Introduction

Atrial fibrillation (AF) is a common type of cardiac arrhythmia with an estimated prevalence of 3.3% in UK adults aged ≥35 years [1]. AF is a risk factor for stroke; patients with AF have a 5-fold increased risk of stroke compared with people without AF [2], and around a quarter of all strokes are attributed to this arrhythmia [3]. In addition, increased levels of mortality, morbidity, and disability with longer hospital stays are observed in stroke patients with AF compared with stroke patients without AF [4,5].

Pharmacological therapy recommended to reduce the risk of stroke in AF includes the use of oral anticoagulants (OACs). The introduction of direct oral anticoagulants (DOACs) for AF since 2012 in the United Kingdom provided a choice of treatment alongside the older OAC class of vitamin K antagonists (VKAs), such as warfarin, which has been available for over 60 years. The VKA OACs require regular monitoring of international normalised ratio (INR) to keep patients in the optimal therapeutic range (typically 2.0 to 3.0) in which risk of both ischemic and bleeding events are minimised [6]. A patient may require dose adjustments to stay within their INR target range. A key measure of quality of warfarin treatment is, therefore, the time in therapeutic range (TTR), which estimates the proportion of time a patient has spent with INR within optimal range. A TTR of 0.75 or greater is often considered as indicating optimal INR control and suggests a patient is spending a high proportion of their time in their INR target range.

ARISTOTLE was a pivotal randomised controlled trial (RCT) of the DOAC apixaban designed to demonstrate noninferiority compared with warfarin in the prevention of stroke or systemic embolism (SE) in patients with AF. The results demonstrated superiority of apixaban over warfarin for both prevention of stroke/SE and safety (major bleeding) [7]. Results in the European Union patient subset from the trial suggested the observed superiority of apixaban might be dependent on how well warfarin therapy was managed in the comparator group [8], an analysis that has not yet been performed outside of trial settings. In the National Institute for Health and Care Excellence (NICE) review of ARISTOTLE, several professional groups noted the TTR of warfarin users in ARISTOTLE may be lower than what is typical in UK clinical practice [9].

Treatment guidelines for DOACs are based on evidence from RCTs; however, it is unclear whether these results extend to patient groups typically excluded from trials such as those with increased bleeding risk or severe comorbidities. While there have been a number of previous studies of DOAC effectiveness using noninterventional data, there remains uncertainty on whether the data sources and methods used have fully accounted for the lack of treatment randomisation and issues such as selection bias and confounding. Comparing results from real-world studies with RCT results is challenging due to differences in patient populations, treatment adherence, and study design. However, reference trial emulation involves use of an existing named RCT to (1) inform observational study design and (2) benchmark results against, providing confidence in validity of the selected observational methods and data [10–13]. The noninterventional analysis methods can then be applied, under a set of assumptions, to reliably estimate effects in groups of patients with AF who would have been excluded from (or underrepresented in) the reference trial [14] such as patients aged >80 that were underrepresented in ARISTOTLE compared with patients with AF in UK clinical practice and patients with increased bleeding risk that were excluded by the trial eligibility criteria.

There is increasing interest in trial emulation using observation data and in the application of recent developments in pharmacoepidemiology methods involving the inclusion of prevalent users. This study used a framework that involved coarsened exact matching to select patients matching the trial population on aggregate, and sampling of prevalent users in a way that avoids selection bias and emulates the process of screening into an RCT, to construct a cohort of patients similar to the target trial population that included both new and prevalent users. This methodological approach could be adapted to a variety of treatments and different therapeutic areas.

This study sought to (1) create an ARISTOTLE-analogous cohort using routinely collected primary and secondary care data in the UK, (2) benchmark results obtained in the ARISTOTLE-analogous cohort with ARISTOTLE results, and (3) explore whether apixaban treatment-effects in clinical practice are influenced by how well warfarin therapy is controlled.

Materials and methods

This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 STROBE Checklist).

Patient selection

Step 1: Application of trial eligibility criteria to patients in CPRD.

We first selected HES-linked patients registered in CPRD Aurum between 1 January 2013 and 31 July 2019, who had at least 6 months between registration and the index date. ARISTOTLE recruited both new (warfarin-naïve) and prevalent (warfarin-experienced) users of warfarin with randomisation stratified on prior warfarin (or other VKA) exposure status (warfarin naïve or experienced). To be classified as warfarin-naïve patients were required to have no evidence of exposure to warfarin or other VKA in the 5 years prior to the index date. To enable selection of a similar cohort of patients in CPRD Aurum (including both new and prevalent users of warfarin), the following process was used in determining index date:

- apixaban users

index date = first prescription of apixaban in the study period;

apixaban user classified as warfarin-naïve or warfarin-experienced at this date

- warfarin users

for new users of warfarin: index date = first prescription of warfarin in the study period;

for prevalent users of warfarin: a pool of potential index dates was selected containing all prescription dates in the study period, with index date selected at the later treatment-history sampling stage (see step 3).

ARISTOTLE eligibility criteria (Table A2 in S1 Appendix) [7] were applied giving a trial-eligible cohort for apixaban users, a trial-eligible cohort of new users of warfarin, and a pool of potential index dates (with all potential index dates kept in regardless of ARISTOTLE eligibility at this stage) for warfarin continuers (prevalent warfarin users).

Step 2: Selection of apixaban trial-analogous patients in CPRD.

We selected a subset of the CPRD Aurum trial-eligible apixaban patients that better matched the ARISTOTLE apixaban participants based on aggregate summaries for the following key ARISTOTLE baseline characteristics:

• Age
• Sex
• Congestive heart failure or left ventricular systolic dysfunction
• Hypertension requiring treatment
• Diabetes mellitus
• Prior stroke/transient ischaemic attack (TIA)/SE
• Level of renal impairment
• Prior VKA/warfarin exposure

To characterise the baseline patient characteristics of ARISTOTLE, we used the key publication of the trial results [7], discussion of trial results by regulatory bodies [8,9,16], and publications on the trial presenting cross-tabulations on key characteristics [17,18].

An ARISTOTLE-analogous cohort of CPRD Aurum apixaban patients was then selected using a modified form of coarsened exact matching [19] (see S1 Appendix for details).

Step 3: Matching of apixaban trial-analogous patients to warfarin trial-eligible patients in CPRD.

To emulate ARISTOTLE, which stratified randomisation on prior VKA exposure status, patients in the CPRD cohort were matched separately within the VKA-naïve and VKA-experienced strata. A 3-step procedure, based on methods proposed by Suissa and colleagues [20] and Webster-Clark and colleagues [21], was used to select and match patients in the VKA-experienced strata while avoiding selection bias; this procedure is summarised in Fig 1 and described in S1 Appendix.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Matching of apixaban trial-analogous patients to warfarin trial-eligible patients *This method has been found in a simulation study [21] to give unbiased results.

CPRD, Clinical Practice Research Datalink; RCT, randomised controlled trial; VKA, vitamin K antagonist.

https://doi.org/10.1371/journal.pmed.1004377.g001

The trial-analogous CPRD Aurum apixaban patients were matched to warfarin CPRD Aurum patients using greedy nearest-neighbour matching on the logit of the PS; a caliper of 0.2 times the standard deviation of the logit of the PS was used for matching as recommended by Austin [22].

The covariates included in the PS models are detailed in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Covariates Included in the PS models.

https://doi.org/10.1371/journal.pmed.1004377.t001

The model resulting in the most balanced cohort was chosen with balance assessed by looking at standardised differences across all variables after matching using a target threshold of 0.05 for the maximum difference allowed for any individual variable. Balance of covariates considered to be most important in predicting outcome were prioritised, namely, age, sex, and stroke risk factors.

Statistical analysis

Methods of analysis.

A prospective protocol was published prior to the analysis detailing the planned analyses ([14]; also in S1 Appendix).

Changes from the planned protocol are described in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Changes from planned analyses.

https://doi.org/10.1371/journal.pmed.1004377.t002

All time-to-event endpoints were analysed using a Cox proportional hazards model, stratified by prior VKA status (experienced, naïve). The effectiveness outcomes were analysed using the intention-to-treat principle, and major bleeding was analysed using an on-treatment censoring scheme. Patients were censored at 2.5 years after index date reflecting typical maximum duration of follow-up in ARISTOTLE. Cluster-robust standard errors were used with pair membership as the clustering variable [24,25]. The proportional hazards assumption was assessed by looking at the log-log of the Kaplan–Meier survival curves and inspection of scaled Schoenfeld residuals plotted against time. Analyses were performed using SAS version 9.4 and R version 4.2.1.

Supplementary analyses.

A protocol planned analysis in the subset of patients deemed adherent (with adherence measured by TTR in the warfarin users and by proportion of days covered by prescriptions in the apixaban users) was planned to assess the impact of adherence on outcomes. The planned analysis was not possible due to the apixaban prescription data not providing a useful measure of adherence. An analysis by INR TTR was performed instead to assess the impact of warfarin control on results with all outcomes analysed by TTR (TTR < 0.75 and TTR ≥ 0.75). Individual predicted TTR based on baseline variables was used for patients missing TTR. In order to perform the TTR analysis while maintaining balance in the baseline covariates, inverse probability treatment weighting (IPTW) was used to rebalance the baseline characteristics, applying stabilised weights to the ARISTOTLE-analogous apixaban users. A similar approach to the main analysis was used with PS models constructed separately for the new users and warfarin-experienced users.

An additional post hoc analysis was performed looking at the proportion of apixaban patients prescribed reduced-dose apixaban along with a comparison of the patients meeting the criteria for dose-reduction against the dose actually prescribed. In this analysis, apixaban dose in the ARISTOTLE-analogous CPRD cohort was assessed and compared against the ARISTOTLE protocol-specified criteria and NICE criteria for reduced apixaban dose. ARISTOTLE specified that participants meeting any 2 of the following criteria assessed at the time of randomisation should have their apixaban dose reduced to 2.5 mg BID: age ≥80 years, body weight ≤60 kg, or serum creatinine ≥1.5 mg/dL. These criteria are equivalent to the NICE guidelines for dose reduction with NICE having an additional criteria indicating reduced dose in those with creatinine clearance 15 to 29 mL/minute.

In addition, to assess the impact of the quality of dose-adjustment in the CPRD cohort on the observed effectiveness of apixaban relative to warfarin, a supplementary post hoc analysis was performed looking at the results in the subset of apixaban patients prescribed the correct dose compared with IPTW rebalanced warfarin comparators.

Sensitivity analyses.

Primary and secondary effectiveness outcomes were also analysed using the on-treatment censoring scheme to investigate whether treatment discontinuation compromises confidence in the effectiveness analyses.

Treatment persistence was defined by looking at longitudinal prescription data for OACs; OAC treatment windows were derived in which gaps > = 6 months between prescription dates were considered as distinct treatment windows. The end of each OAC treatment window was derived as the date of the last prescription of index OAC + the number of days supply given in the last prescription + a grace period of 30 days. In cases of overlapping OAC treatment windows, the date of the first prescription of the subsequent OAC treatment window was used to define the end of the prior OAC window. A prescription for a different OAC from the index OAC treatment was considered as a treatment switch. An ending of index OAC treatment with no subsequent prescription for any other OAC recorded was considered as treatment stop. Gaps of > = 6 months with no subsequent OAC prescriptions recorded were categorised as having stopped OAC treatment.

The set of patients who switched or discontinued treatment during follow-up were examined to ascertain whether selection bias due to attrition may have affected the on-treatment analyses (Table A9 in S1 Appendix).

Apixaban was first launched for AF in the UK in January 2013, with relatively few patients receiving a prescription in the first year it was available; we therefore performed a sensitivity analysis with the start of study period shifted forwards a year to investigate the impact of inclusion of early adopters who may differ from later adopters of a new drug.

Confounding and bias.

In the study period, apixaban was a newly available treatment leading to the possibility of channelling bias [26]. By applying trial eligibility criteria to both treatment cohorts and matching using baseline covariates, we aimed to minimise channelling bias. To handle confounding, treatment arms were matched using propensity score matching (PSM) [27].

Benchmarking results against ARISTOTLE.

The study hypothesis was that results in the CPRD ARISTOTLE-analogous cohort would be comparable to the ARISTOTLE results, as defined by the prespecified benchmarking criteria. A slightly weaker benefit of apixaban versus warfarin was expected based on the weaker benefit seen in the EU subgroup of ARISTOTLE, and an expectation that the quality of warfarin control in UK patients may be higher than that observed in ARISTOTLE.

The benchmarking criteria for considering the results in the trial-analogous CPRD cohort to be comparable with ARISTOTLE were prespecified and published previously [14]:

• The effect size must be clinically comparable with the ARISTOTLE findings; the HR for time to stroke/SE with the HR must be between 0.69 and 0.99. This range is not symmetrical around the ARISTOTLE estimate of 0.79 as it is anticipated the treatment effect in routine clinical care may be weaker than that seen in the optimised setting of a clinical trial.
• The upper limit of the 95% CI for the HR for time to stroke/SE must be less than 1.38 (noninferiority margin used in ARISTOTLE, updated since protocol—see Table 2).

The benchmarking step applied only to the primary effectiveness outcome in the trial-analogous CPRD cohort; results in other groups such as patients underrepresented or excluded from the trial would not necessarily be expected to remain consistent to the RCT results, given the relative risks may differ in these groups. Comparability of other outcomes was to be assessed descriptively with no formal criteria or hypothesis testing used.

Results

Participants

Between 1 January 2013 and 31 July 2019, there were 86,888 people with AF prescribed apixaban and 159,632 prescribed warfarin in HES-linked CPRD Aurum practices (Fig 2). Application of minimum registration period and ARISTOTLE inclusion criteria reduced this to 67,539 apixaban and 139,527 warfarin patients. After applying ARISTOTLE exclusion criteria, there were 41,487 apixaban and 101,159 warfarin patients.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Selection of ARISTOTLE-analogous CPRD Aurum Cohort.

Flow of number of individuals included in the analysis. AF, atrial fibrillation; ALT, alanine transaminase; AST, aspartate transaminase; BMI, body mass index; BP, blood pressure; CPRD, Clinical Practice Research Datalink; HES, Hospital Episodes Statistics; Rx, Prescription; SES, socioeconomic status; ULN, upper limit of normal; VKA, vitamin K antagonist. ^aSevere comorbid condition with life expectancy <1 year or reasons making participation impractical; ^bALT or AST > 2X ULN or Total Bilirubin ≥ 1.5X ULN; ^cPregnant or breastfeeding within 3 years prior. See Table A1 in S1 Appendix for detailed list of inclusion and exclusion criteria. Note: For prevalent warfarin users, trial eligibility only revealed at point of random selection into the cohort for prevalent users. Numbers in figure show maximum theoretical number of warfarin users available should they be selected only at a time they were eligible for the trial.

https://doi.org/10.1371/journal.pmed.1004377.g002

Selecting apixaban patients to match ARISTOTLE on key baseline characteristics yielded 9,120 apixaban patients (3,912 new users and 5,208 prevalent users) available for PSM to 101,159 warfarin patients. For 274 apixaban patients, no match could be found giving a PS matched cohort of 8,846 apixaban and 8,846 warfarin patients.

Application of ARISTOTLE inclusion/exclusion criteria and matching to ARISTOTLE

Applying the ARISTOTLE inclusion/exclusion criteria and matching to ARISTOTLE baseline patient characteristics resulted in a cohort similar to the ARISTOTLE apixaban participants (Table 3); for example, median age was 78 and mean CHADS₂ score 2.4 in CPRD Aurum before applying trial criteria and matching, whereas the median age of 71 and mean CHADS₂ score 2.1 after these steps matched the ARISTOTLE apixaban participants. The ARISTOTLE-analogous apixaban arm matched the trial arm on prior VKA exposure, age, sex, stroke risk factors and CHADS₂ score, and proportion of patients with moderate or severe renal impairment.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Baseline characteristics of patients with AF prescribed apixaban and warfarin in CPRD Aurum compared with ARISTOTLE participants: (i) before and (ii) after applying ARISTOTLE inclusion and exclusion criteria and (iii) after matching to the trial participants.

https://doi.org/10.1371/journal.pmed.1004377.t003

Differences remained on baseline characteristics it was not feasible to match on, namely, ethnicity (95.2% white, 2.4% Asian in CPRD Aurum apixaban versus 82.6% white, 14.4% Asian in ARISTOTLE) and concomitant medications (amiodarone 3.8%, aspirin 5.8%, digoxin 13.9% in CPRD Aurum apixaban users versus amiodarone 11.1%, aspirin 31.3%, digoxin 32.0% in ARISTOTLE apixaban arm). See S1 Appendix for details on matching feasibility.

Main results

The HR for stroke/SE in the PS matched groups was 0.98 (95% CI 0.82,1.19) (Fig 3 and Table A3 in S1 Appendix). This association was consistent with the noninferiority margin (upper limit of the 95% CI less than 1.38) [7] but did not show superiority as predicted by ARISTOTLE [HR 0.79 (95% CI 0.66,0.95)] (Fig 3 and Table A2 in S1 Appendix). The outcome of all-cause mortality also showed noninferiority [Aurum 1.03 (0.93,1.14) versus trial 0.89 (0.80,0.998)] but did not indicate apixaban superiority. Absolute event rates for the primary outcome and components were close to the trial for apixaban—for example [comparing Aurum versus trial], stroke/SE event rate of 1.27%/year versus 1.27%, whereas the warfarin group had a lower event rate compared with ARISTOTLE (stroke/SE event rate of 1.29%/year versus 1.60% and hemorrhagic stroke 0.33%/year versus 0.47%/year) (Fig 3). Mean duration of follow-up in the cohort was 1.8 years in the apixaban arm and 2.2 years in the warfarin arm.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Forest plot showing HRs (dots) and 95% CIs (lines) for apixaban vs.

warfarin. Absolute event rates (%/year) and HR (95% CIs) are presented for key effectiveness outcomes in (i) ARISTOTLE, (ii) CPRD Aurum trial-matched cohort, (iii) CPRD Aurum trial-matched with TTR < 0.75, and (iv) CPRD Aurum trial-matched with TTR ≥ 0.75. Dashed line shows noninferiority margin 1.38 for the upper bound of the 95% CI of the HR used in ARISTOTLE for the primary outcome of stroke or SE. For the analysis by TTR, IPTW was applied to the apixaban users targeting the treatment effect in the warfarin users with TTR < 0.75 and TTR ≥ 0.75. CI, confidence interval; CPRD, Clinical Practice Research Datalink; HR, hazard ratio; IPTW, inverse probability treatment weighting; SE, systemic embolism; TTR, time in therapeutic range.

https://doi.org/10.1371/journal.pmed.1004377.g003

Analysis of apixaban dose-adjustment

The proportion of patients meeting the criteria for reduced dose apixaban (Table 4) was similar between the CPRD ARISTOTLE-analogous apixaban, warfarin, and RCT apixaban groups (4.9%, 4.9%, and 4.7%, respectively). When including the additional NICE criteria of creatinine clearance, 5.1% of apixaban users in the ARISTOTLE-analogous cohort had an indication for reduced-dose apixaban, yet a larger proportion (14.3%) were prescribed reduced dose apixaban implying some patients in CPRD Aurum may have been prescribed the wrong dose and/or information on criteria for dose reduction may have been missing from CPRD Aurum.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Apixaban dose-adjustment in CPRD Aurum compared with ARISTOTLE.

https://doi.org/10.1371/journal.pmed.1004377.t004

A futher analysis of the quality of dose-adjustment in patients in CPRD Aurum (Table 5) indicated 10.5% of patients may have been prescribed an incorrect dose of apixaban at the index prescription based on the data contained in their electronic health records (EHRs). The majority of incorrect dose relating to patients being prescribed reduced-dose apixaban despite not meeting the criteria for dose reduction. A large proportion of patients prescribed an incorrect dose had only 1 dose adjustment criteria (59.6% of those with incorrect dose), suggesting some prescribers may have thought a dose reduction was warranted when only 1 criteria was present. Other possible reasons for the incorrect dose-adjustment observed here may be data on the criteria missing from the EHR (i.e., incorrect ascertainment) or consideration of other medical history that made a prescriber adjust the dose.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Quality of apixaban dose-adjustment in CPRD Aurum ARISTOTLE-analogous cohort.

https://doi.org/10.1371/journal.pmed.1004377.t005

To assess the impact of the quality of dose-adjustment in the CPRD cohort on the effectiveness of apixaban, a supplementary post hoc analysis was performed looking at the results in the subset of apixaban patients prescribed the correct dose (N = 7,921) compared with IPTW rebalanced warfarin comparators. The results in this subset were consistent with the primary results showing apixaban to be noninferior to warfarin (stroke/SE 0.96 [0.78,1.17], death 0.97 [0.87,1.09]) with the results moving slightly closer to those observed in ARISTOTLE.

Safety results

The analysis for safety outcomes is presented in Fig 4 and Table A5 in S1 Appendix; patients on apixaban had a lower risk of major bleeding compared with those on warfarin, HR (95% CI) 0.88 (0.77,1.00), consistent with ARISTOTLE. Analysis by TTR suggested superiority of apixaban for major bleeding in those with TTR <0.75 [0.74 (0.63, 0.86)], whereas apixaban users had a similar risk of major bleeding compared with those with optimal warfarin control (TTR ≥ 0.75) [1.08 (0.90,1.30)].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Forest plot showing HRs (dots) and 95% CIs (lines) for apixaban vs.

warfarin. Absolute event rates (%/year) and HR (95% CIs) are presented for key safety outcomes in (i) ARISTOTLE, (ii) CPRD Aurum trial-matched cohort, (iii) CPRD Aurum trial-matched with TTR < 0.75, and (iv) CPRD Aurum trial-matched with TTR ≥ 0.75. For the analysis by TTR, IPTW was applied to the apixaban users targeting the treatment effect in the warfarin users with TTR <0.75 and TTR ≥0.75. CI, confidence interval; CPRD, Clinical Practice Research Datalink; HR, hazard ratio; IPTW, inverse probability treatment weighting; TTR, time in therapeutic range.

https://doi.org/10.1371/journal.pmed.1004377.g004

Discussion

In our emulation of ARISTOTLE using UK routinely collected healthcare data, we found results that met our predefined criteria for comparability with the trial. We saw noninferiority of apixaban versus warfarin for prevention of stroke or SE, all-cause mortality, and major bleeding but did not see superiority of apixaban versus warfarin for these outcomes as was seen in ARISTOTLE. We found higher TTR in the patients using warfarin in our cohort compared with the warfarin arm of ARISTOTLE (median 0.76 versus 0.66). While our analysis by TTR showed noninferiority of apixaban versus warfarin for our stroke or SE outcome, we observed an increased risk of death on apixaban compared with patients well-controlled on warfarin (TTR ≥ 0.75) but not when compared with those on poorly controlled warfarin (TTR < 0.75). For major bleeding, while apixaban was superior when compared to those on poorly controlled warfarin, there was no difference when compared to those on well-controlled warfarin. We saw evidence suggesting suboptimal dosing of apixaban in our cohort with approximately 10% of patients in the apixaban arm prescribed the reduced dose without meeting the criteria for the reduced dose.

We found the differences in the overall treatment-effect estimates between our cohort and ARISTOTLE may be explained by the lower proportion of Asian patients in our cohort, differences in INR control in the warfarin arm of our cohort compared with ARISTOTLE, and the higher proportion of patients prescribed a reduced dose of apixaban in our cohort compared with ARISTOTLE.

Our findings are consistent with a UK study of ischemic stroke, which compared DOACs with warfarin [29]. A Danish study found similar results to ours for stroke/SE [30], although they found apixaban users had a lower risk of death; a study of US claims data [31] also found apixaban was associated with a lower risk of death. A systematic review and meta-analysis of 16 studies [32] found pooled results for stroke and intracranial haemorrhage that were consistent with ours. One study (in US claims data) also aimed to replicate ARISTOTLE [33,34] and, in contrast to our study, found superiority for apixaban for stroke/SE, which may be linked to population differences such as lower TTR in US patients on warfarin [35] and differences in ethnicity. None of these studies matched to the ARISTOTLE trial participants, included prevalent users, or looked at how warfarin control impacted results. Further details on these studies including design and key results are summarised in Table A10 in S1 Appendix.

A key strength of our study was the use of a framework that sampled prevalent users (the continuing users of warfarin in this study) in a way that avoided selection bias facilitating the construction of a cohort of patients similar to the target trial population, which included both new users of apixaban and warfarin (VKA-naïve) and patients with prior VKA exposure (VKA-experienced) that were randomised to stay on warfarin or switch to apixaban. The use of PSM, stratified by treatment history, enabled us to select a matched cohort well balanced on important covariates. The successful emulation of ARISTOTLE by our study shows that valid treatment effects can be obtained for important outcomes with OACs using noninterventional methods with routinely collected clinical data. Having validated this framework, in future studies, we can look at the effectiveness of OACs in AF patient groups not included or underrepresented in the RCT, such as elderly patients and those at increased bleeding risk. We also recommend future analyses with an extended follow-up period compared with this study to compare the long-term outcomes seen in the noninterventional cohort with projected long-term outcomes from the RCT.

An additional strength of our study was the ability to explore the quality of warfarin treatment in our cohort and the impact of INR control on the treatment effect estimates. Our finding that the benefits of apixaban versus warfarin for some outcomes depended on the quality of INR control in the warfarin arm answers questions raised in the NICE premeeting briefing, which looked at apixaban in the NVAF population and noted the TTR seen in ARISTOTLE “may be lower than what is typical in UK clinical practice” and “apixaban compared with well-controlled warfarin (TTR 75% or more) may not be superior in the long term” [8]. ARISTOTLE presented outcomes by centre (for example, hospital) TTR quartile and did not show a signal of treatment efficacy differing by centre TTR quartile. We were able to use IPTW to estimate the treatment effect in the different warfarin TTR groups and used predicted TTR for warfarin users missing TTR to attempt to limite the risk of selection bias.

While our study aimed to emulate ARISTOTLE using suitable methods, there were several limitations. Some of the criteria assessed for ARISTOTLE eligibility may not be well recorded in CPRD leading to a risk of misclassification. Furthermore, misclassification of ARISTOTLE eligibility criteria and baseline covariates could be differential by treatment in the VKA-experienced patients if criteria such as renal function are more likely to be checked before changing treatment. However, the most important risk factors for the primary outcome of stroke (the components of CHA₂DS₂-VASc stroke risk score) are mostly well recorded in CPRD Aurum and HES.

Our cohort did not attempt to match the trial on the use of concomitant medications in order for our cohort to reflect typical UK prescribing. In ARISTOTLE, 31% of participants were using aspirin and 11% using amiodarone at baseline, whereas in our cohort, only 6% were recorded as using aspirin and 4% amiodarone. Amiodarone potentiates the effects of warfarin, and concomitant use of amiodarone with DOACs is associated with increased risk of major bleeding [36], while concomitant use of aspirin increases the risk of bleeding for both warfarin [37] and DOACs [38]. The difference in concomitant medication usage between our cohort and the trial population may explain some of the observed differences in treatment effects.

A key limitation of our study was the inability to match ARISOTLE on ethnicity, meaning the CPRD Aurum cohort included a low number of patients from Asian and Hispanic groups when compared with the RCT (14.5% of participants in ARISTOTLE were Asian compared to 2.4% in our ARISTOTLE-analogous CPRD cohort). There are known racial differences in the treatment effects of OACs with Asian patients experiencing a higher risk of haemorrhagic stroke and intracranial haemorrhage compared with white patients; in ARISTOTLE, Asian participants experienced double the risk of stroke or SE when on warfarin therapy when compared with white participants [39]. The reasons for the increased risk of bleeding associated with warfarin therapy in Asian patients is hypothesised to be associated with differences in drug metabolism and prevelance of cerebral microbleeds [40]. The difference in proportion of Asian patients between our cohort and ARISTOTLE is therefore likely to explain some of the differences in treatment effects seen and limits the generalisability of our study, with the results of our study of most relevanance to white patients. This limitation on ethnicity arose from the data source used and time period studied (patients with AF in CPRD Aurum 2013–2019), which had a low proportion of Asian patients, likely due to AF being associated with older age combined with a lower prevalence of AF in Asian patients compared with white patients [41]. While CPRD Aurum is largely representative of the UK population in relation to ethnicity [42], diversity is still limited for older individuals. Despite this, CPRD Aurum has shown to be a useful resource for investigating treatment effects in different ethnic groups for indications such as hypertension, which is more prevalent and occurs at a younger age in ethnic minority groups, with similar trial replication methods used to compare antihypertensive treatment effects in underrepresented ethnic groups [13].

The approach our study used for handling missing data on baseline covariates relied on assumptions on the relationship between missingness, treatment, and outcomes, which may not be valid; however, the low proportion of missing data means that this is unlikely to have impacted the results. In the coarsened exact matching step, the choice of variables will have an impact on the resulting cohort selected, meaning a different combination of variables could lead to different results. There is a risk that residual confounding may be present despite the use of PSM. The use of PSM also has the potential to introduce bias by dropping patients from the cohort [19]; however, PSM is well suited to the process of trial emulation including prevalent users, and a low number of apixaban users were dropped due to unsuccessful matching. The inclusion of prevalent users of warfarin in the cohort risks the introduction of selection bias [20,21]; this was avoided by use of a method shown to produce unbiased estimates in a simulation study [21]. We found consistent results between our new and prevalent user strata across multiple outcomes providing reassurance the method used was likely to have successfully avoided selection bias.

Apixaban along with other DOACs were rapidly adopted as preferred first-line OAC in AF during the study period; it was therefore not possible to match on calendar date leading to a difference in follow-up time between the treatment arms in our cohort. A higher proportion of warfarin users switched to alternative OAC during follow-up compared with those prescribed apixaban (16% versus 6%). The impact of this differential switching during follow-up was addressed in the sensitivity analyses. The availability of new alternative treatments during the study period also means there is a risk of channelling bias in that over time the patients still on warfarin are more likely to be those doing well on warfarin. INR control prior to the index date was not included in the PS for the prevalent users due to a high rate of missing data; however, other variables associated with poor INR control were included in the models, and an exploratory post hoc analysis including a variable for poor INR control gave results consistent with the primary results.

Adherence to treatment was difficult to assess in our study due to automatic repeat prescriptions; treatment persistence was more useful in providing a measure of pattern of medicine use over time. In the analysis by TTR, the adherence of patients using apixaban was not accounted for; however, a previous UK study showed apixaban had higher adherence than VKAs [43], meaning we would expect to see better effectiveness outcomes in apixaban. Futhermore, the use of IPTW in the analysis by TTR means predictors of poor adherence are likely to have been balanced between treatments. The analysis of TTR is limited by this being a post-baseline measure available for only 1 treatment arm leading to a risk of selection bias in this analysis—patients with TTR available in the study may be more healthy than those without this measure given that patients have to survive and not be hospitalised to have INR measurements available in CPRD Aurum. The limitation of use of a post-baseline measurement available for 1 treatment arm was also evident in the RCTs of DOACs versus warfarin and is mitigated in our study through the use of IPTW and predicted TTR for patients that were missing TTR (using a model to predict TTR that used INR measurements restricted to the first year of follow-up). Given the risk of selection bias in the analysis by TTR and risk of misclassification of TTR for those missing TTR, these results should be considered exploratory and interpreted with caution. Sensitivity analyses in our cohort using an on-treatment censoring scheme showed evidence of attrition bias. The regular measurement of INR and availability of alternative anticoagulants makes warfarin therapy particularly prone to attrition bias since a patient may be more likely to switch to a DOAC if their INR is frequently out of the optimal range or if they have not been adhering to scheduled INR testing.

To conclude, we found that applying a reference trial emulation approach allowed us to emulate a landmark randomised trial of apixaban versus warfarin using UK noninterventional data, with results meeting prespecified benchmarking criteria based on the reference trial results. This trial emulation method provides valid treatment effect estimates for apixaban compared to warfarin and can be used to determine risks and benefits of AF medications in people treated in routine clinical care. This study demonstrates a successful real-world application of novel methods that have been proposed for the inclusion of prevalent users in observational studies, with the application of an adaptation to mimic the screening process making the method suitable for emulation of RCTs that include prevalent users. These methods could be adapted for emulation of RCTs in other therapeutic areas and for looking at patient groups underrepresented or excluded from RCTs.

The weaker overall treatment benefit observed in our cohort appears to be due to a higher proportion of patients with well-controlled warfarin in the UK clinical context, compared with the trial. Our exploratory analysis by TTR showed similar results for stroke and a greater benefit for apixaban for major bleeding compared with TTR <0.75; conversely, a slightly higher risk of death was observed on apixaban compared with well-controlled warfarin.

Disclaimer

The views expressed in this paper are those of the author and not do not necessarily reflect those of the SFDA or its stakeholders. Guaranteeing the accuracy and the validity of the data is a sole responsibility of the research team.

Supporting information