Abstract:
|
Multiple imputation (MI) is a preferred approach to missing outcome data in longitudinal studies of progressive disease; older techniques, like complete case analysis (CCA) and last observation carried forward (LOCF), can bias toward unduly good outcomes. But the preferred approach for studies of acute illnesses with some recovery, such as acute stroke, is understudied. In these settings, CCA and LOCF may bias toward worse group outcomes than actually occur. Therefore, these methods are viewed as "conservative" and remain often used. Using data from a well-known acute stroke trial, we simulated data missingness and compared 5 handling methods: 1) CCA, 2) worst outcome assigned, 3) LOCF, 4) MI using baseline covariates (BCVs), and 5) MI using BCVs plus later observed outcomes. Imputation methods that ignored post-baseline data showed poor correlation with actual outcomes and reduced study power. LOCF preserved power but biased outcome estimates to worse than actual. MI with BCVs plus interim outcome observations yielded highest power, accuracy, and lack of directional bias. We describe techniques to assess bias and variance of imputation methods in acute illness trials generally.
|