Association of visit-to-visit variability of hemoglobin A1c and medication adherence

Medication nonadherence is widespread, but there are few efficient means of detecting medication nonadherence at the point of care. Visit-to-visit variability in clinical biomarkers has shown inconsistent efficiency to predict medication adherence.

To examine the performance of visit-to-visit variability (VVV) of hemoglobin A1c to predict nonadherence to antidiabetic medications.

In this cross-sectional study using a clinical and administrative database, adult members of a managed care plan at a safety-net medical center from 2008 to 2012 were included if they had ≥ 3 noninsulin antidiabetic prescription fills within the same class and ≥ 3 A1c measurements between the first and last prescription fills. The independent variable was VVV of A1c (within-subject standard deviation of A1c), and the dependent variable was medication adherence (defined by medication possession ratio) determined from pharmacy claims. Unadjusted and adjusted multivariate logistic regression models were created to examine the relationship between VVV of A1c and medication nonadherence. Receiver-operating characteristic (ROC) curves assessed the performance of the adjusted model at discriminating adherence from nonadherence.

Among 632 eligible subjects, mean A1c was 7.7% ± 1.3%, and 83% of the sample was nonadherent to antidiabetic medications. Increasing quintiles of VVV of A1c and medication nonadherence were both associated with increased within-subject mean A1c and younger subject age. The logistic regression model (adjusted for age, sex, race/ethnicity, within-subject mean A1c, number of A1c measurements, number of days between the first and last antidiabetic medication prescription fills, and rate of primary care visits during the study period) showed a nonsignificant association of VVV of A1c and medication nonadherence (OR = 1.19, 95% CI = 0.42-3.38 for the highest quintile of VVV). Adding VVV of A1c to a model including age, sex, and race only modestly improved the C-statistic of the ROC curve from 0.6786 to 0.7064.

VVV of A1c is not a robust predictor of antidiabetic medication nonadherence. Further innovation is needed to develop novel methods of detecting nonadherence.