Data source
This retrospective observational study used the US DOD data from January 1, 2012 to September 30, 2015. The DOD provides health care to over 9.4 million beneficiaries located in all 50 US states and multiple countries globally. Eligible beneficiaries include active duty, activated guard and reserve, retirees, survivors, some inactive guard and reserve, and their family members. Most beneficiaries are retired service members and their family members (5.42 million, 57%), many of whom are Medicare eligible (3.18 million). Beneficiaries remain in the system for an average length of 7.2 years, which is 2–3 times longer than commercial insurance plans. The data repository includes comprehensive datasets providing integrated information about the inpatient, outpatient, ER, and pharmacy claims from the US DOD facility and civilian/private sector care for eligible beneficiaries.
Medical and pharmacy claim coding utilizes the National Drug Code (NDC) coding system, Healthcare Common Procedure Coding System (HCPCS) codes, Current Procedural Terminology (CPT) codes, and the International Classification of Disease, 9th Revision, Clinical Modification (ICD-9-CM).
Study population
This study selected adult patients with ≥1 pharmacy claim for an OAC (warfarin, apixaban, dabigatran, or rivaroxaban) from January 1, 2013 to September 30, 2015 (identification period). Edoxaban was not included in the analysis due to small sample size (N = 131). The first DOAC prescription claim date was defined as the index date for patients with a DOAC claim(s). For those without a DOAC claim, the first warfarin prescription claim date was defined as the index date. The baseline period was defined as one-year before the index date, during which patients had ≥1 medical claim for AF (ICD-9-CM: 427.31) and continuous enrolment [12]. Patients were excluded from the study if they had claims for valvular heart disease, heart valve replacement, dialysis, kidney transplant, end-stage chronic kidney disease, venous thromboembolism, reversible AF, or a pharmacy claim for an OAC during the baseline period, hip or knee replacement within 6 weeks prior to the index date, > 1 OAC claim on the index date, or a pregnancy diagnosis during the study period (Additional file 1: Table S1).
The follow-up period was defined as one day after the index date until the earliest of the following dates: OAC discontinuation date (≥30-day gap between OAC prescriptions), switch to a non-index OAC < 30 days before or after discontinuation, death, end of continuous medical and pharmacy enrollment, or end of study period [13].
Outcome measures
Defined by primary or secondary diagnosis position on inpatient claims, stroke/SE was utilized as the effectiveness outcome measure while major bleeding served as the measure for safety outcomes. Stroke/SE was further classified into ischemic stroke, hemorrhagic stroke, and SE. Major bleeding consisted of intracranial hemorrhage (ICH), gastrointestinal (GI) bleeding, and major bleeding at other key sites. Validated administrative claim-based algorithms as well as published articles were used to derive the stroke/SE and major bleeding code lists. (Additional file 1: Table S2) [14,15,16,17].
Baseline variables
Baseline measurements included patient demographics, comorbidities, medications, hospitalizations during the 12-month baseline period, and clinical risk scores (HAS-BLED [hypertension, abnormal kidney or liver function, stroke, bleeding, age > 65 years, and drugs/alcohol abuse or dependence], Charlson Comorbidity Index [CCI], and CHA2DS2-VASc).
The CHA2DS2-VASc stroke risk score and HAS-BLED bleeding risk scores were calculated (Additional file 1: Table S3 and Table S4) [18, 19]. Note that for the HAS-BLED score, INR and other lab values were unavailable in the data; a modified score (range 0 to 8) was used.
Statistical methods
The design, analytical methods, and presentation of this study were informed by the guidelines for comparative effectiveness research [20, 21].
To assess significant differences for dichotomous variables, Pearson’s Chi-square tests were performed. For continuous variables, student t-tests were used.
To control for potential confounders between comparative cohorts (apixaban vs warfarin, rivaroxaban vs warfarin, and dabigatran vs warfarin), one-to-one propensity score matching (PSM) was used to balance demographics and clinical characteristics and to estimate the average treatment effects in patients with similar characteristics for whom each of the two OACs would be a reasonable treatment choice [22]. The logistic regression for the propensity score calculation included inpatient admissions, baseline medication use, age, gender, US geographic region, CCI score, HAS-BLED score, CHA2DS2-VASc score, stroke and bleeding history, and comorbidities [23]. The nearest neighbor method without replacement with a caliper of 0.01 was used. The balance of baseline patient characteristics was checked based on mean standardized differences with a threshold of 10% [24].
Incidence rates per 100 person-years of stroke/SE and major bleeding in PSM matched cohorts were calculated. To assess the risk of stroke/SE and major bleeding for patients in the matched cohorts, Cox proportional hazards models were utilized. Hazard ratios, 95% confidence intervals (CIs), and p-values were provided. OAC treatment was included as the independent variable, and no other covariates were included in the model because the cohorts were balanced.
Sensitivity analyses
Sensitivity analyses, for the purpose of testing the robustness of the main results, were conducted. In the first of these analyses, cohorts were stratified by dosage of DOACs (standard and reduced) on the index date to assess if the outcomes were altered by DOACs dosage. The post-PSM population was separated per dosage of DOACs on the index date: standard-dose (apixaban 5 mg-warfarin, rivaroxaban 20 mg-warfarin, and dabigatran 150 mg-warfarin) and reduced-dose (apixaban 2.5 mg-warfarin, rivaroxaban 15 mg-warfarin, and dabigatran 75 mg-warfarin). In each matched subgroup by dosage of DOACs, imbalanced baseline variables with standardized difference > 10% were included in the Cox proportional hazards models. The statistical significance of the interaction term between treatment and dose was determined with a cutoff point of p-value = 0.10.
Second, patients who had catheter ablation within 2 months prior to the index prescription and those who had cardioversion 1 month before or after index drug were excluded. After excluding those patients, the balance of the baseline characteristics was checked and variables which were unbalanced were incorporated in the multivariate model. These patients were excluded because they likely had a low risk of stroke and received the OACs for the procedures and not long-term stroke prevention. Third, a sensitivity analysis using the 6-months after the index date as follow-up was also conducted. In this analysis, patients were censored at the earliest of: the OAC prescription discontinuation date, date of switching, date of death, date of disenrollment, end of the study period (September 30, 2015), or 6 months after the index date. This sensitivity analysis allowed the follow-up period to be more balanced between the cohorts. Lastly, a sensitivity analysis using the intent-to-treat method was used, where patients were followed based on the index drug regardless of discontinuation or switch.