Nonparametric Analysis of Covariance and its Role in Noninferiority Clinical Trials

Abstract

Some randomized clinical trials are conducted to compare test treatment, an active reference control, and placebo with the objectives of demonstrating that test treatment is not inferior to reference control, and to show that both test treatment and reference control are better than placebo. We typically want to adjust for covariables that are strongly associated with the response of interest in order to gain variance reduction, to adjust for random imbalances of covariables, and to clarify the degree to which differences between randomized groups are due to treatments rather than other factors associated with response.

Parametric modeling is often used to evaluate the relationship between covariables and the conditional distributions of response given the covariables. There can be concerns, however, about model assumptions, and they are not always straightforward to assess. An alternative is to use a nonparametric method for the primary evaluation of treatment comparisons. The nonparametric method is performed through linear models for (unconditional) differences between treatment groups for means of response criteria (or functions of such means) and covariables jointly with specifications that adjust random differences for means of covariables to zero. This paper discusses the role of nonparametric analysis of covariance in clinical trials to compare test treatment, an active reference control, and placebo with three examples.