Analysis of an ordinal endpoint for use in evaluating treatments for severe influenza requiring hospitalization

Abstract

Background/Aims A single best endpoint for evaluating treatments of severe influenza requiring hospitalization has not been identified. A novel six-category ordinal endpoint of patient status is being used in a randomized controlled trial (FLU-Intravenous Immunoglobulin - FLU-IVIG) of intravenous immunoglobulin. We systematically examine four factors regarding the use of this ordinal endpoint that may affect power from fitting a proportional odds model: (1) deviations from the proportional odds assumption which result in the same overall treatment effect as specified in the FLU-IVIG protocol and which result in a diminished overall treatment effect, (2) deviations from the distribution of the placebo group assumed in the FLU-IVIG design, (3) the effect of patient misclassification among the six categories, and (4) the number of categories of the ordinal endpoint. We also consider interactions between the treatment effect (i.e. factor 1) and each other factor. Methods We conducted a Monte Carlo simulation study to assess the effect of each factor. To study factor 1, we developed an algorithm for deriving distributions of the ordinal endpoint in the two treatment groups that deviated from proportional odds while maintaining the same overall treatment effect. For factor 2, we considered placebo group distributions which were more or less skewed than the one specified in the FLU-IVIG protocol by adding or subtracting a constant from the cumulative log odds. To assess factor 3, we added misclassification between adjacent pairs of categories that depend on subjective patient/clinician assessments. For factor 4, we collapsed some categories into single categories. Results Deviations from proportional odds reduced power at most from 80% to 77% given the same overall treatment effect as specified in the FLU-IVIG protocol. Misclassification and collapsing categories can reduce power by over 40 and 10 percentage points, respectively, when they affect categories with many patients and a discernible treatment effect. But collapsing categories that contain no treatment effect can raise power by over 20 percentage points. Differences in the distribution of the placebo group can raise power by over 20 percentage points or reduce power by over 40 percentage points depending on how patients are shifted to portions of the ordinal endpoint with a large treatment effect. Conclusion Provided that the overall treatment effect is maintained, deviations from proportional odds marginally reduce power. However, deviations from proportional odds can modify the effect of misclassification, the number of categories, and the distribution of the placebo group on power. In general, adjacent pairs of categories with many patients should be kept separate to help ensure that power is maintained at the pre-specified level.