AbstractThe problem of when to include a covariate in the analysis of variance is considered in the special case of organ weight analysis in animal toxicology studies. The covariate is bodyweight prior to death, which may be subject to treatment effects. A simulation study is carried out to compare four rules for deciding whether or not to include the covariate. It is concluded that if there is background information which shows a linear relationship between variate and covariate it is advisable to adjust for the covariate, however weak the relationship may appear to be on the current set of data. Alternative procedures lead to unacceptably high Type I error rates.