## 4.1. Selection of Primary Studies

The electronic literature search identified 1,580 citations. Upon screening the titles and abstracts, 1,485 citations were excluded. These consisted mainly of reviews, non-relevant study designs or research questions, and studies in which interventions were not of interest. A total of 95 potentially relevant publications were retrieved for full-text review, as well as 5 additional references identified through other sources. Of the 100 potentially relevant reports, 72 articles did not meet the inclusion criteria. Therefore, a total of 28 articles^{19–46} reporting results from 12 individual RCTs were included in this review.

To be considered for inclusion, a trial needed to have at least one relevant comparison between two interventions of interest in patients who were eligible for anticoagulant therapy. At least one treatment arm was excluded in 3 of the 12 included trials, as the intervention dosage was not consistent with current recommendations. Placebo-controlled trials were also included, as they provide indirect evidence that can be incorporated in the network meta-analysis.

The trial selection process appears in a PRISMA flowchart (Appendix 8). Included and excluded trials are listed in Appendices 9 and 10, respectively. Note that two studies that were excluded, AVERROES and ACTIVE-A, were included in a sensitivity analysis presented in Appendix 21.

A summary of research available for each treatment strategy included in the MTC is subsequently provided in . A total of nine treatment strategies were included in the MTC. Information retrieved was not sufficient for the following treatments to be included in the NMA:

Summary of Interventions Evaluated.

## 4.2. Study and Patient Characteristics

The CADTH systematic review included 12 individual RCTs (reported in 28 publications);^{19–46} all evaluated the efficacy and safety of NOACs, warfarin, or ASA with or without clopidogrel in patients with AF. However, no direct treatment comparisons were available to assess the relative efficacy of one new oral anticoagulant drug with another. Details regarding trial characteristics are subsequently presented in .

Summary of Included Trials.

Of the 61,050 randomized patients included in this review, 4 large multicentre trials account for 57,284 patients (94%): ARISTOTLE (n = 18,201),^{26–29} RE-LY (n = 18,113),^{35–42} ROCKET-AF (n = 14,264),^{43–45} and ACTIVE-W (n = 6706).^{19–22} Trials recruited patients with AF, most of them with at least one risk factor for stroke; however, patients with recent stroke or TIA were usually excluded. Most trials included SSE as a primary efficacy outcome, while bleeding events was a frequent safety outcome. Follow-up ranged between 12 weeks and 3.5 years. All trials published results between 1989 and 2011.

The risk of stroke increases with age and is also substantial in patients with prior stroke or TIA. Higher-risk patients may also present with congestive heart failure, hypertension, and/or diabetes. All these factors are reflected in the CHADS_{2} score, a commonly used model for risk stratification.^{15,57} A mean CHADS_{2} score was reported in five trials, encompassing the vast majority of patients included in the systematic review: ARISTOTLE (n = 18,201),^{26–29} ARISTOTLE-J (n = 222),^{30} RE-LY (n = 18,113),^{35–42} ROCKET-AF (n = 14,264),^{43–45} and ACTIVE-W (n = 6706).^{19–22} Reported CHADS_{2} scores in these trials were consistent with a high-risk population (CHADS_{2} ≥ 2), with patients from ROCKET-AF showing the highest risk for stroke (mean CHADS_{2} = 3.5) — with 55% of patients with prior stroke or TIA, which is higher than the proportion of these patients included in the other NOAC trials.

Mean age across the included trials ranged from 65 years^{33} to 83 years.^{46} BAFTA^{31} and WASPO^{46} both required patients to be ≥ 75 years. Only four other trials reported age categories^{26,27,30,34–42} and, in these, the proportions of patients > 75 years ranged from 30%^{34} to 40%.^{43–45} All trials included patients of both gender and most were relatively balanced. However, the proportions of patients with a prior stroke or TIA varied substantially across the included trials, ranging from 3% in JAST^{33} to 55% in ROCKET-AF.^{43–45} The same observation can be made regarding the proportions of patients with other concomitant conditions; i.e., congestive heart failure, hypertension, diabetes, and MI. There was also inconsistency in VKA experience among the included populations. VKAs have high inter-individual variability^{10–12} and optimal dosing requires time. Therefore, a newly started VKA treatment may very well result in under- or over-anticoagulation, decreasing efficacy or increasing the risk of hemorrhage.

In ARISTOTLE, a dose-modification algorithm was used to reduce the dose of apixaban to 2.5 mg twice daily to minimize the potential for higher exposure in patients who may be at an inherently higher bleeding risk and to maintain a balance between efficacy and safety in such populations. Doses of 2.5 mg twice daily were used in a subset of patients with two or more of the following criteria: age ≥ 80 years, body weight ≤ 60 kg, or serum creatinine level ≥ 1.5 mg/dL (133 µmol/L).^{26} A reduced dose of apixaban 2.5 mg twice daily was administered to 428 patients in the apixaban group (4.7%).^{26} In ROCKET-AF, the dose of rivaroxaban was reduced to 15 mg once daily, but only if creatinine clearance ranged between 30 ml/min and 49 ml/min.^{44} As a result, a total of 1,474 patients (20.7%) received a reduced dose of rivaroxaban. By contrast, patients in RE-LY were randomized to two different doses of dabigatran (110 mg and 150 mg), without any dose adjustment and regardless of the patient characteristics.

TTR was reported in all but one trial^{24,25} that included a warfarin treatment arm. TTR was ≥ 66% only in BAFTA^{31} and WASPO,^{46} suggesting moderate to poor INR control in the other trials, with mean TTR ranging from 44%^{32} to 64%^{19,35}. The rest of the time, the patients may be at risk of bleeding (INR > 3.0) or at risk of thromboembolism (INR< 2.0).

A summary of patient baseline characteristics is subsequently presented in .

Summary of Patient Baseline Characteristics.

## 4.3. Critical Appraisal of Included Studies

The studies were individually critically appraised and the details are available in Appendix 11. Overall, there was substantial variation in study quality. However, large multicentre trials such as ARISTOTLE (n = 18,201),^{26–29} RE-LY (n = 18,113),^{35–42} and ROCKET-AF (n = 14,264),^{43–45} which account for the vast majority of patients included in the systematic review, appear to be methodologically rigorous. These trials all compared new anticoagulant agents versus warfarin.

There were various levels of quality concerns with other trials evaluating ASA with or without clopidogrel and placebo/no treatment, especially for ARISTOTLE-J,^{30} CAFA,^{32} and PETRO.^{34} These concerns include open-label designs, differences in baseline characteristics across treatment groups, and discontinuation of up to 29% of randomized patients. Lack of information on allocation concealment for these trials precluded a definite judgment on whether patients and investigators could foresee assignment to the treatment group. In addition, these trials were substantially smaller, older, and overall of lower quality than the anticoagulant trials, therefore affecting our level of confidence in the estimates of effect.

An important aspect of anticoagulant studies that include a VKA group is the quality of INR control, which was adequately reported through TTR. The benefit of warfarin therapy is expected to increase with TTR, which gets higher with improved INR control.^{15} However, mean TTR was considered optimal (≥ 66%) only in two trials: BAFTA^{31} and WASPO,^{46} while it was less but still near the optimal value (between 60% and 66%) in four trials: ACTIVE-W,^{19–22} ARISTOTLE,^{26–29} ARISTOTLE-J,^{30} and RE-LY.^{35–42} Mean TTR was sub-optimal (< 60%) in CAFA,^{32} PETRO,^{34} and ROCKET-AF.^{43–45} This suggests moderate to poor INR control, hence favouring the comparator. However, poor INR control could also arguably increase the generalizability of the results, because INR monitoring is often not optimal in clinical practice.^{78} AFASAK^{23} and JAST^{33} did not have a warfarin treatment arm.

Finally, the use of a dose-modification algorithm in ARISTOTLE^{26–29} to reduce the dose of apixaban to 2.5 mg twice daily could theoretically have favoured apixaban in an indirect comparison among the NOACs. A lower dose would be expected to minimize the potential for higher exposures in populations that may be at an inherently higher bleeding risk. The fact that fewer than 5% of all patients were treated with the lower dose of apixaban, however, suggests that the potential impact of this potential bias is likely very small.

## 4.4. Indirect Comparisons

MTCs and pairwise meta-analyses were conducted for each of the six outcomes figuring in . The number of individual RCTs included in the evidence networks varied from 6 to 10 studies, and these collectively included between 58,457 and 60,592 patients. A summary of data available for the MTC is presented as well in . Evidence networks for the NMA are presented in Appendix 12. Appendix 13 presents a summary of results from the NMA, with warfarin as a common comparator, while Appendix 14 reports pairwise comparisons among all interventions. Finally, detailed data incorporated into the models are presented in Appendix 15.

Summary of Evidence Available for the MTC.

All models provided a reasonable fit to the data (Appendix 16) when compared with the unconstrained data points (i.e., values are close to the number of unconstrained data points for all models fitted). In addition, direct estimates of effect sizes aligned closely with the estimated effect sizes derived from the NMA.

The results from the random-effects model were very similar to those of the fixed-effects model for all outcomes (Appendices 16 and 17), with the main difference being the size of the CIs for each point estimate; specifically, the credible intervals were greater for the random-effects model. A detailed comparison of the relative merits of the fixed- versus random-effects model is beyond the scope of this report; results from the fixed-effects model (versus random-effects model) are presented in the main text, for the following reasons:

### 4.4.1. Stroke and Systemic Embolism

There were 9 RCTs (N = 60,424) that reported data for SSE. Results from the NMA are summarized in , but are also presented in Appendix 13.

OR and ARD of All-Cause SSE for Antithrombotic Therapies Relative to Adjusted-Dose Warfarin for Patients With AF, Fixed-Effects NMA^{a}.

Statistical significance for SSE was reached for apixaban and dabigatran 150 mg compared with adjusted-dose warfarin, based on a lower OR. The use of these two agents led to absolute risk reductions ranging from 1 to 6 fewer events per 1,000 patients treated each year with apixaban and from 3 to 9 fewer events with dabigatran. By contrast, low-dose ASA and the combination of clopidogrel plus low-dose ASA appeared statistically significantly less effective than adjusted-dose warfarin at preventing SSE, with absolute results ranging from 4 to 27 more events per 1,000 patients treated each year. No statistically significant difference was detected between warfarin and each of the following interventions: rivaroxaban, dabigatran 110 mg, medium-dose ASA, and no treatment/placebo.

Among the different treatments, dabigatran 150 mg had the highest probability of having the best outcome on SSE (87.5%) and was ranked highest among the treatment options (Table A13.1 in Appendix 13).

Results for pairwise comparisons versus warfarin for SSE are reported in Table A13.1 in Appendix 13. Results for all pairwise contrasts are presented in Table A14.1 in Appendix 14. Estimates of effect relative to warfarin derived from the direct pairwise comparisons aligned closely with those obtained from the NMA in both direction and magnitude — except for low-dose ASA, medium-dose ASA, and no treatment/placebo, for which there was some variation in the OR. However, in no case was there a discrepancy between the direct pairwise comparisons and the NMA in statistical significance of the effect sizes.

As for pairwise comparisons among other treatment options, dabigatran 150 mg was associated with statistically significantly fewer SSE versus dabigatran 110 mg and rivaroxaban. No other differences between the new anticoagulant agents reached statistical significance. The magnitude of the differences among all new anticoagulants was considered relatively small, with OR ranging from 0.83 to 1.33.

There were no statistically significant differences among low-dose ASA, medium-dose ASA, and clopidogrel plus low-dose ASA. However, low-dose ASA and the combination of clopidogrel plus low-dose ASA were statistically significantly less effective at preventing SSE compared with all the anticoagulantdrugs, including adjusted-dose warfarin, dabigatran, apixaban, and rivaroxaban. The effect size for SSE for medium-dose ASA was greater relative to the anticoagulant treatments, but this difference only reached statistical significance versus dabigatran 150 mg.

The point estimates for the Bayesian random-effects NMA were similar to those reported in the fixed-effects NMA, although the credible intervals were wider in the random-effects model (Appendix 17). As a result, some findings that were statistically significant in the fixed-effects model did not retain statistical significance in the random-effects model. This was the case for apixaban and dabigatran 150 mg, as well as low-dose ASA and the combination of clopidogrel plus low-dose ASA, which became non-significant in the random-effects model.

### 4.4.2. Major Bleeding

Data for major bleeding were available from 10 RCTs (N = 60,503). Results from the NMA are summarized in , but are also presented in Appendix 13. Although there were some variations in the definition of major bleeding, most of the trials employed a definition that was consistent with the following: bleeding requiring transfusion of at least two units of red blood cells or the equivalent of whole blood, or associated with a decrease in the hemoglobin level of ≥ 20 g/L, bleeding at a critical site, or fatal bleeding. ICH was typically included in the definition.

Odds Ratio and Absolute Risk Difference of Major Bleeding for Antithrombotic Therapies Relative to Adjusted-Dose Warfarin for Patients With AF, Fixed-Effects NMA.

Statistical significance for major bleeding was reached for apixaban and dabigatran 110 mg compared with adjusted-dose warfarin, with a lower OR. Absolute risk reductions with these agents ranged from two to 13 fewer events per 1,000 patients treated each year. No statistically significant differences in the OR for major bleeding were detected between warfarin and each of the remaining interventions: rivaroxaban, dabigatran 150 mg, clopidogrel plus low-dose ASA, and all ASA dosages.

No treatment/placebo was associated with the largest positive effects estimate with an OR = 0.33 (22 fewer events per 1,000 patients treated each year), but did not reach statistical significance (95% CrI, 0.08 to 1.08). This intervention had the highest probability of having the best outcome for major bleeding (88.0%) and was ranked highest among all treatments (Table A13.2 in Appendix 13). Among the active treatments, apixaban had the highest probability of having the best outcome for major bleeding (10.1%) and was ranked highest among the active treatments.

Results for pairwise comparisons versus warfarin for major bleeding are reported in Table A13.2 in Appendix 13. Results for all pairwise contrasts are presented in Table A14.2 in Appendix 14. Estimates of effects derived from the direct pairwise comparisons aligned closely with those obtained from NMA in both direction and magnitude. In no case was there a discrepancy between the direct pairwise comparisons and the NMA in the statistical significance of the effect sizes.

As for pairwise comparisons, apixaban and dabigatran 110 mg were associated with statistically significantly fewer major bleeding events versus dabigatran 150 mg and rivaroxaban. The OR for apixaban was also statistically significantly lower compared with clopidogrel plus low-dose ASA. There were no significant differences associated with the ASA treatments, both for comparisons among themselves and compared with the anticoagulant treatments.

As in the case of SSE, the point estimates for the Bayesian random-effects NMA of the major bleeding data were similar to those reported in the fixed-effects NMA, although the credible intervals were wider in the random-effects model (Appendix 17). As a result, apixaban and dabigatran 110 mg did not retain statistical significance in the random-effects model.

### 4.4.3. Stroke and Systemic Embolism versus Major Bleeding

Prevention of SSE (i.e., the most important benefit of treatment) must be balanced by the potential for an increased risk of serious bleeding. Plotting the relative effect sizes of the various treatments for these two outcomes, SSE versus major bleeding (which includes ICH) illustrates the overall comparative benefit/risk profile for the nine interventions that we analyzed (). Data in were obtained from the results presented in Tables A13.1 and A13.2 (in Appendix 13) for the ARD for SSE and major bleeding versus warfarin, calculated using a fixed-effects NMA. Examination of the data in suggests that the benefit/risk of the NOACs is positive compared with warfarin (decrease in SSE and/or major bleeding) and largely similar among one another. Comparison of all anticoagulant treatment (including warfarin) to ASA with or without clopidogrel indicates that, whereas the NOACs decrease the risk of SSE and major bleeding, ASA clearly has a less favourable benefit/risk profile and fails to minimize the risk of SSE and/or major bleeding.

ARD for SSE versus MB Relative to Warfarin for Nine Different Interventions for the Reference Case ^{a}.

### 4.4.4. All-Cause Mortality

Ten RCTs (N = 60,592) reported data for all-cause mortality. Results from the NMA for this outcome are presented in Appendix 13, Table A13.3. Results for all pairwise contrasts are presented in Table A14.3 in Appendix 14.

Only apixaban achieved a statistically significant reduction in all-cause mortality compared with adjusted-dose warfarin in the NMA analysis (OR = 0.89 [0.794 to 0.997]). For this outcome, apixaban led to absolute risk reductions versus warfarin ranging from 0.1 to 8 fewer events per 1,000 patients treated each year. Other antithrombotic therapies did not seem to impact mortality: results from both the NMA and pairwise comparison analyses did not find any statistically significant differences among other treatment options. However, the NMA comparison for warfarin versus dabigatran and rivaroxaban yielded results which, although not statistically significant, were similar in terms of magnitude to those observed for apixaban (Table A13.3 in Appendix 13).

### 4.4.5. Extracranial Hemorrhage

There were 9 RCTs (N = 60,428) that reported data for extracranial hemorrhage. Results from the NMA are presented in Appendix 13. As was the case for mortality, only apixaban achieved a statistically significant reduction in extracranial hemorrhage compared with adjusted-dose warfarin in the NMA analysis (OR = 0.8 [0.68 to 0.94]), leading to absolute risk reductions versus warfarin ranging from 2 to 8 fewer events per 1,000 patients treated each year. No other treatment options were associated with a statistically significant reduction in extracranial hemorrhage compared with adjusted-dose warfarin.

Results for pairwise comparisons versus warfarin for major bleeding are reported in Table A13.4 in Appendix 13. Results for all pairwise contrasts are presented in Table A14.4 in Appendix 14. Apixaban also resulted in a statistically significant reduction in extracranial hemorrhage compared with dabigatran 150 mg, rivaroxaban, and medium-dose ASA. Medium-dose ASA was also associated with a statistically significant increase in extracranial hemorrhage relative to placebo/no treatment.

### 4.4.6. Intracranial Hemorrhage

Data for ICH were available from eight RCTs (N = 59,756). Results from the NMA are presented in Appendix 13. All NOACs were associated with statistically significant reductions in ICH compared with adjusted-dose warfarin (OR = 0.42 [0.3 to 0.58] for apixaban, OR = 0.31 [0.19 to 0.47] for dabigatran 110 mg, OR = 0.4 [0.27 to 0.59] for dabigatran 150 mg, and OR = 0.65 [0.46 to 0.92] for rivaroxaban, versus warfarin). Absolute risk reductions were similar between these interventions, ranging from three to seven fewer events per 1,000 patients treated each year. Other antithrombotic therapies did not seem to impact ICH, as there were no other statistically significant differences among treatment options.

Results for pairwise comparisons versus warfarin for ICH are reported in Table A13.5 in Appendix 13. Results for all pairwise contrasts are presented in Table A14.5 in Appendix 14. Dabigatran 110 mg resulted in a statistically significant decrease in ICH compared with rivaroxaban. Clopidogrel plus low-dose ASA was associated with a statistically significant increase in ICH relative to apixaban, dabigatran, rivaroxaban, and no treatment/placebo; however, there is substantial uncertainty surrounding these pairwise comparisons, as expressed by the very large CI.

### 4.4.7. Myocardial Infarction

A total of six RCTs (N = 58,457) reported data for MI. Results from the NMA are presented in Appendix 13. Dabigatran 150 mg was associated with an increase in MI compared with adjusted-dose warfarin that reached statistical significance (OR = 1.41 [1.02 to 1.96]). ARD versus warfarin ranged from 4 to 0.1 more events per 1,000 patients treated each year. Other antithrombotic therapies did not impact MI, as there were no other statistically significant differences among treatment options.

Results for pairwise comparisons versus warfarin for MI are reported in Table A13.6 in Appendix 13. Results for all pairwise contrasts are presented in Table A14.6 in Appendix 14. Dabigatran (both doses), medium-dose ASA, and the combination of clopidogrel plus low-dose ASA were associated with a statistically significant increase in MI compared with apixaban. Medium-dose ASA was also associated with a statistically significant increase in MI compared with rivaroxaban.

### 4.4.8. Subgroup Analyses

#### a. CHADS_{2} Score

##### CHADS_{2} < 2

We conducted a subgroup analysis where we considered patients who had a lower stroke risk (i.e., CHADS_{2} <2). Evidence diagrams for the CHADS_{2} < 2 subgroup analyses are presented in Appendix 12. A breakdown of the proportion of the patients with CHADS_{2} = 0 and CHADS_{2} = 1, for which the treatment recommendations^{3–5} differ, are provided in Appendix 18. However, there were only very few patients for all interventions with CHADS_{2} = 0. Therefore, although this subgroup analysis is labelled CHADS_{2} < 2, almost all included patients had a CHADS_{2} = 1.

There were limited data for ASA for this subgroups analysis, both for CHADS_{2} < 2 and for CHADS_{2} ≥ 2. As a result, we were required to use subgroup data that either only partially aligned with the subgroups (i.e., BAFTA had CHADS_{2} 1–2 versus 0–1) or study-level data from studies consisting of low-risk patients (e.g., CAFA, JAST), although a small proportion of patients (< 25%) had a CHADS_{2} ≥ 2.

**Stroke and Systemic Embolism:**
Table A13.7 in Appendix 13 shows comparisons between warfarin and each NOAC. Whereas dabigatran 150 mg and apixaban were superior to warfarin in the reference case analysis, results differed in the CHADS_{2} < 2 subgroup. No statistically significant differences were observed between warfarin and each evaluated NOAC for SSE. No data were available for rivaroxaban, so it was not included in the subgroup analysis. SSE results for the ASA treatments in the CHADS_{2} < 2 analysis matched those of the reference case; specifically, low-dose ASA and the combination of clopidogrel plus low-dose ASA appeared statistically significantly less effective than adjusted-dose warfarin at preventing SSE, this time with an ARD of 12 and 21 more SSE per 1,000 patients treated each year. There was no statistically significant difference between warfarin and medium-dose ASA.

Table A14.7 in Appendix 14 shows all pairwise comparisons. Dabigatran 150 mg was associated with statistically significantly fewer SSE versus dabigatran 110 mg, as in the reference case analysis. However, there were no other statistically significant differences anymore among the NOACs (no data was available for rivaroxaban).

Other results for this subgroup analysis were consistent with those for the overall population. More precisely, there were no statistically significant differences among low-dose ASA, medium-dose ASA, and clopidogrel plus low-dose ASA. However, low-dose ASA and the combination of clopidogrel plus low-dose ASA were statistically significantly less effective at preventing SSE compared with anticoagulant drugs. Results for medium-dose ASA did not reach statistical significance.

**Major Bleeding:**
Table A13.8 in Appendix 13 shows comparisons between warfarin and each NOAC. The results of the analysis for major bleeding in the CHADS_{2} < 2 subgroup were similar to those of the reference case analysis; statistical significance for major bleeding was reached for apixaban and dabigatran 110 mg compared with adjusted-dose warfarin, while no statistically significant differences in the OR for major bleeding were detected between warfarin and each of the remaining interventions.

As in the reference case analysis, no treatment/placebo had the highest probability among the different treatments of having the best outcome for major bleeding (67.8%) and was ranked highest (Table A13.8 in Appendix 13). Among the active treatments, apixaban had the highest probability among the different treatments of having the best outcome for major bleeding (16.5%) and was ranked highest.

Table A14.8 in Appendix 14 shows all pairwise comparisons. There were no statistically significant differences among the NOACs for major bleeding (no data was available for rivaroxaban). This is in contrast to the reference case analysis, where apixaban and dabigatran 110 mg were superior to dabigatran 150 mg and rivaroxaban in the overall population. As in the reference case analysis, the OR for major bleeding was not different among the ASA treatments (low-dose ASA, medium-dose ASA, and clopidogrel plus low-dose ASA), and between ASA monotherapy and anticoagulants; however, the combination of clopidogrel plus low-dose ASA resulted in more major bleeding than apixaban and dabigatran (both doses).

**Stroke and Systemic Embolism Versus Major Bleeding:** As previously noted for the reference case, plotting the relative effect sizes (regarding ARD versus warfarin) of the various treatments for SSE (benefit) versus major bleeding (risk) illustrates the overall comparative benefit/risk profile for the eight treatments that we analyzed for the CHADS < 2 subgroup (). Examination of the data in suggests that the benefit/risk of the NOACs is positive compared with warfarin (decrease in SSE and/or major bleeding) and closer to the benefit/risk of warfarin than the ASA treatments: whereas the NOACs (except rivaroxaban, which was not included in this analysis) decrease the risk of both SSE and major bleeding, ASA has a less favourable benefit/risk profile and fails to reduce the risk of SSE and major bleeding. These results are not dissimilar to the risk/benefit in the reference case ().

ARD for SSE versus MB Relative to Warfarin for Eight Different Interventions for CHADS_{2} Score < 2.^{a}.

##### CHADS_{2} ≥ 2

We conducted a subgroup analysis where we considered patients who had a higher stroke risk (i.e., CHADS_{2} ≥ 2). Evidence diagrams for the CHADS_{2} ≥ 2 subgroup analyses are presented in Appendix 12. A breakdown of the proportion of the patients with CHADS_{2} ≥ 2 is provided in Appendix 18. There were limited data for ASA for this subgroups analysis, both for CHADS_{2} < 2 and for CHADS_{2} ≥ 2. As a result, we were required to use subgroup data that only partially aligned with the subgroups (i.e., BAFTA had CHADS_{2} 3 to 6 versus 2 to 6).

**Stroke and Systemic Embolism:**
Table A13.7 in Appendix 13 shows comparisons between warfarin and each NOAC. Results for the CHADS_{2} ≥ 2 subgroup were consistent with the reference case analysis: dabigatran 150 mg and apixaban were both statistically significantly superior to warfarin. However, this contrasts with the CHADS_{2} < 2 subgroup, where no statistically significant differences were observed between warfarin and each evaluated NOAC for SSE. No statistically significant difference was detected between warfarin and each of the following interventions: rivaroxaban, low-dose ASA, and the combination of clopidogrel plus low-dose ASA (medium-dose ASA could not be assessed in the CHADS_{2} ≥ 2 subgroup).

As in the reference case and the CHADS_{2} < 2 subgroup analyses, dabigatran 150 mg had the highest probability among the different treatments of having the best outcome for SSE (80.9%) and was ranked highest among the treatments.

Table A14.7 in Appendix 14 shows all pairwise comparisons. These results were consistent with the reference case analysis and the CHADS_{2} < 2 subgroup: superiority of dabigatran 150 mg over dabigatran 110 mg was the only comparison among NOACs reaching statistical significance. Low-dose ASA and the combination of clopidogrel plus low-dose ASA were statistically significantly less effective at preventing SSE compared with anticoagulants. When compared against one another, the combination of clopidogrel plus low-dose ASA led to more SSE events than low-dose ASA alone. No data was available for medium-dose ASA.

**Major Bleeding:**
Table A13.8 in Appendix 13 shows comparisons between warfarin and each NOAC. In the CHADS_{2} ≥ 2 subgroup, only apixaban was statistically significantly superior to warfarin. This contrasts with the reference case and the CHADS_{2} < 2 subgroup analyses, where dabigatran 110 mg was also superior to warfarin. No statistically significant differences in the OR for major bleeding were detected between warfarin and each of the remaining interventions.

Apixaban had the highest probability among the different treatments of having the best outcome for major bleeds (78.5%) and was ranked highest (Table A13.8 in Appendix 13). Note that no treatment/placebo was not included in this subgroup analysis, but had the highest probability of minimizing major bleeding in the reference case and the CHADS_{2} < 2 subgroup analyses. However, apixaban was the highest ranked among active treatment options.

Table A14.8 in Appendix 14 shows all pairwise comparisons. In patients with a CHADS_{2} ≥ 2, apixaban was superior to dabigatran 150 mg and rivaroxaban. This contrasts with the absence of any differences among the NOACs in the CHADS_{2} < 2 subgroup, but is in line with the results of the reference case analysis. There were no statistically significant differences between low-dose ASA and clopidogrel plus low-dose ASA, and between these treatments and the other interventions, as was the case in the reference scenario.

**Stroke and Systemic Embolism Versus Major Bleeding:** To illustrate the overall comparative benefit/risk profile for the seven treatments that we analyzed for the CHADS_{2} ≥ 2 subgroup, we plotted the relative effect sizes (for ARD versus warfarin) of the various treatments for SSE (benefit) versus major bleeding (risk) in . Examination of the data in suggests that the benefit/risk of the NOACs is positive compared with warfarin (decrease in SSE and/or major bleeding) and very similar to the benefit/risk of warfarin. This is consistent with the reference case () and the CHADS < 2 subgroup (). As in the reference case and CHADS < 2 subgroup, low-dose ASA has a less favourable benefit/risk profile than the NOACs; however, the benefit/risk profile for clopidogrel plus low-dose ASA appears to be close to the anticoagulants in this subgroup, which was not the case in the CHADS < 2 subgroup or the reference case. However, the level of confidence in this observation is considerably lower than for the anticoagulant drugs, as the estimated effect sizes for clopidogrel plus low-dose ASA are derived from a single RCT with substantially fewer patients compared with the anticoagulant trials.

ARD for SSE versus MB Relative to Warfarin for Seven Different Interventions for CHADS_{2} Score ≥ 2^{a}.

#### b. Other Subgroups

##### Age and Time in Therapeutic Range

We also conducted subgroup analyses, where results were stratified by age and TTR. Evidence diagrams for subgroup analyses are presented in Appendix 12. Although fewer data were available for these subgroups compared with the CHADS_{2} score, analyses could be performed for all anticoagulant drugs (warfarin and the NOACs). Data on antiplatelet agents were scarce; evaluated interventions were limited to low-dose ASA for the age ≥ 75 years subgroup; the combination of clopidogrel and low-dose ASA for the TTR < 66% subgroup; and medium-dose ASA for the TTR ≥ 66% subgroup.

##### Stroke and Systemic Embolism

**Age:**
Table A13.9 in Appendix 13 shows comparisons between warfarin and each NOAC. The results of the analysis stratified by age for SSE in the ≥ 75 years subgroup were similar to those of the reference case analysis, where dabigatran 150 mg and apixaban were superior to warfarin; however, only dabigatran retained statistical significance in the < 75 years subgroup. SSE results for ASA in the age ≥ 75 years analysis matched those of the reference case; specifically, low-dose ASA was statistically significantly less effective at preventing SSE compared with adjusted-dose warfarin; however, no antiplatelet agents could be included in the evidence network for the age < 75 years subgroup. 4Based on the increased absolute risk reduction versus warfarin, newer anticoagulants seemed to have a greater benefit for stroke prevention in older patients (≥ 75 years) than in a younger population (< 75 years).

Table A14.9 in Appendix 14 shows all pairwise comparisons. In the < 75 years subgroup, dabigatran 150 mg was associated with statistically significantly fewer SSE versus dabigatran 110 mg, as in the reference case analysis; however, there were no statistically significant differences anymore among the NOACs in the ≥ 75 years subgroup. As in the reference analysis, all anticoagulants were significantly superior to low-dose ASA.

**TTR:**
Table A13.11 in Appendix 13 shows comparisons between warfarin and each NOAC. In the TTR< 66% (poorly controlled) subgroup, only dabigatran 150 mg yielded statistically significant results for stroke prevention over warfarin, as well as over rivaroxaban. Decreases in the number of events observed with apixaban did not reach statistical significance, as it did in the reference case. In the TTR ≥ 66% (adequately controlled) subgroup, no statistically significant improvement was observed across all interventions. Of note, dabigatran 150 mg showed substantially greater benefits when compared with warfarin in patients with a poorly controlled INR compared with adequately controlled patients. Table A14.11 in Appendix 14 shows all pairwise comparisons.

##### Major Bleeding

**Age**: Table 13.10 in Appendix 13 shows comparisons between warfarin and each NOAC. As in the reference case analysis, statistical significance for reductions in major bleeding was reached for apixaban and dabigatran 110 mg compared with adjusted-dose warfarin in the < 75 years subgroup. In this younger population, dabigatran 150 mg was also superior to warfarin, which was not the case previously. Table A14.10 shows all pairwise comparisons. There was no difference between warfarin and rivaroxaban; however, rivaroxaban appeared to cause statistically significantly more major bleeding than both doses of dabigatran. In the ≥ 75 years subgroup, apixaban achieved statistical superiority over all other anticoagulant drugs. There was no statistically significant difference between all of the remaining interventions, including between low-dose ASA and anticoagulants.

**TTR**: Table A13.12 in Appendix 13 shows comparisons between warfarin and each NOAC. While apixaban and both doses of dabigatran led to a statistically significant reduction in major bleeding in the poorly controlled subgroup (TTR < 66%), only apixaban retained statistical significance in the adequately controlled subgroup (TTR ≥ 66%). Notwithstanding statistical significance, all NOACs resulted in substantially greater risk reductions in major bleeding versus warfarin in patients with a poorly controlled INR compared with adequately controlled patients, which was more pronounced than for SSE. Table A14.12 in Appendix 14 shows all pairwise comparisons. As in the reference analysis, apixaban was also superior to dabigatran 150 mg and rivaroxaban in both subgroups.

## 4.5. Pharmacoeconomic Evaluation

### 4.5.1. Base Case Analysis

#### a. CHADS_{2} < 2

and provide the results of the base case analysis for the CHADS_{2} score < 2. Dabigatran 150 mg was the most effective treatment for QALYs (6.648), followed by apixaban (6.605). ASA medium-dose (6.182), ASA low-dose (6.150), and clopidogrel plus ASA (5.878) all produced less QALYs than the NOACs. No data were available for rivaroxaban in patients with a CHADS_{2} score < 2; therefore, rivaroxaban was not included in this analysis.

Results of Base Case Deterministic Analysis: CHADS_{2} Score < 2.

Base Case Results Cost-Effectiveness Plane for CHADS_{2} < 2.

The incremental cost per QALY gained for dabigatran 150 mg versus warfarin was $20,845. As dabigatran 110 mg and apixaban produced fewer QALYs than dabigatran 150 mg but at a greater cost, they were all dominated by dabigatran 150 mg. ASA medium-dose, ASA low-dose, and clopidogrel plus ASA were all dominated by one or more of the NOACs ().

#### b. CHADS_{2} ≥ 2

and provide the results of the base case analysis for CHADS_{2} score ≥ 2. Apixaban was the most effective treatment regarding QALYs (5.381), followed by dabigatran 150 mg (5.370). ASA low-dose (4.858) and clopidogrel plus ASA (4.942) all produced less QALYs than the NOACs.

Results of Base Case Deterministic Analysis: CHADS_{2} Score ≥ 2.

Base Case Results Cost-Effectiveness Plane for CHADS_{2} ≥2.

The incremental cost per QALY gained for both apixaban and dabigatran 150 mg versus warfarin was $17,795. The incremental cost per QALY gained for apixaban versus dabigatran 150 mg was $17,799. As dabigatran 110 mg and rivaroxaban produced fewer QALYs than dabigatran 150 mg and apixaban but at a greater cost, both of these treatments were dominated by dabigatran 150 mg and apixaban. As was the case for the CHADS_{2} <2 subgroup, the ASA treatments were not cost-effective; specifically, ASA low-dose and clopidogrel plus ASA were dominated by all the NOACs.

### 4.5.2. Sensitivity Analyses

#### a. Deterministic Sensitivity Analysis

Full details of the results of the univariate sensitivity analyses are presented in Appendices 7 and 19. In addition, parameters for which univariate sensitivity analyses did not substantially alter the results are presented in Appendix 20.

#### b. CHADS_{2} < 2

Threshold analysis found that based for λ = $50,000, apixaban would be more cost-effective than dabigatran 150 mg if the price of apixaban was less than $2.12 per day (a 34% reduction). To be cost saving compared with warfarin, dabigatran 150 mg would have to cost $1.91 per day, dabigatran 110 mg would have to cost $1.25, while apixaban would have to cost $1.61 per day. The ASA treatment would not be cost-effective compared with warfarin at any price.

Results were very sensitive to the time horizon adopted. With a time horizon of ten years, dabigatran 150 mg would be optimal if λ was greater than $52,953. With a time horizon of two years (the typical duration of the RCTs considered within the MTC), dabigatran 150 mg would be optimal, but only if λ was greater than $371,678.

If the relative effects of treatments on non-vascular deaths were included, apixaban would be optimal if λ was greater than $14,915.

Switching from dabigatran 150 mg to dabigatran 110 mg at age 80 is dominated by remaining on dabigatran 150 mg for lifetime.

#### c. CHADS_{2} ≥2

Results were very sensitive to the costs of dabigatran. Threshold analysis found that, based on a λ of $50,000, dabigatran 150 mg would be more cost-effective than apixaban if the price of dabigatran was reduced by 5% (to $3.03 per day). Rivaroxaban would have to be reduced by more than 80% to $0.49 per day to be cost-effective. To be cost saving compared with warfarin, dabigatran 150 mg would have to cost $2.04 per day, dabigatran 110 mg would have to cost $1.47, rivaroxaban would have to cost $1.37 per day, while apixaban would have to cost $2 per day.

Results were sensitive to the time horizon adopted. With a time horizon of two years (the typical duration of the RCTs considered within the MTC), apixaban would only be optimal, if λ was greater than $190,545.

Using event rates from the ROCKET-AF trial had a minor impact, slightly increasing the QALY gains from the new oral anticoagulant drugs leading to lower incremental costs per QALY gained; e.g., $12,544 for dabigatran 150mg and $11,152 for apixaban.

Excluding the effects of treatment on MI decreased the incremental cost per QALY gained for dabigatran 150 mg versus warfarin to $15,563 and increased the ratio for apixaban versus dabigatran 150 mg to $324,407. If the relative effects of treatments on non-vascular deaths were included, rivaroxaban was most effective and would be optimal if λ was greater than $10,259.

Switching from dabigatran 150 mg to dabigatran 110 mg at age 80 is dominated by remaining on dabigatran 150 mg for lifetime.

#### d. Probabilistic Sensitivity Analysis

##### CHADS_{2} < 2

The results of the probabilistic sensitivity analyses for the expected values of costs, effects, and ICERs did not vary significantly from the deterministic base case analysis ().

Results of Probabilistic Analysis: CHADS_{2} Score < 2.

However, the probabilistic sensitivity analysis demonstrates that there is uncertainty associated with conclusions relating to cost-effectiveness, because no single treatment dominated all other treatments over the whole range of cost-effectiveness thresholds (values of λ); i.e., in no case was a single treatment optimal in all replications.

At a λ of $50,000, dabigatran 150 mg was the optimal treatment in 53.9% of replications, apixaban in 24.3%, ASA medium-dose in 8.1%, dabigatran 110 mg in 6.0%, ASA low-dose in 4.7 %, warfarin in 2.9%, and clopidogrel plus ASA in 0%. Results were similar for all values of λ, from $40,000 to $100,000. If λ = $0 (i.e., costs are the only consideration), warfarin was optimal in 57.4% of replications, ASA medium-dose in 24.3%, ASA low-dose in 17.9%, dabigatran 150 mg in 0.3%, and apixaban in 0.1%.

As shown in , dabigatran 150 mg is optimal if a decision-maker is willing to pay at least $20,862 per QALY. The uncertainty associated with these results is illustrated by the cost-effectiveness acceptability curves for these two treatments in . As illustrated in , the probability of warfarin and dabigatran 150 mg being the most cost-effective treatment depends on the willingness-to-pay threshold, such that while warfarin is optimal at lower cost-effectiveness thresholds, dabigatran 150 mg becomes more cost-effective as the willingness-to-pay threshold increases.

Cost-Effectiveness Acceptability Curves for CHADS_{2} < 2.

##### CHADS_{2} ≥ 2

The expected values of costs, effects, and ICERs did not vary significantly from the deterministic base case analysis and the probabilistic analysis (), although dabigatran 150 mg was dominated by apixaban.

Results of Probabilistic Analysis: CHADS_{2} Score ≥ 2.

The probabilistic sensitivity analysis highlights the uncertainty around conclusions relating to cost-effectiveness that were also observed for the CHADS < 2 analysis. As for the CHADS_{2} < 2 analysis, for CHADS_{2} ≥ 2, no single treatment dominated all other treatments over the whole range of cost-effectiveness thresholds (values of λ); i.e., in no case was a single treatment optimal in all replications.

At a λ of $50,000, apixaban was the optimal treatment in 53.2% of replications, dabigatran 150 mg in 39.9%, dabigatran 110 mg in 4.4%, ASA low-dose in 1.5%, warfarin in 1.4%, rivaroxaban in 0.9%, and clopidogrel plus ASA in 0%. Results were similar for all values of λ, from $50,000 to $100,000. If λ = $0 (i.e. costs are the only consideration), warfarin was optimal in 85.7% of replications, ASA low-dose in 13.8%, Dabigatran 150 mg in 0.4%, and apixaban in 0.1%. The cost-effectiveness curves are presented in .

Cost-effectiveness acceptability curves for CHADS_{2} ≥2.

### 4.5.3. Analysis of Variability

#### a. CHADS_{2} Score

For a CHADS_{2} score of 0, dabigatran 150 mg would be optimal if λ was greater than $39,730. Apixaban and dabigatran 110 mg were dominated by dabigatran 150 mg. ASA low-dose, ASA medium-dose, and clopidogrel plus ASA were dominated by warfarin.

For a CHADS_{2} score of 1, dabigatran 150 mg would be optimal if λ was greater than $19,716. Apixaban, dabigatran 110 mg, ASA medium-dose, and ASA low-dose were dominated by dabigatran 150 mg. ASA low-dose, ASA medium-dose, and clopidogrel plus ASA were dominated by warfarin.

For a CHADS_{2} score ≥ 2 with no previous SSE, apixaban would be optimal if λ was greater than $16,543. Dabigatran 150 mg is subject to extended dominance through warfarin and apixaban. Dabigatran 110 mg, rivaroxaban, ASA low-dose, and clopidogrel plus ASA are all dominated by apixaban.

For a CHADS_{2} score ≥ 2 with previous minor stroke, dabigatran 150 mg would be optimal if λ was greater than $14,857. Apixaban, dabigatran 110 mg, rivaroxaban, ASA low-dose, and clopidogrel plus ASA are all dominated by dabigatran 150 mg.

For a CHADS_{2} score ≥ 2 with previous major stroke, apixaban would be optimal if λ was greater than $119,523. Dabigatran 150 mg, rivaroxaban, and dabigatran 110 mg are subject to extended dominance through warfarin and apixaban. ASA low-dose and clopidogrel plus ASA are all dominated by apixaban.

#### b. Age

For patients aged 60, the incremental cost per QALY gained for dabigatran 150 mg versus warfarin is $19,368. Apixaban, rivaroxaban, and dabigatran 110 mg were dominated by dabigatran 150 mg.

For patients aged 70, the ICERs for all therapies versus warfarin were lower than for patients aged 60, implying that treatment is more cost-effective in older patients. Dabigatran 150 mg would be preferred to warfarin if λ was greater than $14,752. Apixaban, rivaroxaban, and dabigatran 110 mg were dominated by dabigatran 150 mg.

For patients aged 80, the ICER versus warfarin was lower for rivaroxaban and apixaban but higher for both doses of dabigatran. Apixaban produced most QALYs at an incremental cost per QALY gained of $19,407. Rivaroxaban, dabigatran 150 mg, and dabigatran 110 mg were dominated by apixaban. ASA low-dose was dominated by apixaban, rivaroxaban, dabigatran 150 mg, and warfarin.

#### c. Time in TherapeuticRrange

In centres with poor INR control (centre-specific TTR < 66%), dabigatran 150 mg was the most effective treatment option, with an incremental cost per QALY gained versus warfarin of $9,993. Clopidogrel plus ASA, rivaroxaban, dabigatran 110 mg, and apixaban were dominated by dabigatran 150 mg. Clopidogrel plus ASA was also dominated by warfarin.

In centres with good INR control (centre-specific TTR > 66%), apixaban was the most effective treatment option, with an incremental cost per QALY gained versus warfarin of $40,023. Rivaroxaban was subject to extended dominance through warfarin and apixaban. Clopidogrel plus ASA, dabigatran 150 mg, and dabigatran 110 mg were dominated by apixaban.