Assessment of bias in the third international stroke trial (IST-3)

The Third International Stroke Trial (IST-3) is the largest randomized controlled trial (RCT) of intravenous alteplase for acute ischemic stroke and contributes substantially to meta-analyses and guidelines. Concerns persist regarding potential bias arising from missing 6-month Oxford Handicap Scale (OHS) score outcomes and from the measurement of the outcome in its predominantly open-label design. We used the publicly available IST-3 dataset. For missing 6-month OHS scores (3.2%), we performed multiple imputation (MI) separately by treatment arm under missing-at-random (MAR) assumptions, followed by a tipping-point sensitivity analysis under missing-not-at-random (MNAR) assumptions. The tipping point was the number of 1-level OHS score upward shifts (δ) in the alteplase arm that resulted in a null effect for ordinal endpoints or a transition from an event to non-event for binary endpoints. Validity of the 6-month OHS score was assessed via Spearman’s rank correlation with baseline National Institutes of Health Stroke Scale scores stratified by outcome assessor (patient, friend, physician) with interpretation anchored to prespecified thresholds. Treatment-effect heterogeneity (HTE) by assessor was evaluated with adjusted regression models and the ICEMAN instrument. Risk of bias was assessed using the Cochrane RoB 2 instrument. Ninety-six participants (3.2%) had missing 6-month OHS scores, and those participants had modestly worse baseline and 7-day prognostic profiles. The OHS < 3 endpoint's multiple imputation estimate included the null. Under an alteplase-arm-specific missing-data assumption, the OHS < 2 endpoint tipped after reclassification of 6 participants, while both ordinal endpoints remained robust. Validity analyses showed lower-than-prespecified correlations for patient-assessed OHS scores across all endpoints, while friend and physician assessor results were indeterminate due to wider confidence intervals. There was moderate credibility of HTE stratified by assessor type, with patient assessors having the largest effect for all endpoints. Risk of bias ratings varied for both domains based on the endpoints. The primary endpoint was most robust (low risk of bias in both domains), while OHS < 2 had a high risk of bias in both domains. The OHS 0–3 vs 4–6 and OHS 0–6 endpoints had a high risk of bias rating only for the measurement of the outcome. In the IST-3 RCT, sensitivity to missing-data assumptions and outcome assessment varied by endpoint. The primary endpoint was robust to bias, while secondary endpoints were not. These findings support cautious, endpoint-specific interpretation of IST-3-derived estimates and highlight the importance of sensitivity analyses and standardized outcome assessment in open-label trials. ISRCTN25765518.