Role of the Proportional Odds Assumption for the Analysis of Ordinal Outcomes in Neurologic Trials

Ordinal scales are categories ordered by their clinical preference, such as the modified Rankin Scale and Glasgow Outcome Scale Extended. They are widely used as key outcome measures in neurologic trials. Compared with binary outcomes, ordinal scales allow a more detailed assessment of the effect of the treatment and provide more statistical power. Typically, the proportional odds (PO) model (or "shift analysis") is used to quantify the treatment effect in a common odds ratio (cOR) and to test the null hypothesis that the treatment has no effect. Clinical researchers and trialists may worry that the PO assumption will not hold and therefore decide to dichotomize the ordinal scale into a binary outcome. Here, we explain that for the purpose of testing the presence of a treatment effect, it is irrelevant whether the PO assumption holds. In fact, pretesting this assumption to decide which test to use for the treatment effect is invalid and can inflate the type I error probability. Although the PO assumption is not relevant for testing, it is definitely relevant for the purpose of describing and summarizing the treatment effect. We suggest a simple graphical check of the PO assumption as more informative than formal testing. If we are satisfied that there is no substantial violation of the PO assumption, it is reasonable to summarize the treatment effect into a single number such as the cOR. Otherwise, no single-number summary measure provides a faithful representation. We illustrate these practical considerations with 3 neurologic trials: the Endovascular Therapy in Acute Anterior Circulation Large Vessel Occlusive Patients with a Large Infarct Core (ANGEL-ASPECT) trial and the Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands ([MR CLEAN], which investigated endovascular therapy in stroke) and the RESCUEicp trial (which investigated decompressive craniectomy in traumatic brain injury). We conclude with recommendations for the statistical workflow. Undue concerns about the PO assumption should not deter researchers from using ordinal scales in neurologic trials. We provide an R package <i>CORPlot</i> to facilitate the graphical check of the PO assumption.