Non-parametric analysis of covariance for confirmatory randomized clinicaltrials to evaluate dose-response relationships

In confirmatory randomized clinical trials that are designed to compare mul tiple doses of a test treatment with a control group and with one another, there are often statistical issues regarding compound hypotheses and multip le comparisons which need to be considered. In most cases the analysis plan needs a clear specification for the proposed order for conducting statisti cal tests (or for managing the overall significance level), which statistic al methods will be used, and whether adjustment for covariates will be perf ormed. There are several benefits of specifying non-parametric analysis of covariance (ANCOVA) for performing the primary confirmatory analyses. Only minimal assumptions are needed beyond randomization in the study design, wh ereas regression model based methods have assumptions about model fit for w hich departures may require modifications that are incompatible with a full y prespecified analysis plan. Non-parametric methods provide traditionally expected results of ANCOVA; namely, a typically small adjustment to the est imate for a treatment comparison (so as to account for random imbalance of covariates between treatment groups) and variance reduction for this estima te when covariates are strongly correlated with the response of interest. T he application of non-parametric ANCOVA is illustrated for two randomized c linical trials. The first has a (3 x 4) factorial response surface design f or the comparison of 12 treatments (that is, combinations of three doses of one drug and four doses of a second drug) for change in blood pressure; an d the second example addresses the comparison of three doses of test treatm ent and placebo for time-to-disease progression. This clinical trial has co mparisons among treatments made for a dichotomous criterion, Wilcoxon rank scores and averages of cumulative survival rates. In each example, the non- parametric covariance method provides variance reduction relative to its un adjusted counterpart. Copyright (C) 2001 John Wiley & Sons, Ltd.