The impact of data source on real-world medication adherence and exposure measures: From prescription to sold.

Abstract

Medication adherence is central to evaluating treatment effectiveness, reducing preventable complications, and informing value-based care. The proportion of days covered (PDC) is widely used but inconsistently calculated across studies because of variations in data sources and methods. Prescription timestamps (written, filled, and sold) capture different stages of the medication journey; however, their impact on adherence measures has not been systematically compared. To evaluate how different prescription data sources and calculation methods influence medication adherence estimates in real-world practice. We conducted a retrospective cohort study of cardiovascular prescriptions (2009-2019) using longitudinal transaction data. Each prescription included written, fill, and medication sold (pickup) dates. Adults with at least 12 months of follow-up and 3 or more prescriptions in the same class (angiotensin-converting enzyme inhibitors, angiotensin II receptor blockers [ARBs], calcium channel blockers, or diuretics) were included (N = 7,226). Two adherence measures were calculated: (1) exposure PDC, reflecting raw coverage without gap adjustments, and (2) adherence PDC, which excluded prolonged gaps using dynamic patient-specific thresholds. PDC was calculated using 5 source definitions (written, fill, sold, written-sold, and fill-sold). Adherent classification agreement at the 0.80 threshold was assessed using cross-tabulations and visualized with Sankey diagrams. Patients contributed more than 230,000 fills. Median fill and sold lags were both 0 days, and the 90th percentile delays were longer for fills (29 days) than for sold (3 days). Exposure PDC varied substantially across sources (median = 0.747-0.794; <i>P</i> < 0.001), whereas adherence PDC was uniformly high (0.959-0.977; <i>P</i> < 0.001). Written dates produced the lowest exposure PDC because of inflated observation windows and gaps, and sold dates yielded the highest. Across drug classes, ARBs had the highest adherence (exposure median = 0.850; adherence = 0.958) and diuretics the lowest (0.720; 0.952). At the 0.80 threshold, adherence classification was stable for adherence PDC (>94% concordance across sources) but highly sensitive for exposure PDC, with more than 15% of patients switching classification when written dates were used instead of fills. Adherence measurement is not fixed but shaped by data source and methodology. Both timestamp choice and gap-handling substantially influence adherence estimates and patient classification. Our dynamic, patient-specific approach to gap adjustment yielded stable metrics, and exposure-based measures remained more source-sensitive. Standardizing and transparently reporting adherence definitions is essential to avoid misclassification. Future work should evaluate the prognostic value of gap-adjusted vs exposure PDC in predicting cardiovascular outcomes to guide more meaningful quality measures and value-based care.