Litter effects: Comments on Golub and Sobin's 'Statistical modeling of litter as a random effect in mixed models to manage 'intralitter likeness''

摘要

The importance of litter effects (clustering of variance among offspring in rodents) has been known for decades. The standard approach was to treat the entire litter as a unit or to select one male and one female from each litter to prevent oversampling. These methods work but are imperfect. Treating the litter as a whole fails to use valuable interindividual differences among offspring, and selecting representative pups fails to use all the data available. Golub and Sobin [https://doi.org/10.1016/j.ntt.2019.106841] address this using a better method. They show that using litter as a random factor in mixed linear models resolves this conundrum. As they demonstrate, such models control for litter clustering by partitioning litter variance from error variance. This reduces error variance and increases the power of F-tests of the independent variable(s). In our experience, this is the optimal solution. But as good as mixed linear models are when used with litter as a random factor, if other aspects of the experimental design are not appropriate, this cannot compensate for threats to validity from small sample sizes, dams not strictly randomly assigned to groups, repeated measure covariance structures not appropriately modeled, interactions not properly sliced, or a posteriori group comparisons not controlled for multiple comparisons. Appropriate handling of litter is only one consideration of experimental design and statistical analysis that when used in combination lead to valid, reproducible data.