Comparison of Imputation Strategies for Incomplete Longitudinal Data in Life-Course Epidemiology.

This article focuses on comparing the performance of three multiple imputation (MI) methods—normal linear regression (NORM), predictive mean matching (PMM), and variable-tailored specification (VTS)—for handling incomplete longitudinal data in life-course epidemiology, using real data from the Health and Retirement Study (HRS). The study induced missingness under nine scenarios combining different missing data mechanisms (missing completely at random, missing at random, and missing not at random) and proportions (10%, 20%, 30%) to evaluate bias, root mean square error, and computation time in estimating the effect of elevated depressive symptoms on mortality. Results showed that while all MI methods performed similarly under missing completely at random and missing at random mechanisms, PMM consistently yielded the lowest root mean square error, competitive computation times, and fewer implementation challenges, making it a favorable approach for imputing life-course exposure data. The study also found that MI methods did not fully recover true effect estimates under missing not at random scenarios but still outperformed complete-case analyses. These findings provide practical guidance for researchers addressing missing data in longitudinal epidemiologic studies.