| Summary: | Background: The description of adherence based on medication refill histories relies on the estimation of continuous medication availability (CMA) during an observation period. Thresholds to distinguish adherence from non-adherence typically refer to an aggregated value across the entire observation period, disregarding differences in adherence over time. Sliding windows to divide the observation period into smaller portions, estimating adherence for these increments, and classify individuals with similar trajectories into clusters can retain this temporal information. Optimal methods to estimate adherence trajectories to identify underlying patterns have not yet been established. This simulation study aimed to provide guidance for future studies by analyzing the effect of different longitudinal adherence estimates, sliding window parameters, and sample characteristics on the performance of a longitudinal clustering algorithm.Methods: We generated samples of 250–25,000 individuals with one of six longitudinal refill patterns over a 2-year period. We used two longitudinal CMA estimates (LCMA1 and LCMA2) and their dichotomized variants (with a threshold of 80%) to create adherence trajectories. LCMA1 assumes full adherence until the supply ends while LCMA2 assumes constant adherence between refills. We assessed scenarios with different LCMA estimates and sliding window parameters for 350 independent samples. Individual trajectories were clustered with kml, an implementation of k-means for longitudinal data in R. We compared performance between the four LCMA estimates using the adjusted Rand Index (cARI).Results: Cluster analysis with LCMA2 outperformed other estimates in overall performance, correct identification of groups, and classification accuracy, irrespective of sliding window parameters. Pairwise comparison between LCMA estimates showed a relative cARI-advantage of 0.12–0.22 (p < 0.001) for LCMA2. Sample size did not affect overall performance.Conclusion: The choice of LCMA estimate and sliding window parameters has a major impact on the performance of a clustering algorithm to identify distinct longitudinal adherence trajectories. We recommend (a) to assume constant adherence between refills, (b) to avoid dichotomization based on a threshold, and (c) to explore optimal sliding windows parameters in simulation studies or selecting shorter non-overlapping windows for the identification of different adherence patterns from medication refill data.
|
|---|
