Comparison of analytical techniques for multiple endpoints in clinical trials

Background: Multiple study endpoints often present such challenges as skewed distributions, high correlation, random missingness, null treatment effects in some of the endpoints, and, in the case of studies of traumatic head injury, high numbers of deaths.

Purpose: This paper discusses strategies for analyzing data with multiple endpoints for a positive treatment effect when comparing the treatment group to the placebo group.

Methods: We explored three kinds of multiple endpoint strategies: testing endpoints individually using a Bonferroni correction (Chi-Square test, Mann-Whitney U test, Student's t-test), creating a composite outcome measure from the multiple endpoints using O'Brien's rank sum method, and incorporating all the intact endpoints into one statistical model (logistic regression, ordinal logistic regression, or linear regression) by using GEE to account for correlation among the endpoints. We used simulation to assess the power of each of these methods under various scenarios of correlation, missingness, null treatment effects, and death rates.

Results: O'Brien's rank sum composite endpoint was the method most robust to high numbers of outcome measures with high correlation, missingness, and a null effect in one or two out of four measures. Logistic regression with GEE was the method most robust to high rates of death among study subjects.

Conclusions: When the scientific question is regarding a treatment effect across all measures, then O'Brien's rank sum method is robust to the most situations. When interpretability of the results regarding a treatment effect across all measures is critical, then logistic regression with GEE may be the technique of choice.