Statistical Practice of Ordinal Outcome Analysis in Neurologic Trials: A Literature Review

Abstract

Background and Objectives Ordinal scales are widely adopted as outcome measures in neurologic randomized controlled trials (RCTs). We aimed to evaluate which statistical methods have been used to test and estimate treatment effects from ordinal outcomes in recent RCTs across a range of acute neurologic diseases. Methods We searched PubMed for RCTs in 5 acute monophasic neurologic diseases (stroke, traumatic brain injury [TBI], subarachnoid hemorrhage [SAH], meningitis, and Guillain-Barré syndrome [GBS]) published in high-impact journals between January 1, 2015, and November 1, 2023. Trials had to report on an ordinal scale as the primary or secondary efficacy outcome. Two independent reviewers performed study screening and data extraction. We evaluated the results to determine how investigators (1) addressed the ordinal nature of outcomes, (2) assessed and reported key assumptions, (3) used longitudinal measurements, and (4) adjusted for prognostic covariates. Results We reviewed 70 RCTs for treatment evaluations in stroke (n = 36), TBI (n = 13), SAH (n = 10), meningitis (n = 7), and GBS (n = 4). In 46 of 70 trials (66%), investigators retained the full ordering information, commonly analyzed using a proportional odds model (33/46 trials, 72%). The proportional odds assumption was not addressed in 23 of 33 trials (62%). In 22 of 70 trials (31%), the ordinal outcome was dichotomized, with notable variation in the cut-point for each ordinal scale. In 41 of 70 trials (59%), the ordinal outcome was assessed at multiple time points while some form of longitudinal data analysis was performed in only 3 of these 41 studies (7%). The time point chosen for analysis varied within neurologic conditions. Covariate adjustment was reported in 48 of 70 studies (69%). Discussion There is a large variation in the current practice of analyzing ordinal outcomes in neurologic trials. Dichotomization and focus on a single time point are common; therefore, information contained in the rank ordering of the outcome and repeated measurements is not fully used. Further research needs to clarify the balance between maximizing the statistical power, making assumptions, and allowing for straightforward interpretations in approaches that make more effective use of the outcome data.