Handling Missing Data in Longitudinal Rehabilitation Research: A Methodological Demonstration With Functional...

This study compared the effects of 3 different approaches to handling missing data (listwise deletion of participants with missing data, mean imputation, and full information maximum likelihood [FIML]) when predicting functional independence trajectories over 10 years in older adults after traumatic brain injury (TBI).

Twenty-three TBI Model Systems (TBIMS) inpatient rehabilitation facilities in the United States.

Adults who sustained a complicated mild, moderate, or severe TBI at age 60 years or older and needed inpatient rehabilitation. They had to meet all eligibility criteria and have one or more functional independence measure (FIM) scores at 1, 2, 5, or 10 years post-TBI from the TBIMS national database.

Retrospective analysis of observational data using hierarchical linear models.

FIM total scores at 1, 2, 5, and 10 years post-TBI.

Different missing data approaches led to drastically different findings. Model comparisons supported a quadratic effect of time only in the listwise deletion model and found no other significant predictors. Linear trajectories were found in the mean imputation and FIML models. For both these models, older age, underrepresented minority status, unemployment at injury, longer posttraumatic amnesia duration, and pre-injury limitations all predicted lower overall FIM trajectories. However, when compared with the mean imputation model, the FIML-estimated b-weights were larger with smaller P-values. Years of education significantly predicted higher overall FIM trajectories in the mean imputation model but not the FIML model, likely because of the artificial shrinking of the estimated b-weight standard errors in mean imputation. History of mental health treatment predicted lower FIM trajectories only in the FIML model.

These findings show that it is critical to use appropriate modern methods to handle missing data because the method can affect outcome trajectory shape and identification of relevant predictor variables. Using older methods for handling missing data, such as listwise deletion, greatly reduces predictive ability, resulting in less generalizability and imprecision in longitudinal rehabilitation research.