Analyzing multiple endpoints in a confirmatory randomized clinical trial—an approach that addresses stratification, missing values, baseline imbalance and multiplicity for strictly ordinal outcomes

Abstract

Confirmatory randomized clinical trials with a stratified design may have ordinal response outcomes, ie, either ordered categories or continuous determinations that are not compatible with an interval scale. Also, multiple endpoints are often collected when 1 single endpoint does not represent the overall efficacy of the treatment. In addition, random baseline imbalances and missing values can add another layer of difficulty in the analysis plan. Therefore, the development of an approach that provides a consolidated strategy to all issues collectively is essential. For a real case example that is from a clinical trial comparing a test treatment and a control for the pain management for patients with osteoarthritis, which has all aforementioned issues, multivariate Mann-Whitney estimators with stratification adjustment are applicable to the strictly ordinal responses with stratified design. Randomization based nonparametric analysis of covariance is applied to account for the possible baseline imbalances. Several approaches that handle missing values are provided. A global test followed by a closed testing procedure controls the family wise error rate in the strong sense for the analysis of multiple endpoints. Four outcomes indicating joint pain, stiffness, and functional status were analyzed collectively and also individually through the procedures. Treatment efficacy was observed in the combined endpoint as well as in the individual endpoints. The proposed approach is effective in addressing the aforementioned problems simultaneously and straightforward to implement.

Clinical Study Design to Assess Both Short- and Long-term Efficacy in Addition to Group Sequential Test on Safety

“…Please note here the difference of the two means is tested. One may also consider testing the probability of better outcome for one treatment than the other, 12 or fitting a logistic regression to compare the binary variable of the SHS change being less than or equal to 0. el , H el , and H s at the interim and final, respectively. Table 1 shows possible sets of nominal significance levels at the interim and final safety tests.…”

“…Please note here the difference of the two means is tested. One may also consider testing the probability of better outcome for one treatment than the other, 12 or fitting a logistic regression to compare the binary variable of the SHS change being less than or equal to 0.…”