Comparing algorithms for composite measures of intra-disease multiple medication adherence: The case of diabetes

Background: Adherence to multiple medications (i.e., separate dosage forms) intended for a disease can be measured by different single estimators, termed as composite estimators of intra-disease multiple medication adherence: 80% days covered (a) by at least one medication (“at least one”); (b) by both medications (“both”); (c) by each medication measured separately (“all”); and (d) computing an average of the individual medication adherence estimates (“average”). Objectives: (a) Assess different composite adherence estimators regarding their ability to predict healthcare utilization; (b) compare and contrast composite estimators. Methods: Using MarketScan 2002–2003 data, 6043 nonelderly patients who filled separate prescriptions of sulfonylurea [SU] and thiazolidinedione [TZD] were identified. Adherence was measured by the proportion of days covered (PDC) over periods of 90 days, 30 days, and cumulatively over such periods. Cox proportional hazards models analyzed all-cause and diabetes-related emergency room (ER) visits as the outcome variables. Results: All composite measures predicted hazards of all-cause or diabetes ER visits (P < 0.001) and each measure showed statistically significant discriminatory power (concordance statistics from 0.55 to 0.58). Cox regression was performed multiple times in which composite estimators measured on a continuous scale (e.g., ‘average’) were dichotomized using several cut-points. In the majority of cases (≤3 out of 8 times in analyses of ER outcomes), optimal results did not occur when the dichotomization cut-point was set at 80%. Conclusions: Each composite estimator showed the fundamental quality of a good measure. Although ‘average’ and ‘all’ approaches offer ease of measurement, there was no clear trend in superiority of one measure over the others. Clinical and practical considerations should dictate the choice of measure.