Study design and data source

We used the target trial framework and STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) checklist to conduct and report a retrospective cohort study using the SEER registry linked to the Medicare database (cancer sites: breast, prostate, and lung) from 2011–2019 [19, 20]. The SEER registry contains patient demographics, primary tumor site, tumor characteristics, and cancer stage at diagnosis, treatment, and follow-up of cancer patients across the US [21]. The Medicare data add to SEER data health care services utilization (medical claims, procedures, and prescriptions) [22]. Table 1 summarizes the protocol for target trial and emulation procedure. The study design and study timeline are illustrated by Figure S1.

Table 1 Protocol for a target trial and emulation procedure using the SEER-Medicare databaseStudy sample and eligibility criteriaStudy sample

We included individuals aged ≥ 66, newly diagnosed non-valvular atrial fibrillation (NVAF) between January 1, 2012 and December 31, 2019, defined as any International Classification of Disease-9th Revision-Clinical Modification (ICD-9-CM) codes 427.31 or 427.32 or any International Classification of Disease-10th Revision-Clinical Modification (ICD-10-CM) codes I48.xx in any position on one Medicare inpatient claim or on two outpatient claims at least 7 days but < 1 year apart [23]. We retained patients with breast, lung, or prostate cancer—the most common cancer types with AFib—from the SEER file at any time before the initial AFib diagnosis (ICD-O-3 codes C50.0-C50.9 for breast; C34.0, C34.1, C34.2, C34.3, C34.8, C34.9, C33.9 for lung; C61.9 for prostate cancer). Patients were required to continuously enroll in Medicare part A, B, D, and without Medicare Advantage or Health Maintenance Organization (HMO) for at least 12 months before initial NVAF diagnosis.

Exclusion criteria

We adapted exclusion criteria from clinical trials [24, 25]. In addition, patients were excluded if they had any other indication than NVAF, contraindication to OACs or had the event of interest shortly before cohort entry: (1) any OAC use during the 12 months baseline period, (2) presence of valvular diseases, repair, or replacement, venous thromboembolism, or joint replacement during the 12 months baseline period, (3) any stroke within 14 days before first NVAF diagnosis, (4) major surgery (i.e., hip fracture, cardiac surgery) or critical bleeding within 30 days before first NVAF diagnosis, (5) renal impairment stage 5 or end-stage renal diseases during the 12 months baseline period. All ICD codes for identification of these conditions can be found in Table S1, Supplementary materials.

Treatment strategies and assignments

In the hypothetical target trial, eligible individuals were randomly assigned to one of the following 5 treatment strategies: (Regimen 1) initiated OAC when CHA2DS2-VASc ≥ 1, (Regimen 2) initiated OAC when CHA2DS2-VASc ≥ 2, (Regimen 3) initiated OAC when CHA2DS2-VASc ≥ 4, (Regimen 4) initiated OAC when CHA2DS2-VASc ≥ 6, and (Regimen 5) never initiated OAC (reference group). In the emulation of target trial, cloning, censoring, and weighting approach were used to mimic the randomization [26]. OAC prescriptions (including warfarin and dabigatran, apixaban, rivaroxaban, edoxaban) were identified from Medicare Part D Prescription Drug Event (PDE) files using national drug code (NDC)) [27]. CHA2DS2-VASc scores were computed from Medicare claims during12 months before AFib diagnosis and monthly during follow-up, based on a composite of conditions including congestive heart failure (1 point), hypertension (1), age ≥ 75 (2 point), diabetes mellitus (1 point), prior stroke, TIA, or thromboembolism (2 point), vascular disease (e.g. peripheral artery disease, myocardial infarction, aortic plaque) (1 point), age 65–74 years (1 point), and sex category (1 point) [15].

Follow-up

The follow-up started at the initial NVAF diagnosis (index date) and ended at the occurrence of a study outcome, the end of administrative censoring (12 months after baseline), death (all-cause deaths from the SEER and Medicare files via the variables of “Date of Death Flag”), loss to follow-up (the earliest of 30 days after the end of continuous Medicare Part A, B, or D enrollment or enrollment in an HMO), or December 31, 2019, whichever came first.

Outcomes

The outcomes of interest were ischemic stroke and major bleeding. We defined major bleeding and ischemic stroke using validated algorithms defined by ICD-9-CM and ICD-10-CM codes in the primary diagnosis from Medicare medical claims files [28, 29, 30].

Covariates

Covariates selected from prior literature were adjusted in the analysis [24, 28, 31]. Time-fixed baseline covariates were extracted within 12-month period prior to first AFib diagnosis, including: demographics (index age, sex, race/ethnicity, calendar year, geographical region, urbanicity), socioeconomic factors (household median income, percentage of household with education level below high school, and Medicaid eligibility), comorbidity risk scores (CHA2DS2-VASc, HAS-BLED, and Comorbidity Scores SEER-Medicare version 2021 by NCI) [32], individual comorbidities (asthma/chronic obstructive pulmonary disease, hematological disorders, dementia, depression, thrombocytopenia, acute kidney disease (AKD), peptic ulcer disease), cancer characteristics (time from cancer diagnosis to the onset of AFib, cancer type, cancer stage, tumor grade, active cancer status [28, 31]), cancer treatment (radiation, and cancer-directed surgery, and potentially interacting antineoplastic agents), and medication history (angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers, calcium channel blockers, beta blockers, antiarrhythmic medications, diuretics, statin, pump proton inhibitors, and serotonin reuptake inhibitors). Socioeconomic factors such as household income and education level are available at the aggregate area level. If patients had more than one type of cancer before AFib diagnosis, we retained the most recent cancer diagnosis. Cancer treatments were obtained from diagnosis codes or procedures codes within 30 days before AFib diagnosis [33]. Due to high proportion of missing values, other cancer characteristics such as number of regional nodes examined, tumor size, TMN classification, and other cancer-type specific characteristics (i.e., hormone receptor status (HR) and human epidermal growth factor receptor 2 (HER2) for breast cancer or histologic type for lung cancer) were used for descriptive purpose but not adjusted in the models [34]. The following time-varying covariates were extracted monthly after AFib onset, including CHA2DS2-VASc score, HAS-BLED score, thrombocytopenia, AKD, radiation, cancer-directed surgery, and use of potentially interacting treatment with OACs. These variables may change over time and has an impact on outcomes and the OAC prescription in each month [35, 36]. All diagnosis codes, drug codes, and procedure codes for covariate ascertainment were described in Table S1, Supplementary materials. We used multiple imputation algorithms (fully conditional specification with logistic regression for categorical variables and predictive mean matching for continuous variables) to impute missing values (urbanicity, cancer summary stage, percentage of residents living below poverty, and percentage of non-high school graduates—Table S2, Supplementary Materials) [37].

Causal contrast

We computed the observational analog of per-protocol (PP) effects because cloning-censoring-weighting approach was used [38]. Those who were not compliant to their assigned treatment regimes were censored during follow-up.

Statistical analysis

Descriptive statistics such as mean and standard deviation (SD) for continuous variables, frequency count and percentage for categorical variables were used to describe the study sample. We quantified the incidence rates of ischemic stroke and major bleeding for each treatment strategy. In the main analysis, cloning-censoring-weighting procedure was used to estimate the treatment effect of 5 treatment strategies [26, 38]. Briefly, we created 5 copies for each individual’s person-time data, then assigned each copy to 5 treatment strategies. At baseline, replicates with baseline CHA2DS2-VASc score that did not comply with their assigned strategy were removed from the dataset. Next, replicates whose data were no longer consistent with their assigned strategy during follow-up were censored. To adjust for potential confounding during follow-up, unstabilized time-varying censoring weights were used. Cumulative weights at each time points due to protocol violation are the product of inverse probability of weights for treatment initiation (IPTWs) and inverse probability of censoring weights (IPCWs) due to loss to follow-up (See Technical Appendix). Total weights were truncated at 99th percentile to avoid extreme weights. To estimate the treatment effects for 5 strategies, we fitted a weighted pooled logistic regression estimated by generalized estimating equations (GEEs) with robust variance estimators. We obtained summary hazard ratios (HRs) with 95% confidence interval (95% CIs) and created weighted survival curves comparing four active treatment strategies with the reference strategy. Statistical analyses were conducted using SAS 9.4 (SAS Institute, Inc., Cary, NC, USA).

Subgroup analyses and sensitivity analyses

We conducted the following subgroup analyses: cancer type (breast, lung, prostate), cancer status at baseline (active, history), cancer stage (in situ, local, regional, and distant), and tumor grade (I, II, and III). In addition, a series of sensitivity analyses were conducted. First, we extended follow-up time to 36 months to explore long-term outcomes for each treatment strategies. Second, since metastatic cancer patients were removed from randomized control trials due to their short live expectancy, we excluded them in this sensitivity analysis [24, 25]. Third, we removed individuals with thrombocytopenia at baseline, since these patients are at elevated risk of bleeding and may not eligible for OAC initiation [39, 40]. Fourth, we further truncated weights at 95th percentile to test the robustness of the treatment effects to the presence of extreme weights [41, 42].