Background: Medication adherence is critical to patient outcomes, can decrease patient mortality and has been identified as an important indicator of medication use quality by the Pharmacy Quality Alliance (PQA). The PQA has endorsed “Proportion of Days Covered” (PDC) as the primary method of measuring adherence. Although easy to calculate, the PDC is a deterministic metric that cannot capture the complexity of such a dynamic phenomenon. To better capture heterogeneity in medication adherence, group-based trajectory modelling (GBTM) is increasingly being proposed in the pharmacoepidemiology literature as a better alternative.
Objectives: To compare the performance of GBTM at capturing treatment adherence compared to PDC and the nonparametric longitudinal K-means (KML) model using simulated data. As a real-world application, the three methods will be compared in their ability to discover subgroups of medication adherence to oral anticancer therapy among women with breast cancer using a large German claims dataset.
Methods: Variables were simulated for time-varying treatment, baseline and time-varying covariates under three trajectory models for medication adherence combining a cat’s cradle and rainbow effects. Scenarios with different proportions of medication adherence, different functional forms of the treatment, as well as sample sizes (n=5000, 1000 and 500) were tested. The number of follow-up periods (6, 8 and 10 months) and number of trajectory groups (3, 4 or 5) were chosen based on the identification conditions of GBTM. The performance of GBTM, PDC and KML were compared using absolute bias, standard deviation of the estimates, accuracy of adherence prediction, relative bias, and relative variance.
Results: GBTM outperformed PDC and KML at capturing all medication adherence patterns that were tested, resulting in lower relative bias, even under model misspecification. For the scenario with three trajectory groups, the relative bias was at least 1.5 higher for PDC and 1.9 higher for KML compared to a correctly specified GBTM. When there was no clear distinction between the trajectory groups, all three methods yielded biased estimates of medication adherence; even in this case, GBTM yielded the lowest bias. For 6 months of follow-up, the accuracy of adherence prediction was around 1.0 for GBTM, 0.96 for KML and 0.94 for PDC; the accuracy tends to decrease with longer follow-up periods.
Conclusions: GBTM was able to better capture patient heterogeneity across complex patterns of medication adherence when compared to PDC and KML. To our knowledge, this is the first quantitative comparison of GBTM with other approaches for modelling medication adherence. Future work will focus on assessing the clinical relevance of subgroup trajectories identified by these models.
.jpg "Alind Gupta, PhD photo")
