Missing data are frequent in Alzheimers disease (AD) trials due to the age of participants and the nature of the disease. This can lead to bias and decreased statistical power. We assessed the level and causes of missing data in a 2-year randomised trial of an AD patient management program (PLASA study), and conducted sensitivity analyses on the primary endpoint (functional decline), using various methods for handling missing data: complete case, LOCF, Z-score LOCF, longitudinal mixed effects model, multiple imputation. By 2 years, 32% of the 1131 subjects had dropped out, with the commonest reasons being death (28% of dropouts) and refusal (22%). Baseline cognitive and functional status were predictive of dropout. All sensitivity analyses led to the same conclusion: no effect of the intervention on the rate of functional decline. All analyses demonstrated significant functional decline over time in both groups, but the magnitude of decline and between-group (intervention versus usual care) differences varied across methods. In particular, the LOCF analysis substantially underestimated 2-year decline in both groups compared to other methods. Our results suggest that data were not “missing completely at random”, meaning that the complete case method was unsuitable. The LOCF method was also unsuitable since it assumes no decline after dropout. Methods based on the more plausible “missing at random” hypothesis (multiple imputation, longitudinal mixed effects models, z-score LOCF) appeared more appropriate. This work highlights the importance of considering the validity of the underlying hypotheses of methods used for handling missing data in AD trials.
Keywords: Alzheimer's disease, randomised controlled trials, clinical trial methodology, missing data, dropout, data interpretation, statistical methods, Cox regression, ADCS-ADL, MAR assumption