Methods/Statistics
Impact of mortality on methods for handling missing data in longitudinal studies: an application to the study of frailty trajectories Anna* Anna Siefkas Rienna Russo Ariela Orkaby Yuan Ma A. Heather Eliassen Marian T. Hannan Jungwun Lee Sebastien Haneuse
Longitudinal studies are subject to missing data such as unit nonresponse, attrition, and item nonresponse. While inverse probability of censoring weights (IPCW) and multiple imputation (MI) are well-established, their appropriateness in longitudinal aging studies is not obvious due to often high mortality, which leads to truncated data that cannot logically be imputed.
We performed a simulation study to assess the impact of mortality on methods for handling missing data in longitudinal studies of frailty. We simulated two complete data sets, each including 3 baseline covariates (one continuous, two binary) and 10 binary frailty index (FI) items at 9 timepoints. In the first, all timepoints were observed for all participants (i.e. no mortality). In the second, we used the value of the FI to predict mortality and truncated some trajectories due to death. To each dataset, we also added unit nonresponse, attrition, and item nonresponse to FI items. Missingness was missing at random conditional on covariates and all other FI items. We compared the performance of methods commonly used in frailty research: complete case analysis, carry forward imputation, setting missing items to 0, prorating the FI denominator, and MI using generalized linear mixed effects models. Attrition was addressed with IPCW. In the data without mortality, FI trajectories were modeled with linear mixed effects models. In the data with mortality, models were partly conditional on death to avoid extrapolating trajectories past the point of mortality.
When mortality was absent, MI recovered model parameters with negligible bias. When mortality was present, all methods demonstrated bias, particularly for the intercept term; average FI was underestimated in imputed data with mortality.
Standard missing data methods should be used with caution in settings with high mortality. Novel methods are needed to address missing data in the presence of mortality, especially in complex settings of aging research.