A Dynamic Panel Model of Prescription Drug Adherence

Research Objective:  Management of chronic illness often depends on patient adherence to a regular regimen of care. An extensive literature reveals that adherence is difficult to predict or to explain, and considerable resources are targeted to improve adherence. We extend previous analyses of adherence behavior to a dynamic framework to consider the impact of unobserved heterogeneity and state dependence (habit or habit persistence) as well as observed patient characteristics.

Study Design:  We use a longitudinal panel data framework and estimate quarterly dynamic probit models of adherence allowing for correlated effects, state dependence and initial conditions (adherence measured in the initial quarter). Multivariate models included age, gender, location, health status, income, benefit plan design, and time.

Population Studied:  We analyzed quarterly adherence to statin medications for 4,943 continuously-enrolled employees of six large firms from January 2009 through June 2012 who filled a prescription for an HMG-CoA reductase inhibitor (i.e., statin) for long-term management of blood cholesterol. Adherence was defined as a Percent of Days Covered greater than or equal to 80%.

Principal Findings:  In any given quarter about two-thirds of employees were adherent, however, when examining the data on a longitudinal basis, employees made transitions between adherent and non-adherent states with 76.1% adherent in some quarters and non-adherent in others. Of those who were non-adherent in a previous quarter, 76.6% continued to be non-adherent in the next quarter. Adherence was also persistent, and 82.1% of adherent employees were adherent in the next quarter. Dynamic multivariate models revealed that when the employee was adherent in the previous quarter the probability of adherence in the current quarter increased by 18.3 percent (p<0.01) and when the employee was adherent in the initial quarter the probability of adherence increased by 22.5 percent (p<0.01), respectively. Other factors, such as gender, exerted much less influence on adherence. For example females were 2.2 percent less likely to be adherent than males (p=0.002). In addition, the relationship between income and adherence was small and positive (p<0.01).

Conclusions:  Understanding the determinants of adherence is important for chronic condition management and to develop appropriate interventions to promote adherence. State dependence, as captured by the level of adherence in the initial and previous quarter, was highly significant, suggesting that there is strong habit or persistence in adherence from one quarter to the next. In comparison, observed patient characteristics, the focus of many previous interventions, played a much smaller role in understanding adherence behavior.

Implications for Policy or Practice:  Our study reveals that both adherence and non-adherence are persistent. Policies focusing on patient behavior by reinforcing habits, as well as patient characteristics, may yield additional improvements in patient outcomes.