Missing data are common in longitudinal epidemiological studies. Single imputation methods, such as last observation carried forward (LOCF), are easy to implement but underestimate uncertainty in imputed values. Multiple imputation (MI) better reflects uncertainty but is more computationally intensive and may encounter challenges in longitudinal studies in the presence of mortality, resulting in partially imputed datasets. In this study, we compared LOCF and MI for imputing depressive symptoms, body mass index (BMI), and smoking status in a community-based cohort of older adults three years after baseline. Bootstrapping (200 samples) was used for confidence intervals. For MI, two imputed datasets were generated within each bootstrap sample.

Participants included 11,942 adults aged ≥65 from the ASPREE trial. Depressive symptoms (CESD-10), BMI, and smoking status were assessed annually. For each variable, mean differences under LOCF and MI were examined in the total sample, among all participants with imputed data, and among participants with imputed data in both LOCF and MI datasets, using t-tests and chi-squared tests.

Prior to imputation, 1,009 participants (8.5%) were missing depressive symptom data, 1,445 participants (12.1%) were missing BMI, and 957 participants (8.0%) were missing smoking status. In the total sample, the LOCF dataset, compared to MI, had higher mean depressive symptoms and a lower proportion of current smokers. Among all participants with imputed data, the LOCF dataset had higher mean depressive symptoms, lower mean BMI, and a higher proportion of current smokers. Restricting analyses to participants with imputed data in both datasets, the LOCF dataset had higher mean depressive symptoms and BMI.

Meaningful differences were observed between the LOCF and MI datasets, with LOCF resulting in higher mean depressive symptoms. It is important to consider the differential impact of imputation methods when interpreting results from longitudinal studies.