Data from the Medstat MarketScan database (11 million covered lives from 45 large employers), which contains detailed information on medical conditions, insurance coverage, and payments for inpatient, outpatient, and prescription drug services, were used for this study. Continuously enrolled adults over the age of 18 who were prescribed any oral anti-hyperglycemic agents over a 24-month period from January 1, 2000, to December 31, 2001 were eligible for the analysis.
As the FDC evaluated in this study became commercially available in August 2000, the study sample was determined using a 2-step process. First, individuals prescribed metformin or sulfonylurea (the antidiabetic drug category that includes glyburide) or both before July 2000, were identified. Then, those prescribed both metformin and sulfonylurea concurrently (either separately or the FDC) after August 2000 were included in the final sample.
The date of the first claim for metformin and sulfonylurea (either separately or in the FDC) between September 1, 2000 and December 31, 2001 was defined as that subject's index date. Each patient was then followed for 180 days after the index claim. Drug adherence was measured by medication possession ratio (MPR), the proportion of days on which a patient had medication available. To calculate the MPR, each day in the 180-day follow-up period was evaluated as ‘covered’ or ‘not covered’ by a prescription fill or refill. If all days were ‘covered’ by a prescription, then adherence was 100%. This MPR algorithm is similar to the adherence measure described by Bryson and colleagues.9
The days on which a patient prescribed 2 drugs had only one available was included in the analysis; the MPR was reduced by 50% for these partial adherence days.
Propensity Score Method
To compare adherence rates between patients prescribed a single-pill FDC and those taking the same two drugs but with 2 separate prescriptions, one must recognize that in practice, patients were not randomly assigned into the 2 treatment groups. Therefore, differences in observed covariates in the 2 groups may exist, and these differences could lead to biased estimates of treatment effects. The propensity score method has been commonly used to reduce such biases in observational studies.10
A propensity score is the probability of being assigned to the treatment, given a set of observed covariates. Individuals are matched or grouped into strata based on their score. Once the propensity scores and covariates within each stratum are balanced between treatment groups, the treatment assignment within each stratum can be functionally regarded as random.8
The propensity score method is a 2-stage approach. In the first step, the probability of using FDC is estimated by logistic regression model, adjusting the covariates and their interaction terms to balance the propensity score and covariate distributions within each stratum. In the second stage, stratification matching is used to estimate the average treatment effect.11
The average treatment effect is the weighted average of the adherence differences between FDC users and nonuser across the strata.
Six predictors for the switch to FDC were chosen based on adherence literature: demographics (age, gender), geographic region (east, north central, west and south), employment status (hourly worker, union worker, retiree, and dependent), health insurance characteristics (average drug co-payment, type of plan including fee-for service, HMO, PPO, and POS), health service utilization during the study period, and comorbidities. The utilization covariate includes inpatient and outpatient use, number of medications, the percentage of brand name medications, the days supplied per medication refill, and the number of refills during the follow-up period. A binary variable was included to distinguish whether the patient took one or both of the medications (metformin and/or sulfonylurea) at baseline. Adherence rates in the baseline period before FDC was calculated and included as a predictor of treatment assignment. Statistical analyses were performed with STATA 9 (College Station, Texas).
To assess whether duration of treatment impacted the findings, the follow-up period was changed to 90 and 120 days. In addition, an alternative method, a fixed-effect model that controls for time invariant unobserved factors, was used.