(154) Comparison of Approaches to Handle Missing Data for Estimating Longitudinal Patterns of Anticholinergic and Sedative Drug Load Among Older Adults

Publication Number: 828

Background: Long-term exposure to drugs with anticholinergic and sedating properties is associated with worse cognitive and physical function in older adults, therefore, it is important to identify subgroups who may be at increased risk of prolonged exposure at high doses. Estimating longitudinal trajectories of use may be impacted by how missing data are addressed.

Objectives: To assess whether estimates of longitudinal trajectories in use of these drugs (collectively assessed via the Drug Burden Index (DBI score)) and trajectory membership are affected by the missing data approach.

Methods: We used a 0.2% nationwide, random sample of US Medicare beneficiaries, age 66+ years with fee-for-service coverage (A, B, D) during 2014-2016. Our cohort included initiators of DBI drugs, using a 1-year lookback to exclude prevalent users. After initiating, we estimated each patient’s average daily score each month (12 scores per patient). We estimated 1-year DBI score trajectories using group-based trajectory modeling (GBTM) and k-means clustering (KMC) methods, allowing for 2-5 groups/clusters. DBI score could be missing due to death, administrative censoring, or insurance disenrollment. To assess the impact of missing data, we compared a complete case analysis approach (all 12 scores needed) versus each trajectory estimator’s default approach (only 1+ score needed for GBTM, 2+ scores needed for KMC).

Results: Among the 10,993 eligible patients, 6% died, 31% were administratively censored, and 3% disenrolled. In both longitudinal trajectory methods, the complete case analysis approach suggested that 3 trajectory patterns captured heterogeneity in DBI scores over time: a very small group (2%) with persistently high DBI scores (DBI>1); a small group (18%) with stable lower DBI scores (DBI≈0.5); and a large group (80%) with quickly decreasing lower DBI scores (DBI < 0.5). Results from the default approaches were qualitatively the same as the complete case analysis. Graphical displays will accompany these results.

Conclusions: In this study, with high prevalence of missing data, longitudinal trajectories were similar across competing approaches for handling missing data. Future work will explore the impacts of missing data and the potential for selection bias when characterizing longitudinal patterns of other drug classes.