Handling of Missing Outcome Data in Acute Stroke Trials: Advantages of Multiple Imputation Using Baseline and...

In acute stroke randomized trials, missingness of final functional outcome data reduces study power and potentially biases findings of treatment effect. Best methods for handling missing outcome data have not been well delineated for diseases with monophasic onset and subsequent improvement, like acute stroke.

We simulated data missingness in the public dataset of the landmark, second NINDS-tPA trial, by randomly removing 5%-25% of actual values for the 3-month modified Rankin Scale (mRS) of global disability. We evaluated 5 missing data-handling methods: complete case analysis (CCA), worst case imputation (WCI), last observation carried forward (LOCF), multiple imputation using baseline covariates only (MI-B), and multiple imputation using baseline and postbaseline observations (MI-BP).

With the original trial's 333 patients, tissue plasminogen activator was associated with 3-month disability benefit, both for mRS dichotomized at 0-1 (P = .014) and shift analysis (P = .035). Distance (root mean square error) of imputed from actual mRS values was best for LOCF (1.17) and MI-BP (1.28), intermediate for MI-B (1.89) and worst for WCI (3.77). Directional bias (mean difference) was least for MI-BP (.01) and MI-B (-.16), intermediate for LOCF (-.37), and worst for WCI (-3.22). Preservation of formally positive results was greatest for MI-BP and LOCF (preserved at all missingness rates), intermediate for CCA and MI-B (preserved only with missingness <10%-20%), and least for WCI (preserved only with missingness <5%-20%).

For acute stroke trials, multiple imputation using baseline and postbaseline observations is an advantageous approach to missing outcome data-handling, yielding high accuracy, reduced directional bias, and greater preservation of study power.