ABSTRACT
The appropriate statistical evaluation of reproduction/development studies, particularly multi-generation assays [200], is rather complex. The most important feature of reproductive assays is the correlation between the litter mates within a dam. The simpler teratogenicity study will be used here for demonstration purposes, where pregnant females are treated with the test compound in a randomized design including several dose groups and a negative control. At a certain stage of gestation the females are sacrified and several reproductive parameters such as number of pre-implantations, implantations, alive or dead pups, malformations, and pup weights are estimated for each pup within each female. Statistically two major problems exist: i) modeling the sub-unit litter mates within the randomized experimental unit female, so-called per-litter analysis (recommended by the ICH-guideline [200]) and, ii) modeling the possible competition between early loss/mortality and malformation of the pups. The first problem is split methodologically into a mixed model for the continuous endpoint pup weight and into a mixed model for correlated proportions for the incidence endpoints. The second problem is even more complex and will be discussed only briefly in Sections 5.3.2 and 5.4. Finally, female-specific endpoints occur also, such as number of corpora lutea. Their analysis can be easily performed by nonparametric Dunnett/Williams-type procedure (see Section 1.2.6)
Features:
• Per-litter analysis for continuous endpoints, such as pup weight
• Per-litter analysis for proportions, such as malformation rate
• Analysis of different-scaled multiple endpoints, such as pup weight and malformation rate
• Analysis of female-specific endpoints, such as corpora lutea
• Analysis of behavioral data
in
A dose-related weight reduction of the pups within the litters can be interpreted as a teratogenic effect. A Williams-type trend test in the mixed model with either adjusted p-values or simultaneous confidence intervals will be used. Instead of using the common linear model estimates, a mixed model containing both fixed effects (dose) and random effects (litter) is appropriate. It is the preferred approach for modeling correlated continuous data, such as repeated measures, technical replicates, paired organs, and within-litter dependencies [305, 391]. As a data example we use a textbook example [390] where in a randomized design 30 female rats were randomly assigned to receive control, low, or high dose of an experimental compound.
