This was a retrospective observational study using administrative data from Population Data BC (PopDataBC) [5] covering 1996 to end of 2019 for the entire population of British Columbia (BC), Canada (~ 5 million residents). The data was from the following databases: Medical Services Plan (MSP) [6] containing outpatient visits, the Discharge Abstract Database (DAD) [7] containing hospitalizations, the Consolidation File [8] containing demographics such as sex, place of residence, and registration with the provincial healthcare plan, and the Vital Statistics Database [9] containing date and primary cause of death. Information on all medications dispensed outside of hospital between January 1996 and December 2019 was retrieved from PharmaNet and linked to the other datasets. [10]

We created an incident cohort of adult patients ≥ 18 years of age with non-valvular AF. Using the algorithm validated by Navar et al. with a positive predictive value of 95.7% [11] we included individuals who had ≥ 3 recorded visits in MSP or DAD related to AF or atrial flutter, with at least one of the three recorded visits being AF-specific (ICD-9: 427.31; ICD-10: I48; Supplemental Appendix A). At least two of the visits had to occur within one year.

Next, using ATC codes from PharmaNet records, we captured all prescriptions for OACs available in the study jurisdiction (warfarin, dabigatran, apixaban, rivaroxaban, edoxaban). Individuals were excluded if they did not have continuous public medical insurance coverage during the 1-year prior to their first OAC prescription fill, or if their first OAC prescription occurred before 1997 to permit 1-year baseline covariate ascertainment, or after Jan 2019 to ensure ≥ 1 year follow-up before end of the data. The date on which the first OAC prescription was filled after the first AF diagnostic code (or within 60 days before the first AF code), was referred to as the “index date” (Fig. 1).

Overview of the study design

To ensure incident AF cases and incident OAC use, those with an AF code or OAC prescription fill during the 3 years prior to their AF diagnosis date or their index date, respectively, were excluded. To improve specificity, we then excluded individuals with indications for OAC other than non-valvular AF (e.g. venous thromboembolism, rheumatic valve disease; Supplemental Appendix A) based on codes appearing any time before their index date. We also excluded patients with two different OACs filled on the same index date. Finally, because ≥ 2 OAC prescription fills were required to measure adherence, patients with only one OAC prescription fill were excludedPatients were followed from the index date until the end of their follow-up time, defined as December 2019, date of death, or discontinuation of their public medical insurance enrolment, whichever came first.

Patient characteristics were measured during the 1-year prior to the index date (the “baseline period”).

To quantify days supply of OAC, the number of days of medication dispensed to the patient was calculated for every prescription fill for every patient (Supplemental Appendix D). For patients on DOACs, whose dosing regimens are relatively fixed, daily dose and days supply was obtained directly from PharmaNet. For warfarin patients whose daily dose changes frequently in response to their International Normalized Ratio (INR) values, the Random Effects Warfarin Days’ Supply (REWarDS) method was used to estimate patient’s daily dose and days supply for Eq. 1.[12]

In order to study the execution and persistence phases of adherence, 90-day consecutive time windows (the standard length of dispensing in the study jurisdiction) were created for each patient starting from index date until the end of their follow-up time (Fig. 2, Scenario A). The primary measure of adherence for this study was proportion days covered (PDC). PDC for each 90-day consecutive window was calculated using standard approaches [13]. If a patient permanently discontinued their medication, all consecutive windows after their last supply ran out were assigned PDCs of zero until the end of their follow-up period (Fig. 2A). This approach was based on clinical guideline recommendations for life-long OAC therapy for AF patients (even after experiencing a bleed) [2]. In cases of oversupply, the date of refill for the second prescription was adjusted to when the medication supply for the previous refill was estimated to run out (Fig. 2B). Patient’s average adherence over follow-up was obtained by calculating the mean of PDC values of all their follow-up windows. The mean PDC was reported for the cohort over the entire follow-up. PDC was also analyzed dichotomously with nonadherence defined as PDC < 0.8 and sensitivity analyses conducted with thresholds of 0.7 and 0.9.

Visual representation of PDC calculation using consecutive windows of 90 days throughout the follow-up period. Scenario A: The end of the blue bars denotes the times at which the patients’ supply is expected to run out. The PDC for each window is calculated by dividing the days’ supply in each window by 90 days (the length of the window). Scenario B: Oversupply has occurred in windows 3 and 5. The dates of the prescription fills have been adjusted to when the medication supply for the previous refill is estimated to run out

Since PharmaNet does not contain information on medications administered during hospitalizations, for our primary analysis, hospitalization periods were removed from both the denominator and numerator when calculating PDC. A sensitivity analysis was conducted with patients assumed to be 100% adherent during periods of hospitalization (Supplemental Appendix C). The Intraclass Correlation Coefficient (ICC) was used to compare the PDCs calculated by the two methods.

Generalized mixed effect linear regression models were used to identify factors associated with adherence. This was limited to the DOAC era, (after October 2010, when the first DOAC was marketed in Canada). Given the repeated data structure (multiple PDC measurements per person), the generalized estimating equation (GEE) technique, with PDC as the response variable, was used to account for the dependencies among data. Drug class (VKA or DOAC) was treated as a time-varying variable to allow for switches during follow-up and examination of patients’ current OAC. Out-of-pocket cost was also a time-varying variable. Other variables were fixed in time. Follow-up time was added to the regression model as an offset term to account for highly variable follow-up times among patients. The “exchangeable” correlation structure was found to best fit our data.

Candidate covariates were sex, age, years since OAC initiation, individual components of the CHA2DS2-VASc [14] stroke and modified HAS-BLED bleeding risk scores [15], weighted Charlson comorbidity index [16], number of concomitant medications (with polypharmacy defined as ≥ 5 concurrent medications, not including OAC on index date) [17], neighbourhood income quintile, out of pocket cost per day supply, and number of drug class switches during follow-up. Time was measured as number of years since the index date. After controlling for confounders, potential interactions between drug class and time, number of drug class switches, history of a major bleed, history of stroke, neighbourhood income quintile, and out-of-pocket prescription cost incurred by patient were evaluated. Statistically significant interaction terms were included in the model. Selection of the initial list of candidate covariates was guided by availability in the data, previous literature [4] and our clinical experience. We used variable selection analysis during the modelling step to develop the final list. To avoid the limitations of traditional variable selection methods such as stepwise (forward or backward) algorithms or selection based on p-values, we used a mix of variable inflation factor (VIF) to avoid collinearity/multicollinearity among the covariates and quasi-likelihood criterion (QIC) to select the best set of covariates with low VIF.

Secular trends in adherence were also analyzed and stratified by OAC drug class.

Analyses were done in R 4.0.3 (R Foundation for Statistical Computing, Vienna, Austria). Database access was approved by the data stewards and the study protocol was approved by the University of British Columbia Clinical Research Ethics Board (approval number: H17-02420).