Dealing with missing delirium assessments in prospective clinical studies of the critically ill: a simulation study and reanalysis of two delirium studies

Background: In longitudinal critical care studies, researchers may be interested in summarizing an exposure over time and evaluating its association with a long-term outcome. For example, the number of days a patient has delirium (i.e., brain dysfunction) during their critical care stay is associated with the presence and severity of long-term cognitive problems. In large pragmatic trials and multicenter observational studies, particularly when electronic medical record data is used, the information on daily exposure status may be available at some time points and not at others. Model-based multiple imputation is a well-established, widely adopted method to deal with missing data. But the uncertainty around multiple imputation for summary exposure variables is whether the imputation is to be performed at the summary level or at the daily assessment level.

Methods: We compare the following approaches to imputing and summarizing partially missing longitudinal data: 1) active imputation, where we impute the summary; 2) passive imputation, where we impute the daily missing data, and then compute the summary; 3) ad hoc methods where we assume all missing time points have the a) most or the b) least extreme value; and 4) complete case analysis where only participants with complete data are analyzed. These methods were applied under different missingness mechanisms, varying proportions of missingness, and association of missingness with an auxiliary variable using simulations that closely mirrors real-life critical care data to be relevant to real-world clinical practice. The performance of the approaches were compared using bias of the estimated coefficients, standard error of the estimate and coverage. We also apply these imputation strategies to two datasets in critical care.

Results: Simulations show that all methods performed comparably when the proportion of missingness was small, indicating that in such instances, the gain over using any imputation model is minimal. But as the proportion of missingness increases, the passive imputation approach provides efficient and less biased estimates under the missingness at random and missingness completely at random mechanism.

Conclusions: For longitudinal data where a summary exposure is of interest, we recommend practitioners adopting the passive imputation strategy.