Borrow, don't breed: How historical controls can cut concurrent animal numbers.

Recommendation 013/2026 of the German Federal Institute for Risk Assessment (BfR) considers a reduction of concurrent control animals possible when historical control data demonstrate stability. However, the recommendation provides no method for quantifying stability or determining the magnitude of reduction. We present a framework that addresses this gap for designs whose control group size follows from a power calculation on pre-specified continuous endpoints, each assessed separately. Designs whose group size is prescribed by a test guideline, as in standard repeated-dose toxicity studies, fall outside its scope. The framework offers two approaches scaled to the available expertise. Both begin with a shared planning step that estimates how much the control-group mean of an endpoint varies across studies of the same design and assigns that endpoint, not the laboratory, to one of three stability zones (stable, moderate, or unstable). In the stable zone, a frequentist approach maintains 80% power while accepting a modestly larger minimum detectable effect and requires no Bayesian expertise. A full Bayesian approach achieves a greater reduction at constant power and effect size, provided concurrent and historical controls agree, but requires statistical expertise. We demonstrate, using 89 artificial historical control animals across 8 studies, that the frequentist approach saves 12 of 37 control animals per study and the Bayesian approach up to 32, compared with zero under the status quo. Naive pooling, the intuitive but incorrect alternative, can inflate the false positive rate from 5% to over 30%. A freely available planning tool (C*Power) supports the planning step.