Bleeding Risk with Apixaban vs Rivaroxaban: A Reference Trial Emulation Predicting the Results of COBRRA-VTE and COBRRA-AF using US Healthcare Claims

Abstract

Background

Direct oral anticoagulants (DOACs) such as apixaban and rivaroxaban, have transformed the management of thromboembolic disorders. RCTs were initiated in 2017 (COBRRA-VTE) and 2021 (COBRRA-AF) to directly compare apixaban and rivaroxaban on the risk of bleeding in patients with venous thromboembolism (VTE) and atrial fibrillation (AF). We aimed to emulate the design and measurements of these ongoing trials and use US healthcare insurance claims data to predict their findings.

Methods

In cohort studies designed to emulate COBRRA-VTE (NCT05264168) and COBRRA-AF (NCT05256797) in three US healthcare insurance claims databases (Medicare, Optum Clinformatics ^®, Merative MarketScan), the risk of bleeding with apixaban was evaluated relative to the risk with rivaroxaban. Pre-treatment confounding factors were adjusted for via 1:1 propensity score matching. The primary outcome was a composite of major bleeding and clinically relevant non-major bleeding (defined using ICD-9 and ICD-10 diagnosis codes) recorded in the inpatient setting. Results across databases were pooled using fixed effects meta-analyses. Protocols were pre-registered on clinicaltrials.gov before inferential analyses were conducted.

Results

In the COBRRA-VTE emulation (N = 14,023 matched pairs, mean age = 67, 55% female), we observed a 31% decreased risk of bleeding for apixaban compared to rivaroxaban (incidence rate per 100 person-years apixaban: 12.8, rivaroxaban: 19.1; HR= 0.69; 95% CI=0.59 to 0.79). For the COBRRA-AF emulation (N = 176,990 matched pairs, mean age = 74, 50% female), we similarly observed a 31% decreased risk (incidence rate per 100 person-years apixaban: 6.4, rivaroxaban: 9.2; HR= 0.69; 95% CI=0.66 to 0.71) for apixaban compared to rivaroxaban.

Conclusions

In database studies that closely emulated the design and measurements of the COBRRA-VTE and COBRRA-AF trials, a substantially lowered risk of bleeding with apixaban was observed, consistent with the results of COBBRA-VTE (published after completion of our analyses), and indirect comparisons from other trials.

Introduction

Venous thromboembolism (VTE) and atrial fibrillation (AF) are the two most common health conditions that are managed with oral anticoagulation therapies¹. For decades, the standard of care was vitamin K antagonists (VKA) such as warfarin², and the introduction of direct oral anticoagulants (DOACs) in 2010 transformed the management of thromboembolic disorders. Apixaban and rivaroxaban target Factor Xa and have demonstrated improved safety and efficacy compared to VKAs in multiple studies.^3–7 These two drugs are the most frequently prescribed DOACs for VTE and AF.^8–10

In the absence of head-to-head RCTs comparing apixaban with rivaroxaban, network meta-analyses of RCTs and several non-randomized cohort studies conducted using healthcare databases consistently observed clinically meaningful lower risk of major bleeding for apixaban versus rivaroxaban.^11–16 Despite this body of evidence, investigators funded by the Ottawa Hospital research institute launched two RCTs to compare the safety of apixaban and rivaroxaban on the risk of major and clinically relevant non-major bleeding among patients with venous thromboembolism, COBRRA-VTE in 2017¹⁷ and atrial fibrillation COBRRA-AF in 2021.¹⁸ Both trials were ongoing at the time of this analysis (October 2024). The results from COBRRA-VTE were published in 2026¹⁹, while COBBRA-AF is expected to complete in December 2027.

The increasing availability of routinely collected and population-representative patient-level healthcare claims data provides opportunities to evaluate the comparative effectiveness and safety of drugs in non-randomized studies. By applying modern study design and analysis approaches, researchers can generate evidence that complements and expands upon the findings from RCTs, enabling a better understanding of the effects of drugs in diverse clinical settings,²⁰ including frail older adults,²¹ patients with kidney disease,²² or those at elevated risk of bleeding.^23–24 Moreover, robust evidence generated from well-designed non-randomized database studies can contribute to evidence generation regarding clinical questions which may not be answered with RCTs. The RCT-DUPLICATE (Randomized Controlled Trials Duplicated Using Prospective Longitudinal Insurance Claims: Applying Techniques of Epidemiology) initiative systematically conducted emulations of 32 RCTs that were meant for regulatory submissions to evaluate when and how database studies can come to causal conclusions, using concordance with RCT results as a benchmark.^25–30

As part of the RCT-DUPLICATE initiative, we emulated the design and measurements of the COBRRA-VTE and COBRRA-AF trials using US healthcare claims databases. The goal was to predict findings of the 2 COBRRA trials that were ongoing, without publicly available results at the time the study protocols were pre-registered.

Methods

Data sources

Data sources used for this study are available from the Centers of Medicare & Medicaid Services (CMS), Optum’s de-identified Clinformatics^® Data Mart Database (Clinformatics^®), and Merative Marketscan. Access restrictions apply to these data, which were used under license for this study. For both COBRRA-VTE and COBRRA-AF emulations, we used three U.S. healthcare claims data sources: two commercial health insurance claims databases, Optum’s de-identified Clinformatics^® Data Mart Database (Clinformatics^®) (2004 – 2020) and IBM MarketScan (2003 – 2019), as well as subsets of Medicare Parts A, B, and D (2009 – 2017) including all patients dispensed an oral anticoagulant. Each data source contained de-identified information for covered healthcare encounters of patients enrolled in participating healthcare insurance plans. The available data included demographic factors (age, sex), start and end dates of enrollment, dispensed medications with dates, dose, and days supply, and procedures and medical diagnoses with place of service and associated service dates. The date of death was captured with high completeness in Medicare data from the master beneficiary summary file or the vital status file. In the two commercial databases, death was captured from the social security death master index and in-hospital deaths. The study was approved by the Mass General Brigham Institutional Review Board (IRB # 2018P002966).

Study Design, outcome, and follow-up.

We identified groups of patients who were at least 18 years old and initiated apixaban or rivaroxaban starting from the apixaban approval date for the indication of VTE (August 21^st, 2014, for COBRRA-VTE) and AF (December 28^th, 2012, for COBRRA-AF) until end of availability of data in each database. The new-user active comparator cohort design used in this emulation has various strengths when analyzing secondary healthcare data which have been described in detail elsewhere.^31–33 New use was defined by a dispensing of apixaban or rivaroxaban after a washout of at least 180 days with medical and drug coverage and no dispensing of any DOAC or warfarin. The cohort entry date was the day of the first filled prescription of apixaban or rivaroxaban meeting both washout criteria and all other eligibility criteria. Patients were included if they had inpatient or emergency department diagnosis of venous thromboembolism within 14 days before or on cohort entry (COBRRA-VTE) or diagnoses of atrial fibrillation in any care setting within 180 days prior to or on cohort entry (COBRRA-AF). These conditions were identified using the International Classification of Diseases, Ninth Revision, and 10^th Revision (ICD-9 and ICD-10) codes (Table S1). Patients with any of the following characteristics in the 180 days before or on cohort entry were excluded: previous anticoagulation therapy, stage 4 or 5 chronic kidney disease, ESRD, dialysis, occurrence of pregnancy, cancer, or any other contraindication to anticoagulation (Figure 1). The eligibility criteria applied to the cohort were based on those of the COBRRA trials (Table S1).

Figure 1. Schematic of the study design.

Footnote – Rx = Prescription; INLC = Inclusion; EXCL = Exclusion.

The primary outcome was a composite of major bleeding and clinically relevant non-major bleeding using ICD-9 and ICD-10 diagnosis codes occurring in any position in the inpatient setting. All outcome definitions were based on previously validated algorithms with reported positive predictive values ranging from 86% to 97%.^{34, 35} We modified the major bleeding algorithm in our study by adding clinically relevant non-major bleeding, defined by inpatient ICD diagnosis codes for non-critical sites without transfusion (Table S2). In line with the RCTs, the COBRRA-VTE and COBRRA-AF emulations assessed the outcome in the first 3 months and 12 months follow-up period after initiation of anticoagulation therapy, respectively. The primary analysis used an “as treated” approach regarding follow up wherein study follow-up began the day after cohort entry and continued until the end of the study period, the end of continuous health plan enrollment, the occurrence of a study outcome, discontinuation of the index medication (with a 10-day allowable grace period and risk window), a switch to the comparator, initiation of another anticoagulant, admission to nursing home, or death. The duration of the index exposure was captured using the date of dispensation and the days supply. A 10-day gap was allowed between the end of a prior dispensation and a subsequent one to allow for non-adherence and potential continued biologic effect of exposure after the last pill is taken. Similarly, the duration of exposure was extended 10 days beyond the last dispensation in a sequence. This approach emulates a per-protocol analysis in the RCTs and minimizes exposure misclassification. For the secondary as-started analysis, emulating an intention-to-treat analysis of the RCTs, we did not censor for medication discontinuation, switching, or addition of other anticoagulants.

Statistical analysis

Within each database, we matched apixaban initiators with rivaroxaban initiators using 1:1 propensity score (PS) matching without replacement on their estimated PS using a nearest neighbor approach with a caliper of 1% to balance on >100 pre-exposure patient characteristics. Propensity scores were estimated via a multivariable logistic regression model incorporating all measured baseline covariates to predict the probability of receiving apixaban versus rivaroxaban. The covariates were measured during the six months before and including the cohort entry date and included demographics, calendar time, comorbidities, use of medications, indicators of healthcare utilization as well as indexes of overall health state such as combined comorbidity score and claims-based frailty index.^36–37 All covariates included in the PS appear in Table S3 and Table S4. Data analyses were conducted in each database separately and results were pooled using fixed effects meta-analyses.

After propensity score matching, we estimated incidence rates of the primary outcomes for each treatment group including their 95% Confidence intervals (CIs) and fit a Cox proportional hazard models to estimate hazard ratios (HRs) with their 95% CIs. Additionally, we plotted the cumulative incidence of major and clinically relevant non-major bleeding (Figure 2). Trial emulations were conducted using the validated Aetion Evidence Platform^{38, 39} with supplemental programming using SAS 9.4 and Cran R v4.1.1.

Figure 2. Cumulative incidence of the composite major bleeding and clinically-relevant non-major bleeding outcome before and after propensity score matching for both COBRRA-VTE and COBRRA-AF emulations.

A new-user active comparator design with propensity score (PS) matching was used to estimate treatment effects across two emulations:

• COBRRA-VTE (i): In the PS-matched population, curves diverged early, with apixaban showing a lower incidence over a 90-day follow-up period (N = 14,023 per group).
• COBRRA-AF (ii): In the PS-matched population, the lower cumulative incidence for apixaban was maintained over a 365-day follow-up period (N = 176,990 per group).

The results consistently indicate that apixaban was associated with a lower cumulative incidence of outcome events relative to rivaroxaban in both unmatched and PS-matched pooled populations.

Development of the study protocol

Several feasibility and data checks were conducted during the development of the study protocol, prior to pre-registration on clinicaltrials.gov.⁴⁰ These checks included assessing whether there was at least equal statistical power in the database study as in the relevant COBRRA trial for detecting superiority of the primary end point. The statistical power was estimated from the total number of event counts without stratifying by treatment group. Additionally, we assessed the pre- and post-matching balance on baseline covariates between the study arms by calculating standardized differences, reviewing the overlap in the PS distributions of two arms, and the c-statistics before and after PS matching (see appendix for Tables S3 & S4, Tables S5 & S6). All details of these data and feasibility checks are included in the study protocols that were registered at clinicaltrials.gov before the inferential analyses were started (NCT05264168 registered on 2022-03-03, NCT05256797 registered on 2022-02-25).

To contextualize our findings, we compared results for the primary outcome of major and clinically relevant non-major bleeding in our study to existing network meta-analyses of RCTs and the results of the COBRRA-VTE trial, which were published after completion of the analyses. Agreement between the trial and database study estimates was evaluated using three prespecified binary metrics previously used in the RCT-DUPLICATE initiative⁴¹:

1. Statistical agreement, defined as both the database study and trial estimates and their 95% confidence intervals lying on the same side of the null;

2. Estimate agreement, defined as the estimate of the database study falling within the 95% confidence interval of the trial;

3. Standardized difference agreement, defined as an absolute standardized difference |Z| < 1.96, with $\text{Z}=\frac{{\widehat{Ѳ}}_{RCT}-{\widehat{Ѳ}}_{RWE}}{\sqrt{{\widehat{\sigma }}_{RCT}^2+{\widehat{\sigma }}_{RWE}^2}}$) where ${\widehat{Ѳ}}_{RCT}$ and ${\widehat{Ѳ}}_{RWE}$ are the treatment effect estimates with associated variances ${\widehat{\sigma }}^2$.

RESULTS

COBRRA-VTE emulation

In the COBRRA-VTE emulation, we identified 41,337 eligible patients with VTE, initiating either apixaban (N=17,913) or rivaroxaban (N=23,424). In the unmatched cohort, patients initiating apixaban were older than rivaroxaban initiators (Mean age (SD): 68.0 (14.3) years vs. 64.5 (14.6) years) and they had a slightly higher mean combined comorbidity score (Mean score (SD): 2.82 (2.28) vs 2.09 (2.04)). The PS-matched population had 14,023 matched pairs with well-balanced baseline characteristics (Table 1 and Table S3) and similar PS distributions with c-statistics in all three databases moving from 0.7 to 0.5 after matching (Table S5). Baseline characteristics of the matched cohorts for each database are included in Table S7. The median (IQR) follow-up time was 79 (38 – 90) days. Details of follow-up time and censoring reasons in each databases are presented in Table S8.

Table 1:

COBRRA-VTE and COBRRA-AF trial emulation – Selected^* Baseline Characteristics of Patients after 1:1 Propensity Score Matching.

Patient characteristics	COBRRA-VTE Emulation			COBRRA-AF Emulation
	Apixaban N=14,023	Rivaroxaban N=14,023	St. Diff	Apixaban N= 176,990	Rivaroxaban N= 176,990	St. Diff
Age, mean (SD), years	66.77 (14.25)	66.77 (13.85)	0.00	73.88 (9.75)	73.91 (9.65)	0.00
Age categories; n (%)
18 – 54	2,964 (21.1%)	2,927 (20.9%)	0.00	7,970 (4.5%)	7,885 (4.5%)	0.00
55 – 64	2,244 (16.0%)	2,276 (16.2%)	0.01	18,361 (10.4%)	18,391 (10.4%)	0.00
65 – 74	3,864 (27.6%)	3,884 (27.7%)	0.00	62,360 (35.2%)	62,366 (35.2%)	0.00
>= 75	4,951 (35.3%)	4,936 (35.2%)	0.00	88,299 (49.9%)	88,348 (49.9%)	0.00
Males sex; n (%)	6,315 (45.0%)	6,273 (44.7%)	0.01	88,972 (50.3%)	88,932 (50.2%)	0.00
Race; n (%)
White	8,209 (58.5%)	8,254 (58.9%)	0.01	127,949 (72.3%)	127,937 (72.3%)	0.00
Black	1,377 (9.8%)	1,334 (9.5%)	0.01	7,011 (4.0%)	7,016 (4.0%)	0.00
Hispanic	355 (2.5%)	340 (2.4%)	0.01	4,162 (2.4%)	4,140 (2.3%)	0.01
Region; n (%)
Northeast	2,488 (17.7%)	2,516 (17.9%)	0.01	33,033 (18.7%)	32,987 (18.6%)	0.00
South	6,178 (44.1%)	6,135 (43.7%)	0.01	69,779 (39.4%)	69,835 (39.5%)	0.00
Midwest	3,123 (22.3%)	3,188 (22.7%)	0.01	40,027 (22.6%)	40,214 (22.7%)	0.00
West	2,213 (15.8%)	2,165 (15.4%)	0.01	33,957 (19.2%)	33,746 (19.1%)	0.00
Burden of comorbidities in past 180 days
Combined comorbidity score†; mean (SD)	2.48 (2.11)	2.48 (2.12)	0.00	1.89 (1.86)	1.90 (1.86)	0.01
Frailty score‡; mean (SD)	0.19 (0.07)	0.19 (0.07)	0.00	0.18 (0.06)	0.18 (0.06)	0.00
CHADS2 score; mean (SD)	0.76 (0.97)	0.75 (0.95)	0.01	1.22 (1.18)	1.22 (1.18)	0.00
Cardiovascular Comorbidities observed in past 180 days; n (%)
Diabetes mellitus without mention of complications	2,816 (20.1%)	2,820 (20.1%)	0.00	40,852 (23.1%)	40,863 (23.1%)	0.00
Hypertension	9,575 (68.3%)	9,565 (68.2%)	0.00	141,665 (80.0%)	141,738 (80.1%)	0.00
Hyperlipidemia	6,677 (47.6%)	6,648 (47.4%)	0.00	91,604 (51.8%)	91,776 (51.9%)	0.00
Acute Myocardial Infarction	663 (4.7%)	653 (4.7%)	0.00	3,895 (2.2%)	3,909 (2.2%)	0.00
ACS/unstable angina	460 (3.3%)	469 (3.3%)	0.00	3,729 (2.1%)	3,737 (2.1%)	0.00
Old Myocardial Infarction	575 (4.1%)	569 (4.1%)	0.00	6,393 (3.6%)	6,490 (3.7%)	0.01
Stable angina	327 (2.3%)	347 (2.5%)	0.01	5,820 (3.3%)	5,860 (3.3%)	0.00
Coronary atherosclerosis and other forms of chronic ischemic heart disease	2,343 (16.7%)	2,319 (16.5%)	0.01	42,366 (23.9%)	42,366 (23.9%)	0.00
Previous cardiac procedure (CABG or PTCA or Stent) + History of CABG or PTCA	766 (5.5%)	755 (5.4%)	0.00	10,845 (6.1%)	10,930 (6.2%)	0.00
Ischemic stroke	892 (6.4%)	879 (6.3%)	0.00	18,172 (10.3%)	18,262 (10.3%)	0.00
Hemorrhagic stroke	28 (0.2%)	25 (0.2%)	0.00	258 (0.1%)	275 (0.2%)	0.03
Transient Ischemic Attack (TIA)	243 (1.7%)	244 (1.7%)	0.00	6,538 (3.7%)	6,588 (3.7%)	0.00
Other cerebrovascular disease	417 (3.0%)	397 (2.8%)	0.01	5,483 (3.1%)	5,492 (3.1%)	0.00
Late effects of cerebrovascular disease	334 (2.4%)	325 (2.3%)	0.01	4,227 (2.4%)	4,294 (2.4%)	0.00
Heart failure (CHF)	1,995 (14.2%)	1,947 (13.9%)	0.01	32,588 (18.4%)	32,693 (18.5%)	0.00
Peripheral Vascular Disease (PVD) or PVD Surgery	1,095 (7.8%)	1,099 (7.8%)	0.00	11,198 (6.3%)	11,285 (6.4%)	0.00
Other cardiac dysrhythmia	2,069 (14.8%)	2,065 (14.7%)	0.00	132,861 (75.1%)	132,733 (75.0%)	0.00
Cardiac conduction disorders	1,119 (8.0%)	1,116 (8.0%)	0.00	13,819 (7.8%)	13,707 (7.7%)	0.00
Other CVD events	3,814 (27.2%)	3,840 (27.4%)	0.00	53,825 (30.4%)	53,919 (30.5%)	0.00
Cardiovascular medications used in past 180 days; n (%)
ACE inhibitors	3,359 (24.0%)	3,345 (23.9%)	0.00	54,498 (30.8%)	54,500 (30.8%)	0.00
ARBs	2,203 (15.7%)	2,269 (16.2%)	0.01	41,663 (23.5%)	41,567 (23.5%)	0.00
Alpha blockers	1,227 (8.7%)	1,186 (8.5%)	0.01	20,488 (11.6%)	20,444 (11.6%)	0.00
Beta blockers	3,907 (27.9%)	3,920 (28.0%)	0.00	119,643 (67.6%)	119,713 (67.6%)	0.00
Loop Diuretics	1,877 (13.4%)	1,876 (13.4%)	0.00	37,654 (21.3%)	37,713 (21.3%)	0.00
Other diuretics||	367 (2.6%)	384 (2.7%)	0.01	7,906 (4.5%)	7,895 (4.5%)	0.00
Use of nitrates	512 (3.7%)	513 (3.7%)	0.00	9,650 (5.5%)	9,594 (5.4%)	0.00
Use of statins	4,933 (35.2%)	4,928 (35.1%)	0.00	90,874 (51.3%)	90,893 (51.4%)	0.00
Use of other lipid-lowering drugs	598 (4.3%)	605 (4.3%)	0.00	7,385 (4.2%)	7,287 (4.1%)	0.01
Calcium channel blockers	3,023 (21.6%)	3,064 (21.8%)	0.00	67,972 (38.4%)	68,026 (38.4%)	0.00
Digoxin	66 (0.5%)	63 (0.4%)	0.01	13,781 (7.8%)	13,826 (7.8%)	0.00
Anti-arrhythmic	55 (0.4%)	65 (0.5%)	0.01	31,224 (17.6%)	31,299 (17.7%)	0.00
Measures of healthcare utilization during past 180 days
No. of Hospitalizations; mean (SD)	1.23 (0.60)	1.23 (0.60)	0.00	0.33 (0.58)	0.33 (0.59)	0.00
No. of Emergency Department (ED) visits; mean (SD)	2.34 (3.99)	2.31 (5.08)	0.01	0.87 (1.70)	0.88 (2.90)	0.00
No. of Office visits; mean (SD)	8.02 (9.06)	8.00 (13.09)	0.00	9.75 (9.43)	9.71 (15.73)	0.00
No. of distinct medications; mean (SD)	9.00 (5.27)	9.01 (7.85)	0.00	8.62 (4.38)	8.63 (7.28)	0.00
No. of Cardiologist visits; mean (SD)	2.13 (3.67)	2.12 (5.28)	0.00	4.50 (5.23)	4.49 (9.39)	0.00
No. of Internal medicine/family medicine visits; mean (SD)	9.51 (10.98)	9.53 (16.90)	0.00	7.74 (10.20)	7.74 (16.60)	0.00
Bone mineral density; n (%)	442 (3.2%)	450 (3.2%)	0.00	6,290 (3.6%)	6,356 (3.6%)	0.00
Colonoscopy; n (%)	511 (3.6%)	523 (3.7%)	0.01	7,253 (4.1%)	7,263 (4.1%)	0.00
Flu vaccine; n (%)	2,950 (21.0%)	2,981 (21.3%)	0.01	46,584 (26.3%)	46,627 (26.3%)	0.00
Mammogram; n (%)	1,211 (8.6%)	1,196 (8.5%)	0.00	16,443 (9.3%)	16,518 (9.3%)	0.00
PSA test or Prostate exam for DRE; n (%)	1,237 (8.8%)	1,244 (8.9%)	0.00	22,517 (12.7%)	22,417 (12.7%)	0.00
Pap smear; n (%)	497 (3.5%)	501 (3.6%)	0.01	3,860 (2.2%)	3,877 (2.2%)	0.00
Copay per prescription#; mean (SD), $	16.90 (29.02)	17.07 (46.83)	0.00	79.94 (99.03)	79.79 (180.57)	0.00
Copay for pharmacy cost**; mean (SD), $	169.97 (337.36)	171.10 (478.78)	0.00	183.69 (303.14)	183.19 (507.28)	0.00
Insurance type; n (%)
Commercial	3,944 (28.1%)	3,961 (28.2%)	0.00	23,524 (13.3%)	23,518 (13.3%)	0.00
Medicare	2,695 (19.2%)	2,678 (19.1%)	0.00	41,113 (23.2%)	41,119 (23.2%)	0.00
Low-income indicator††; n (%)	472 (3.4%)	451 (3.2%)	0.01	4,108 (2.3%)	4,142 (2.3%)	0.00

^*Full tables of baseline characteristics provided in Supplemental Tables S3 and S4

^†36. Gagne JJ, Glynn RJ, Avorn J, Levin R, Schneeweiss S. A combined comorbidity score predicted mortality in elderly patients better than existing scores. J Clin Epidemiol. 2011;64(7):749–759.

^‡37. Kim DH, Schneeweiss S, Glynn RJ, Lipsitz LA, Rockwood K, Avorn J. Measuring frailty in Medicare data: development and validation of a claims-based frailty index. J Gerontol A Biol Sci Med Sci. 2018;73(7):980–987.

^||Other diuretics include: Eplerenone, spironolactone and triamterene

^#Average copay amount per prescription for each patient

^**Total copay amounts for all pharmacy-related services (based on all prescription claims) for each patient

^††This indicates beneficiaries who receive the Low-income subsidy, including those enrolled in Medicare Advantage Prescription Drug (MAPD) plan as well those enrolled in an MAPD plan and are eligible for both Medicare and Medicaid (Dual eligible).

In the primary as-treated analysis, the incidence rate for the composite major bleeding and clinically relevant non-major bleeding outcome was 12.8 per 100 person-years (320 events) among apixaban initiators and 19.1 per 100 person-years (440 events) among rivaroxaban initiators. We observed a 31% decreased hazard (HR = 0.69; 0.59 to 0.79) or 6.3 fewer events (RD −6.3; −8.6 to −4.0) per 100 person-years of apixaban use versus rivaroxaban (Table 2 and Figure 3). In the individual databases the results were: Optum Clinformatics^®: HR=0.71 (0.52 to 0.97), Marketscan: HR=0.90 (0.62 to 1.31) and Medicare: HR=0.65 (0.54 to 0.78) (Table S9). In the intention-to-treat (ITT) analysis that followed patients for up to a year, the matched population had 14,079 patients in the apixaban and rivaroxaban arms, and we observed a 29% decreased risk of major or clinically relevant non major bleeding [HR= 0.71 (0.62–0.81)] (Table S10).

Table 2.

Event rates and effect estimates for as-treated analyses of Primary outcome^* and two secondary outcomes ^†, ^‡

Primary outcome: Major and clinically relevant non-major bleeding (in any position) after 1:1 PS-matching
Trial Name	Apixaban			Rivaroxaban
	N	Events	Incidence Rate / 100 Person-years (95% CI)	N	Events	Incidence Rate / 100 Person-years (95% CI)	Rate Difference / 100 Person-years (95% CI)	Adjusted HR (95% CI)
COBRRA-VTE	14,023	320	12.8 (11.5, 14.3)	14,023	440	19.1 (17.4, 21.0)	−6.3 (−8.6, −4.0)	0.69 (0.59, 0.79)
COBRRA-AF	176,990	4,865	6.4 (6.2, 6.6)	176,990	6,647	9.2 (9.0, 9.5)	−2.8 (−3.1, −2.5)	0.69 (0.66, 0.71)
Secondary outcome I: Major bleeding (in any position) after 1:1 PS-matching
Trial Name	Apixaban			Rivaroxaban
	N	Events	Incidence Rate / 100 Person-years (95% CI)	N	Events	Incidence Rate / 100 Person-years (95% CI)	Rate Difference / 100 Person-years (95% CI)	Adjusted HR (95% CI)
COBRRA-VTE	14,023	118	4.7 (3.9, 5.6)	14,023	184	7.9 (6.8, 9.2)	−3.2 (−4.7, −1.8)	0.59 (0.47, 0.75)
COBRRA-AF	176,990	1,860	2.4 (2.3, 2.5)	176,990	2,891	4.0 (3.8, 4.1)	−1.6 (−1.7, −1.4)	0.60 (0.56, 0.64)
Secondary outcome II: (i) Recurrent VTE for COBRRA-VTE & (ii) Stroke for COBRRA-AF after 1:1 PS-matching
Trial Name	Apixaban			Rivaroxaban
	N	Events	Incidence Rate / 100 Person-years (95% CI)	N	Events	Incidence Rate / 100 Person-years (95% CI)	Rate Difference / 100 Person-years (95% CI)	Adjusted HR (95% CI)
COBRRA-VTE	14,023	663	27.0 (25.0, 29.1)	14,023	694	30.5 (28.3, 32.9)	−3.5 (−6.6, −0.5)	0.91 (0.82, 1.02)
COBRRA-AF	176,990	695	0.9 (0.8, 1.0)	176,990	782	1.1 (1.0, 1.2)	−0.2 (−0.3, −0.1)	0.94 (0.84, 1.04)

^*Major and clinically relevant non-major bleeding (in any position).;

^†Major bleeding (in any position);

^‡Recurrent VTE for COBRRA-VTE & Stroke for COBRRA-AF after 1:1 PS-matching

Figure 3. Effect estimates of COBRRA trial emulations vs network meta-analyses and COBRRA-VTE comparing apixaban to rivaroxaban for the primary composite bleeding outcome.

Analyses indicate that the risk of bleeding was lower among apixaban initiators compared to rivaroxaban initiators in both COBRRA-VTE and COBRRA-AF emulations. These findings were seen to be consistent when compared against network meta-analyses and the COBRRA-VTE trial results.

CI - confidence interval; HR - hazard ratio.

COBRRA-AF emulation

In the COBRRA-AF emulation, after applying the eligibility criteria, we identified 495,701 eligible patients with AF, initiating either apixaban (N=291,020) or rivaroxaban (N=204,681). In our cohort, compared to rivaroxaban initiators, patients initiating apixaban were slightly older (Mean age (SD): 75.2 (9.8) years vs. 73.2 (9.9) years) and had a higher mean combined comorbidity score (Mean score (SD): 2.14 (2.04) vs 1.84 (1.83). In the PS-matched population, there were 33,730, 30,907 and 112,353 matched pairs within the Optum Clinformatics^®, Marketscan and Medicare databases, respectively. Baseline characteristics of the matched cohorts for each database are included in Table S11. In the pooled cohort (N=176,990 apixaban and rivaroxaban matched pairs) baseline characteristics were well-balanced (Table 1 and Table S4) with similar PS distributions and the c-statistics in the Optum Clinformatics^® and Marketscan databases moved from 0.7 to 0.5 after matching; they moved from 0.6 to 0.5 after matching in the CMS database (Table S6). The median and interquartile range (IQR) for follow-up time was 98 (38 – 222) days. Details of follow-up time and censoring reasons in each databases are presented in Table S12.

In the primary as-treated analysis, the incidence rate for the composite major bleeding and clinically relevant non-major bleeding outcome was 6.4 per 100 person-years (4,865 events) among apixaban initiators and 9.2 per 100 person-years (6,647 events) among rivaroxaban initiators. Thus, we observed a 31% decreased hazard (HR = 0.69; 0.66 to 0.71) or 2.8 fewer events (RD −2.8; −3.1 to −2.5) per 100 person-years with exposure to apixaban versus rivaroxaban (Table 2 and Figure 3). Risk reductions for major and clinically relevant non major bleeding were consistent across the individual databases: Optum Clinformatics^®: HR=0.77 (0.70 to 0.84), Marketscan: HR=0.71 (0.64 to 0.80) and Medicare: HR=0.68 (0.66 to 0.72) (Table S9). In the intention-to-treat (ITT) analysis that followed patients for up to a year, there were 177,423 matched pairs, and we observed a 26% risk reduction [HR= 0.74 (0.72–0.77)] (Table S13).

For the primary outcome of major and clinically relevant non-major bleeding in our study, our results were consistent with existing network meta-analyses of RCTs, showing a substantially lowered risk of bleeding events for patients treated with apixaban (Figure 3). When compared to the results of the COBRRA-VTE trial (RCT: 0.46 (0.33 to 0.65), database: 0.69 (0.59 to 0.79), the findings were consistent in showing a strong protective effect for apixaban and our results met the statistical significance agreement metric. However, because the observed effect was stronger in the trial than what was observed in clinical practice, and the point estimate of the database study was slightly outside of the tight confidence intervals for the trial, neither the estimate agreement metric nor the standardized difference agreement metric was met.

DISCUSSION

We emulated the designs of two ongoing randomized trials that directly compare apixaban to rivaroxaban regarding the risk of major bleeding and clinically relevant non-major bleeding in patients with venous thromboembolism (COBRRA-VTE) and atrial fibrillation (COBRRA-AF)to predict the results of these trials. Our studies were conducted using non-randomized healthcare claims database analyses. Before pre-registering the study protocols, we demonstrated evidence of adequate statistical power and ability to achieve balance on all relevant confounders. Such data checks were conducted to understand the data quality and completeness and to assess our ability to control confounding before engaging in the actual prediction of the two ongoing COBRRA trials. We observed a one-third lower risk of major and clinically relevant non-major bleeding in apixaban initiators compared to rivaroxaban initiators with VTE and AF.

Although database studies are recognized as an approach to generate evidence regarding the safety and efficacy of medications in routine clinical practice, some stakeholders continue to be skeptical as to whether these types of studies can generate valid causal conclusions. The view that studies that are not randomized are intractably biased^42–44 is a consequence of the prevalence of demonstrably flawed database studies resulting in difficulty judging the quality of a given database study^45–47. Such circumstances lead to situations such as the randomized head-to-head DOAC comparisons in the COBRRA trials where there is a plethora of evidence from network meta-analyses and observational studies with low risk of bias that consistently demonstrate the reduction in risk of bleeding for apixaban versus rivaroxaban.

Our database study predicting the results of the COBRRA-VTE trial provided evidence of substantially lowered risk of major bleeding for apixaban, similar to the trial. The results of COBRRA-AF are not yet available for comparison. However, our database study results are also consistent with existing network meta-analyses of RCTs that also observed strong protective effects of apixaban versus rivaroxaban. For the primary outcome of major and clinically relevant non-major bleeding, two network meta-analyses of RCTs comparing apixaban to rivaroxaban in patients with atrial fibrillation^11,12 and two network meta-analysis in patients with VTE^13,14 obtained effect estimates of similar magnitude to our COBRRA emulations.

In spite of this body of evidence, as noted by COBRRA trial investigators on clinicaltrials.gov, “there have been no direct head-to-head comparisons of these two anticoagulants, meaning comparative safety data is not available. Increasing use of DOACs … highlight the need for a comparison trial …”⁴⁸ (emphasis added). The risk of major and clinically relevant non-major bleeding from rivaroxaban in the ROCKET-AF⁴⁹ trial was 14.9% per year and assuming that the relative risk of bleeding for apixaban versus rivaroxaban is approximately 0.66, as has been consistently shown by multiple network meta-analyses and non-randomized database studies, this would result in a risk of approximately a 9.8% per year for apixaban. This translates to an estimated number needed to harm of 20 and a projected 75 excess bleeding events in the COBRRA-AF trial.

Regulatory authorities are increasingly considering non-randomized studies as part of a continuum of types of study designs that can be used together in evaluating drug safety and effectiveness.^50–52 Such approaches can encourage the appropriate use of non-randomized study designs to complement evidence from RCTs, accelerate evidence development, and help reduce research waste.

There are limitations to the specific trial emulations reported in this study. First, there are differences in measurement and ascertainment of study parameters between those collected during the trial as compared to definitions applied in the database study. In the absence of a publicly available protocol or other information about how major and clinically relevant non-major bleeding was defined for the COBRRA trials, our team had to make a decision and chose to base our outcome algorithm on an algorithm previously used by FDA investigators that was based on inpatient diagnoses in critical sites or inpatient diagnoses in non-critical sites with procedure codes for transfusion (Table S2)^{34, 35}. Although this algorithm was found to have a high positive predictive value, it has various limitations in capturing major bleeding such as requiring a primary hospital discharge diagnosis and thus not detecting some bleeding events coded as a secondary diagnosis. Also, the specific sites and bleeding definition may or may not align with the definitions used by the COBRRA investigative team. Prior work has demonstrated the sensitivity of results for studies of direct oral anticoagulants to the definition of major bleeding⁵¹. Second, as with all observational research, residual confounding due to some unmeasured characteristics cannot be completely ruled out. However, our prespecified data checks demonstrated that the new-user active comparator study design equipped with PS-adjustment was able to balance >100 clinically relevant variables, substantially reducing concerns of confounding by unmeasured covariates.

CONCLUSION

In this large cohort study designed to predict the findings of two ongoing DOAC trials (COBRRA-VTE and COBRRA-AF) before their completion, we observed that apixaban was meaningfully safer than rivaroxaban with regards to a composite bleeding outcome in patients with AF and VTE. The observed decreased risk of bleeding for apixaban compared to rivaroxaban is consistent with results from COBBRA-VTE and indirect comparisons of other RCTs. When the results of the COBRRA-AF trial become available, we will assess the degree to which our emulation was successful in predicting the trial findings and attempt to understand any observed divergence in results.

Supplementary Material

Tables S1 – S13

Key Points Summary.

What is Known:

• Published network meta-analyses and non-randomized database studies, suggest that apixaban has a one-third lower risk of major and clinically relevant non-major bleeding compared to rivaroxaban in patients with AF and VTE.

• Two RCTs comparing apixaban to rivaroxaban to assess the risk of major bleeding and clinically relevant non-major bleeding are currently ongoing.

What the Study Adds:

• Using the RCT-DUPLICATE design, data, and analytics framework, we conducted database studies to closely emulate the designs and measurements of the ongoing COBRRA-VTE and COBRRA-AF trials to predict their results.

• Our results were concordant with previously published network meta-analyses and nonrandomized database studies.

• Our study highlights the role of well-designed database studies using fit-for-use data to generate timely results that complement RCT evidence.

Transparency statement:

This study was pre-registered prior to conducting analyses at CinicalTrial.gov (NCT05264168 and NCT05256797). The initial protocol and subsequent amendments are also included in the ClinicalTrial.gov submission. All analyses for these trial emulations were conducted using the validated Aetion Evidence Platform and the complete reports are available at https://osf.io/dvne9/. The study used de-identified data covered by data use agreements and received approval from the Mass General Brigham Institutional Review Board (IRB # 2018P002966). Access to the de-identified source data may be granted to authorized researchers via application and licensing with the data providers: Medicare [resdac@umn.edu, https://resdac.org/research-identifiable-files-rif-requests], Optum’s de-identified Clinformatics^® Data Mart Database (Clinformatics^®) [connected@optum.com, https://www.optum.com/business/solutions/life-sciences/real-world-data/claims-data.html] and Merative Marketscan [https://www.merative.com/contact?]. The manuscript is aligned with guidance from the STrengthening the Reporting of OBservational Studies in Epidemiology (STROBE) checklist.

Non-standard Abbreviations and Acronyms:

DOACs

Direct oral anticoagulants

VTE

Venous thromboembolism

AF

Atrial fibrillation

HR

Hazard ratio

VKA

Vitamin K antagonists

CMS

Centers of Medicare & Medicaid Services

ESRD

End stage renal disease

PS

Propensity score

HR

Hazard ratio

CI

Confidence Intervals