What is Ge-Bu’s opinion?
- Two sophisticated observational studies found an increased risk of both haemorrhages and thromboembolic complications among rivaroxaban users compared to apixaban users.
- Patients with atrial fibrillation using rivaroxaban were also found to have increased mortality rates.
- Based on these two large studies, a causal relationship between the increased risk of complications or mortality and the use of rivaroxaban appears to be a clear possibility.
- When deciding on a new treatment with DOACs for atrial fibrillation or venous thromboembolism, starting with apixaban might be advantageous.
- As for current users of rivaroxaban, the evidence is as yet insufficient to consider switching to apixaban.
- The results of large randomised studies which are currently underway will in the future require a new, definitive evaluation of the choice between these two drugs.
Rivaroxaban vs. apixaban
Since their introduction, the direct acting oral anticoagulants (DOACs) have become very popular. They are mainly prescribed for the prevention of thromboembolic events in patients with atrial fibrillation or for secondary prevention after deep-vein thromboembolism. Rivaroxaban (Xarelto®) and apixaban (Eliquis®) are the most commonly used DOACs in the Netherlands.1 Together, they make up slightly over half of all prescribed anticoagulants in the Netherlands, which means that overall, they are used about one and a half times as often as vitamin K antagonists. Rivaroxaban and apixaban are both reversible inhibitors of activated factor X and are referred to in Dutch guidelines as equivalent options, both relative to each other and in comparison with vitamin K antagonists and other DOACs.2,3 The other DOACs authorised in the Netherlands are edoxaban and dabigatran. Both are used far less, and dabigatran also has a different mode of action, namely inhibition of activated thrombin. The rest of the present article is exclusively concerned with comparing apixaban and rivaroxaban, as two large, high-quality observational studies have recently been published on them. Actually, some studies have been published which suggest that the effectiveness and adverse effects of dabigatran lie somewhere in between those of apixaban and rivaroxaban.4,5 A detailed discussion of these studies, with an evaluation of their quality, is however beyond the scope of the present article.
Larger plasma level variations with rivaroxaban?
Rivaroxaban and apixaban have comparable half-lives in older adults (11 to 13 hours for rivaroxaban and about 12 hours for apixaban).6,7 Among younger adults, the half-life of rivaroxaban is even shorter, 5 to 9 hours. Nevertheless, rivaroxaban is dosed once daily, while apixaban is dosed twice daily. Research has shown that as a result of this, as expected, the plasma concentrations of apixaban tend to vary less than those of rivaroxaban.8,9 Shortly after intake, rivaroxaban reaches higher peak levels, which might increase the risk of haemorrhagic complications. In addition, rivaroxaban has lower plasma levels in the period before the next dose is taken, which might increase the risk of thromboembolic complications. In theory therefore, rivaroxaban could carry a greater risk of both complications compared to apixaban. This argumentation can, however, also be reversed. Longer-lasting low levels of rivaroxaban could result in a lower risk of haemorrhage, while a higher peak level is more effective in dissolving incipient blood clots.
Evidence from research
Rivaroxaban and apixaban have not yet been directly compared in a randomised study. Trials currently underway include the randomised COBRA Study (Comparison of Bleeding Risk between Rivaroxaban and Apixaban for the treatment of acute venous thromboembolism, NCT03266783). Its results are, however, not expected until 2024 at the earliest.10 Some observational studies have found evidence of a higher incidence of haemorrhagic and thromboembolic complications associated with the use of rivaroxaban, compared to apixaban, among patients with venous thromboembolism.11,12 One of these studies found an increased mortality rate among apixaban users compared to rivaroxaban users.11 However, these studies were of limited size, which means that the differences observed could also be attributable to chance. Furthermore, they were probably less reliable because patients were included only during a brief period after apixaban was first marketed. This creates the risk that patients only received apixaban for special reasons, precluding a comparison with those receiving rivaroxaban.
Results were recently published of two large, well-designed and conducted observational studies comparing the effectiveness and adverse effects of rivaroxaban and apixaban among patients with atrial fibrillation and patients with venous thromboembolism.9,13 The study among patents with atrial fibrillation found that both the effectiveness and safety of rivaroxaban were significantly poorer than those of apixaban.9 Apixaban also appears to be more effective and safer than rivaroxaban for the secondary prevention of thromboembolic events.13
Rivaroxaban vs. apixaban for atrial fibrillation
An American retrospective cohort study evaluated the data of patients with atrial fibrillation aged 65 years and older who were enrolled in the Medicare social insurance programme.9 The primary outcome measure was the difference between apixaban and rivaroxaban in the incidence of the composite endpoint of major ischaemic complications (cerebral infarction or systemic embolism) and haemorrhagic complications (intracranial or fatal extracranial bleeding). Patients were included when they received the first standard dosage of 20 mg rivaroxaban once daily, or 5 mg apixaban twice daily, or a reduced dosage of 10 mg rivaroxaban once daily or 2.5 mg apixaban twice daily for patients with impaired kidney function.
The study included patients diagnosed with atrial fibrillation in the last 90 days before the reference date. These patients had to have an indication for the use of rivaroxaban or apixaban and should not have reached any of the study endpoints within the past 30 days. The occurrence of the endpoints from the reference date onwards was then recorded. For each individual patient in this group, a propensity score was calculated based on 208 variables that were expected to be associated with both the outcome and the choice of medication. This propensity score was then used for an ‘inverse probability of treatment weighting’ (IPW), after which the balance of potential confounders in the weighted population was checked. After the balance had been approved, the hazard ratios and differences in incidence for the primary outcome measure were estimated for the weighted population. Further details about this study method are presented in the Background Information below.
A total of 2,078,642 patients were screened, 581,451 of whom were eventually included, 227,572 rivaroxaban users and 353,879 apixaban users. The incidence of the composite endpoint for both major ischaemic and major haemorrhagic complications was 16.1 per 1000 person years in the rivaroxaban group and 13.4 per 1000 person years in the apixaban group. The difference in incidence was 2.7 (95 % confidence interval [CI] 1.9 to 3.5) per 1000 person years; the hazard ratio was 1.18 (1.12 to 1.24). Table 1 presents the results for the individual endpoints. Despite the small numbers, the differences in incidence among the patients who used a reduced dosage (10 mg rivaroxaban once daily or 2.5 mg apixaban twice daily) appeared to be even larger than among those using the standard dosage. The incidence of the primary composite endpoint among the users of the reduced rivaroxaban dosage was 6.4 (4.1 to 8.7) per 1000 person years higher than among the users of the reduced apixaban dosage, while the corresponding difference among the users of the standard dosage was 1.8 (1.0 to 2.6) per 1000 person years. This difference can possibly be explained by the fact that the patients using the reduced dosages are more vulnerable than those using the standard dosage, and therefore are at higher risk of the endpoint used in this study.
Although the differences expressed per 1000 person years seem small, the total impact is considerable. This is due to the combination of a very large groups of users and the usually chronic use. The number needed to harm (NNH) for the total group of users for the composite endpoint is 37 for 10 years of treatment. This means that if 37 patients are treated for 10 years with rivaroxaban, there will be one complication more than if these people had been treated with apixaban. In the Netherlands, there were about 152,000 rivaroxaban users in the first half of 2021.1 For the Dutch situation, therefore, this corresponds to about 400 additional complications a year. The NNH for the users of the standard dosage is 56 for 10 years of treatment, while the NNH for the users of the reduced dosage is 4 for 10 years of treatment.
Tabel 1. Incidence, incidence difference and hazard ratio for the various endpoints for rivaroxaban vs. apixaban
Incidence/1000 person years
Incidence/1000 person years
Incidence difference (95% CI)
Hazard ratio (95% CI)
2.7 (1.9 – 3.5)
1.18 (1.12 – 1.24)
Major thromboembolic complications
1.1 (0.5 – 1.7)
1.12 (1.04 – 1.20)
Major haemorrhagic complications
1.6 (1.1 – 2.1)
1.26 (1.16 – 1.36)
3.1 (1.8 – 4.5)
1.06 (1.02 – 1.09)
Mortality from thromboembolic and haemorrhagic complications
1.2 (0.8 – 1.6)
1.34 (1.21 – 1.48)
CI: confidence interval
Rivaroxaban vs. apixaban for venous thromboembolism
An American retrospective cohort study examined the data of patients with deep vein thrombosis or pulmonary embolism who were insured with a commercial health insurer.13 The primary effectiveness outcome was recurrent venous thromboembolism (deep vein thrombosis of pulmonary embolism); the safety outcomes were gastro-intestinal or intracranial haemorrhages requiring hospitalisation.
The study included patients aged 18 years and over with a first diagnosis of deep vein thrombosis or pulmonary embolism in the past 30 days, who started using rivaroxaban or apixaban for the first time and had been enrolled with the insurer for at least 12 months. For each individual patient in this group, a propensity score was calculated, based on 45 variables which were expected to be potentially associated with both the outcome and the choice of medication. This propensity score was used to match apixaban users with rivaroxaban users. The researchers did not use exact matching but matched within a certain bandwidth (calliper matching). They then checked the balance of the potential confounders in the matched population. If this was judged to be sufficient, they then estimated the incidence and hazard ratios for the primary outcome measure in the matched population. Further explanation of this method is provided in the Background Information below.
Of the 260,622 patients screened, 49,900 were eventually included, 21,613 of whom used rivaroxaban and 28,287 used apixaban. For each group, 18,618 of them could be matched.
In this matched group of 37,236 patients, 475 apixaban users (8.9 per 100 person years) and 595 rivaroxaban users (11.4 per 100 person years) developed a recurrent venous thromboembolism. The hazard ratio for recurrent venous thromboembolism in apixaban users, compared to rivaroxaban users, was 0.77 (95% CI 0.69 to 0.87) and the difference in the risk increased from 0.006 (0.005 to 0.011) in the first two months to 0.011 (0.011 to 0.013) in the first six months.
A total of 386 apixaban users (7.2 per 100 person years) and 577 rivaroxaban users (11.0 per 100 person years) had a gastro-intestinal or intracranial haemorrhage requiring hospitalisation. The hazard ratio for bleeding events in apixaban users compared to rivaroxaban users was 0.60 (0.53 to 0.69) and the difference in the risk increased from 0.011 (0.010 to 0.011) in the first two months to 0.015 (0.013 to 0.015) in the first six months.
No ‘number needed to harm’ (NNH) was calculated for this study, as the difference in the risk changed too much over time, making an extrapolated estimate of the NNH over a longer period unreliable.
New-user cohort design
Both of the studies discussed above used a new-user cohort design, which means that only patients who were starting the study drug for the first time could be included in the cohort. Past research had also studied current-user cohorts, a design in which all patients using the drug at the time the study starts are included. This was, however, found to create a bias, as has become clear in a study of the effect of postmenopausal hormone therapy. This study found contradictory results between observational studies on the one hand and a large randomised study on the other.14
In response to this controversy, it was concluded that the observational studies had been biased, as the current users represented a selection from the patients who had tolerated the drug well over a longer period, without serious adverse effects. Patients who were liable to develop complications automatically disappeared from the study. This bias could be prevented by looking at the same database again, and this time only selecting new users. The results were then found to be comparable to those of the randomised study.14 This design also resembles more closely a randomised study, since in a randomised study all patients also start using the study medication for the first time.
The new-user cohort design is therefore nowadays the new standard for observational studies of drug use. Analyses in current-user cohorts are now regarded as unreliable and should be avoided.15 A disadvantage of this development is that the results of studies using the new-user cohort design cannot be simply applied to current users, as current users in routine practice represent a selection consisting of those users who have tolerated the drug well over a longer period. This results in additional uncertainty about the usefulness and necessity of switching users who have had no complications while using rivaroxaban over a longer period to another drug.
Confounding arises when patients with different prognoses for the outcome are also given different treatments. Conventional adjustment for confounding often makes use of a regression model in which the probability of the outcome is estimated on the basis of the treatment and the confounders. Both of the studies discussed here make use of a propensity score to adjust for confounding.
A propensity score provides, for each patient, an individual estimate of the propensity, that is, the ‘tendency’ to receive a particular treatment. This estimate is made using a logistic regression model, based on all measured potential confounders. A propensity score can then be used in different ways to adjust for confounding. The quality and completeness of this adjustment for confounding obviously depends on the quality and completeness of the model with which the propensity score is estimated.
Inverse probability of treatment weighting
Inverse probability of treatment weighting (IPW) is a special form of standardisation, which uses the propensity score to determine the weights that must be used for standardisation. The propensity score indicates the individual probability of treatment, given the presence or absence of the confounders. In IPW, each patient’s data is weighted in the analyses by the inverse of the probability of the patient receiving the treatment they actually received. Thus, patients who were given rivaroxaban are weighted by 1/[probability of rivaroxaban], while patients who received apixaban are weighted by 1/[probability of apixaban].
As a result, a patient who, according to all confounders present, has a high probability of rivaroxaban, is therefore given a lower weight if they actually receive rivaroxaban. However, if this patient receives apixaban, the probability of apixaban is used. This probability is then very small, as it is 1-[probability of rivaroxaban], since all patients receive one of the two drugs. As a result, this patient is given a very high weight.
IPW thus assumes that there are ‘rare patients’ and ‘less rare patients’. The rare patients are those who receive a treatment for which they had a low probability according to their propensity score. Since the propensity score is determined by the confounders, this means that these patients have a confounder profile that occurs in few patients in the treated group. Giving these rare patients a higher weight results in a weighted population in which all confounder profiles are equally frequent. Doing this for both the rivaroxaban and apixaban groups also results in two weighted groups which at baseline have comparable confounder profiles and hence have a comparable prognosis for the outcome in question, thus adjusting for the confounding. Of course, the quality and completeness of this adjustment is once again determined by the quality and completeness of the model that is used to estimate the propensity score.
Matching is an attempt to find for each patient who uses rivaroxaban another patient who uses apixaban and who, at baseline, has roughly the same prognosis for the outcomes studied. This simulates, at the outset of the study, the starting situation in a randomised study, as two groups are created with a comparable prognosis for the outcome. The difference is that these two groups are not determined by chance, but are deliberately composed in a meticulous manner.
Matching can be done on the basis of individual confounders, but is often also done using a propensity score. In this method, each patient given rivaroxaban is matched with another patient who had the same probability of receiving rivaroxaban, but did actually receive apixaban. The individual confounders do then not necessarily have to be the same in each matched pair, as long as they have an equal probability of receiving either treatment.
The advantage of this approach is that a table showing the baseline characteristics of the matched population can make it immediately clear whether the main variables after matching are indeed equally distributed. A potential disadvantage is that it may not be possible to find a suitable match for each patient. In that case, one may use ‘calliper matching’, whereby the matching is not exact, but remains within a previously set bandwidth, the ‘calliper’. This allows more matched pairs to be formed, but carries the risk that the less precise matching leads to a residual disbalance between the groups. The chosen width of the calliper is therefore of crucial importance. If it is too narrow, many patients will become ineligible, as they cannot be matched. If it is too wide, some confounding will remain after matching. If too many patients become ineligible, this not only results in a loss of power, but also in a loss of representativeness and generalisability.
The observational studies discussed in this article may be affected by ‘residual confounding’, which means that after adjustment for confounding, a certain degree of confounding still remains. Residual confounding can arise in two different ways.
Firstly, the statistical methods used may not have resulted in complete adjustment for confounding. This can happen, for example, if after adjustment for the difference between patients over and under 65 years of age, there is still confounding due to age within the individual age groups, because the chosen categories have been too coarsely divided. Another cause may be that a confounder is added to a model as a linear parameter whereas the association to be adjusted is not linear. This may happen, for instance, if age is added to a model as a continuous variable, whereas the association with the outcome follows a U- or of S-shaped curve.
Secondly, in some cases not all confounders are known and measured. In such cases the confounders are impossible to adjust for, and the confounding due to them remains.
Both studies discussed here used sound techniques to adjust for confounding. In addition, both studies have adjusted for many potential confounders. Finally, both studies checked whether after adjustment (by weighting or matching) the confounders were well-balanced between the groups, which indeed proved to be the case. This balance indicates that the techniques they used were correctly applied, so sufficient adjustment for confounding due to the variables used has been achieved. This excludes the first source of potential residual confounding. However, this is the only aspect that can be shown by the balance. The fact that the techniques used have been correctly applied unfortunately does not guarantee that there is no confounding left, as the second source of residual confounding remains. There is no way to be absolutely sure that all potentially important confounders have been adjusted for. It does appear, however, that both studies have adjusted for the main potential confounders. Hence, the presence of major residual confounding does not seem likely.
Details of the studies discussed
Cohort: US Medicare recipients aged 65 years or over
Primary endpoint(s): composite endpoint of the incidence of large ischaemic (cerebral infarction or systemic embolism) and haemorrhagic (intracranial or fatal extracranial bleeding) complications
Secondary endpoints: nonfatal intracranial bleeding, total mortality
Setting and duration: 1 January 2013 (first year in which apixaban was reimbursed by Medicare) to 30 November 2018
Inclusion criteria: patients with a diagnosis of atrial fibrillation in the past 90 days and a first retrieved prescription for rivaroxaban or apixaban, having been enrolled in Medicare for at least 1 year, and having had at least one contact in that year
Main exclusion criteria: oral anticoagulants within the past year or a terminal illness
Groups compared: rivaroxaban, normal dosage (20 mg once daily) or reduced dosage (10 mg once daily) compared with apixaban, normal dosage (5 mg twice daily) or reduced dosage (2.5 mg twice daily)
Number of patients included: 581,451 patients (227,572 rivaroxaban and 353,879 apixaban)
Duration of treatment: median 174 days (interquartile range: 62 to 397 days)
Statistics: 208 variables were used to determine a propensity score, in order to estimate the probability of treatment of individual patients. This propensity score was then used to carry out an inverse probability of treatment weighting. In the resulting balanced population, hazard ratios and differences in incidence per 1000 person years were estimated
Limitations: potential residual confounding due to factors not measured; censoring for discontinuation of medication; no conclusions possible about biological mechanism; limited generalisability due to inclusion of specific population; no information available about correct or incorrect intake of medication
Funding: National Heart Lung and Blood Institute (NHLBI)
Trial registration: none
Analyses carried out by: authors, independent of sponsor
Conflicts of interest: 5 of the 10 authors (no relation with rivaroxaban or apixaban)
Cohort: insured with commercial health insurer
Primary endpoint(s): for effectiveness: recurrent venous thrombosis (deep vein thrombosis or pulmonary embolism); for safety: gastro-intestinal or intracranial haemorrhage requiring hospitalisation
Secondary endpoints: none
Setting and duration: 1 January 2015 to 30 June 2020
Inclusion criteria: patients with a diagnosis of venous thrombosis within 30 days before first treatment with rivaroxaban or apixaban, and having been with the insurer for at least 1 year
Main exclusion criteria: anticoagulants or a diagnosis of deep venous thrombosis or pulmonary embolism during the year before the present diagnosis
Groups compared: rivaroxaban uses were compared with apixaban users
Number of patients included: 37,236 patients (18,618 per group after matching)
Duration of treatment: median 102 days (interquartile range: 30 to 128 days)
Statistics: 45 variables were used to determine a propensity score, in order to estimate the probability of treatment of individual patients. This propensity score was then used to match as many apixaban users with a rivaroxaban user, using calliper matching. In the resulting balanced population, incidences per 100 person years and hazard ratios were estimated and survival curves constructed
Limitations: potential residual confounding due to factors not measured; only serious outcomes (requiring hospitalisation) included; limited generalisability due to inclusion of specific population, which became even smaller after matching; no information available about correct or incorrect intake of medication
Trial registration: none
Analyses carried out by: not reported
Conflicts of interest: 3 of the 4 authors (2x with apixaban manufacturer and 1x with rivaroxaban manufacturer)
- Stichting Farmaceutische Kengetallen. Available at: https://www.sfk.nl/publicaties/PW/2021/rivaroxaban-stoot-acenocoumarol-van-de-troon. Accessed 17 May 2022.
- Nederlands Huisartsen Genootschap. NHG-Standaard Atriumfibrilleren. September 2017, latest revision September 2017 (under review). Available at: https://richtlijnen.nhg.org/standaarden/atriumfibrilleren#samenvatting-richtlijnen-beleid. Accessed 17 May 2022.
- Nederlands Huisartsen Genootschap. NHG-Standaard Diepveneuze trombose en longembolie. Lateste revision January 2021 (under review). Available at: https://richtlijnen.nhg.org/standaarden/diepveneuze-trombose-en-longembolie#samenvatting-richtlijnen-beleid. Accessed 17 May 2022.
- Rutherford OW, Jonasson C, Ghanima W, Söderdahl F, Halvorsen S. Comparison of dabigatran, rivaroxaban, and apixaban for effectiveness and safety in atrial fibrillation: a nationwide cohort study. Eur Heart J Cardiovasc Pharmacother. 2020 Apr 1;6(2):75-85. doi: 10.1093/ehjcvp/pvz086.
- Noseworthy PA, Yao X, Abraham NS, Sangaralingham LR, McBane RD, Shah ND. Direct Comparison of Dabigatran, Rivaroxaban, and Apixaban for Effectiveness and Safety in Nonvalvular Atrial Fibrillation. Chest. 2016 Dec;150(6):1302-1312. doi: 10.1016/j.chest.2016.07.013.
- Zorginstituut Nederland. Farmacotherapeutisch Kompas. Rivaroxaban. Available at: https://www.farmacotherapeutischkompas.nl/bladeren/preparaatteksten/r/rivaroxaban#overdosering. Accessed 17 May 2022.
- Zorginstituut Nederland. Farmacotherapeutisch Kompas. Apixaban. Available at: https://www.farmacotherapeutischkompas.nl/bladeren/preparaatteksten/a/apixaban. Accessed 17 May 2022.
- Toorop MMA, van Rein N, Nierman MC, Vermaas HW, Huisman MV, van der Meer FJM, et al. Inter- and intra-individual concentrations of direct oral anticoagulants: The KIDOAC study. J Thromb Haemost. 2022 Jan;20(1):92-103. doi: 10.1111/jth.15563.
- Ray WA, Chung CP, Stein CM, Smalley W, Zimmerman E, Dupont WD, et al. Association of Rivaroxaban vs Apixaban With Major Ischemic or Hemorrhagic Events in Patients With Atrial Fibrillation. JAMA. 2021 Dec 21;326(23):2395-2404. doi: 10.1001/jama.2021.21222. Erratum in: JAMA. 2022 Apr 5;327(13):1294.
- NIH. ClinicalTrials.gov. Comparison of Bleeding Risk Between Rivaroxaban and Apixaban for the Treatment of Acute Venous Thromboembolism (COBRRA). Available at: https://clinicaltrials.gov/ct2/show/NCT03266783. Accessed 17 May 2022.
- Sindet-Pedersen C, Staerk L, Pallisgaard JL, Gerds TA, Berger JS, Torp-Pedersen C, et al. Safety and effectiveness of rivaroxaban and apixaban in patients with venous thromboembolism: a nationwide study. Eur Heart J Cardiovasc Pharmacother. 2018 Oct 1;4(4):220-227. doi: 10.1093/ehjcvp/pvy021.
- Dawwas GK, Brown J, Dietrich E, Park H. Effectiveness and safety of apixaban versus rivaroxaban for prevention of recurrent venous thromboembolism and adverse bleeding events in patients with venous thromboembolism: a retrospective population-based cohort analysis. Lancet Haematol. 2019 Jan;6(1):e20-e28. doi: 10.1016/S2352-3026(18)30191-1.
- Dawwas GK, Leonard CE, Lewis JD, Cuker A. Risk for Recurrent Venous Thromboembolism and Bleeding With Apixaban Compared With Rivaroxaban: An Analysis of Real-World Data. Ann Intern Med. 2022 Jan;175(1):20-28. doi: 10.7326/M21-0717.
- Hernán MA, Alonso A, Logan R, Grodstein F, Michels KB, Willett WC, et al. Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart disease. Epidemiology. 2008 Nov;19(6):766-79. doi: 10.1097/EDE.0b013e3181875e61.
- Ray WA. Evaluating medication effects outside of clinical trials: new-user designs. Am J Epidemiol. 2003 Nov 1;158(9):915-20. doi: 10.1093/aje/kwg231.