Investigating the Potential for Bias When Using a Widely Accepted Medication Adherence Measure to Predict Mortality

Summary measures of medication adherence, such as the proportion of days covered (PDC), are often used to analyze the association between medication adherence and various health outcomes. We hypothesized that PDC and similar measures may lead to biased results in some situations when used to estimate the association between adherence and the outcome event (e.g., mortality). Thus, the objective was to determine the conditions under which PDC and similar measures might produce biased estimates of the association between adherence and mortality and to review methods to avoid such bias.

Simulation study and analysis of data from a large retrospective cohort study.

We conducted a comprehensive simulation to compare how adherence estimates varied-using prescription-based (end point was final date that medications were on hand or date of death) and interval-based (end point based on fixed calendar interval or date of death) PDC denominators-when deaths occurred either during or after the adherence exposure period. We then made similar comparisons using data from a retrospective study that included comprehensive measures of medication refill activity and clinical outcomes of 207,841 patients with diabetes mellitus who were prescribed one or more oral antidiabetic medications. When deaths occurred within the adherence exposure period, substantial bias in adherence estimates was possible regardless of the PDC denominator type.

Investigators using PDC or similar proxy measures should carefully consider the temporal relationship between adherence exposure and clinical outcomes when the outcome event affects the adherence measurement.