Alternative tumour necrosis factor inhibitors (TNFi) or abatacept or rituximab following failure of initial TNFi in rheumatoid arthritis: the SWITCH RCT

Tumour necrosis factor inhibitors

Tumour necrosis factor inhibitor drugs [etanercept (Enbrel^®; Pfizer, New York City, NT, USA), infliximab (REMICADE^®, Janssen Pharmaceutical, Beerse, Belgium), adalimumab (HUMIRA^®; Abbott, now AbbVie, North Chicago, IL, USA), certolizumab pegol (CZP) (CIMZIA^®; UCB, Brussels, Belgium) and golimumab (Simponi^®; Janssen Pharmaceutical)] in combination with MTX produce better outcomes in RA than in placebo or treatment with MTX alone. 31–37 TNFi drugs, however, differ in several respects:

1. molecule type [chimeric (mouse–human) monoclonal antibody (infliximab), fully human monoclonal antibody (adalimumab, golimumab), pegylated Fab fragment of a humanised monoclonal antibody (CZP) and a TNF receptor fusion protein (etanercept)]38

2. target (etanercept binds both TNF and another cytokine, lymphotoxin alpha)39

3. binding affinity to TNF

4. mechanism of drug action40–42

5. clinical administration (intravenous vs. subcutaneous).

Non-tumour necrosis factor inhibitors

Following the development of TNFis, recognition of other key cytokines and immune cells in RA pathogenesis43 led to the development of additional bDMARDs: rituximab (MabThera; Roche, Basel, Switzerland), a chimeric anti-CD20-depleting monoclonal antibody,44 tocilizumab (Actemra^®; Roche), an IL-6 receptor monoclonal antibody45 and abatacept (Orencia^®; Bristol-Myers Squibb, New York City, NY, USA), a recombinant fusion protein T-cell co-stimulation blocking agent. 46 All of these bDMARDs demonstrated significant benefits compared with placebo and MTX in MTX-inadequate response44,47,48 and TNFi-inadequate response49–51 groups, respectively.

Non-response to tumour necrosis factor inhibitor

In the context of first-line bDMARD TNFi failure, NICE guidance recommends using rituximab as second-line bDMARD. 59 Switching to alternative TNFi, abatacept or tocilizumab, is permitted only if patients have had an inadequate response to rituximab57 or are intolerant of rituximab57,59 or if rituximab is contraindicated. 57,59 This is in the absence of any trial data demonstrating that rituximab is more appropriate than the alternative bDMARDs. Of note is the fact that this technology appraisal guidance applies to bDMARD use with background MTX. For individuals who are unable to take MTX, TNFi switching is permitted. This guidance has not been comprehensively updated following approval of tocilizumab and abatacept as first-line bDMARDs.

It is the absence of robust trial data to support NICE’s guidance regarding the process to follow in the event of failure of initial TNFi treatment (discussed below), effectively limiting the treatment choice to rituximab, which we recognise is not effective for all individuals, that is the basis for the SWITCH study.

Observational studies

Several early-phase uncontrolled studies and an initial small randomised study suggested benefit in switching between TNFi agents. 60–70 A report of extremely high responses on alternative TNFi agent in specific subgroups of patients62 also indicated the potential value and the need to explore this approach further. A literature review71 documented 29 reports on switching from one TNFi to another in RA, with the data largely indicating benefit of switching from a first TNFi to a second, with switching for secondary non-response likely to be more effective. A systematic review reported similar findings. 72

Randomised controlled trials

No randomised controlled trials (RCTs) of switching between adalimumab, etanercept and infliximab have been conducted in patients in whom these three established TNFi have failed. The rationale and argument for switching between different TNFi drugs, however, were further strengthened by a large, international, multicentre, randomised, Phase II study73 that investigated 461 patients who had previously received and either failed or were intolerant of one or more TNFis. Patients were randomised to either golimumab (50 mg or 100 mg every 4 weeks) or placebo. American College of Rheumatology 20 (ACR20) response rates at week 14 were significantly higher in the golimumab groups than in the placebo group (35% and 38% vs. 18%, respectively). More recently, two RCT studies,74,75 one with open-label evaluation,74 have demonstrated significant efficacy of CZP in prior TNFi treatment failures. In a small but first prospective RCT,74 patients in whom an initial TNFi was stopped because of secondary non-response (i.e. the initial response to the TNFi was lost) were randomised to 12 weeks of either CZP (n = 27) or placebo (n = 10), followed by an open-label CZP 12-week period. The primary end point was the proportion of patients reaching an ACR20 response by week 12, observed in 61.5% of patients in the CZP group, compared with 0% of patients in the placebo group. Placebo patients who were switched blindly to CZP attained similar results seen with CZP in weeks 0–12. As this result was highly significant, study inclusion was terminated after entry of 33.6% of the originally planned 102 patients. The REALISTIC study, a 28-week Phase IIIb study, assessed safety and maintenance of response to CZP in a diverse population of RA patients, stratified by prior TNFi exposure, concomitant MTX use and disease duration.

Randomised controlled trials

Randomised controlled trials49–51 and their long-term extension studies (LTEs)76–78 have demonstrated the benefits of non-TNFi bDMARDs over placebo/MTX following TNFi-inadequate response.

The randomised evaluation of long-term efficacy of rituximab in RA (REFLEX) study evaluated the efficacy of rituximab versus placebo in patients receiving MTX who had failed at least one TNFi. 49 Significantly more rituximab-treated patients than placebo-treated patients achieved ACR20, American College of Rheumatology 50 (ACR50), American College of Rheumatology 70 (ACR70) and moderate to good European League Against Rheumatism (EULAR) responses at week 24. Of note, despite a significant reduction in Disease Activity Score in the rituximab group, the mean DAS28 at week 24 was still high, at 5.1 units (a reduction of 1.83 from 6.9 at baseline). 49 In the LTE study (and thus a selected subgroup), rituximab showed sustained effects on joint damage progression. 76 The ATTAIN (A Therapeutic Trial of Afatinib In the Neoadjuvant Setting) study51 compared the efficacy of abatacept and placebo/MTX in patients with TNFi-inadequate response and found that significantly more patients in the abatacept group achieved ACR20, ACR50 and ACR70 responses, impressive quality-of-life results and improvement in DAS28 (a reduction in Disease Activity Score of > 1.2 in 70% in the abatacept group vs. 18.2% in the placebo group)51 and patients continued to maintain these improvements throughout the 2-year LTE study. 77 The RA study in anti-TNF failures (RADIATE) study compared the efficacy of tocilizumab (8 mg/kg or 4 mg/kg) plus MTX to placebo plus MTX in patients with one or more TNFi-inadequate responses. 50 At week 24, more patients in the tocilizumab groups than in the placebo group achieved ACR20, ACR50 and ACR70 responses and good or moderate EULAR responses. 50 The efficacy of tocilizumab was maintained for up to 4.2 years during the LTE study. 78

Observational studies

A number of observational studies have compared clinical response after switching to either rituximab or alternative TNFi in patients who failed initial TNFi treatment. 79–82 As summarised below, most have suggested better efficacy on switching to rituximab, although there are also reports of equivalent clinical responses in patients who switched to either alternative TNFi or rituximab following failure of one or more TNFi therapies. 83,84 These observational studies, however, had several design limitations, such as small sample sizes,79 selection bias,79–82 pooling all causes of TNFi failure81 and missing data,79–82 although they tried to address these issues by calculating propensity scores and using multivariable analysis techniques. 79–82

Analysis of patients with RA in the Swiss Clinical Quality Management in Rheumatic Diseases RA registry (SCQM-RA),79,80 who had treatment failure with at least one TNFi and were switched either to alternative TNFi or to one cycle of rituximab, showed that switching to rituximab may be more effective than switching to another TNFi. Furthermore, when the motive for switching was ineffectiveness of the TNFi, patients who received rituximab achieved a significantly better improvement in DAS28 at 6 months than patients who received alternative TNFi. 80 However, when the reason for switching was other causes, the improvement in DAS28 was similar in the two groups. 80 The same registry also reported that rituximab was as effective as alternative TNFi in preventing joint erosions in patients who had previous treatment failure with a TNFi. 83 A study of 1300 patients with RA on the British Society of Rheumatology (BSR)’s Biologics Register who had a failed response to their first TNFi treatment and were switched to a single cycle of either rituximab or alternative TNFi found that patients who switched to rituximab had better EULAR responses and were more likely to achieve improvement in their Health Assessment Questionnaire (HAQ) scores. 81 More recently, a global observational real-life study (SWITCH-RA)82 also showed that among patients with RA who failed to respond to, or were intolerant of, a single previous TNFi, those who were switched to rituximab achieved significantly better clinical responses at 6 months than those patients who were switched to a second TNFi. However, further subgroup analysis showed that these differences were observed only in seropositive [to either or both of rheumatoid factor (RF) and anti-citrullinated peptide antibody (ACPA)] patients who switched because of lack of efficacy of the first TNFi. 82

An observational study from the US Consortium of Rheumatology Researchers of North America (CORRONA) cohort85 reported clinical effectiveness of abatacept versus subsequent TNFi in patients with RA following one or more TNFi drug failures, using propensity scoring to reduce bias attributable to systematic prescribing practices. Six- and/or 12-month response outcomes with change in disease activity, remission rates based on the Clinical Disease Activity Index (CDAI) and modified DAS28 (mDAS28) and American College of Rheumatology (ACR) response rates were evaluated. The main analysis included all patients regardless of the reason for switching in the main analysis, with inadequate response to prior TNFi addressed in a sensitivity analysis. For the primary outcome (minimum clinically important difference in the change in CDAI score of 4.3) and the secondary outcomes, no differences between the two treatments were recorded.

Head-to-head comparisons

Gottenberg et al. 86 reported a 52-week pragmatic open-label RCT that randomised 300 patients who did not respond to a first TNFi to receive either alternative TNFi or an alternative-mechanism bDMARD (abatacept, rituximab or tocilizumab). This was a superiority trial, with the primary outcome of a good or moderate EULAR response at 6 months. Primary outcome was achieved in a significantly greater proportion of patients in the non-TNFi group, with 70% achieving a good or moderate EULAR response, compared with 52% in the second TNFi group. Similar differences were observed from week 12 and persisted at week 52, with, in addition, significantly better low disease activity and remission rates. Although an instructive and randomised study, the multiple treatment options included within the non-TNFi group limit the extent to which these data can inform on which specific targeted agent should be considered.

In contrast, a recent preliminary report from a Dutch randomised trial of 144 patients with RA who had failed a first TNFi and were randomised (1 : 1 : 1) to receive alternative TNFi, abatacept or rituximab, showed that there was comparable improvement in the DAS28, HAQ scores and Short Form questionnaire-36 items (SF-36) outcome measures over a 12-month period. 87 Rituximab therapy was the most cost-effective of the three (although this finding may not be true in other countries with different health-care provision and pricing structures). 87 Further studies with larger sample sizes and inclusion of tocilizumab in the treatment options are needed to confirm these results.

Primary objective

The primary objective was to determine whether or not an alternative-mechanism TNFi or abatacept is non-inferior to rituximab in disease response at 24 weeks post randomisation.

Secondary objectives

The secondary objectives were:

• to compare alternative TNFis and abatacept with rituximab for disease response, quality of life, toxicity and safety over 48 weeks

• to undertake an evaluation of the cost-effectiveness of switching patients to alternative TNFi, abatacept or rituximab

• to compare structural and bone density outcomes for abatacept and alternative TNFis with those for rituximab over 48 weeks using plain radiography and bone densitometry score.

Exploratory objectives

The exploratory objectives were:

• to determine the optimal sequence of treatments by assessing whether or not the response to the second treatment in patients with RA is affected by which initial TNFi the patients failed treatment on (TNFi monoclonal or TNFi receptor fusion protein)

• to evaluate whether or not the response to the second treatment (alternative TNFi, abatacept or rituximab) is affected by whether or not the patient was a primary (no initial response) or secondary (loss of an initial response) response failure to their initial TNFi therapy

• to ascertain whether or not seropositive (to either or both of RF and ACPA) and seronegative patients with RA behave differently in their response and disease outcome measures across the three treatment arms, particularly with respect to rituximab.

Recruitment

Patients were approached during standard clinic visits for the management of their RA, or were identified by waiting lists, registries or reviews of case records, and sent a personalised letter inviting them to participate. Patients were provided with verbal and written details about the trial and had as long as they required to consider participation. Assenting patients provided written consent before being registered into the trial and formally assessed for eligibility. Patients at Chapel Allerton Hospital also had the option of giving informed consent for blood and tissue samples to be taken for the SWITCH trial biobank for future scientific research. The participant information sheet and consent forms are provided in Appendix 1.

Abatacept

Abatacept is a selective T-cell co-stimulation blocking agent that is a fusion protein composed of the Fc region of the immunoglobulin G1 (IgG1) fused to the extracellular domain of cytotoxic T-lymphocyte-associated protein 4 (CTLA-4).

Alternative tumour necrosis factor inhibitors

Rituximab (control)

A genetically engineered chimeric (human–murine) monoclonal antibody against the B-cell protein marker CD20 (clusters of differentiation 20).

Efficacy of rituximab to placebo was established in a similar patient population in the REFLEX study. 49

Table 1 illustrates the treatment regimen including mode of administration and dose, for each of the three treatment arms. The intervention period was 48 weeks, achieved via treatment regimens administered for a minimum of 24 weeks.

Screening and baseline assessments

Following written informed consent and prior to any trial-related procedures, patients were registered into the study. All patients had a screening assessment within 4 weeks prior to the baseline assessment (and, when applicable, the assessment was repeated at the baseline assessment) to establish eligibility. The clinical assessment included a medical history, a physical examination, which included measurements of height, weight and vital signs, electrocardiography (ECG), chest radiography and a screen for TB (if not performed within specified time window prior to the screening visit). In addition, a 28 swollen joint count (SJC) and tender joint count (TJC) were performed (Figure 1), and blood and urine tests [haematology, blood chemistry, C-reactive protein (CRP) level test, erythrocyte sedimentation rate (ESR), serological tests, hepatitis B and C screen, a pregnancy test and urinalysis] were undertaken. At the baseline assessment, a further blood test was undertaken to assess glucose levels and lipid profiles.

FIGURE 1.

Manikin showing joints to be included in the SJC and TJC. MCPJ, metacarpophalangeal joint; PIPJ, proximal interphalangeal joint.

At screening and baseline assessments, patients completed a Global Assessment of Arthritis using a visual analogue scale (VAS) and reported the extent of their early-morning stiffness. The clinician assessed the Global Disease Activity using a VAS. At the baseline assessment, patients completed a Global Assessment of Pain VAS and an assessment of their general health using a VAS.

Intervention and observational phase assessments

Randomised patients attended clinic assessment visits at weeks 12, 24, 36 and 48 in the interventional phase and at weeks 60, 72, 84 and 96 in the observational phase. Patients allocated to the subcutaneous TNFi or abatacept therapies had additional standard assessment for safety purposes (usually week 4) in line with local practice. At the Leeds Chapel Allerton Hospital site, biological samples from patients consenting to the SWITCH trial biobank substudy were collected prior to commencement of the trial treatment and at weeks 2, 4, 12, 24 and 48 or at the time of early discontinuation, and stored for future research. See Appendix 5, Tables 33–35, for the schedule of events for rituximab, infliximab and subcutaneous bDMARDs.

Primary outcome measure

The primary outcome measure was the absolute reduction from baseline in DAS28 at 24 weeks post randomisation. DAS28 is a measure of disease activity in RA. 103,104 The composite score is calculated as a function of the number of tender and swollen joints (total 28 joints), the ESR and the patient’s global assessment of their arthritis measured using a VAS (see Appendix 6, Box 1).

Secondary outcome measures

The following outcomes were measured over 48 weeks (at each of the visit time points).

Analysis

Formal analyses were conducted using a two-sided 5% level of significance, with exception of the primary analysis, which used a one-sided 2.5% level of significance. All analyses were performed using SAS version 9.4 (SAS Institute Inc., Cary, NC, USA). The statistical analysis plan is provided in Appendix 7.

The trial planned for one interim analysis to allow for early stopping of either abatacept or alternative TNFi if inferiority compared with rituximab was demonstrated. This analysis would have taken place once 50% of patients (i.e. 239) had reached 24 weeks of follow-up. Following the early termination of the trial, there was no basis for an interim analysis.

Patient populations

All patients recruited into the trial were included in the analysis using ‘intention to treat’ (ITT) and analysed according to the randomised allocation.

A per protocol (PP) population analysis was also undertaken; patients who deviated from the protocol or failed to comply with the required treatment regimen were excluded (see Appendix 7 for further details of exclusions). For the analysis of the primary outcome measure, non-inferiority needed to be demonstrated in both the ITT and PP populations in order to infer non-inferiority.

The complete-case population included all patients with complete data for the relevant outcome measure summarised.

The safety population included all patients who received at least one dose of treatment and is summarised by treatment received.

Missing data

Multiple imputation by chained equations was used to impute missing values at the component level for the DAS28 and ACR20,119 under the assumption that the data were ‘missing at random’. The minimisation factors of disease duration, RF/ACPA status and non-response category and the values of the component end points at all visits from baseline to week 48 (for DAS28 components) and to week 24 (for ACR components) were included in the imputation models. Centre was not included because of the small number of patients recruited in each centre.

Imputations were performed separately for each of the three treatment groups; 27 imputed data sets were created for each treatment group for the primary end point DAS28 and 22 for the secondary end-point ACR20, corresponding to the maximum percentage of missing data across all end points and time points. Predictive mean matching was used to select a value to impute from the three observed values closest to the fitted value in each imputation. 120 The composite end points were then derived in each of the fully imputed data sets (see Appendix 7 for further detail).

Primary outcome measure

A mixed-effects linear regression model was fitted to the primary end point, absolute reduction in the DAS28 at 24 weeks, with covariates corresponding to the minimisation factors and treatment group; centre was fitted as a random effect.

Algebraic representation for the mixed-effects linear regression model is:

Parameter estimates from the mixed-effects models across each of the fully imputed data sets were combined using Rubin’s rules. 119,121 Estimates of each treatment effect and corresponding 95% CIs and p-values were reported in relation to the predefined non-inferiority margin of –0.6 units on the reduction in the DAS28 at 24 weeks post randomisation.

The primary analysis model was fitted to the ITT, PP and complete-case populations. Non-inferiority in both the ITT and PP patient populations was required in order to conclude non-inferiority. A sensitivity analysis was completed on the ITT population with an additional covariate, baseline DAS28, fitted to the primary analysis model.

Key secondary outcome measures

Additional secondary outcome measures

All additional secondary outcome measures, including further measures of disease activity and quality of life were summarised by treatment group and compared informally using descriptive statistics. The predefined subgroup analyses to evaluate the treatment modification effect of RF/ACPA status, initial TNFi group failed on and non-response category on the DAS28 were summarised by treatment group. In addition, treatment compliance, toxicity and safety were summarised.

Generalisability of the patient population randomised

Summaries of the clinical and demographic variables collected on the non-registration logs for patients considered for enrolment, registered but not randomised, and for patients randomised are provided in Appendix 9, Table 44. Age distribution and sex of registered, consented and non-randomised patients were similar. The proportion of patients who were RF seropositive or ACPA positive was also similar between patients registered but not randomised and those patients who were randomised; however, there was a large proportion of patients for whom RF and ACPA status was unknown among non-registered patients, making an assessment of generalisability on RF/ACPA status difficult.

Rituximab

All 40 patients randomised to rituximab received at least one infusion of rituximab. By week 12, all infusions had been given in line with the protocol, although infusions for four patients had been delayed (because of patient choice, clinician examination delaying first dose, AE and out-of-range pre-treatment tests). A total of 35 (87.5%) patients were known to be at least 80% compliant with treatment up to week 24.

Abatacept

All 41 patients randomised to receive abatacept received at least one injection of abatacept. A total of 29 patients (70.7%) were known to be at least 80% compliant with treatment up to week 24.

Alternative tumour necrosis factor inhibitor

Forty out of 41 patients (97.6%) randomised to the alternative TNFi treatment arm received their allocated treatment. A total of 31 patients (75.6%) were known to be at least 80% compliant with their randomised treatment up to week 24.

Minimisation factors

The median disease duration was 6.7 years (range 0.4–43.5 years) (see Table 2). Seventy-seven patients (63.1%) were secondary non-responders and 100 patients (82.0%) were RF seropositive or ACPA positive.

Demographics

Ninety patients (73.8%) had a current comorbidity, with the most frequently reported being hypertension (42 patients; 34.4%), osteoarthritis (33 patients; 27.0%) and hypercholesterolaemia (25 patients; 20.5%) (see Table 3).

Treatment history

Seventy patients (57.4%) had previously failed to respond to a monoclonal antibody TNFi agent. The most common first TNFi agent used was etanercept (52 patients; 42.6%), followed by adalimumab (28 patients; 23.0%) and then CZP (25 patients; 20.5%) (see Table 4).

A total of 24 patients (19.7%) had received some form of steroid or corticosteroid within 4 weeks of screening, with oral prednisolone the most frequently reported (22 patients;18.0%). Sixty patients (49.2%) reported receiving NSAIDs within 4 weeks of screening, with naproxen, diclofenac and ibuprofen most frequently reported. The most common non-bDMARDs used prior to participating were sulfasalazine (91 patients; 74.6%), hydroxychloroquine (84 patients; 68.9%) and leflunomide (26 patient;, 21.3%) (see Appendix 11).

Baseline disease activity and individual component measures

The mean number of TJCs and SJCs at baseline was 15.9 (SD 7.3) and 9.5 (SD 6.2), respectively. The median ESR was 26.0 mm/hour (quartiles 11.0 mm/hour and 43.0 mm/hour) and the median CRP level was 6.0 mg/l (quartiles 5.0 mg/l and 18.0 mg/l) (see Table 5).

The mean DAS28 at baseline was 6.1 units (SD 1.1 units), with 76.2% patients having high disease activity (see Table 5 and Appendix 12, Table 59). The mean CDAI score was 38.3 (SD 13.3), with 82.0% patients categorised as having high disease activity (see Table 5 and Appendix 12, Table 63). Furthermore, the mean SDAI score was 40.2 (SD 14.0) and 77.9% patients were categorised as having high disease activity (see Table 5 and Appendix 12, Table 65).

The median Physician Global Assessment Disease Activity score was 66.0 (quartiles 57.0 and 79.0).

Baseline patient-reported outcomes

The median global assessment of arthritis, general health and pain scores, as rated by the patient, were 71.0 (quartiles 56.0 and 83.0), 59.0 (quartiles 47.0 and 70.0) and 71.0 (quartiles 58.0 and 81.0), respectively (see Table 6). The median HAQ-DI score was 1.9 (quartiles 1.5 and 2.1), the median RAQoL score was 22.0 (quartiles 15.0 and 25.0) and the median HADS score was 15.0 (quartiles 10.0 and 21.0) (see Table 6).

Comparability of baseline characteristics between groups

The baseline characteristics were balanced across the three treatment groups, with the following notable exceptions. A higher percentage of females were randomised to abatacept (95.1%, compared with 80.5% and 75.0% randomised to alternative TNFi and rituximab, respectively). A greater proportion of patients in the alternative TNFi or abatacept arms were current or past smokers (70.7% and 68.5%, respectively) than in the rituximab arm (47.5%). A higher proportion of patients on abatacept had osteoarthritis, whereas more patients on rituximab had a history of hypercholesterolaemia. A greater proportion of patients on alternative TNFi and abatacept had a history of depression and also a history of thyroid dysfunction. Furthermore, a slightly greater proportion of patients on the alternative TNFi or rituximab had a history of diabetes (see Table 3). A slight imbalance in the RF/ACPA status was apparent, whereby a greater proportion of patients on alternative TNFi were seropositive than those taking abatacept and rituximab (see Table 2). The slight imbalances observed are consistent with the random allocation to the treatment group and result from the small sample size.

Tables 47–51 in Appendix 10 summarise the baseline characteristics for the PP population.

Primary outcome

Intention-to-treat patient population

Tables 7 and 8 and Figure 5 present the results of the primary end-point analysis for the ITT patient population.

TABLE 7 - Primary end-point analysis: multivariable linear regression analysis for a reduction in the DAS28 at 24 weeks post randomisation – ITT patient population

CCP, cyclic citrullinated peptide.

Effect	Estimate	95% CI	p-value
Intercept	1.50	0.72 to 2.28	< 0.001
Randomised treatment: alternative TNFi vs. rituximab	0.30	–0.45 to 1.05	0.436
Randomised treatment: abatacept vs. rituximab	0.04	–0.72 to 0.79	0.927
RF/anti-CCP seropositivity: both seronegative vs. either seropositive	–0.17	–1.00 to 0.66	0.690
Disease duration: ≥ 5 years vs. < 5 years	–0.05	–0.70 to 0.60	0.883
Non-responder type: secondary vs. primary	–0.45	–1.11 to 0.21	0.181

TABLE 8 - Adjusted mean reduction in the DAS28 and the corresponding difference between alternative TNFi, abatacept and rituximab at 24 weeks post randomisation – ITT patient population

One-sided p-value corresponding to the one-sided 97.5% test for non-inferiority of alternative TNFi/abatacept compared with rituximab using the lower margin of –0.6 units; the one-sided p-value needs to be < 0.025 to conclude non-inferiority with rituximab.

Treatment arm	Adjusted mean DAS28 reduction at week 24 (95% CI)	Treatment comparison	Difference in mean DAS28 reductions (experimental – rituximab 95% CI)	Probabilitya (δ > –0.6)
Rituximab (n = 40)	1.17 (0.56 to 1.77)	–	–	–
Alternative TNFi (n = 41)	1.47 (0.85 to 2.08)	Alternative TNFi vs. rituximab	0.30 (–0.45 to 1.05)	0.0094
Abatacept (n = 41)	1.20 (0.62 to 1.78)	Abatacept vs. rituximab	0.04 (–0.72 to 0.79)	0.0493

FIGURE 5.

Primary end-point analysis: difference in mean reduction in the DAS28 at 24 weeks post randomisation between alternative TNFi, abatacept and rituximab – ITT population analysis. The dotted line at zero is the null difference value and the thick dashed line denotes the non-inferiority margin.

For the comparison between alternative TNFi and rituximab, the lower limit of the 95% CI for the difference in the mean reduction in the DAS28 lies above the predefined, non-inferiority limit of –0.6 units; the difference in the mean reduction in DAS28 (alternative TNFi – rituximab) at 24 weeks post randomisation was 0.3 units (95% CI –0.45 to 1.05 units). Therefore, alternative TNFi was non-inferior to rituximab in the ITT patient population.

For the comparison between abatacept and rituximab, the lower limit of the 95% CI lay just below the predefined non-inferiority limit; the difference in mean reduction in the DAS28 at 24 weeks (abatacept – rituximab) was 0.04 units (95% CI –0.72 to 0.79 units). Therefore, abatacept was not shown to be non-inferior to rituximab in the ITT patient population.

Summaries of the missing values for the component end points of the DAS28 by treatment group are presented in Appendix 13, Tables 69 and 70.

Sensitivity analysis

Table 9 and Figure 6 present results of the sensitivity analysis on the primary end point for the ITT patient population, adjusting for the baseline DAS28.

TABLE 9 - Sensitivity analysis: adjusted mean reduction in DAS28 and corresponding difference between alternative TNFi, abatacept and rituximab at 24 weeks post randomisation – ITT patient population

One sided p-value corresponding to one-sided 97.5% test for non-inferiority of alternative TNFi/abatacept compared with rituximab using the lower margin of –0.6 units; the one sided p-value needs to be < 0.025 to conclude non-inferiority with rituximab.

Treatment arm	Adjusted mean DAS28 reduction at week 24 (95% CI)	Treatment comparison	Baseline-adjusted difference in mean DAS28 reductions (experimental – rituximab 95% CI)	Probabilitya (δ > 0.6)
Rituximab (n = 40)	1.07 (0.53 to 1.60)	–	–	–
Alternative TNFi (n = 41)	1.60 (1.05 to 2.15)	Alternative TNFi vs. rituximab	0.54 (–0.12 to 1.20)	< 0.001
Abatacept (n = 41)	1.11 (0.60 to 1.62)	Abatacept vs. rituximab	0.05 (–0.61 to 0.70)	0.026

FIGURE 6.

Sensitivity analysis on the primary end point: difference in the mean reduction in the DAS28 at 24 weeks post randomisation between alternative TNFi, abatacept and rituximab.

For the comparison between alternative TNFi and rituximab, the lower limit of the 95% CI for the difference in the mean reduction in the DAS28 lies above the predefined non-inferiority limit of –0.6 units; the difference in mean reduction in the DAS28 (alternative TNFi – rituximab) at 24 weeks post randomisation was 0.54 units (95% CI –0.12 to 1.20 units). Therefore, again alternative TNFi was non-inferior to rituximab in the ITT patient population.

For the comparison between abatacept and rituximab, the lower limit of the 95% CI lay just below the predefined non-inferiority limit; the difference in the mean reduction in the DAS28 at 24 weeks (abatacept – rituximab) was 0.05 units (95% CI –0.61 to 0.70 units). Therefore, there was only marginal evidence that abatacept was non-inferior to rituximab in the ITT patient population.

Per protocol population

A total of 41 patients were included in the PP population with 13 (31.7%), 14 (34.1%) and 14 (35.0%) in alternative TNFi, abatacept and rituximab treatment groups, respectively. Owing to the small number of patients in the PP population, only the primary outcome has been analysed.

Tables 10 and 11 and Figure 7 provide the results of the primary end-point analysis for the PP population.

TABLE 10 - Primary end-point analysis: multivariable linear regression analysis for a reduction in the DAS28 at 24 weeks post randomisation – PP population

CCP, cyclic citrullinated peptide.

Effect	Estimate	95% CI	p-value
Intercept	2.07	0.91 to 3.23	0.001
Randomised treatment: alternative TNFi vs. rituximab	–0.58	–1.72 to 0.55	0.312
Randomised treatment: abatacept vs. rituximab	–0.15	–1.27 to 0.98	0.796
RF/anti-CCP seropositivity: both seronegative vs. either seropositive	–0.73	–1.96 to 0.50	0.245
Disease duration: ≤ 5 years vs. > 5 years	–0.05	–1.10 to 1.01	0.930
Non-responder type: secondary vs. primary	–0.05	–0.99 to 0.88	0.908

TABLE 11 - Adjusted mean reduction in the DAS28 and corresponding difference between alternative TNFi, abatacept and rituximab at 24 weeks post randomisation – PP population

One-sided p-value corresponding to the one-sided 97.5% test for non-inferiority of alternative TNFi/abatacept compared with rituximab using the lower margin of –0.6 units; the one-sided p-value needs to be < 0.025 to conclude non-inferiority with rituximab.

Treatment group	Adjusted mean DAS28 reduction at week 24 (95% CI)	Treatment comparison	Difference in mean DAS28 reductions (experimental – rituximab 95% CI)	Probabilitya (δ > –0.6)
Rituximab (n = 14)	1.66 (0.77 to 2.55)	–	–	–
Alternative TNFi (n = 13)	1.07 (0.11 to 2.03)	Alternative TNFi vs. rituximab	–0.58 (–1.72 to 0.55)	0.489
Abatacept (n = 14)	1.51 (0.70 to 2.31)	Abatacept vs. rituximab	–0.15 (–1.27 to 0.98)	0.216

FIGURE 7.

Primary end-point analysis: difference in the mean reduction in the DAS28 at 24 weeks post randomisation between alternative TNFi, abatacept and rituximab – PP population.

For the comparison of alternative TNFi versus rituximab, the lower limit of the 95% CI for the difference in the mean reduction in the DAS28 was below the predefined non-inferiority limit of –0.6 units; the difference in the mean reduction in the DAS28 at 24 weeks post randomisation (alternative TNFi – rituximab) was –0.58 units (95% CI –1.72 to 0.55 units). Therefore, non-inferiority of alternative TNFi to rituximab was not demonstrated in the PP population and so we cannot conclude that alternative TNFi is non-inferior to rituximab.

Similarly, for the comparison of abatacept versus rituximab, the lower limit of the 95% CI for the difference in the mean reduction in the DAS28 was below –0.6 units; the difference in the mean reduction in the DAS28 at 24 weeks post randomisation (abatacept – rituximab) was –0.15 units (95% CI –1.27 to 0.98 units). Therefore, non-inferiority of abatacept to rituximab was not demonstrated in the PP population. Hence, a conclusion of non-inferiority of abatacept to rituximab was not reached in both analyses (although in an underpowered cohort).

Complete-case analysis population

The primary end-point analysis was also conducted on the population of all patients with complete data at both baseline and week 24. Table 12 and Figure 8 provide the results of the primary end-point analysis in the complete-case analysis population.

TABLE 12 - Adjusted mean reduction in the DAS28 and corresponding difference between alternative TNFi, abatacept and rituximab at 24 weeks post randomisation – complete-case population

One-sided p-value corresponding to the one-sided 97.5% test for non-inferiority of alternative TNFi/abatacept compared with rituximab using the lower margin of –0.6 units; the one-sided p-value needs to be < 0.025 to conclude non-inferiority with rituximab.

Treatment group	Adjusted mean DAS28 reduction at week 24 (95% CI)	Treatment comparison	Difference in mean DAS28 reductions (experimental – rituximab 95% CI)	Probabilitya (δ > –0.6)
Rituximab (n = 32)	1.14 (0.44 to 1.85)	–	–	–
Alternative TNFi (n = 36)	1.24 (0.57 to 1.91)	Alternative TNFi vs. rituximab	0.10 (–0.71 to 0.91)	0.044
Abatacept (n = 34)	1.10 (0.47 to 1.74)	Abatacept vs. rituximab	–0.04 (–0.86 to 0.79)	0.090

FIGURE 8.

Primary end-point analysis: difference in the mean reduction in the DAS28 at 24 weeks post randomisation between alternative TNFi, abatacept and rituximab – complete-case population.

For both alternative TNFi and abatacept, the lower limit of the 95% CI for the true difference in the mean reduction in the DAS28 lay just below the predefined non-inferiority limit of –0.6 units. The difference in the mean reduction in the DAS28 at 24 weeks post randomisation for alternative TNFi compared with rituximab was 0.10 units (95% CI –0.71 to 0.91 units) and for abatacept relative to rituximab was –0.04 units (95% CI –0.86 to 0.79 units). Therefore, non-inferiority of alternative TNFi and abatacept to rituximab was not demonstrated in the complete-case analysis population.

Summary statistics for the DAS28 at baseline, week 24 and the corresponding reduction for the complete-case population is summarised in Appendix 12, Table 58.

Exploratory subgroup analysis

Table 13 presents the least squares means and corresponding 95% CIs of the reduction in the DAS28 at week 24 by RF/ACPA seropositivity status, initial TNFi type and non-responder status to an initial bDMARD for the ITT population; corresponding summary statistics for the complete-case population are provided in Appendix 12, Table 53. The results of the subgroup analysis are not sufficiently precise to draw definitive conclusions and, therefore, only informal comparisons between treatment groups were made. All results should be interpreted cautiously.

Secondary outcomes

Disease Activity Score of 28 joints over time

Table 14 provides the parameter estimates for the model for the DAS28 up to week 48. Table 15 presents the adjusted DAS28 and corresponding difference from rituximab at each time point up to week 48. Figure 9 provides a graphical representation of the adjusted means and corresponding 95% CIs from the model. The covariance matrix is provided in Appendix 12, Table 55.

TABLE 14 - Estimated coefficients from a multivariable covariance pattern model for the DAS28 over 48 weeks post randomisation – ITT patient population

Effect	Parameter estimate	95% CI	p-value
Intercept (rituximab at week 12)	2.72	1.48 to 3.96	< 0.0001
Baseline DAS28	0.38	0.21 to 0.55	< 0.0001
Randomised treatment
Alternative TNFi vs. rituximab	–0.18	–0.72 to 0.35	0.503
Abatacept vs. rituximab	–0.04	–0.57 to 0.48	0.870
RF/ACPA status: both seronegative vs. either RF/ACPA seropositive	0.19	–0.32 to 0.69	0.475
Years since diagnosis: ≥ 5 years vs. 0–4 years	–0.16	–0.57 to 0.25	0.440
Non-response type: secondary vs. primary	0.13	–0.27 to 0.53	0.524
Visit
24 weeks	–0.09	–0.54 to 0.36	0.701
36 weeks	–0.14	–0.52 to 0.24	0.477
48 weeks	–0.29	–0.75 to 0.17	0.218
Interaction effect for
Alternative TNFi
At 24 weeks	–0.32	–0.96 to 0.33	0.338
At 36 weeks	–0.53	–1.11 to 0.05	0.071
At 48 weeks	–0.20	–0.88 to 0.47	0.552
Abatacept
At 24 weeks	–0.01	–0.65 to 0.63	0.976
At 36 weeks	0.05	–0.51 to 0.60	0.871
At 48 weeks	0.10	–0.55 to 0.75	0.755

TABLE 15 - Adjusted mean (95% CI) DAS28 and comparisons of alternative TNFi and abatacept with rituximab by assessment time point over 48 weeks

Treatment arm	Adjusted mean (95% CI)	Difference in adjusted means (95% CI)	p-value
12 weeks
Rituximab	5.09 (4.69 to 5.48)	–	–
Abatacept	5.03 (4.63 to 5.43)	–0.04 (–0.57 to 0.48)	0.870
Alternative TNFi	4.89 (4.47 to 5.32)	–0.18 (–0.72 to 0.35)	0.503
24 weeks
Rituximab	4.99 (4.50 to 5.47)	–	–
Abatacept	4.93 (4.46 to 5.41)	–0.05 (–0.71 to 0.60)	0.872
Alternative TNFi	4.49 (3.99 to 4.99)	–0.50 (–1.16 to 0.16)	0.137
36 weeks
Rituximab	4.94 (4.49 to 5.39)	–	–
Abatacept	4.94 (4.49 to 5.39)	0.00 (–0.60 to 0.61)	0.995
Alternative TNFi	4.22 (3.75 to 4.70)	–0.71 (–1.32 to 0.11)	0.022
48 weeks
Rituximab	4.79 (4.28 to 5.29)	–	–
Abatacept	4.84 (4.38 to 5.31)	0.06 (–0.59 to 0.71)	0.859
Alternative TNFi	4.40 (3.88 to 4.92)	–0.39 (–1.04 to 0.27)	0.249

FIGURE 9.

Adjusted mean DAS28 over the 48-week intervention period and the comparisons of alternative TNFi and abatacept with rituximab over 48 weeks. Adjusted mean DAS28 over the 48-week intervention period (top, left axis). Estimated differences between each intervention arm and rituximab at each time point (bottom, right axis).

There was no evidence of a treatment effect for abatacept compared with rituximab at any of the time points (see Table 15). This analysis showed significant evidence of a difference between alternative TNFi and rituximab at week 36 (–0.71 units, 95% CI –1.32 to –0.11 units; p = 0.022), although this difference was not maintained at week 48.

From 24 weeks post randomisation, the 95% CIs for the mean DAS28 in the alternative TNFi group exclude values greater than 4.99, suggesting that the mean DAS28 for this group is lower than the DAS28 threshold for high disease activity (i.e. DAS28 > 5.1 units). For abatacept and rituximab the estimated 95% CIs for the mean DAS28 include values corresponding to high disease activity at all time points.

Reduction in the Disease Activity Score of 28 joints of ≥ 1.2 units over time

The frequency of patients achieving a DAS28 response over 48 weeks for the complete-case population is provided in Appendix 12, Table 54.

Parameter estimates for the model are given in Table 16. Table 17 shows the odds of achieving DAS28 ≥ 1.2 units in each group and the odds ratios relative to rituximab at each time point up to week 48. Figure 10 provides a graphical representation of the fitted values from the model. The covariance matrix is provided in Appendix 12, Table 56.

TABLE 16 - Estimated coefficients for the multivariable covariance pattern model for the DAS28 response over 48 weeks post randomisation: ITT patient population

Effect	Odds ratio (95% CI)	p-value
Intercept		< 0.001
Baseline DAS28	2.23 (1.65 to 3.02)	< 0.001
Randomised treatment
Alternative TNFi vs. rituximab	1.07 (0.38 to 3.02)	0.904
Abatacept vs. rituximab	1.22 (0.44 to 3.39)	0.701
Seropositivity: RF/ACPA both seronegative vs. either RF/ACPA seropositive	0.74 (0.33 to 1.65)	0.464
Years since diagnosis: ≥ 5 years vs. 0–4 years	1.56 (0.82 to 2.97)	0.178
Non-response type: secondary vs. primary	0.96 (0.51 to 1.81)	0.888
Visit
24 weeks	0.98 (0.45 to 2.13)	0.963
36 weeks	1.17 (0.49 to 2.77)	0.726
48 weeks	1.77 (0.64 to 4.90)	0.269
Interaction effect for
Alternative TNFi
At 24 weeks	2.37 (0.81 to 6.94)	0.116
At 36 weeks	1.79 (0.56 to 5.72)	0.327
At 48 weeks	1.32 (0.33 to 5.33)	0.699
Abatacept
At 24 weeks	0.95 (0.33 to 2.76)	0.930
At 36 weeks	0.92 (0.28 to 2.98)	0.888
At 48 weeks	0.75 (0.19 to 2.93)	0.675

TABLE 17 - Adjusted odds of achieving a DAS28 response and the corresponding odds ratios over time

Treatment arm	Adjusted odds ratio of response (95% CI)	Adjusted odds ratio of response vs. rituximab (95% CI)	p-value
12 weeks
Rituximab	0.65 (0.29 to 1.42)	–	–
Abatacept	0.79 (0.38 to 1.65)	1.22 (0.44 to 3.39)	0.701
Alternative TNFi	0.69 (0.32 to 1.50)	1.07 (0.38 to 3.02)	0.905
24 weeks
Rituximab	0.64 (0.29 to 1.39)	–	–
Abatacept	0.74 (0.35 to 1.55)	1.16 (0.42 to 3.24)	0.771
Alternative TNFi	1.61 (0.75 to 3.46)	2.53 (0.90 to 7.07)	0.077
36 weeks
Rituximab	0.76 (0.33 to 1.74)	–	–
Abatacept	0.85 (0.40 to 1.81)	1.12 (0.37 to 3.42)	0.839
Alternative TNFi	1.44 (0.66 to 3.12)	1.91 (0.65 to 5.60)	0.240
48 weeks
Rituximab	1.15 (0.49 to 2.71)	–	–
Abatacept	1.05 (0.50 to 2.19)	0.91 (0.30 to 2.73)	0.869
Alternative TNFi	1.61 (0.72 to 3.62)	1.40 (0.47 to 4.19)	0.543

FIGURE 10.

Adjusted odds of achieving a DAS28 response and odds ratios of a response over 48 weeks. Covariance pattern model of a DAS28 response over time. Unstructured covariance pattern. Random centre effect excluded, baseline score adjusted. Estimated odds of achieving a DAS28 response at follow-up (top, left axis). Odds ratios of response of alternative TNFi and abatacept to rituximab at each time point (bottom, right axis).

There was no evidence of a treatment effect for either alternative TNFi or abatacept compared with rituximab after adjusting for the minimisation factors and baseline covariates, at any of the time points (see Table 17).

American College of Rheumatology 20 response at week 24

Table 18 provides the parameter estimates for the model and Table 19 provides the adjusted odds ratios for achieving ACR20 response at 24 weeks post randomisation. Figure 11 provides a graphical representation of the fitted values from the model.

TABLE 18 - Multivariable logistic regression model for the ACR20 response at 24 weeks post randomisation: ITT patient population

Parameter	Odds ratio	95% CI	p-value
Intercept			0.062
Randomised treatment
Alternative TNFi vs. rituximab	2.06	0.77 to 5.53	0.150
Abatacept vs. rituximab	1.19	0.44 to 3.21	0.736
RF/ACPA status: both seronegative vs. either seropositive	1.39	0.48 to 3.99	0.539
Disease duration: ≥ 5 years vs. < 5 years	1.73	0.73 to 4.12	0.216
Non-responder type: secondary vs. primary	0.55	0.24 to 1.26	0.161

TABLE 19 - Adjusted odds of an ACR20 response at 24 weeks post randomisation: ITT patient population

Treatment group	Adjusted odds ratio of response (95% CI)	Adjusted odds ratio of response vs. rituximab (95% CI)	p-value
Rituximab	0.43 (0.20 to 0.94)
Abatacept	0.51 (0.24 to 1.08)	1.19 (0.44 to 3.21)	0.736
Alternative TNFi	0.89 (0.41 to 1.96)	2.06 (0.77 to 5.53)	0.150

FIGURE 11.

Odds ratios of an ACR20 response relative to rituximab at 24 weeks post randomisation: ITT patient population.

There was no evidence of a difference in the odds of achieving an ACR20 response at 24 weeks post randomisation in either intervention compared with rituximab (OR 2.06, 95% CI 0.77 to 5.53 and OR 1.19, 95% CI 0.44 to 3.21 for alternative TNFi and abatacept, respectively).

Summaries of the missing values for the component end points of the ACR20 by treatment group are presented in Appendix 13, Tables 69 and 70.

Additional secondary outcomes

Clinical assessment of disease activity over 48 weeks

Appendix 12, Tables 54 and 57–65, summarise measures of disease activity over 48 weeks by randomised treatment group. Summary statistics for each of the secondary outcomes over the observational period (weeks 60–96) are provided in Appendix 12, Tables 58–65. A brief description of the findings is given below. There were too few patients to make firm inferences from any of these analyses.

American College of Rheumatology 20, American College of Rheumatology 50 and American College of Rheumatology 70 response

The proportion of patients on alternative TNFi who achieved an ACR20 response appeared to increase over time, reaching 54.8% by week 48, although, among patients on abatacept, the proportion decreased to 35.5% by week 48 and, among those on rituximab, it had increased by week 48 to 42.9% after an initial reduction at week 24 (27.0%), potentially reflecting the need for a repeat cycle of rituximab in the second 6-month period. The proportion of patients achieving an ACR50 response generally increased over time on alternative TNFi and rituximab and plateaued after week 24 for abatacept. The proportion of patients achieving an ACR70 fluctuated over time across all three treatment groups; overall, only 11.8% achieved an ACR70 by week 48 (see Appendix 12, Table 57).

Disease Activity Score of 28 joints and Disease Activity Score of 28 joints response category

The mean reduction in the DAS28 over 48 weeks was greater among patients on alternative TNFi than among those on rituximab, whereas a similar mean reduction in the DAS28 was apparent in the abatacept and rituximab groups [alternative TNFi mean at 48 weeks, 1.6 units (SD 1.64 units); abatacept mean at 48 weeks, 1.4 units (SD 1.38 units); and rituximab mean at 48 weeks, 1.2 units (SD 1.49 units)] (see Appendix 12, Table 58).

The proportion of patients who achieved DAS28 low disease activity or remission at 24 weeks was similar in the alternative TNFi and rituximab groups, but lower in the abatacept group (19.5%, 20.0% and 14.6%, respectively). The proportion of patients achieving remission at week 24 showed a similar pattern (alternative TNFi, 9.8%; rituximab, 10.0%; abatacept, 7.3%). Furthermore, in the alternative TNFi group, this figure continued to increase, reaching 26.8% by week 48; in contrast, in the abatacept and rituximab groups, it fell, to 7.3% and 7.5%, respectively (see Appendix 12, Table 59). The proportion of patients reaching remission at week 48 showed a corresponding increase in the alternative TNFi group (12.2%), but in the abatacept and rituximab groups, this fell to 4.9% and 5.0%, respectively.

European League Against Rheumatism response

The proportion of patients with a good EULAR response at 24 weeks was similar in the alternative TNFi and rituximab groups, at 19.5% and 17.5%, respectively, but was lower in the abatacept group, at 12.2%. Furthermore, in the alternative TNFi group, the proportion of patients achieving a good response generally increased over time, reaching 26.8% at week 48, whereas in the abatacept and rituximab groups, a reduction to 4.9% and 5.0%, respectively, was observed (see Appendix 12, Table 60).

American College of Rheumatology/Boolean remission

A minority of patients achieved ACR/Boolean remission status over the 48 weeks. At 24 weeks, two patients (4.9%) on alternative TNFi and one patient (2.4%) on abatacept reached ACR/Boolean remission, compared with none on rituximab. By week 48 the frequency remained very low, with only one patient (2.4%) on alternative TNFi and a further one patient (2.4%) on abatacept reaching this status (see Appendix 12, Table 61).

Clinical Disease Activity Index

Similar improvements over 48 weeks were observed in CDAI in the alternative TNFi and rituximab groups,, with a lower improvement observed in the abatacept group [median improvement of 19.3 (quartiles 5.7 and 28.8), 20.3 (quartiles 5.3 and 32.3) and 14.1 (quartiles 5.9 and 29.2) respectively] (see Appendix 12, Table 62).

Simplified Disease Activity Index

Like the CDAI, improvement in the SDAI over 48 weeks was similar in the alternative TNFi and rituximab groups, but lower in the abatacept group [median improvement of 20.1 (quartiles 7.9 and 27.2), 20.1 (quartiles 5.3 and 34.0) and 13.7 (quartiles 6.3 and 31.2) respectively] (see Appendix 12, Table 64).

Patient-reported outcomes over 48 weeks

Figures 12–15 summarise the patient-reported outcomes of the HAQ-DI, RAQoL and HADS and Figures 16–18 summarise the Patient Global Assessment of Pain, Patient Global Assessment of Arthritis and Patient Assessment of General Health VAS scores over 48 weeks by treatment group. Further summaries over the 96 weeks are provided in Appendix 12, Tables 66 and 67.

FIGURE 12.

Box and whisker plot: HAQ-DI score over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

FIGURE 13.

Box and whisker plot: RAQoL over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

FIGURE 14.

Box and whisker plot: HADS anxiety scale over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

FIGURE 15.

Box and whisker plot: HADS depression scale over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

FIGURE 16.

Box and whisker plot: Patient Global Assessment of Pain over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

FIGURE 17.

Box and whisker plot: Patient Global Assessment of Arthritis over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

FIGURE 18.

Box and whisker plot: Patient Assessment of General Health over 48 weeks by treatment group. X inside the bars corresponds to the mean value; X outside the bars corresponds to outlier values. B/L, baseline.

Health Assessment Questionnaire Disability Index, Rheumatoid Arthritis Quality of Life and Hospital Anxiety and Depression Scale

Overall, there was an improvement in the HAQ-DI over the 48-week period across all three treatment groups. The median HAQ-DI score decreased to 1.5 (quartiles 1.1 and 1.9) in the alternative TNFi group, to 1.6 (quartiles 1.0 and 2.1) in abatacept group and to 1.7 (quartiles 1.1 and 2.1) the rituximab group (see Figure 12). Similarly, a general improvement in the RAQoL was observed over time, with no notable differences between treatment groups [median score of 19.0 (quartiles 9.0 and 23.0) was reported at 48 weeks in the alternative TNFi group, 17.5 (quartiles 11.4 and 24.0) in the abatacept group and 19.5 (quartiles 12.0 and 25.0) in the rituximab group] (see Figure 13). Small improvements in the HADS scores over the 48-week period were observed in the alternative TNFi and abatacept groups [median scores for anxiety and depression of 6 (quartiles 3.0 and 11.0) and 4.5 (quartiles 2.0, 9.0), respectively, in the alternative TNFi group and of 7.0 (quartiles 4.0 and 10.0) and 5.0 (quartiles 3.0 and 7.0), respectively, in the abatacept group], whereas no notable improvement was apparent in the rituximab group [median scores of 8.0 (quartiles 5.0 and 12.0) and 5.5 (quartiles 4.0 and 9.0), respectively] (see Figures 14 and 15).

Patients’ global assessment of pain, arthritis and general health

A marked improvement in the patients’ global assessments of pain was apparent in all three treatment groups by 12 weeks post randomisation; thereafter small fluctuations in median pain scores over time were observed (see Figure 16). The median pain VAS scores at 48 weeks were 49.0 (quartiles 22.0 and 66.0), 51.5 (quartiles 19.0 and 72.0) and 57.0 (quartiles 34.0 and 67.0) in the alternative TNFi, abatacept and rituximab groups, respectively; the observed higher median score for rituximab may in part be explained by the higher median score observed at baseline.

Similarly, an initial marked improvement in the patients’ global assessment of their arthritis by 12 weeks post randomisation was apparent across all three treatment groups (see Figure 17). In the alternative TNFi arm, this improvement continued to 36 weeks, followed by a slight deterioration by 48 weeks [median 47.0 (quartiles 22.0 and 69.0)]. In comparison, following the initial 12-week improvement, in the abatacept and rituximab groups there was a slight deterioration in patients’ global assessment of their arthritis by 48 weeks [median 55.5 (quartiles 25.0 and 68.0) and 55.0 (quartiles 35.0 and 70.0), respectively].

Overall, an improvement in patients’ general health was observed over time, with no notable difference between treatment groups; the median score at 48 weeks was 48.0 (quartiles 26.0 and 63.0), 49.0 (quartiles 27.0 and 67.0) and 52.0 (quartiles 31.0 and, 64.0) in the alternative TNFi, abatacept and rituximab groups, respectively (see Figure 18).

Serious adverse events

A full listing of the SAEs and SSARs is included in Appendix 14, Tables 71–74. Ten SAEs were reported in nine patients. Of these, three events in three patients were considered to be related to trial medications and classed as SSARs. There were no suspected unexpected serious adverse reactions (SUSARs).

One patient in the TNFi arm, receiving infliximab, had a SSAR of autoimmune hepatitis but recovered with sequelae. Two patients who received abatacept experienced SSARs (one experienced angiooedema and recovered and one contracted pneumonia and died) and a further two patients who received abatacept experienced SAEs (one experienced chest pain/epigastric pain, which remained unresolved, and one contracted a chest infection but recovered). Finally, three patients who received rituximab each experienced a SAE: one patient developed malignant melanoma and subbsequently died; one suffered a broken coccyx bone, attributable to collapsing, but recovered with sequelae, and a third experienced abdominal pain, but later recovered. A further patient who was randomised to receive rituximab experienced two SAEs occurring prior to first infusion: a flare of RA and left basal pneumonia, from both of which the patient recovered.

As summarised above, two patients died following the development of a SAE. One patient receiving rituximab had developed malignant melanoma, while a second patient developed pneumonia, considered by the treating physicians to be related to abatacept treatment, as well as another illness.

No pregnancies were reported in any of the trial patients or in any partners of the trial patients.

Adverse events

Overall, 243 non-SAEs were reported in 90 patients (Table 20). Twelve events resulted in a permanent cessation of treatment. Table 20 summarises the number of non-SAEs reported by treatment received. A listing of all AEs is provided in Appendix 14, Table 75.

A total of 10 patients experienced one or more AE or SAE that resulted in a permanent cessation of treatment: four patients on alternative TNFi (9.8%), two on abatacept (4.9%) and four on rituximab (10.0%).

Aim and end points

The primary aim of this analysis was to assess the cost-effectiveness of the use of abatacept or alternative TNFi compared with the current practice of rituximab in patients with RA who have failed treatment with an initial TNFi. The primary end point was the cost per quality-adjusted life-year (QALY) gained. The methods used for this within-trial analysis were guided by the recommendations from the NICE methods guide. 123

Perspective and time frame

The study adopted a NHS and Public Social Services perspective for cost evaluation, but a broader societal perspective was adopted for secondary analysis to incorporate costs to patients and productivity costs. Costs and benefits for the base-case analysis were calculated for the study period of 48 weeks. As the time frame of the trial was < 1 year, discounting of the costs and benefits was not required.

Measurement of effectiveness

This analysis used the QALY as the main outcome measure. QALYs are a generic measure of health state that take account of both quality and length of life such that 1 QALY is equal to 1 year in full health. 124

Health-related quality of life was measured using the EuroQol 5 Dimensions, 3 levels (EQ-5D-3L). The EuroQol 5 Dimensions (EQ-5D) is a commonly used generic measure of health-related quality of life and is NICE’s preferred outcome measure for cost-effectiveness analyses. 123 The questionnaire comprises five domains: mobility, self-care, usual activities, pain/discomfort and anxiety/depression. Each domain consists of three levels: no problems, some problems and severe problems. 125

The EQ-5D-3L was administered at baseline and at follow-up visits in weeks 12, 24, 36 and 48. Responses were converted to health state utility values using the UK general population time trade-off tariff values. 126

Measurement of costs

Health-care resource utilisation data were collected using patient self-reported questionnaires covering primary care [e.g. general practitioner (GP) and nurse visits] and secondary care (e.g. hospital stays/visits) resource use over the trial period (see Appendix 2). The questionnaires also captured personal costs to patients related to RA (e.g. travel to/from hospital and cost of aids) and any impact the disease had on their income over the trial period. The resource-use questionnaires were completed at weeks 12, 24, 36 and 48 by the research teams at the participating centres and were supplemented by case report forms (CRFs) capturing data on hospital inpatient or outpatient visits. When there were discrepancies between the CRFs and the patient-completed forms, the CRFs were given precedence. Unit costs for health service staff and resources were obtained from the Personal Social Services Research Unit (PSSRU) report in 2015, entitled Unit Costs of Health and Social Care 2015,127 and NHS Reference Costs 2014 to 2015. 128 For medications, a unit cost per treatment received was assigned. The Commercial Medicines Unit’s electronic market Information Tool (eMit) was used to cost the drugs when possible. 129 However, when drugs were not listed on eMit, costs were taken from the British National Formulary (BNF). 130 When unit costs were not available, targeted literature searches were used to provide the relevant costs which were inflated to 2015 prices (pounds sterling) using an online inflator. 131 Unit costs are presented in Appendix 15, Table 76.

Missing data

The base-case analysis was conducted using only complete cases. That is, patients were included in the analysis if they had no missing resource use data as well as no missing quality-of-life data. In the case of resource use, no missing data were defined as a resource use form having been completed at all time points. For the resource use questionnaires, if a patient recorded that they had used a form of health care (e.g. GP visit) but did not record the number of visits, the mean number of visits was imputed. For quality of life, complete data were defined as a completed quality-of-life questionnaire returned at each time point.

Sensitivity analyses were conducted using imputed data so that all patients were included in the analysis. Two imputation methods were explored: mean imputation and multiple imputation. For the mean imputation, when a QALY value for a given time point was missing, the mean of the non-missing QALYs for the trial arm at that time point was imputed. The same approach was taken to impute missing cost data for resource use. For the multiple imputation, costs for each follow-up and total QALYs were imputed by chained equations using predictive mean matching. 133 Forty-five data sets were imputed (reflecting the percentage of incomplete cases), which were then combined using Rubin’s rules. 119,134

Analysis

The primary analysis was a cost-effectiveness analysis of the three relevant treatment arms of the trial. A complete-case analysis was the primary method for analysing the trial data and an ITT analysis was undertaken as a sensitivity analysis.

Resource use and costs were quantified and analysed using analysis of variance and independent sample t-tests. Owing to the small sample size and, subsequently, the potential violation of the underlying normality assumption when using t-tests, the robustness of the results was checked using a non-parametric bootstrap.

Health-state utilities were used to calculate QALYs using an area under the curve approach:

where EQ-5D_Baseline, EQ-5D₁₂, EQ-5D₂₄, EQ-5D₃₆ and EQ-5D₄₈ are the EQ-5D scores at baseline, week 12, week 24, week 36 and week 48, respectively; 0.230769 represents 12 weeks out of 52 for each time period:

Total costs and QALYs for each arm of the trial were calculated. For the secondary analysis, a wider cost perspective was adopted to include the total costs incurred by the patients.

Incremental cost-effectiveness ratios (ICERs) were calculated. 135 An ICER represents the additional cost per QALY gained for each intervention compared with the next best alternative and is calculated as follows for treatment A relative to treatment B:

where Cost_A and Cost_B are the mean costs and QALY_A and QALY_B are the mean QALYs for groups A and B. An intervention was judged to be cost-effective using the lower limit of the NICE acceptance threshold of £20,000 per incremental QALY (λ = £20,000) as the decision rule for the analysis. 123

The level of sampling uncertainty around the ICER was determined using non-parametric bootstrapping (with replacement) to generate 10,000 estimates of incremental costs and benefits. These were then plotted on the cost-effectiveness plane to visualise the uncertainty around the mean incremental costs and effects. The expected ICERs for the primary analysis were estimated from the means of bootstrapped cost and outcome distributions.

Net monetary benefit (NMB) values were also calculated. Net benefit (NB) combines cost-effectiveness and willingness to pay to give an explicit monetary valuation of the health outcome. It is calculated by rearranging the ICER calculation and incorporating a proposed willingness-to-pay threshold value per QALY. 135 The expected value of the NMB was calculated for each treatment. Treatments with positive NMBs provide more health benefit than is displaced by the associated opportunity costs and should be adopted. The treatment with the highest positive NMB is the most cost-effective. 135 The probability that the treatments were cost-effective was evaluated by generating estimates of NMB for a range of cost-effectiveness thresholds (λ). This analysis was presented as a cost-effectiveness acceptability curve. 136,137 The cost-effectiveness acceptability curve provides decision-makers with useful information regarding the risk of making a wrong decision; however, the decision to fund or not fund a treatment should be made on the expected value of the NMB.

Net monetary benefit is derived for each patient as:

Sample

Of the 122 patients recruited to the trial, 70 patients with complete resource use data and EQ-5D results (25 rituximab, 24 abatacept and 21 alternative TNFi) were included in the base-case analysis.

Baseline characteristics of the 70 patients analysed in the complete case are presented in Table 21 (see Table 3 for baseline characteristics of the ITT population). In all treatment arms more than two-thirds of the patients were female. The average weight was slightly lower in the abatacept group than in the other treatment groups. Fewer patients in the alternative TNFi arm were non-smokers and a higher percentage were past smokers than in the other treatment arms. There was some variation between arms in baseline EQ-5D scores, with the alternative TNFi group having the highest scores. However, the difference between the scores was not statistically significant.

Quality-of-life data

The mean (SD) EQ-5D scores for each trial arm, at each time point, are presented in Table 24. There was an apparent increase in mean EQ-5D score from baseline to week 12 in all treatment groups. Although there were small fluctuations in scores, this initial improvement was more or less maintained up to week 48 in all arms of the trial. All three treatment groups show an increase in EQ-5D score from baseline to week 48. Despite the apparent slightly higher baseline EQ-5D score for the alternative TNFi group, statistical tests indicated that there was no statistically significant difference in EQ-5D scores at baseline (p = 0.2592). This was explored further in a sensitivity analysis adjusting for baseline differences in EQ-5D score.

Table 25 shows the mean EQ-5D change scores between baseline and each of the follow-up time points. Statistical analysis using analysis of variance indicated that the variation among groups in the changes in EQ-5D scores was not statistically significant (p = 0.7071).

Missing data

Seventy-four patients had complete EQ-5D scores across all time points. The remaining 48 had EQ-5D scores missing for at least one of the time periods. All 122 patients completed resource use questionnaires in the first two time periods (12 and 24 weeks), but the response rate dropped in the following two time periods to 98 forms returned (79%) at 36 weeks and 89 forms returned (71%) at 48 weeks.

Cost-effectiveness results

The costs and outcomes from the observed data are presented in Table 26. As the sample size of the observed data is small, the average costs and effects from the non-parametric bootstrap provide a more accurate and robust estimate of the distribution of the population costs and effects. Table 27 shows the costs and QALYs gained for each treatment arm, the incremental costs and QALYs of the relevant comparisons and the resulting ICERs calculated from the bootstrap sample. The treatments are arranged in order of increasing cost so that the incremental costs and QALYs refer to the comparison between the neighbouring treatments in the table. The abatacept treatment group had the highest QALYs gained over the trial period, although the rituximab group had the lowest. The mean total cost was highest for the abatacept treatment group and lowest for the rituximab treatment group.

The results suggest that switching to alternative TNFi would be cost-effective compared with rituximab, as QALY gains are higher and costs are only slightly higher, leading to an ICER value of £5332.02 per QALY gained. This is well below the NICE acceptance threshold (λ = £20,000), which indicates that switching to alternative TNFi would be a cost-effective treatment option. Conversely, the abatacept group has much higher costs and only marginal gains in QALYs compared with the alternative TNFi treatment group. This results in an ICER value of £253,967.96 per QALY gained, indicating that switching to abatacept compared with switching to alternative TNFi drug would not be cost-effective, as this ICER value is well above the NICE cost/QALY threshold. However, these results should be interpreted very cautiously given the small number of cases and the substantial probability of a very small QALY difference, resulting in a divisor close to zero. Moreover, the SD is likely to be underestimated.

Figure 20 shows the joint distribution of incremental costs and incremental effects in the cost-effectiveness plane for alternative TNFi compared with rituximab and alternative TNFi compared with abatacept. The wide spread of the ‘clouds’ shows the high degree of uncertainty around the central results. This is to be expected given the small sample sizes.

FIGURE 20.

Scatterplot on the cost-effectiveness plane: alternative TNFi vs. rituximab and abatacept vs. TNFi.

Net benefit

The NMB for each treatment group is presented in Table 28. Given the decision rule, the NMB results indicate that rituximab and abatacept are not cost-effective treatments, although alternative TNFi has a positive NMB, indicating that it is cost-effective. These results suggest that switching to rituximab following an initial TNFi failure, as recommended in the current NICE guidance, is not the most efficient use of NHS resources. The probability that the treatments are cost-effective is presented in the cost-effectiveness acceptability curve shown in Figure 21. This shows that rituximab has the highest probability of being cost-effective for very low threshold values but that at threshold values > £6000 alternative TNFi has the highest probability of being cost-effective. At a £20,000 threshold, there is a 74.5% probability that alternative TNFi is cost-effective. Abatacept is never the most likely to be cost-effective and has a low probability of cost-effectiveness even at high threshold values.

FIGURE 21.

Cost-effectiveness acceptability curve of an TNFi, rituximab and abatacept.

Sensitivity analysis

The costs and benefits from the observed data for each scenario explored in the sensitivity analyses are presented in Table 29. For each scenario the observed data were used to conduct a non-parametric bootstrap to provide the cost-effectiveness results, which are presented in Table 30.

Given that only complete cases were used for the base-case analysis, sensitivity analyses using mean imputation and multiple imputation were conducted, so that all randomised patients could be included. In each case, this does not alter the conclusion drawn from the base-case analysis, as alternative TNFi dominates all other treatment arms and is the optimal treatment option. We note that in the case of the mean imputation the increase in precision of these estimates is spurious, as the uncertainty attributable to imputation of a single value is ignored. Moreover, the imputed value did not take the characteristics of the missing cases into account; however, these are accounted for in the multiple imputation. Changing the administration of MTX to subcutaneous injection and adjusting for baseline differences also supports the results of the base-case analysis, with alternative TNFi remaining the most cost-effective treatment. Variances of population distributions are known to be underestimated in small samples, which is exacerbated by the high level of skewness in cost data, and we note that these sensitivity analyses have a large effect on the SD of the estimates (i.e. standard error). Consequently, we should be cautious when interpreting these results.

Secondary analyses

Secondary analyses incorporating costs using a wider social perspective were also conducted. As such, the costs to the patient as well as health-care provider costs were included. Initially, this perspective was explored using complete cases: those patients with complete health-care cost and quality-of-life data (rituximab, n = 25; abatacept, n = 24; and alternative TNFi, n = 21). In the complete-case analysis, those observations without complete cost and quality-of-life data were excluded. In addition, the wider social cost perspective was explored using imputed health-care cost and quality-of-life data in a similar way to the primary analysis. This enabled the inclusion of the whole sample in the analysis. The total costs and QALYs for each analysis, obtained from the observed data, are presented in Table 31. The observed data were used to conduct a non-parametric bootstrap to provide the cost-effectiveness results which are presented in Table 32. Including societal costs does not alter the conclusions drawn from the primary analyses, with alternative TNFi continuing to be the most cost-effective treatment.

Value of information analysis

There are 690,000 people with RA in the UK. 3,141 The population EVPI, at the NICE cost-effectiveness threshold value of lambda (λ = £20,000), is £129,227,589. The population EVPI for other values of lambda is plotted in Figure 22. This indicates that it would be highly valuable to the NHS to reduce the current uncertainty regarding the effectiveness of alternative TNFi compared with rituximab in the management of RA.

FIGURE 22.

Expected value of perfect information: population.

Primary outcome

Owing to the early termination of the trial by the funders, the ‘SWITCH’ study was substantially underpowered in its objective to demonstrate non-inferiority of either alternative TNFi or abatacept to rituximab in terms of a reduction in the DAS28 at 24 weeks post randomisation. In the context of the low number of patients, alternative TNFi was non-inferior to rituximab in the ITT patient population, in that the estimate of the treatment effect excluded the non-inferiority margin of –0.6 units, but non-inferiority was not demonstrated in the PP population, which was our prespecified requirement for demonstrating non-inferiority overall. Non-inferiority of abatacept to rituximab was not demonstrated in either patient population; the 95% CI did not exclude the non-inferiority margin of –0.6 units and is therefore a plausible value.

As the trial was not permitted to recruit the target number of patients, the main interpretation of the results is based on the estimated treatment effect and corresponding precision. In the ITT patient population, the estimated mean difference in the reduction in the DAS28 after 24 weeks between alternative TNFi and rituximab was 0.30 units (95% CI –0.45 to 1.05 units); the corresponding estimate for the difference between abatacept and rituximab was 0.04 units (95% CI –0.72 to 0.79 units). Hence, the treatment effect in the ITT patient population is estimated with a precision of ± 0.75 units (corresponding to the half-width of the 95% CI).

Following exclusion of patients according to the prespecified criteria, the number of patients included in the PP population resulted in the treatment effect for both interventions compared with rituximab being estimated with very low precision. The estimated treatment effect for alternative TNFi compared with rituximab was –0.58 units (95% CI –1.72 to 0.55 units) and for abatacept was –0.15 units (95% CI –1.27 to 0.98 units) for the DAS28 at 24 weeks. Therefore, the treatment effect was estimated with a precision of ± 1.13 units, so there is large uncertainty in the estimate of the treatment effect in the PP population.

Exploratory subgroup analysis

Subgroup effect estimates in this underpowered study should be interpreted cautiously; however, this component of the study was particularly novel and important clinically.

The suggestion of an association between negative serological status and poorer response to rituximab was consistent with meta-analyses,3–84 a recent RCT of the first-line use of a bDMARD (as opposed to following first TNFi failure as with the ‘SWITCH’ study) in which non-inferiority of rituximab compared with a TNFi in seropositive patients142 was observed and observational registry data. 82 The dependency of individual CCG receptiveness to secure such agreements, however, increases the potential for regional inconsistency in prescribing options and approach. Complete SWITCH data could have pushed for inclusion in future technology appraisals.

Primary and secondary non-response to an initial TNFi is well recognised and an important marker of the mechanisms for treatment failure, with secondary non-response suggesting a pharmacokinetic basis for failure and primary non-response suggesting the wrong target. 143 The results suggest that primary non-response may benefit from a change in class of bDMARD, whereas secondary non-response can be circumvented by use of alternative TNFi; these are consistent with other published literature, albeit from uncontrolled cohort studies.

Secondary outcomes

Rheumatoid arthritis is a chronic disease that in the main requires long-term DMARD therapy. RCTs usually comprise a primary end point at week 24 (or earlier) that provides guidance only on short-term outcome whereas, in practice, reassurance on the maintenance (durability) of response is equally relevant. The secondary outcomes at weeks 36 and 48 go some way to providing this additional context. In the ‘SWITCH’ study, there was evidence of a greater improvement in disease activity (via the DAS28) at week 36 for alternative TNFi than for rituximab, although this difference was not maintained at week 48. There was no evidence of a difference in the DAS28 between abatacept and rituximab at any time point. However, when assessing disease activity in terms of the odds of achieving a DAS28 response (≥ 1.2 units), there was no evidence of a difference for either intervention compared with rituximab at any of the time points. Moreover, there was no evidence of a difference in the odds of achieving an ACR20 response at 24 weeks post randomisation for either intervention relative to rituximab.

In addition to demonstrating a reduction in the DAS28, the overall disease activity status and setting of a target of at least low disease activity (if not remission) is now established as an important goal, which is coined as part of a ‘treat to target’ management approach of RA. 144 In the ‘SWITCH’ study, the proportions of patients on alternative TNFi and rituximab who achieved DAS28 low disease activity or remission at 24 weeks were similar, whereas the proportion of patients who achieved low disease activity or remission on abatacept was lower. Furthermore, among patients on alternative TNFi, the proportion achieving low disease activity or remission continued to increase to week 48, whereas among those on abatacept and rituximab this proportion fell between 24 and 48 weeks.

Functional and quality-of-life patient-reported outcome measures remain important indicators of patient well-being. Overall, a general improvement in HAQ-DI, RAQoL and the Patient Global Assessment of General Health was apparent over time, with no notable differences between treatment groups. There was a marked initial improvement in the Patient Global Assessment of Pain and Patient Global Assessment of Arthritis at 12 weeks across all three treatment groups. Small improvements in the HADS anxiety and depression scores over the 48-week period were observed in patients treated with alternative TNFi or abatacept, whereas no notable improvement was apparent in those receiving rituximab.

The safety profile was similar for all three treatments. Ten SAEs were reported in nine patients, of which three events in three patients were considered to be related to trial medications. There were no SUSARs reported. Two patients died, in both cases following the development of a SAE (one each in the rituximab and abatacept groups). Ten patients experienced toxicity resulting in a permanent cessation of treatment (four patients on alternative TNFi, two on abatacept and four on rituximab).

The most common protocol deviations related to receiving steroid treatment within 6 weeks of an end-point assessment, not being compliant with treatment to 24 weeks and receiving additional contraindicated treatment, all of which were likely to have contributed to a less conservative estimate of the mean treatment effect relative to the ITT patient population.

Strengths and weaknesses

The principal strength of the ‘SWITCH’ study design was its emphasis on evaluation of a more defined and refined patient population, in keeping with the overall ambitions of the medical community for a more precise, tailored approach to medicine. This in itself, however, posed its own challenges, such as in recruitment. To date, almost all TNFi failure studies (RCTs and observationa studiesl) have included any cause of failure (inefficacy and toxicity/intolerance), limiting the strength in application of the data on an individual patient level and also the potential mechanistic insights that can be drawn from clinical studies (‘reverse translation’). The SWITCH study permitted the enrolment of only patients in whom TNFi had been found to be ineffective; although this is the predominant reason for failure, it will have limited the eligible patient pool. In addition, evaluating only patients on concomitant MTX (to recognise MTX synergy with bDMARD) will have limited the available recruitment pool further. In hindsight, accepting a less precise approach by including all patients (any cause of TNFi failure and TNFi with/without MTX combination) with sensitivity analysis adjusting for MTX combination would have been more pragmatic and reduced some of the challenges in recruitment.

Nevertheless, the scientific design and rigorous conduct of this, albeit small, trial means that the SWITCH study will contribute to the evidence base for future research in RA, including in meta-analyses, and, hopefully, encourage future study design to address the factors such as those included in our exploratory subgroup analysis.

The obvious major weakness was the early termination, which resulted in a small number of patients recruited into the study. This naturally led to uncertainty in providing definitive conclusions. Despite this, the study identifies some indicators of response, which, although not definitive, provide support for, in particular, alternative TNFi rather than rituximab as a second-line bDMARD therapy in the management of RA. Furthermore, although an exploratory outcome, further diminished with the small sample size, the absence of a response to rituximab in the seronegative population is consistent with the published meta-analyses in efficacy trials. 84 These results, thus, further support that the hypothesis that an alternative bDMARD to rituximab may be preferable if a patient is seronegative. Moreover, if secondary non-response has resulted from a first TNFi, then it may also be more appropriate to consider a second TNFi.

It is important to emphasise, however, that no single treatment (or sequence) is appropriate for all patients and no single study will be able to address this completely. Although these and other complementary data may not provide definitive evidence on tailoring therapy, they do provide initial evidence to support a clinical judgement and increase the chance of treatment success on which subsequent studies can build. This approach to stratification of treatment would represent an advance to the current status of generally prescribing rituximab to all patients, thereby dismissing the potential benefits of switching to alternative therapies in patients who fail TNFi.

A further limitation of the design was the open-label nature of the study. Although blinding patients and treating clinicians to the allocated intervention would have reduced the risk and impact of any assessment bias, it would have been impractical to implement: each of the seven distinct bDMARDs in the SWITCH study involved a different route of delivery and dosing regimen, thus requiring multiple dummy infusions (necessitating additional inpatient attendance) and injections to maintain the blind. Such a treatment schedule was given careful consideration by our PPI advisor, and it was concluded that it would have imposed considerable burden on patients and was unethical, potentially either reducing recruitment further or increasing the rate of attrition throughout the study. However, the fact that all patients received an active therapy may have attenuated any bias introduced by the lack of blinding.

When the SWITCH study was designed, it included all three classes of bDMARD available for evaluation as second-line therapies when taken in combination with MTX, thereby reflecting the full range of therapies available. After the trial design and initiation, tocilizumab was approved as a first-line bDMARD and, hence, represented an additional option for the TNFi-inadequate response RA patient cohort. In addition, NICE has also approved the use of tocilizumab as a monotherapy, broadening the available therapies when given as monotherapy (without concomitant MTX).

The trial design allowed a pragmatic approach with clinician choice of the monoclonal antibody if the patient previously received etanercept. Moreover, a fourth monoclonal antibody, golimumab, was introduced into the alternative TNFi arm during the recruitment phase, thereby allowing further clinician choice across all available TNFis and ensuring further generalisability of the trial results to clinical practice. Although this pragmatic design may have introduced some heterogeneity into the results, it reflected current practice in the NHS and ensured that the EULAR guidance, recommending discussion between patient and treating clinician in the choice of bDMARD, was adhered to; this was considered important in order to maximise recruitment.

Principal findings

The trial-based cost-effectiveness analysis indicated that the current NICE guidance to switch to rituximab after initial TNFi failure was not the most cost-effective use of NHS resources. Instead, switching to alternative TNFi drug following an initial failure was the most cost-effective treatment option. Switching to alternative TNFi was more costly, but provided more QALYs than rituximab over a 48-month period. Moreover, switching to abatacept was not cost-effective compared with switching to alternative TNFi, as the associated costs were much higher and the QALY gain was small.

All three trial arms showed an increase in mean EQ-5D score over the 12-month trial period, with an initial increase from baseline to 12 weeks that was more or less maintained to 48 weeks. The small (non-significant) differences in QALYs resulting from these treatments could relate to the benefits being measured only over the trial period of 48 weeks, meaning that longer-term benefits could not be considered. In the primary analysis costs from a health and social services perspective and a wider social perspective were highest for the abatacept group and lowest for the rituximab group, but only slightly lower than the costs for the alternative TNFi group. These costs-effectiveness results were driven by the (statistically significant) difference in costs between abatacept and the other treatment groups. These results may change in longer follow-up if the side-effect profiles of the treatments are different when the drugs are used for prolonged periods, or effectiveness persists for different durations.

Base-case cost-effectiveness results were not sensitive to assumptions in a small number of deterministic sensitivity analyses. In line with the NICE reference case that indicates that costs to patients may be included in cost-effectiveness evaluations, results from a secondary analysis including the costs incurred by patients remained consistent, indicating that switching to alternative TNFi is cost-effective but switching to abatacept is not. However, these results should be treated with caution because of the small sample size, for which estimation of asymmetrical cost distributions is particularly difficult.

An estimate of the value of perfect information suggested that further research, in order that robust economic decisions are made, is worth in the order of £129M. This suggests that the early termination of the study by the funders may have imposed extremely large costs (either financial or in terms of health forgone) on the NHS, through allowing a high degree of decision uncertainty to persist. It is acknowledged, however, that, even had the trial achieved its expected sample size, it is unlikely to have addressed all uncertainty.

Strengths and weaknesses of the economic analysis

The main strength of this analysis lies in the randomised controlled design of the study, which enabled the collection of high-quality data over the 48-week time horizon of the trial that were subsequently used in this analysis.

The small sample size achieved led to greater uncertainty in the conclusions drawn from the analyses and limited the scope of the analyses that could be performed. Only complete cases were used for the primary analysis. This meant that an even smaller sample was used for the primary analysis and the results must be treated with caution because of the potential for bias and the high likelihood of overestimation of the level of precision in the estimates of cost-effectiveness. The value of perfect information estimate indicates that further research is likely to be of value. Given the relatively short duration of follow-up, consideration of longer-term outcomes would be beneficial as part of any future research.

Meaning of the study

The costs associated with TNFi drugs and abatacept are relatively high because of the larger number of treatments that are given weekly or fortnightly (see Table 1), compared with the costs of rituximab, which is given in weeks 1 and 2 with a potential repeat at 6 months. However, when all health economic data were taken into account, to give the full costs incurred of the treatments over 48 weeks from a health and social services perspective, the difference in overall costs between the TNFis used in this trial and rituximab is quite small. The overall costs associated with abatacept are much higher. With only a small, non-significant, difference in the change in EQ-5D scores between trial arms, it is these differences in costs that drive the results and show that the use of alternative TNFi following an initial TNFi failure could be viable as a cost-effective treatment option alongside the currently approved treatment, rituximab.

Unanswered questions and further research

Although the analysis conducted here provides useful insights into the costs and effects of the treatment options in the trial period of 48 weeks, further research is required to provide evidence on the cost-effectiveness for alternative TNFi over a longer time horizon beyond 48 weeks. This could involve the development of an economic model that would also allow an extension to the value of information analysis conducted here to include the expected value of partial perfect information. This would provide a clearer idea of the areas around which further research should be conducted. Additional research would also be beneficial using a larger sample size to reduce the uncertainty in the conclusions drawn from the analysis and to ensure that a representative sample of the patient group is captured.

The existing evidence base has been limited to numerous small trials, whereas the ambition of the SWITCH randomised trial was to deliver a large-scale definitive trial that would have represented a paradigm shift in the RA community, delivering the largest RA pragmatic trial undertaken in the UK. Early termination of the SWITCH trial limits the conclusions that may be drawn and is therefore considered a lost opportunity to obtain definitive evidence on cost-effectiveness of either treatment option. However, the data presented may be used in meta-analysis in future research.