Statistical modeling with litter as a random effect in mixed models to manage "intralitter likeness".

"Intralitter likeness," the possibility that the shared genetics and/or maternal environment in multiparous species causes strong similarity for outcome variables in littermates, violates a core statistical assumption, that of observation independence, when littermate outcomes are analyzed. Intralitter likeness has been of major concern to investigators for several decades. Despite consensus and guidance, many research reports in the rodent literature continue to ignore intralitter likeness. A historical review of the literature revealed that the long-preferred solution was to include litter as an effect in statistical models. Limitations in software development and computing capacity prior to 1990, however, appear to have led researchers and guidance authorities to endorse instead the method of using one value per litter. Here, the history of discussions regarding intralitter likeness in developmental neurotoxicological research is reviewed; growing knowledge regarding the biological bases and significance of intralitter likeness is discussed; principles underlying the use of litter as a random effect in mixed models are presented; statistical examples are provided illustrating the advantages and critical importance of including litter as a random effect in mixed models; and results using all data points (all pups from all litters) with litter as a random effect, are compared to results based on random selections of representative littermates. Mixed models with litter included as a random effect have distinct advantages for the analysis of clustered data. Modern computing capacity provides ready accessibility to mixed models for all researchers. Accessibility however does not preclude the need for appropriate expertise and consultation in the use of mixed (hierarchical) models.