Abstract TP43: Evaluating the Generalizability of the Common Odds Ratio: A Demonstration using Two Trials of Endovascular Therapy

Abstract

Introduction: The modified Rankin Scale (mRS) is an ordinal outcome, ranging from 0 (no symptoms) to 6 (death). Recent literature has demonstrated a preference for statistical analysis which maintains the ordinal nature of the scale. The proportional odds (PO) model assumes a single common odds ratio (OR) for all order-preserving dichotomizations of the outcome, but the validity of this assumption cannot be known in advance. The aim of this study is to determine if the common OR was a close approximation of cutpoint-specific treatment effects calculated using the more general cumulative logit model in two endovascular treatment trials. Methods: DEFUSE 3 and IMS III, both trials of endovascular therapy (EVT), are used as examples. Ordinal logistic regression was applied via the cumulative logit model with and without the PO assumption. Effect estimates and corresponding 95% confidence intervals are presented in the figure. Results: In both IMS III and DEFUSE 3, the PO assumption was not rejected. In IMS III, the common OR of 1.20 in favor of EVT is reflective of the cutpoint-specific ORs ranging from 1.1 (for mRS 0-2 and 0-3) to 1.6 (for mRS=0). In DEFUSE-3, the unadjusted PO model yields a common OR of 2.77 in favor of EVT. The cutpoint-specific ORs all favor EVT, but the magnitude of the treatment effect ranges from 1.3 for no symptoms (mRS 0) to 4.0 for functional independence (mRS 0-2). Conclusions: Cutpoint-specific point-estimates of the treatment effect can vary substantially, even when the proportional odds assumption is not rejected. This can be observed when the sample size is small and the test for proportionality of the odds has low power, as is the case in DEFUSE 3 (n=182). In this scenario the common OR may lack generalizability across cutpoints.