Statistical tests for experimental transplant survival.

Abstract

Statistical analysis of experimental transplant survival data presents some difficulties. The data cannot be assumed to be normal, the variance differs between different groups, and in some groups a proportion of animals show indefinite survival. At the same time, groups are typically small. In view of these consideration, nonparametric tests such as the Mann-Whitney test (unpaired Wilcoxon) or its derivatives are often used. Analysis of clinical survival data is more often carried out using the log rank test, or proportional hazards models. These tests may give rise to different conclusions on the same data. Published data (1) were analyzed using the Mann-Whitney test (as performed by the original authors; censoring is treated as rejection. Markees et al. do not state how they deal with this.) and the log rank test; this dataset contains a number of groups larger than often reported. At the same time, the variances were compared by performing the Mann-Whitney test on transformed data obtained by taking the square of the deviation from the group mean; this test may be conservative, as censored data are ignored when present in both groups. The analysis was performed using Minitab with a macro written for the log rank test. The P-values are presented for three different comparisons of groups using all of these tests.Table Both the log rank and Mann-Whitney tests use ranks, but treat them in different ways; other rank-based tests exist (2). It is reported that Mann-Whitney-type tests are more powerful if the effect is expressed as a reduction in early hazard (3); log rank tests are sensitive to a proportional change in hazard. There is no general advantage to one or the other unless one has some knowledge of the likely effect on hazards; it might, for instance, be suggested that the log rank test would be more powerful for continuing effects such as tolerance. The test should be decided in advance and not chosen to suit the results. It should also be recognized that a change in variance is sufficient to cause us to reject the “null hypothesis,” even if it is difficult to interpret, particularly in the presence of censoring. Further progress in this matter, and more efficient analysis of experimental data, might come from more thorough statistical modeling of such experiments. Phil McShane1 Nuffield Department of Surgery; John Radcliffe Hospital; Headington, Oxford, OX3 9DU UK