Sample size estimation for pilot animal experiments by using a Markov Chain Monte Carlo approach.

Abstract

The statistical determination of sample size is mandatory when planning animal experiments, but it is usually difficult to implement appropriately. The main reason is that prior information is hardly ever available, so the assumptions made cannot be verified reliably. This is especially true for pilot experiments. Statistical simulation might help in these situations. We used a Markov Chain Monte Carlo (MCMC) approach to verify the pragmatic assumptions made on different distribution parameters used for power and sample size calculations in animal experiments. Binomial and normal distributions, which are the most frequent distributions in practice, were simulated for categorical and continuous endpoints, respectively. The simulations showed that the common practice of using five or six animals per group for continuous endpoints is reasonable. Even in the case of small effect sizes, the statistical power would be sufficiently large (≥ 80%). For categorical outcomes, group sizes should never be under eight animals, otherwise a sufficient statistical power cannot be guaranteed. This applies even in the case of large effects. The MCMC approach demonstrated to be a useful method for calculating sample size in animal studies that lack prior data. Of course, the simulation results particularly depend on the assumptions made with regard to the distributional properties and effects to be detected, but the same also holds in situations where prior data are available. MCMC is therefore a promising approach toward the more informed planning of pilot research experiments involving the use of animals.