Amyotrophic lateral sclerosis (ALS) clinical trials suffer a large proportion of drop-outs. Ignoring missing data can lead not only to underpowered tests, but also to selection bias. Current strategies for handling not at random missing data have several limitations. To determine the most effective approach, we compared the standard procedures with the pattern mixture model, using the data from a randomized dose-finding trial on lithium for the treatment of ALS, which reported a high rate of drop-outs (68.4%). We evaluated the ALS Functional Rating Scale-Revised (ALSFRS-R) profile using mixed effect models on different reference populations (1. Intention-to-treat, 2. "Completers", 3. Last observation carried forward, 4. "0-imputation"). All four strategies have limitations on account of: 1. Violation of the "missing completely at random" assumption of the mixed model; 2. Underpowered results on selected patients; 3. Underestimation of the time effect on ALSFRS-R decline and misuse of the assumption that those who discontinued could not get worse; 4. Overestimation of the time effect on ALSFRS-R decline and misuse of the assumption that those who discontinued could not have scores different from zero. The pattern mixture models fitted better than models that did not consider the missing data pattern effect (p = 0.006 and p = 0.0002). Pattern mixture model thus seem superior and we recommend its use to obtain more accurate estimates even when the information is missing.
- Amyotrophic lateral sclerosis
- Informative missing data
- Pattern mixture model
ASJC Scopus subject areas
- Pharmacology (medical)