On Wilcoxon Comparisons for Multisite Glinical Trials

SUMMARY New rank-based methods for analyzing data from multisite clinical trials are presented ·in the context of "mixed" linear models. In contrast to current rank methods, the new procedures test for a drug main effect in the presence of a random drug by site interaction (or drug by investigator interaction when there is only one investigator per site). Analogous procedures are also provided for the "fixed effects" situation, and comparisons are made with current methods. The rationale for an analysis which assumes random investigator effects is described. 1 1. Introduction Clinical trials to compare a new drug with a standard often involve the participation of a number of investigators at different sites. For each investigator (or site), subjects are assigned randomly to the new drug or standard, so that the experimental design can be viewed as a randomized block design with each treatment (drug) replicated several times in each block. For simplicity we assume that there is only one investigator at each site so that a block may be described as either an investigator or a site. In assessing performance of a new drug, questions of interest relate to whether there is an overall difference between drugs across investigators (a drug main effect) and whether the difference is reasonably consistent (no substantial drug by investigator interaction). Using "linear model" terminology and standard practice, investigator effects should be regarded as "fixed" only if inferences about drug performance are to be limited to just those investigators in the study. The more likely scenario is that it is of interest to make inferences about drug performance for a larger population of clinicians. In that case the investigator effects would be viewed as random even though selection of participating investigators is not at random. This is because inferences based on investigators as the relevant level of replication are more likely to reflect potential performance in a broader population than are procedures which ignore the components of variation due to clinicians. To test for a drug main effect without ruling out a drug by investigator interaction, it is thus necessary to specify whether the first-mentioned "fixed effects" model or the latter "mixed" model is appropriate. In this article we consider rank-based alternatives to the analysis of variance (ANOYA) F tests for analyzing data from multisite clinical trials, with emphasis on the mixed model situation. Our work was prompted by the realization that currently popular methods …