their studies. A recent article drew attention to a little-known exact algebraic expression for the VPC/ICC for the special case of the two-level random-intercept Poisson model. In this article, we make a substantial new contribution. First, we derive exact VPC/ICC expressions for more flexible negative binomial models that allows for overdispersion, a phenomenon which often occurs in practice. Then we derive exact VPC/ICC expressions for three-level and randomcoefficient extensions to these models. We illustrate our work with an application to student absenteeism.Translational AbstractMultilevel models (random effects, mixed-effects or hierarchical linear models) are now a standard generalization of conventional regression models for analyzing clustered and longitudinal data in the social, psychological, behavioral, and medical sciences. A natural first step in any multilevel analysis is to report the degree of clustering in the response because it is the assumed presence of clustering which is the fundamental motivation for fitting multilevel models. Confirming that there is a statistically significant degree of clustering is not enough. One must additionally communicate the practical importance of clustering and this is done by reporting variance partition coefficients (VPCs) and intraclass correlation coefficients (ICCs). Different VPC and ICC expressions have been proposed and are now widely used for multilevel continuous, binary, ordinal, and nominal response models. In contrast, there has been almost no work for multilevel count response models. In this article, we propose new VPC and ICC expressions for both Poisson and negative binomial count response models. We then derive VPC/ICC expressions for three-level and random-coefficient extensions to these models. We illustrate all our work with an application to student absenteeism.