In randomized controlled trials, ordinal outcomes are commonly used as primary endpoints to measure the efficacy of interventions. Conventionally, the ordinal outcome has been analyzed on the original scale, by first creating a binary measure, or by treating the outcome as a continuous variable. It has been suggested that analyzing the data on the original ordinal scale provides the most powerful test for detecting a treatment effect. However, there are situations when dichotomizing the ordinal outcome yields higher power. In this thesis, we review the conventionally used statistical methods for analyzing ordinal outcomes, and apply these methods to simulated hypothetical trials.The simulated hypothetical trials are definedbased on different distributional assumptions for the control arm. To test for a treatment effect, we apply the cumulative logit model to the ordinal outcome and the Fisher;;s exact test to the dichotomized outcome. The power to detect a treatment effect is compared across these two methods for each control arm distribution and a variety of treatment effect sizes. The power to detect the treatment effect depends on the control arm distribution and the anticipated treatment effect.We found that dichotomizing the ordinal outcome can yield higher power to detect a treatment effect if the difference in the proportion of patients at a single level of the ordinal outcome is large and the remaining differences are spread over the remaining categories as opposed to shifted into a single category.
【 预 览 】
附件列表
Files
Size
Format
View
Statistical Considerations for the Analysis of Ordinal Outcomes in Randomized Controlled Trials
PDF
download
878987797
028-85220240
