Abstract Body: Background: Number needed to treat (NNT) is an intuitive measure of treatment benefit/harm to both clinicians and patients. For ordinal endpoints, NNTs derived by dichotomously collapsing scales are easily calculated but almost always underestimate the amount of benefit a treatment confers. Several methods have been described to derive NNT for ordinal outcomes. We aimed to directly compare the performance of each of the most widely described methods of ordinal NNT derivation.

Methods: NNT was calculated for eight stroke clinical rials/meta-analyses purposively selected to include small, moderate and large treatment effects and diverse treatment types, including thrombolysis (NINDS tPA, IST3), thrombectomy (HERMES, AURORA, Large Core EVT), surgery for AIS (Hemicraniectomy), surgery for ICH (ENRICH), and blood pressure control (OPTIMAL BP). We derived benefit per hundred treated (BPH, same as absolute risk reduction) and NNT using four Permutation Methods (CMH test with tied pairs ignored, divided in half, assigned by NIHSS outcome, and by better than model expectations) and three Joint Outcome Table Methods (Multi-sampling, Min-Max, and Expert Panel).

Results: The net BPH and NNT values derived for each trial using each method are shown in the Table. Overall, the magnitude of treatment effect estimates rendered by Permutation methods were substantially lower than by Joint Outcome methods Table methods. In 4 of 8 trials, Permutation method estimates were lower than dichotomous estimates – a biologic impossibility. Among the automatic Joint Outcome Table Methods, the repetitive Multi-sampling technique yielded mildly lower estimates than the algorithmic Min-Max technique.

Conclusions: Permutation (Cochran–Mantel–Haenszel) ordinal NNT derivation methods systematically underestimate treatment benefit magnitude due to bias towards the null from calculation-related nondifferential misclassification. Joint Outcomes Table methods may be preferred for future clinical trials that report net NNT for ordinal outcomes.