A Practical Guide to Choosing Multiple-Comparison Tests for Efficacy Analysis in Small-Sample Patient-Derived Xenograft Studies

Patient-derived xenograft (PDX) studies are a cornerstone of preclinical oncology, yet the multiple comparison procedures (MCPs) used for tumor-volume endpoints are often selected by software default rather than by evidence from PDX-like study designs. Many classical MCPs were developed for large size, approximately normally distributed samples, whereas PDX studies typically include only 3-10 mice per arm. We benchmarked 15 all-pairwise MCPs, eight many-to-one MCPs, and a longitudinal linear mixed model (LMM) using 4,123 real PDX efficacy experiments. Type I error was evaluated using 3,000 simulation replicates per cell across a (3-7 arms) x (3-10 mice/arm) grid under normal and lognormal null distributions. Empirical power was summarized by Hedges' g effect-size bins and examined according to per-arm sample size, arm count, normality, and homogeneity of variance. Fisher's LSD, Duncan's test, and Student-Newman-Keuls produced inflated Type I error rates under the normal null. Among the calibrated all-pairwise tests, Tukey's HSD showed the highest empirical power, while Conover's all-pairs procedure was the strongest nonparametric alternative. As expected, many-to-one tests were more powerful than their all-pairwise counterparts. Raw tumor volume also provided higher power than log-transformed tumor volume, and the linear mixed model was the only approach able to detect drug effects when effect sizes were small. These findings support removing Fisher's LSD, Duncan's test, and SNK from PDX analysis plans, using Tukey's HSD for all-pairwise comparisons and Dunnett's test for vehicle-versus-treatment designs, both on raw tumor volume data, and considering LMMs when longitudinal data are available and model diagnostics support their use.