Objective: To assess the performance of standard techniques for handling missing data in longitudinal studies under a variety of missing data scenarios. We demonstrate our analysis via extensive simulation studies and the applications to Parkinson’s Progression Markers Initiative (PPMI) study.

Background: Missing data is common in longitudinal studies, where the missingness may occur in longitudinal outcomes, covariates or both. Motivated by the PPMI study where research quality MRI measures are missing for 62% of patients, we carry out a systematic study assessing the performance of practical methods for handling missing data in longitudinal analysis. The missing data may emerge in longitudinal outcomes, covariates or both.  Moreover, the reasons of missingness may depend on observed outcomes, covariates or both. Hence, there is a need to find the most appropriate method for each specific missing data scenario, leading to accurate statistical inference in real data analysis.

Method: We compare available case analysis (ACA) and a few well-developed multiple imputation (MI) methods, including fully conditional specification (FCS), classification and regression tree (CART) and multilevel MI (e.g., PAN). The underlying analysis model is linear mixed-effects model (LMM). For each missing data scenario, we compare the percentage bias, relative efficiency and coverage of probability of the considered methods, and accordingly give our recommendation based on these statistics. We applied these methods to the PPMI study to investigate the association between baseline MRI volume and cognitive decline based on Montreal Cognitive Assessment (MoCA).

Results: We find that different methods are suitable to different missing data scenarios. MI does not always outperform ACA. ACA is likely to produce biased estimates when the missingness of covariates depends on observed outcomes. Among the MI methods, FCS is preferred to PAN although the latter accounts for a more complex imputation procedure. In the application to the PPMI study, we do not observe significant difference in the estimation results between ACA and FCS, but PAN produces some biased estimates.

Conclusion: Our study can guide the selection of appropriate methods for handling missing data in longitudinal studies. Thus, it will help improve statistical power and deliver accurate inference in biomedical studies.

« Back to MDS Virtual Congress 2021

MDS Abstracts - https://www.mdsabstracts.org/abstract/statistical-methods-for-dealing-with-missing-data-in-longitudinal-studies-with-application-to-parkinsons-disease-cognition-research/