Tumour necrosis factor-a inhibitors for ankylosing spondylitis and non-radiographic axial spondyloarthritis: a systematic review and economic evaluation

Prognosis

Axial spondyloarthritis is a painful, progressive form of inflammatory arthritis. It mainly affects the spine but can also affect other joints, tendons and ligaments. Other areas such as the eyes and bowel can also sometimes be involved in non-radiographic and radiographic (AS) forms of axSpA. 9 The pain and stiffness of axSpA adversely affects optimal daily functioning. These symptoms are a result of a combination of reversible components of the disease, such as inflammation and flares, and irreversible components, such as syndesmophytes and vertebral bridging (bony deposition). 10 Most patients with AS develop the first symptoms at 25–45 years of age. 11 Progression of the disease is variable and difficult to predict. 12 There is often a delay of many years between patients first noticing symptoms and the diagnosis of axSpA being received. Many people with axSpA have AS, with evidence of bony deposition as well as inflammation. In later-stage AS, joints and bones may fuse together, a process that can occur over a long period of time and cause restricted movement. The functional impairment because of inflammation and/or bony deposition can have a profound effect on health and quality of life, and lead to withdrawal from active employment, with resultant adverse financial consequences; the burden of disease is greater in more socially deprived patients. 13 The prognosis is poor, although there is some evidence that deterioration plateaus in well-established AS. 14 Paradoxically, early disease (nr-AxSpA) may be less readily diagnosed and patients offered fewer treatment options even though it can be as, or even more, debilitating that established AS. 15

Ankylosing spondylitis is associated with an increased risk of death; it is estimated that patients have a standardised mortality ratio (SMR) of ≥ 1.5. The increased risk appears to be greater in men, with one study reporting a statistically significant increase in SMR of 1.63 in men but no significant increase in women (SMR 1.38) with AS. 16 This study found that, after correcting for age, sex, disease duration and pre-existing cardiovascular disease, independent predictors of increased mortality were elevated CRP level, diagnostic delay, not using NSAIDs and work disability. According to British Society for Rheumatology (BSR) guidelines, the excess mortality is mainly accounted for by cardiac valvular disease, amyloidosis and fractures. 17 nr-AxSpA affects approximately equal numbers of men and women, but it is more likely that men will develop AS. 18

Epidemiology

Currently, only limited epidemiological data are available for axSpA defined according to ASAS criteria. For AS, the prevalence is thought to be around 0.25% in European populations. 19 It is around three times more common in men than in women. 20 A recent study published in the USA reported an estimated AS prevalence of 0.52–0.55%, and the prevalence of axSpA as approximately 1.0–1.4%. 21 The proportion of nr-AxSpA among patients with axSpA is estimated to be between 20% and 80%. 22 Each year in the UK an estimated 2% of patients in a general practice will present with back pain and up to 5% of these will show features of AS. 23

Measurement of disease

There are a number of components and measures of disease activity in axSpA;24 a patient’s health-related quality of life (HRQoL) is determined by both by physical functioning and by disease activity. In turn, physical function is determined by spinal mobility and disease activity, and spinal mobility is determined by structural damage and inflammation of the spine. 24 In nr-AxSpA, a patient may have significant inflammation but no detectable structural damage; in AS, a patient may have both significant inflammation and structural damage; and in late AS, there may be less inflammation but extensive structural damage.

The main tools used for the assessment of various components of the disease are listed in Table 1.

Placebo response

The term ‘placebo effect’ can be used to describe different types of ‘effect’ but it generally encompasses one or more of three different meanings. First, there is the temporal (before–after) change after placebo medication, in which the effects of a placebo intervention cannot be distinguished from the natural course of the disease or regression to the mean. Second, there is the causal effect of placebo intervention associated with the treatment ritual, and, finally, there is the effect of all the psychological processes involved in the interaction between doctor and patient. 26 For the placebo-controlled trials in AS and nr-AxSpA these non-pharmacological components can be assumed to act equally in the anti-tumour necrosis factor (TNF) and placebo arms. Results from the placebo arms measure the non-pharmacological effects and the difference between the anti-TNF and placebo arms measures the pharmacological effect. All three components of the placebo effect could be important to consider when evaluating trials in this assessment, although once the trial treatment periods have ended, it is likely that the effect of the natural course of the disease becomes the most important factor of any ‘placebo’ effect. Estimated cost-effectiveness ratios and associated policy decisions may be sensitive to assumptions regarding the mechanism underlying placebo responses. 27

The natural course of disease activity in AS is known to vary over time with exacerbations, or flares, being common. In a study of flares in patients with AS, clinically relevant changes in Bath Ankylosing Spondylitis Disease Activity Index (BASDAI; but not in function) were noted during minor/localised flares (which occurred in 59% of patients in any given week). Although major/generalised flares were less common (reported in 12% of patients in any given week) they were associated with clinically relevant changes in both disease activity and function. 28 Pain is a key component of BASDAI and the ASAS responder outcomes; a Cochrane systematic review of placebos for all clinical conditions found that placebo interventions can influence patient-reported outcomes, especially pain (and nausea). 29 The authors also concluded that it was difficult to distinguish patient-reported effects of placebo from biased reporting, and that the effect on pain varied from negligible to clinically important, even among trials with low risk of bias.

Management of disease

Short- and long-term treatment goals for axSpA include minimising pain and stiffness, maintaining function and posture, arresting disease progression and maintaining quality of life and ability to work. Current conventional therapy for axSpA includes acute anti-inflammatory treatment with NSAIDs and physiotherapy and exercise.

Conventional therapy for AS is limited to NSAIDs (despite very limited supporting clinical trial evidence)30 and recommendations regarding appropriate physical activity. Other statements in the ASAS/EULAR (European League Against Rheumatism) recommendations for the management of AS include analgesics such as paracetamol and opioid-like drugs that may be considered for residual pain. Glucocorticoid injections into the direct site of inflammation (but not systemic) may be of benefit. The use of disease-modifying antirheumatic drugs (DMARDs, such as methotrexate and sulfasalazine) has been all but abandoned after evidence of lack of benefit. The cornerstone of non-pharmacological treatment of patients with AS is patient education and regular exercise; home exercises are effective. Physical therapy with supervised exercises, land- or water-based, individually or in a group, should be preferred, as these are more effective than home exercises. Patient associations and self-help groups may be useful. A Cochrane review of 11 trials concluded that the current best available evidence suggests that physiotherapy is beneficial for people with AS, but that it is still not clear which treatment protocol, duration and intensity should be recommended in the management of AS. 31 Physiotherapy is universally recommended32 but variable in practice.

Biologic drugs are the only treatment shown to be efficacious in the treatment of symptoms and signs of disease activity in axSpA and AS. Current National Institute for Health and Care Excellence (NICE) and BSR guidance recommends treatment with the anti-TNFs adalimumab, etanercept and golimumab in adults with active (severe) AS only if certain criteria are fulfilled, but it does not recommend infliximab for AS. 17,33

Inclusion criteria

Two reviewers independently screened all titles and abstracts. Full manuscripts of any titles/abstracts that were relevant were obtained where possible and the relevance of each study assessed by two reviewers according to the criteria below. Any discrepancies were resolved by consensus and, when necessary, a third reviewer was consulted. Studies available only as abstracts were included.

Searches

The following databases were searched for relevant clinical effectiveness and cost-effectiveness research:

• MEDLINE

• EMBASE

• Cumulative Index to Nursing and Allied Health Literature Plus

• Science Citation Index

• ClinicalTrials.gov

• Cochrane Central Register of Controlled Trials

• Cochrane Database of Systematic Reviews

• Database of Abstracts of Reviews of Effects

• International Prospective Register of Systematic Reviews (PROSPERO)

• Health Technology Assessment Database

• Conference Proceedings Citation Index – Science

• National Guidelines Clearinghouse

• NHS Evidence

• NHS Clinical Knowledge Summaries

• NHS Economic Evaluation Database (NHS EED).

The terms for search strategies were identified through discussion within the research team, by scanning the background literature and browsing the MEDLINE medical subject headings. No date or language limits were applied. As several databases were searched, some degree of duplication resulted. To manage this issue, the titles and abstracts of bibliographic records were imported into EndNote bibliographic management software (version X7, Thomson Reuters, CA, USA) to remove duplicate records. Databases were searched from inception, with most of the searches being performed in June or July 2014. The full search strategies used in each database, together with the search dates, are listed in Appendix 1.

Data extraction

Data relating to study design, outcome results and quality were extracted by one reviewer using a standardised data extraction form and independently checked for accuracy by a second reviewer. Disagreements were resolved through consensus, and, when necessary, a third reviewer was consulted. Data from studies with multiple publications were extracted and reported as a single study. Data were also extracted from the manufacturer’s submissions when they were not available from other sources. 34–37 Clinicaltrials.gov records and relevant US Food and Drug Administration (FDA) or European Medicines Agency reports were also used to extract any missing data. When data could only be estimated from graphs, the estimates used in the previous assessment report38 were used when available. In the light of the multidomain outcomes which incorporated pain scores (the ASAS and BASDAI outcomes), it was decided that pain scores on their own would not be extracted.

This report contains reference to confidential information provided as part of the NICE appraisal process. This information has been removed from the report and the results, discussions and conclusions of the report do not include the confidential information. These sections are clearly marked in the report.

Critical appraisal

The quality of RCTs was assessed using the Cochrane risk of bias tool,39 with additional assessments made for baseline imbalance of important prognostic indicators. 40 The relevant prognostic and treatment response indicators were identified from both published research and clinical advice. The risk of bias assessments were performed by one reviewer, and independently checked by a second. Disagreements were resolved through consensus, and, when necessary, a third reviewer was consulted. Open-label extension studies were evaluated based on the imputation methods and patient withdrawal criteria used.

Methods of data synthesis

This section describes the data set construction and meta-analyses conducted for the different outcomes individually. Chapter 5 provides detailed evidence synthesis methods that incorporate different outcomes within one analysis and presents clinical outcome estimates appropriate for the economic model.

Results of the data extraction in terms of study characteristics and quality assessment are presented in tables and summarised narratively. Results of open-label studies, drug survival and switching studies and natural history studies were also summarised narratively. As several of the RCTs were placebo-controlled up to 24 weeks, only time points beyond 24 weeks were evaluated in the open-label studies. AE data from the RCTs were pooled when enough data were identified; otherwise, the AE data and the other studies relating specifically to AEs were summarised narratively.

Clinical effectiveness data were synthesised using Bayesian meta-analysis methods. The main analysis was of outcomes reported from 10 to 16 weeks. A sensitivity analysis was done of outcomes reported from 24 to 30 weeks.

Quantity and quality of research available

The electronic database searches identified 2284 references. After screening titles and abstracts, full copies of 198 papers were assessed for inclusion in the review. Three trials of axSpA populations were excluded because results were not available separately for the AS and nr-AxSpA populations. 41–43 One study of adalimumab appeared likely to be eligible but was excluded as it was only available as a ClinicalTrials.gov record, without any results or further study details. 44 One excluded study was an ongoing trial of golimumab (called GO-AHEAD). 45

Twenty-eight eligible RCTs were identified, with 24 being suitable for data synthesis. Three etanercept trials were not suitable for data synthesis because the study durations were only 6 weeks,46–48 and one infliximab trial was unsuitable because a (currently) unlicensed dose (3 mg/kg) had been studied. 49 The Barkham 2009 trial50,51 of infliximab in nr-AxSpA patients (see Table 2) was included in the clinical efficacy section because, even though infliximab is not currently licensed for patients with nr-AxSpA, the dose used in this trial was the same as that licensed for AS. Furthermore, there was no reason to think it could not be considered in the same class as the other anti-TNFs when treating a nr-AxSpA population. The results of the trial therefore had the potential to be useful to help inform the relative efficacy of anti-TNFs for nr-AxSpA.

Of the 17 RCTs in which participants were studied beyond the randomised phase (i.e. in open-label studies), 71 additional full publications or conference abstracts were identified. Figure 1 illustrates the flow of studies through the review process.

FIGURE 1.

Flow chart showing the number of studies identified and included.

Clinical effectiveness results: efficacy results from randomised controlled trials

Individual results for all 24 trials are presented in Appendix 4.

Long-term efficacy results from open-label extensions of randomised controlled trials

Of the 24 included RCTs, 17 reported data from an open-label extension phase. Results for all studies are presented in Appendix 8. Considerable effort has been put into patient follow-up in anti-TNF trials with the result that data up to 5 years are available (there are data up to 8 years for infliximab but these included an involuntary treatment break which is not discussed further). The longest follow-up durations in patients with AS by anti-TNF are adalimumab 260 weeks, etanercept 264 weeks, infliximab 156 weeks, golimumab 268 weeks and certolizumab pegol 96 weeks. However, the data were reported across numerous publications and in various formats. Results were reported as observed, as completer analyses, using imputation (and rarely LOCF) for non-responders and LOCF for missing continuous data, but these related to differing populations (at varying time points): all patients randomised, all patients who took active drug at any point in the study or all patients who took active drug just during the open-label phase. The follow-up protocols were not clearly reported, with stopping rules unclear, but it appears that not all non-responders discontinued therapy. Therefore, the results may not reflect clinical practice should response be required for treatment continuation.

Table 15 presents the results based on non-responder imputation (NRI) analyses for the main studies when these results could be extracted. For AS the results show that across all the anti-TNFs after approximately 2 years of treatment, around half of patients are still achieving a good level of response to therapy. The results for golimumab look particularly strong with around 60% of all randomised patients achieving ASAS 40 and BASDAI 50 after 5 years. However, this is probably not reflective of clinical practice, as many of the normal weight patients took the 100-mg dose of golimumab rather than the 50-mg dose: the licence permits the use of 100-mg dose only in patients with a body weight of more than 100 kg who do not achieve an adequate clinical response after three or four doses. The equivalent results for adalimumab and etanercept are approximately 30% and 50%, although it is unknown if the difference may be because of variations in follow-up protocols rather than true treatment difference.

The long-term follow-up for nr-AxSpA patients (Table 16) shows continued high proportions of responders. At 1 year around half of patients are achieving an ASAS 40 or BASDAI 50 level response and with certolizumab this is maintained at 2 years and with adalimumab at 3 years.

When the long-term data are presented as observed or completer analyses, the long-term results are similarly good; withdrawal rates are not high and a high proportion of those who remain on treatment continue to achieve a good response, see the example data available from one trial of adalimumab and one of certolizumab pegol (Table 17).

At long-term follow-up mean final values or mean change from baseline for BASDAI, BASFI and BASMI, when reported, were generally maintained at clinically meaningful levels.

For adalimumab, data from the large ATLAS trial61 showed that mean changes from baseline at 1, 2 and 3 years remain stable and clinically meaningful at around –3.7 units for BASDAI and at around –2.9 units for BASFI. Similarly, the mean final value for BASMI remains at a level indicative of clinically significant treatment benefit (3.1 to 3.7 units). At 5 years the mean final values are BASDAI 1.8 units, BASFI 2.1 units, and BASMI 3.7 units. Clearly these results relate only to those patients who have remained on adalimumab in the long-term (40% of those who started adalimumab). They do, however, demonstrate continued benefit in a significant proportion of patients.

For certolizumab, results for these outcomes are available up to 96 weeks. At this time point the mean BASDAI and BASFI are indicative of clinically significant treatment benefit (both around 3 units).

The long-term data from Calin et al. 83 for etanercept, with 81 patients at 2 years and 59 (73%) remaining at 5 years also report mean BASDAI and BASFI scores of around 3.

From GO-RAISE90 at 2 years for those who took golimumab throughout the trial and follow-up (n = 138), median BASDAI score was around 3 and median BASFI score was around 2. These values are from a LOCF analysis of all patients randomised to golimumab 50 mg.

For infliximab, the Braun et al. 98 and follow-up studies found, from 1 to 3 years, a stable mean BASDAI score of around 2.6, a stable mean BASFI score of around 3 and a stable mean BASMI score of around 2.7.

Overall, the reported data (although not particularly robust) do indicate that significant proportions of patients continue to derive real benefit from continued use of anti-TNFs. There is nothing to indicate any difference between them.

Almost no data were available regarding radiographic progression of bony disease in patients with AS. Furthermore, it should be noted that radiographic changes and progression of these take many years to appear and radiography is an insensitive tool by which to evaluate the progression of AS. Therefore, evidence, particularly that from relatively short-term studies, has to be interpreted with caution. The limited evidence includes mSASSS change from baseline, reported for golimumab from the GO-RAISE90 study at 4 years (208 weeks): 1.3 units (SD 4.1 units) based on the 111 of 138 patients randomised to 50 mg. As results from untreated cohorts suggest a progression rate of 2 units/2 years, a rate of 1.3 units (or even 2 units) over 4 years seems beneficial. For further discussion of this issue see Effect of anti-tumour necrosis factors on radiographic progression. MASES was reported only for adalimumab from ATLAS;61 in patients remaining on therapy at 2 years the mean change from baseline was 2.2 units (n = 217).

For nr-AxSpA patients long-term data for the continuous outcomes was limited to 1 year’s follow-up. For adalimumab, data were available from only one small study (Haibel 2008,52 n = 46): BASDAI change from baseline 2.8 units (95% CI 2.1 to 3.6 units); BASFI change from baseline 2 units (95% CI 1.4 to 2.6 units); BASMI change from baseline –0.4 units (95% CI –0.7 to –0.04 units); and MASES change from baseline of 0.9 units (95% CI –0.02 to 1.9 units). In addition, of 26 patients with magnetic resonance images at baseline and 52 weeks’ follow-up, none showed a change in sclerosis or in erosions. For etanercept, data were available on 205 patients randomised to etanercept or placebo and then on long-term etanercept (Dougados 201476): [commercial-in-confidence (CiC) information has been removed]. For certolizumab, LOCF analysis at 96 weeks (n = 97) gave a BASDAI final value score of 3.0, and a BASFI score of 2.4. Overall, the 1-year results in nr-AxSpA patients are similar to each other and also reflect those seen in AS patients. Again, the short-term nature of this follow-up relative to the 8–10 years over which radiographic changes develop must be borne in mind.

Findings from anti-tumour necrosis factor patient registry studies

Clinical effectiveness results: adverse events

Review of natural history of ankylosing spondylitis and non-radiographic axial spondyloarthritis

In order to get some understanding of what happens to patients who, although eligible for anti-TNF therapy for their AS or nr-AxSpA, do not receive it, we conducted a rapid review of relevant literature. This was not a systematic review but one that started with the library of papers found by the main searches for RCTs of the anti-TNFs and then followed relevant citations to papers on AS and axSpA in patients not receiving an anti-TNF. Potentially relevant papers were those that reported on the pattern of disease, AS or nr-AxSpA or axSpA, without treatment with anti-TNFs over time. This process identified a number of relevant registries: OASIS,118 Scotland and Ireland Registry for Ankylosing Spondylitis (SIRAS), Devenir des Spondylarthropathies Indifferenciées Récentes, Esperanza, Spanish Registry of spondyloarthritis, German Spondyloarthritis Inception Cohort (GESPIC) and St Mary RheumaToid Arthritis (SMART). Additional searches of MEDLINE were conducted using these specific registry names. All relevant studies identified through this process are presented in Table 27.

The studies collectively explore the associations between the various components of axSpA: disease activity, structural damage and spinal mobility. The exploration of the ASSERT trial baseline data24,124 reveals that HRQoL as determined by SF-36 physical and mental components, is determined by BASFI and BASDAI; BASFI is determined by BASDAI, mSASSS and BASMI (spinal mobility); and BASMI is independently determined both by irreversible spinal damage (mSASSS) in late disease and reversible spinal damage (MRI) in early disease.

The studies identified that from a clinical practice and patients’ point of view disease progression in terms of BASFI, a measure of the patient’s functional ability, is very important. A number of studies on the disease progression of AS have been based on the European OASIS cohort118 (a consecutive cohort, started in 1996, although there were no further specific eligibility criteria); the total cohort numbers 217 patients. One of these, a study by Landewe et al. ,10 demonstrated that physical function impairment (BASFI) is independently affected by both disease activity (BASDAI) and bony progression, usually assessed using mSASSS despite this being a measure of bony growth in the spine only (and not in the sacroiliac joints). Other studies by Ramiro124,125 have demonstrated that radiographic progression, increases on average by around 2 mSASSS units every 2 years. 124,125 However, this progression is highly variable; the average patient with inactive disease [Ankylosing Spondylitis Disease Activity Score (ASDAS) 0] would progress by 5 mSASSS units over 12 years compared with a patient with ‘very active disease’ (ASDAS 4) who would have 19 units of progression. 124 In addition, of 68 patients who were followed for 12 years, 18% had no progression on mSASSS. 125 The variability is also demonstrated by the results based on a different cohort: a single German clinic (n = 146). 12 Baseline characteristics were similar to those in the OASIS cohort118 (see Table 27). Mean follow-up was 3.8 years (SD 1.7 years) and mean mSASSS change was 1.3 units/year (SD 2.5 units/year) with a range of 0–22.8 mSASSS units. Thirty-four (23%) patients showed no progression.

There is evidence that BASDAI is relatively constant over time. An analysis of data from a UK registry, SIRAS, demonstrated that patients stratified into high or low disease activity (BASDAI) remain in their separate groups over many (12) years. 149 Data on the long-term pattern of patient function (BASFI) in patients not being treated with anti-TNFs are more scarce. A cohort study, from a single centre in England, provided data on 69 patients followed over 10 years [two data points: at baseline (1998) and at 10 years (2008)]. 14 The assessment of BASDAI confirmed that it remains relatively constant [mean at baseline 4.1 units (SD 2.5 units) and after 10 years 4.4 units (SD 2.7 units) (p = 0.36)]. Patient function was assessed using the Revised Leeds Disability Questionnaire (RLDQ) rather than BASFI, but provided evidence of deteriorating function over time: mean RLDQ at baseline was 10.4 (SD 8.3), and after 10 years was 13.6 (SD 10.9) (p = 0.002). Analysis of longitudinal data from the SMART (Bath, UK) data set (n = 223) found that BASFI increased over time by 0.035 units/symptom year. 150 In patients with baseline BASDAI of ≥ 4 (those that would be treated with anti-TNFs and 68% of the total cohort) the rate of BASFI increase was 0.039 units/symptom year. Estimates of the rate of change in BASFI over time were also reported in a cost-effectiveness modelling study. 152 The data were from patients who were captured in two surveys at two time points 1992/1994 and November 2002 approximately 8 years apart (n = 1100). The estimate of annual BASFI progression was 0.07 points, but when only patients with BASDAI score of ≥ 4 were included in the analysis, BASFI progression was estimated as 0.054. It was reported that data from a cohort of 493 patients who had been followed up for more than 3 years generated similar findings; the number was not actually reported for the whole survey but was 0.059 for the BASDAI score of ≥ 4 subgroup.

Natural history data from patients with nr-AxSpA are even more scarce than those for AS patients, with no long-term data identified. A comparison of AS and nr-AxSpA patients from a cohort of 100 consecutive patients (Herne clinic, Germany) (AxSpA n = 100, nr-AxSpA n = 44, AS n = 56) found that slightly higher proportions of AS patients met pre-specified cut-off points of disease severity than did nr-AxSpA patients, but the differences were statistically significant only for ASDAS, CRP level, mSASSS and the number of inflamed lesions; the proportion of males was also statistically significantly different. 153 The results are given in Table 27. The difference for BASFI was very close to statistical significance.

A larger cross-sectional study of the GESPIC cohort [n = 462 patients with axSpA (AS or nr-AxSpA)] also found differences between AS and nr-AxSpA patients. 18 When AS (≤ 5 years) and nr-AxSpA were compared, there were statistically significant differences in Physician Global Assessment, BASFI (3.1 in AS vs. 2.5 in nr-AxSpA), BASMI (1.9 in AS vs. 1.1 in nr-AxSpA), spinal mobility and lateral spinal flexion, CRP level and ESR, and all radiographic measures (mSASSS 4.9 in AS vs. 1.4 in nr-AxSpA). mSASSS was statistically significantly worse in males versus females and between CRP level > 6 and < 6, although it is unclear whether or not this is a meaningful cut-off point for CRP level.

In two longitudinal studies of progression in nr-AxSpA,154,155 also using the GESPIC cohort, progression in terms of sacroiliitis and in terms of radiographic progression in the spine (mSASSS) was slightly more rapid in AS than in nr-AxSpA but not statistically significantly so. Raised CRP level at baseline was a predictor of both measures of progression in AS but only for sacroiliitis in nr-AxSpA. The presence of syndesmophytes was predictive of higher progression rates as assessed by mSASSS in both AS and nr-AxSpA. Of the 95 patients with nr-AxSpA, 11 (11.6%) fulfilled the modified New York criteria for AS after 2 years of follow-up. A review of the burden of illness in nr-AxSpA157 cited this (11.6%) progression rate along with a 10% rate over 2 years and a 24% rate over 10 years. However, the 10-year rate was derived from a broader, more heterogeneous population than the GESPIC cohort: patients had undifferentiated spondyloarthropathies with over half not having inflammatory low back pain. 156,158 The GESPIC study recruited only patients with axSpA (AS or nr-AxSpA).

Studies of disease progression in nr-AxSpA focus on aspects of the disease that can be assessed through imaging techniques: radiography or MRI. This may appear reasonable given the subjective, patient-questionnaire basis of the BASFI score.

Finally, there is evidence that as well as being progressive, the course of AS includes flares. A study based on the population of a trial comparing probiotic and placebo treatment in AS found that the overall flare rate was 71.4 per 100 person-weeks; the major flare rate was 12/100 person-weeks and the minor flare rate was 59.4/100 person-weeks. 28 BASDAI and BASFI varied with type of flare: mean BASDAI scores were 5.5 (major flare), 3.1 (minor flare) and 2–2.5 (flare free), and mean BASFI scores were 5.5 (major flare), 3.1 (minor flare) and 2.5–3.5 (flare free). A pilot study used the SMART cohort (Bath, UK) to investigate the pattern of disease and impact of disease flares. 159 Of the 114 patients, 96% reported experiencing flares. Flare duration varied by patient: days (40%), weeks (30%) and months (30%). Fifty per cent of patients reported flares on a background of symptoms, while 26% reported gradually developing and resolving flares after which symptoms were worse than before the start of the flare. These patterns were associated with higher BASFI scores. Around 20% reported flares with no symptoms between. A small proportion (7%) reported gradually developing and resolving flare with periods of no symptoms.

In summary, the available studies indicate that in AS and nr-AxSpA disease activity (BASDAI) is fairly stable over time and does not generally progress, although it can be at a high (severe) level early in the disease. Patients function (as assessed by BASFI) does deteriorate over time, although the course is not constant or predictable. BASFI is determined by both disease activity and bone neo-formation; progression of BASFI score over time is driven by progression of bony disease as assessed by imaging scores such as mSASSS, or the presence of syndesmophytes. Best estimates of yearly disease progression rates without anti-TNF therapy are around 1.0 mSASSS units and 0.035–0.07 BASFI units. Information on the natural history of nr-AxSpA is relatively sparse. While disease progression appears to be faster in AS, patients with nr-AxSpA can have severe disease activity and hence poor function.

Summary of randomised controlled trial results

The quality of the trial evidence was generally high; most studies were unlikely to have produced results which were biased. For both the AS and nr-AxSpA populations, the results of the meta-analyses demonstrated that anti-TNFs produce statistically significant and clinically relevant benefits to patients in terms of improving function and reducing disease activity. The common class-effect model used may have underestimated the uncertainty in the effect estimates. Although there is a possibility that infliximab is more effective than other TNF inhibitors at least at 12 weeks, there is no strong evidence to support this. For the disease activity, function and responder outcomes, the class-efficacy estimates were consistently slightly smaller for nr-AxSpA than for AS, most noticeably for BASFI and BASDAI 50. Statistical heterogeneity was more apparent in the nr-AxSpA analyses than in the AS analyses. This may be a result of both clinical heterogeneity in the nr-AxSpA trials (such as variation in CRP levels or the proportion of MRI positive patients) and the fact that fewer studies were available for analysis. In the light of the statistical heterogeneity across the nr-AxSpA trials, both the reliability of the nr-AxSpA-pooled estimates and their true relevance to patients seen in clinical practice is questionable.

The US FDA reanalyses of two key nr-AxSpA trials further emphasised the heterogeneity in the nr-AxSpA population. Results for an adalimumab trial in nr-AxSpA patients suggested reduced efficacy in a centrally diagnosed nr-AxSpA population than in a locally diagnosed population and that the treatment benefit in the whole trial population may have been driven by benefit in patients who actually had AS, not nr-AxSpA. Conversely, in a certolizumab pegol trial which recruited both populations, the efficacy findings were consistent across the AS and nr-AxSpA subpopulations, regardless of the discrepancy in local or central pelvic radiograph readings.

Long-term efficacy

The longest follow-up durations in patients with AS by anti-TNF were 5 years for adalimumab, 5 years for etanercept, 3 years for infliximab, around 5 years for golimumab and nearly 2 years for certolizumab pegol. The results showed that across all the anti-TNFs after approximately 2 years of treatment, around half of patients still achieved a good level of response to therapy. At 5 years around 60% of golimumab patients, 50% of etanercept patients and 30% of adalimumab patients still achieved a good treatment response. However, the long-term studies were not as well-reported as the RCTs, and their results were derived from less reliable data; it is therefore unknown if these are true treatment differences or a result of differences in follow-up protocols, and/or imputation and analysis methods.

The long-term follow-up for nr-AxSpA patients showed a continued high proportion of responders. At 1 year around half of patients on adalimumab, etanercept or certolizumab still achieved an ASAS 40 or BASDAI 50 level response. With certolizumab this is maintained at 2 years and with adalimumab at 3 years.

When the long-term data are presented as observed or as completer analyses, the long-term results are similarly good: withdrawal rates are not high and those patients who remain on treatment continue to achieve a good response.

For all anti-TNFs, at long-term follow-up mean final values or mean change from baseline for BASDAI, BASFI and BASMI, when reported, were generally maintained at levels indicative of clinically significant treatment benefit for those patients with AS and those with nr-AxSpA.

Four studies reported on radiographic disease progression over 2 years of follow-up in terms of mSASSS in patients taking adalimumab, infliximab, etanercept and golimumab. All four open-label, uncontrolled follow-up studies found that mSASSS increased by a mean of around 0.9 over 2 years. Three of these studies compared their rates with those from the OASIS cohort118 (of patients not taking an anti-TNF) and found no difference. In conclusion, there is no real evidence for the impact of anti-TNF treatment on radiographic disease progression; a beneficial effect cannot be assumed, nor, given the short-term nature of the follow-up and the insensitivity of radiography as a tool for the evaluation of disease progression in AS, can one be discounted. There are some data to suggest an identifiable benefit from around 4 years but results from ongoing long-term studies should help to clarify this issue.

Registry data demonstrate that around 60% of patients with AS treated with a first anti-TNF will still be taking their therapy at 2 years, with median drug survival of 3.1 years (based on Danish registry n = 1436). Sequential treatment with anti-TNFs can be worthwhile but the drug survival response rates and benefits are reduced with second and third anti-TNFs, with the proportion of BASDAI 50 responders falling approximately 10% with each subsequent anti-TNF and the median BASDAI and BASFIs achieved increasing (worsening). The lower efficacy of a second anti-TNF relative to a first is reflected in lower median drug survival and proportion of patients remaining on therapy at 2 years. Interestingly, despite a further reduction in response and efficacy with a third anti-TNF, drug survival does not fall further, suggesting that patients may be allowed to, and be prepared to, continue with a less than optimally effective anti-TNF at this stage in their treatment history.

Adverse effects

Data from large systematic reviews, which included patients with a wide range of diseases, suggest that, in the short term, anti-TNFs as a group are associated with significantly higher rates of serious infections, tuberculosis reactivation, non-melanoma skin cancer, total AEs and withdrawals because of AEs, when compared with control treatments. Specifically, infliximab is associated with significantly higher rates of total AEs and withdrawals because of AEs, and that certolizumab pegol is associated with significantly higher rates of serious infections and SAEs. Analyses from the present review showed etanercept to be statistically significantly more likely to result in an injection/infusion site reaction compared with infliximab, although analysable data on such reactions were not reported for the other three anti-TNFs. Evaluations of longer-term data are more scarce, although they suggest similar safety profiles across anti-TNFs. Data from the open-label studies included in this review also do not suggest that there are important differences between treatments, other than a higher incidence of injection site reactions following treatment with etanercept. These open-label data are, however, limited by the small sample sizes and non-randomised study designs.

Natural history

The available studies indicate that in AS and nr-AxSpa disease activity (BASDAI) is fairly stable over time and does not generally progress, although it can be at a high (severe) level early in the disease. Patient function (as assessed by BASFI) does deteriorate over time, although the course is not constant or predictable. BASFI is determined by both disease activity and bony disease; progression of BASFI over time is driven by progression of bony disease as assessed by imaging scores such as mSASSS, or the presence of syndesmophytes. Best estimates of yearly disease progression rates without anti-TNF therapy are around 1.0 mSASSS units and 0.035 to 0.07 BASFI units. Information on the natural history of nr-AxSpA is relatively sparse. While disease progression appears to be faster in AS, patients with nr-AxSpA can have severe disease activity and hence poor function.

Overall conclusions

• For both the AS and nr-AxSpA populations the results of the meta-analyses demonstrated that anti-TNFs produce statistically significant and clinically important benefits to patients in terms of improving function and reducing disease activity. The efficacy estimates were consistently slightly smaller for nr-AxSpA than for AS.

• In AS, although there is a little variation in treatment effects and it is possible that infliximab may be more effective than other anti-TNFs at 12 weeks, the evidence for this is not strong and it is plausible that anti-TNFs may have a common class effect, with the treatments being equally effective.

• Statistical heterogeneity was more apparent in the nr-AxSpA analyses than in the AS analyses. This may be because of both clinical heterogeneity in the nr-AxSpA trials and the fact that fewer studies were available for analysis. In the light of this heterogeneity, both the reliability of the nr-AxSpA-pooled estimates and their true relevance to patients seen in clinical practice is questionable.

• Effectiveness was maintained over time. About 50% of patients maintained a benefit at 2 and 5 years.

• Evidence for an effect of anti-TNFs on radiographic disease progression was limited. The relatively short-term follow-up available to date and the insensitivity of radiography as an imaging tool precluded the drawing of firm conclusions regarding the role of anti-TNFs in preventing or delaying the progression of AS; there are some data to suggest an identifiable benefit from around 4 years, but results from ongoing long-term studies should help to clarify this issue.

• Sequential treatment with anti-TNFs can be worthwhile in patients with AS but the drug survival response rates and benefits are reduced with second and third anti-TNFs.

Methods

An initial systematic search was undertaken in the NHS EED using a combination of technology names and disease terms. Further searches were undertaken in MEDLINE and EMBASE for modelling and utility studies using disease terms only (as known references were not identified from the initial search in NHS EED). Only full economic evaluations that compared two or more options and consider both costs and consequences (including cost-effectiveness, cost–utility and cost–benefit analyses) were included in the review of existing economic literature. No language and date limits were initially applied, although eligibility of studies was subsequently restricted to those reporting results which were specific to the UK. Full details of the search strategies used are reported in Appendix 1.

In addition, as part of the current multiple TAs process, each manufacturer submitted de novo evidence on the cost-effectiveness of the anti-TNFs in line with their indications for the treatment of AS and nr-AxSpA. 34–37 These submissions are reviewed and the findings compared with those found in the review of previously published studies.

Results of review of existing cost-effectiveness evidence

The combined searches retrieved 210 citations. A total of six UK studies reporting on the cost-effectiveness of anti-TNFs for the treatment of AS were identified. No previously published studies were identified for patients with nr-AxSpA.

Four of these studies were industry-funded assessments of the following anti-TNFs: infliximab (Kobelt et al. 2004152 and Kobelt et al. 2007160 both funded via an unrestricted grant by Schering-Plough), etanercept (Ara et al. 161 funded by Wyeth pharmaceutical P.C.) and adalimumab (Botteman et al. 162 funded by Abbott Laboratories). The three studies published in 2007 are largely based on the economic analyses originally submitted by the manufacturers to NICE as part of the previous multiple TA (TA143). 17 As the earlier publication by Kobelt has been superseded by the 2007 publication, only the latter publication is further considered in this review. The remaining two UK studies were publications of the assessments and/or critiques undertaken by the independent Assessment Group/Evidence Review Group (ERG) for infliximab, etanercept and adalimumab for TA14338 and golimumab for TA233. 163 Therefore, a total of five studies met the inclusion criteria and are included in this review.

The following sections provide a narrative discussion of each publication. The quality assessment of these studies is shown in Appendix 10. A single critique section is used to highlight the key issues and potential limitations of existing published cost-effectiveness evidence. These issues are then revisited with respect to the de novo analyses submitted by the manufacturers considering how these key issues and potential limitations have been addressed in the two separate indications. The final section highlights the remaining issues and uncertainties and provides the basis for informing the development of a separate independent analysis of the cost-effectiveness of anti-TNFs for AS and nr-AxSpA relevant to informing decisions for the NHS.

Assessment of published cost-effectiveness studies

Summary and critique of published cost-effectiveness studies

No previously published studies were identified which assessed the cost-effectiveness of anti-TNFs for nr-AxSpA. Consequently, the de novo submissions provided by the manufacturers provide the only existing evidence which can be considered to inform decisions for the NHS. Of the previously published UK cost-effectiveness study identified, there appear marked differences between the results of the industry-funded assessments and the results from the independent assessment by LRiG. Importantly, the results of the independent critique and exploratory reanalysis by the ERG for TA23333 also appear potentially less favourable than the industry-funded published assessments. Although the DSU review of models submitted as part of TA14317 has reconciled many of the key differences and highlighted the key assumptions, a number of key uncertainties remain. The remainder of this section provides an overview of the issues and uncertainties identified based on existing published studies and the DSU reports. This summary provides an important basis for considering the extent to which the de novo submissions provided by the manufacturers for this appraisal have adequately addressed these.

All existing models are based on similar two part structures:

• initial-response period (short-term model used to determine initial response rate)

• post-response period (longer-term model used to characterise natural history of disease (i.e. without anti-TNFs) and impact of anti-TNFs (while on therapy and when therapy is stopped).

All models use changes in BASDAI and/or BASFI to quantitatively model the short- and longer-term costs and quality-of-life implications (using QALYs) of the use of anti-TNFs versus CC alone.

Although there are differences between the modelling of the initial response period, existing models are broadly comparable being based on an assessment around 12 weeks (and potentially at 24 weeks as well) using a particular variant of existing BSR guidelines. Patients receiving anti-TNFs who meet the response criteria at the 12-/24-week assessment are continued on anti-TNFs. Anti-TNFs are withdrawn in non-responders at the 12-/24-week assessment point and patients subsequently receive CC alone.

However, there are marked differences between existing studies in relation to the modelling of the post-response period and the assumptions used. This period is often separated into different time intervals allowing different assumptions to be made regarding the effect of anti-TNFs (i.e. initially improving with time in responders but then later ‘levelling off’ or even deteriorating over a longer-term time horizon relative to CC). An important difference between existing models is the timing of this ‘levelling off’ period and assumptions employed over a longer time horizon. The differences in approaches and the timing of this ‘flattening off’ period are also closely linked to the data used, that is whether or not the changes in BASDAI/BASFI used in the model are restricted to the 12- to 24-week data from RCT evidence reported during the double-blind phase (Kobelt et al. 2007160 and Ara et al. 2007161) or also incorporate longer-term data from the open-label extensions. Studies which use change in BASDAI/BASFI data directly in the model, from the double-blind phase, appear to use shorter ‘levelling off’ periods than studies using data from the open-label extension phase (Botteman et al. 2007,162 McLeod et al. 200738 and Armstrong et al. 2007163).

Those studies incorporating an open-label extension typically assume continuing changes in BASDAI/BASFI for responders to anti-TNFs versus non-responders/CC beyond the initial 12/24-week period. Importantly, none of the studies using open-label extension data appear to provide any discussion of the potential for selection bias (e.g. related to the initial consent for patients to participate and/or agree to switch treatments as well as ongoing selection issues concerning retention over a longer period) and how these should be considered and/or adjusted for in the economic model. However, the implication of this is important, as the assumption being made by several models appears to incorporate an assumption of an increasing effect of anti-TNFs in responders over time (i.e. in terms of continuing improvements in BASDAI/BASFI), which does not appear to be adequately justified or related to any underlying clinical/pharmacological mechanism. In the absence of the counter-factual (i.e. comparable data in patients who did not participate or were subsequently withdrawn from the open-label study) it is unclear whether the apparent increasing effect is simply a function of the selection issue or is a real effect of the anti-TNFs. Importantly, those studies which only use data from the double-blind periods of RCTs often cite the open-label data as providing supportive evidence regarding the maintenance of the effects observed at 12/24 weeks but do not use it to support an assumption of an increasing effect over time.

The longer-term impact on costs and utilities beyond the initial response period are subsequently quantified by estimating separate BASDAI/BASFI ‘trajectories’ for different patient categories. The three main categories are:

1. CC

2. non-responder to anti-TNFs at 12-/24-week assessment

3. initial responder to anti-TNFs at 12-/24-week assessment.

The ‘trajectory’ for patients who are responders to anti-TNFs at the initial 12-/24-week assessment are further separated into (1) the period up to the point that anti-TNFs are subsequently withdrawn (i.e. because of loss of efficacy or AEs) and (2) the period post TNF-α inhibitor withdrawal.

After the ‘levelling off’ period for BASDAI, the majority of existing studies assume BASDAI is constant over the longer term, that is the BASDAI of responders to anti-TNFs is assumed to be lower than the equivalent BASDAI value (lower disease activity) applied to CC/non-responders and a constant difference is assumed to be retained until patients discontinue. At the point of discontinuation of anti-TNFs, patients subsequently revert back to the same value assumed for CC/placebo and non-responders to anti-TNFs at 12/24 weeks. Hence, any improvement in BASDAI is assumed to dissipate immediately or within a short period (3–6 months) after discontinuation of anti-TNFs.

All existing studies model BASFI as a linearly increasing function over the longer term for non-responders/CC, that is a constant rate of change is subsequently applied which is used to characterise the impact of disease progression on functional ability, typically a worsening of 0.07 (0–10 scale) units per annum. Again, the same assumptions applied to BASDAI for non-responders to anti-TNFs are applied to BASFI, that is beyond 12/24 weeks non-responders are assumed to follow an identical BASFI ‘trajectory’ as that of CC/placebo patients. By contrast, patients who respond to anti-TNFs are typically assumed not to ‘progress’ further in terms of functional disability, or progress at a lower rate than CC patients, while continuing to receive anti-TNFs. Hence the difference in individual mean BASFI scores increases over time in existing economic models between patients who continue to receive anti-TNFs and non-responders/CC.

The only study which employs a markedly different approach to the modelling of BASDAI and BASFI for responders is the study undertaken by the previous independent assessment group (LRiG) for TA143. 17 Instead, LRiG applied a quadratic function to the BASDAI and BASFI scores of responders. This approach assumed that the difference compared with CC was decreasing (initially) with time, that is, over time, the differences in BASDAI/BASFI would slowly reduce in responders and eventually be the same as for CC. While the logical problems of applying a quadratic function over a longer period were recognised by the authors (i.e. function begins to increase after a particular period) and was addressed using a series of logical restrictions (i.e. BASDAI/BASFI score constrained to be the same or better than CC), the clinical ‘face’ validity of this approach also appears questionable in the context of longer-term projections which are required for appropriate assessments of cost-effectiveness.

Another key difference between existing studies relates to the assumptions made concerning the subsequent trajectory of BASFI for patients who withdraw from active treatment. Given that BASFI is linearly increasing with time for CC, the assumption of the subsequent BASFI trajectory is potentially an important driver of cost-effectiveness. This is often referred to as ‘rebound’. Typically, two scenarios are used:

• Rebound equal to gain: when patients fail therapy (after initially responding), their BASFI deteriorates by the same amount by which it improves when they responded to therapy.

• Rebound back to natural history/CC: when patients fail therapy (after initially responding), their BASFI deteriorates to the level and subsequent trajectory it would have been had they not initially responded to therapy.

In the absence of evidence on the magnitude of any rebound, these alternative scenarios represent the ‘best-case’ and ‘worst-case’ scenarios possible. In other words, the reality regarding rebound is likely to be somewhere between these two scenarios which should, therefore, be seen as the limits.

The implications of the different rebound scenarios are clearly illustrated in Figures 2 and 3. Studies which are based on assumptions of rebound equal to gain incorporate an ongoing benefit of anti-TNFs in patients in whom therapy is subsequently withdrawn after an initial response. Hence, such an assumption is more optimistic than assuming no continuing benefit at the point treatment is withdrawn.

FIGURE 2.

Illustration of the scenario of rebound equal to gain.

FIGURE 3.

Illustration of the scenario of rebound to CC.

Although the impact of discontinuation in patients who initially respond is clearly an important issue, the assumptions underpinning the subsequent trajectories of patients who are non-responders at 12/24 weeks to anti-TNFs are rarely explicitly justified. The most common assumption applied is that non-responders during the initial period follow the same subsequent trajectory for BASDAI/BASFI as CC/placebo patients beyond the 12-/24-week assessment point. However, the appropriateness of this assumption does not appear to have been discussed in existing studies. Essentially, for this assumption to hold, the initial response to anti-TNFs has to be independent of baseline patient characteristics, such that response to treatment is effectively a random process. However, if response is not independent of patient characteristics, the implication is that responders/non-responders to TNF-α inhibitors may be systematically different from each other. This has implications for the appropriateness of current assumptions being applied to non-responders at 12/24 weeks and subsequent responders who later withdraw. For example, all other things being equal, if patients with more severe disease (high BASDAI/high BASFI) were more likely not to respond, then assuming that the non-responders at 12/24 weeks follow the same trajectory as the ‘average’ CC/placebo patient is likely to be optimistic towards the anti-TNFs (and vice versa for if less severe patients are more likely to respond). Hence, rather than following the trajectory of an ‘average’ placebo/CC patient, a non-responder may actually follow a different trajectory, that is that of an equivalent more/less severe CC patient. Inevitably, the impact of different patient characteristics is likely to be more complex than the simplistic scenarios presented above.

As previously noted, all models use changes in BASDAI and/or BASFI to quantitatively model the short- and longer-term costs and quality-of-life implications (using QALYs) of the use of anti-TNFs versus CC. The justification for using these measures appears largely driven by the existence of external sources of costs and health utility estimates which can be directly linked to these measures and not to others (e.g. BASMI, ASDAS, mSASSS, etc.). Hence current models appear more of a function of the data which are available to link to costs and utilities rather than being based on a clear underlying biological or clinical process. This raises more general conceptual concerns regarding existing models and also regarding the generalisability of findings in an AS population to the separate nr-AxSpA population.

The use of BASDAI/BASFI per se is perhaps not the most significant issue, as in the absence of alternative mapping functions to costs and/or utilities it is unclear how to estimate longer-term costs and QALYs without ultimately linking to these measures. However, it is concerning that the majority of existing studies do not appear to link the data and assumptions applied to these measures to any coherent clinical underpinning regarding differences between population characteristics and the effect of anti-TNFs. Consequently, ‘progression’ over time is currently modelled entirely via changes in BASFI, as BASDAI is assumed to remain constant. However, no attempt is made to justify why BASFI increases, the rate at which it increases and how this rate might differ across different groups as well as the impact that anti-TNFs might have (i.e. any effect on BASFI which may be independent of the effect on BASDAI).

Modelling ‘progression’ implicitly (i.e. employing natural history estimates of the rate of change of BASFI from external studies) rather than explicitly (i.e. attempting to explain how BASFI evolves over time in relation to inflammatory and other processes and how these may differ within populations and across the AS and nr-AxSpA groups) has led to a series of implicit/evidence-free assumptions. These include:

• No change in BASFI while receiving anti-TNFs (i.e. assuming implicitly that these act as disease modifiers and that while patients respond and continue to receive them, further deterioration in functional progression is completely prevented).

• Lower BASFI changes while receiving anti-TNFs (i.e. assuming that anti-TNFs do not completely halt further deterioration in functional progression but that the rate of progression is reduced relative to progression on CC).

• Similar natural history rates of change in BASFI across different subgroups and populations (i.e. assuming that rate of change in BASFI is independent of time and/or patient characteristics).

Similar conceptual concerns were also highlighted by the NICE DSU in their work to support TA143, noting that in inflammatory arthritis a clearer conceptual relationship is assumed between disease activity, radiographic progression and physical functioning, such that changes in physical functioning can be more clearly related to different processes and evidence for the anti-TNFs on each separate process. In highlighting these issues, the DSU cited emerging longer-term data reported for anti-TNFs based on measures of radiographic progression (mSASSS) in AS. Although this evidence was not formally included in their analyses, the evidence was cited to indicate that an assumption of no further progression while on anti-TNFs for AS was potentially optimistic based on emerging longer-term radiographic progression data.

Importantly, the only UK study published since the NICE DSU review did subsequently use a less favourable assumption concerning the impact of anti-TNFs on functional progression (BASFI). The assumption used by the manufacturer for golimumab37 assumed that the longer-term rate of change in BASFI for responders who continued on treatment would be 50% of that assumed for CC/non-responders. Although this assumption is a significant departure from the base-case assumptions applied within previous industry-funded studies, no justification appeared to be identified to support this by Armstrong et al. 163 in the review of the manufacturer’s submission.

In summary, there appear to be significant differences between the cost-effectiveness results reported in existing UK published studies. Many of these differences appear largely because of differences in data sources (i.e. double-blind period vs. open-label extensions), subsequent assumptions and estimates related to the magnitude and duration of the differences in BASDAI and BASFI measurements between responders and non-responders in the short to medium term (i.e. the ‘levelling off’ period) and then longer term in relation to assumptions concerning BASFI progression, and issues around ‘placebo’ effect and the withdrawal of anti-TNFs. Some of the main differences between existing studies have been highlighted in a separate review by the NICE DSU. However, while this review is helpful in identifying the impact of parameter and structural assumptions, it does not provide a basis for informing which assumptions appear most justified based on existing data and clinical understanding of the progression of AS and the impact of anti-TNFs. It is also concerning that many of the existing studies are based on CiC data and hence lack transparency regarding specific inputs and assumptions.

To date, only two UK studies have attempted to assess the cost-effectiveness of the alternative anti-TNFs. One of these studies, McLeod et al. ,38 assumed that the alternative treatments were identical in terms of clinical effectiveness and hence only considered differences in the acquisition, administration and monitoring costs. The justification provided by the authors was based on the lack of statistically significant differences across key outcome measures based on indirect comparisons. The other study, Armstrong et al. ,163 assumed differences in the clinical effectiveness of the alternative anti-TNFs based on a separate MTC. However, differences between the anti-TNFs appeared sensitive to the studies included and the specific outcomes considered. Hence different conclusions could be drawn concerning the most ‘efficient’ intervention depending on the analysis considered. However, the magnitude of differences in clinical effect and QALYs remained small and the clinical and economic value of this might appear questionable.

There are conceptual concerns surrounding all existing models relating to the subsequent projection of BASDAI and BASFI over a longer time horizon which are required in order to generate more appropriate lifetime estimates of costs and QALYs required for cost-effectiveness assessments. The speculative nature of these projections was highlighted as a significant concern by the previous independent assessment group (LRiG) and hence their longer-term results were presented as exploratory scenarios. However, it appears that all existing models are largely based on implicit approaches and assumptions, and lack a clearer conceptual basis which might help to more appropriately inform parameter estimates and structural assumptions, and facilitate a more evidence-based assessment of the potential longer-term impact of anti-TNFs.

The following sections present a summary of the de novo submissions provided by the manufacturers34–37 for the separate AS and nr-AxSpA indications. Brief overviews of the manufacturers’ submissions for AS and nr-AxSpA are provided alongside a summary of the base-case cost-effectiveness results. This is followed by a more in-depth comparison of key parameter and structural assumptions across the manufacturers and the separate indications. The issues and concerns regarding existing published studies are used as the basis for a more critical assessment of these submissions and the extent to which these concerns have been adequately addressed and key uncertainties which still remain have been highlighted is investigated.

It should be noted that although fully incremental results were routinely presented by each manufacturer, there were differences between manufacturers in terms of how the results were presented and also whether or not the correct calculations based on dominance and extended dominance were included. Consequently, the fully incremental ICER tables reported are based on our own calculations to ensure accuracy and consistency between the various manufacturer results tables.

Overview of AbbVie (adalimumab) model

The economic model presented by AbbVie34 compared the cost-effectiveness of adalimumab versus conventional therapy and other licensed anti-TNFs for nr-AxSpA and AS. Separate state-transition models were developed for the two indications separately based on the ASAS guidelines for the use of anti-TNFs. All patients were assumed to take conventional therapy/background therapy (e.g. NSAIDs) during the modelled horizon and also receive one of the licensed anti-TNFs or placebo (conventional therapy only). Specifically, patients were assumed to stay on therapy as long as they had an adequate therapeutic response (i.e. ASAS 40 for nr-AxSpA and ASAS 20 for AS) and patients were assumed to discontinue therapy when insufficient response occurred. Discontinuations due to AEs or reasons other than therapeutic failures were also included.

The model consists of a short-term component (first 12 weeks) and a longer-term component to estimate lifetime cost-effectiveness (40 years). In common with previously published models, the model was based on the estimation of BASDAI and BASFI scores over time. The model used the available long-term open-label extension data of trials of adalimumab (up to 156 weeks in ABILITY-158 for nr-AxSpA and 260 weeks in ATLAS61 for AS, Figures 4 and 5) as well as including assumptions beyond these study durations to inform the life-time cost-effectiveness results. To avoid extrapolating life-time improvement by applying a functional form to the BASDAI/BASFI data, the manufacturer applied the mean observed BASDAI and BASFI scores until the last available data point and carried forward the last observed values to the end of horizon.

FIGURE 4.

Observed mean BASDAI and BASFI scores for adalimumab ASAS 20 responders in the licensed population from ATLAS61 (AS).

FIGURE 5.

Observed mean BASDAI and BASFI scores for adalimumab ASAS 40 responders in the licensed population from ABILITY-158 (nr-AxSpA).

Response rates and other select treatment efficacy end points were based on a separate systematic review and network meta-analysis. In the base case, ASAS 40 for nr-AxSpA and ASAS 20 for AS were used to define clinical response at week 12, based on the primary outcome measures from the clinical trials of adalimumab. In the base-case analysis, placebo responders at week 12 were assumed to lose response and return to baseline disease severity. Patients who subsequently withdrew from TNF-α treatment at any time point were also assumed to return to baseline disease severity (rebound equal to gain). Longer-term discontinuation was assumed to be time-dependent and was based on a log-normal parametric distribution from the separate open-label RCTs adjusting for subsequent loss of response.

In the base-case model, the BASFI score for all patients not on TNF-α inhibitor treatment increases in a linear fashion by 0.084 (scale 0–10) per year in patients with nr-AxSpA, in line with the evidence from the ABILITY-1 trial,58 in which each additional year of baseline symptom duration was reported to be associated with a significant (+0.084; p = 0.0005) increase in baseline BASFI score, adjusting for the age of onset (age at first reported axial SpA symptom) to control for the age effect on functional damage. An estimate of +0.056 was applied to patients with AS based on applying a similar approach to the ATLAS trial,61 adjusting for age at disease diagnosis. Hence, a higher BASFI progression was applied to patients not on anti-TNFs in the nr-AxSpA population compared with the AS population.

The BASDAI and BASFI scores were used jointly to estimate quality of life associated with AS, using the relationship observed between the utility scores (measured in HUI-3) and the BASDAI and BASFI scores in the ATLAS trial. 61 Observed EQ-5D scores were mapped to BASDAI and BASFI for the relationship in the base case for nr-AxSpA from ABILITY-1. 58

The relationship between BASDAI and costs, derived from a reanalysis of the OASIS data, was applied in the base case. Costs of drug, administration, initiation and monitoring, and AEs were also included. Discounting was applied at 3.5% for both costs and outcomes. SMRs of 1 and 1.5 were assumed for nr-AxSpA and AS, respectively. Uncertainty surrounding results was addressed using probabilistic sensitivity analyses (PSAs).

Base-case results from AbbVie (adalimumab) model

The main base-case ICER results from the manufacturer are summarised in Table 36 for the AS population. From a NHS perspective, the base-case cost per QALY gained versus CC ranged from £16,391 per QALY (adalimumab) and £44,448 per QALY (infliximab).

Table 37 reports the results based on the fully incremental analysis. In the manufacturer base-case analysis, certolizumab and etanercept were ruled out by extended dominance. The ICER of adalimumab was £16,391 per QALY compared with CC. The ICER of the next more costly (and non-dominated) TNF-α inhibitor was £238,500 per QALY for the comparison between infliximab and adalimumab.

The main base-case ICER results from the manufacturer and fully incremental analysis are summarised in Tables 38 and 39 for the nr-AxSpA population. The ICERs versus CC ranged from £12,866 (certolizumab) to £13,288 per QALY (adalimumab). In the fully incremental comparison, adalimumab was extendedly dominated and hence the ICER for certolizumab versus CC is the only ICER reported (£12,866).

The manufacturer reported more favourable ICERs versus CC in the nr-AxSpA population compared with the AS population. This appears largely driven by two inputs: (1) the lower BASDAI/BASFI scores assumed for responders based on ABILITY-158 (compared with ATLAS61) and (2) the higher annual BASFI progression rate assumed for non-responders/CC in the nr-AxSpA population (0.084 vs. 0.056).

Overview of UCB (certolizumab) model

The economic model presented by UCB34 compared the cost-effectiveness of certolizumab with conventional therapy and other licensed anti-TNFs for nr-AxSpA and AS. Separate Markov cohort models were developed for the two indications separately based on the subpopulations of the RAPID-axSpA trial. 64 Separate analyses were argued to be necessary given that the comparators differed for each subpopulation. Analyses performed for the AS subpopulation consisted of all patients with AS from the RAPID-axSpA study,64 including those who were anti-TNF therapy experienced or therapy naive. The nr-AxSpA subpopulation consisted of anti-TNF therapy-naive patients only, as there were no anti-TNF therapy-experienced patients in this subpopulation.

The analyses used a lifetime time horizon in the base case. An alternative time horizon of 20 years was tested in a scenario analysis. A NHS and PSS perspective was used and an annual discount rate of 3.5% was applied to costs and outcomes. All costs are reported at 2013 values.

The model consists of a short-term component and a longer-term component to estimate lifetime cost-effectiveness. The duration of the short-term component varied between the models used for the AS and the nr-AxSpA subpopulations based on the response end point assumed. Response was assessed at 24 weeks in the AS subpopulation which was argued by the manufacturer to be in accordance with clinical practice as indicated key British opinion leaders. For the nr-AxSpA subpopulation, response assessment was assumed at 12 weeks, as comparator data were available only at that time point. In their base case, the manufacturer used ASAS 20 to determine response in line with the primary outcome measure in the RAPID-axSpA. 64 However, it should be noted that ASAS 20 response at week 12 was the primary outcome in the RAPID-axSpA trial. Hence, although the measure of response used is in accordance with the primary outcome of the RAPID-axSpA trial, the differential timing of this applied across the separate populations clearly deviates from this. This has potential issues because at week 16 patients were allowed an ‘early escape’ from placebo and hence results at week 24 used for the AS subpopulation are no longer based on the original randomised population.

Assessment in Ankylosing Spondylitis 20 response rates for certolizumab and relative treatment effects for the other anti-TNFs were derived based on a separate systematic review and MTC. The base-case model inputs applied in the manufacturer’s submission34 are replicated (and associated footnotes) in Tables 40 and 41.

The MTC was also used to determine change in baseline BASFI and BASDAI scores. The base-case inputs for change from baseline in BASFI and BASDAI at week 24 for the AS subpopulation reported by the manufacturer are replicated in Tables 42 and 43. The manufacturer noted that the mean change from baseline reported in the tables is that observed per trial arm, which includes both the ASAS 20 responders and non-responders in each arm. In order to determine the change in BASFI and BASDAI for responders alone, the manufacturer used the equation:

This approach assumed that the change in BASFI (and BASDAI) score among ASAS 20 non-responders is equal to that of the CC arm. Thus, the equation is used to algebraically solve for change in BASFI (and BASDAI) score among ASAS 20 responders. The manufacturer stated that:

As an example for the AS subpopulation base case, the change in BASFI among ASAS 20 responders for CZP [certolizumab pegol] is: [academic-in-confidence (AiC) information removed]. Thus, in this example, the actual change from baseline in AS responders to CZP is [AiC information removed]. The same approach was used for change from baseline for BASDAI. This approach, where the change from baseline for BASDAI and BASFI is calculated among responders only, is consistent with previous evaluations pharmacoeconomic evaluations conducted for AS

Manufacturer’s submission, pp. 69–7034

The manufacturer base-case inputs for change from baseline in BASFI and BASDAI at week 12 for the nr-AxSpA subpopulation are replicated in Tables 44 and 45.

The manufacturer’s submission34 assumed no change in BASDAI and BASFI for CC during the response period. The manufacturer justified this assumption with reference to evidence from RAPID-axSpA,64 ATLAS61 and ABILITY-158 studies, although no specific data were reported to support this.

These change scores are assumed to be maintained for BASDAI as long as a patient continues to receive an anti-TNF. For AS patients on CC, an additional annual increase of 0.07 points (scale 0–10) in BASFI is assumed and justified by the manufacturer according to the assumptions deemed reasonable by a previous NICE committee. Hence, while the change scores are assumed constant, the absolute difference between patients receiving anti-TNFs and CC is increasing over time given the underlying progression assumed for BASFI for patients receiving CC. The assumption of no progression in BASFI for patients receiving anti-TNFs is not explicitly discussed within the manufacturer’s submission, neither are separate results provided for alternative assumptions.

The same annual rate (0.07) in BASFI progression for CC is also applied to the nr-AxSpA subpopulation. In addition, it is assumed that some nr-AxSpA patients may progress to AS during their course of treatment. The manufacturer’s model adopts an estimate for disease progression for the nr-AxSpA subpopulation based on a German cohort of axSpA patients, GESPIC. In this cohort, the rates and predictors of radiographic spinal progression over 2 years were estimated based on mSASSS. In total, 7.4% of the 95 nr-AxSpA patients were reported to show spinal radiographic progression, which was defined as a worsening of mSASSS by ≥ 2 units over 2 years. As this 7.4% progression represents a proportion, it was converted to a rate for use in the economic model, assuming an exponential distribution through the following formula:

The manufacturer’s submission34 is not explicit about how this additional aspect of progression subsequently alters the BASDAI/BASFI trajectories within the nr-AxSpA model. However, examination of the electronic model submitted by the manufacturer reveals that once patients are assumed to show spinal radiographic progression, they effectively become AS patients by picking up the same absolute values of BASDAI and BASFI (on and off treatment) applied in their AS subpopulation model. The justification for this approach and the values subsequently assigned are not formally discussed by the manufacturer and the validity of the approach appears questionable (i.e. given other differences, e.g. disease duration, severity of radiographic disease etc., that may differ between the two populations even after radiographic progression has occurred in the nr-AxSpA subpopulation).

Patients who subsequently withdrew from TNF-α treatment at any time point were assumed to revert back to the same trajectory as CC over a 6-month period (i.e. rebound back to CC/natural history). A constant annual rate of discontinuation of 7% was assumed for all anti-TNFs over the longer-term period in both the AS and nr-AxSpA populations. The estimate of 7% applied to the AS subpopulation was justified by citing the rate apparently assumed by the NICE committee for TA143 and the lack of long-term evidence more generally. This estimate was referred to earlier in the review section of our report when the additional analyses undertaken by the NICE DSU were considered (see Data extraction). Identical assumptions for discontinuation rates were assumed for the nr-AxSpA subpopulations, although no justification was provided by the manufacturer.

The BASDAI and BASFI scores were used jointly to estimate quality of life in both subpopulations based on EQ-5D data collected in the RAPID-axSpA64 study. Data from patients having EQ-5D, BASDAI and BASFI scores available at baseline and at weeks 12 and 24 were used to estimate a relationship between utility and the BASDAI and BASFI scores. Utilities were subsequently converted using a logistic transformation with the justification based on possible floor and ceiling effects, as they are bounded by 0 and 1. Without access to the original data, it is not possible to determine the impact of this transformation, although it should be noted that EQ-5D is not bounded by 0 (i.e. negative values are possible). The manufacturer used a repeated-measures logistic regression to model the relationship between utility and the BASDAI and BASFI scores.

The relationship between BASFI and costs, derived from the OASIS study and used by the previous independent assessment group in TA143,17 was applied in the base case. Costs of drug, administration, initiation and monitoring were included. The costs and HRQoL of AEs were not included. Discounting was applied at 3.5% for both costs and outcomes. Uncertainty surrounding outcomes was addressed using PSA.

Base-case results from UCB (certolizumab) model

The main base-case ICER results from the manufacturer35 are summarised in Table 46 for the AS population, together with a fully incremental comparison of ICERs in Table 47. The ICERs versus CC ranged from £16,647 per QALY (certolizumab) and £42,671 per QALY (infliximab). In the fully incremental analysis, certolizumab dominated (i.e. less costly and more expensive) all other TNF-α treatments apart from infliximab. However, it should be noted that the costs of certolizumab are based on a patient access scheme (PAS) which has been proposed but is not yet formally agreed with the Department of Health and NICE. Results without the PAS were not reported by the manufacturer. UCB will make certolizumab pegol (Cimzia,^® USB Pharma) available free of charge to all NHS patients for the first 3 months of therapy, at which point clinical response should be clear. Only after this 3-month stage will the NHS be charged for continuing to use this therapy.

The ICER of certolizumab was £16,647 per QALY versus CC and the ICER for infliximab was £113,871 (vs. certolizumab).

The main base-case ICER results from the manufacturer35 are summarised in Table 48 for the nr-AxSpA population, together with a fully-incremental comparison of ICERs in Table 49. In contrast to the results for AS, there was a more marked difference between the ICERs of the alternative anti-TNFs and CC. The ICERs versus CC ranged from £15,615 (certolizumab) to £50,692 per QALY (etanercept). The higher differential ICERs appears to be largely a result of the more heterogeneous trials included in the MTC for the nr-AxSpA populations and a higher differential effect assumed for certolizumab vis-à-vis the other alternative anti-TNFs compared with the AS population. Importantly, other manufacturers (Pfizer) argue that the results for certolizumab in this population may be confounded by population characteristics which could invalidate the indirect comparison of certolizumab versus the other comparator treatments in the current nr-AxSpA MTC. In the fully incremental analysis, certolizumab dominated adalimumab and etanercept.

Overview of Pfizer (etanercept) model

The economic model submitted by Pfizer36 compared the cost-effectiveness of etanercept versus conventional therapy and other licensed anti-TNFs for AS, nr-AxSpA and a combined population (axSpA). The results for the combined population are not summarised in this review but are reported separately in the manufacturer’s submission. The model is based on a lifetime time horizon and costs and benefits are discounted at an annual rate of 3.5%. The reference year for costs was reported to be 2014.

The model was based on a patient-level simulation model based on a discrete event simulation (DES). The analysis was conducted from a NHS/PSS perspective. Data to populate the model were derived from key clinical trials for etanercept, the results of a clinical systematic review, MTC and, in a separate analysis presented for the nr-AxSpA population, a match-adjusted indirect comparison (MAIC). The model structure was reported to be developed in accordance with current OMERACT (Outcome Measures in Rheumatology) guidance and was constructed around the BASDAI and the BASFI in line with other published studies.

The AS population was defined based on current NICE guidance in TA14317 and TA233. 33 The nr-AxSpA population was defined based on the scope issued by NICE and was defined by the manufacturer as people with severe axSpA without radiographic evidence of AS but with objective signs of inflammation, whose disease has responded inadequately to, or who are intolerant to, NSAIDs.

An important aspect of the submission for the nr-AxSpA population was an attempt to adjust analyses for differences in the baseline patient characteristics between the trials included. The manufacturer reported that:

The clinical systematic review identified that the baseline characteristics of nr-AxSpA patients within the randomised controlled trials of certolizumab pegol and adalimumab were heterogeneous, and potentially differed in characteristics that could act as treatment effect modifiers. Furthermore, the populations of these trials also included sizable proportions of AS patients who were originally classified as nr-AxSpA on the basis of a difference between centralised and localised readings of X-rays.

To address the differences in the proportions of AS patents in the trials due to reclassification upon central assessment, analyses were conducted using match adjusted indirect comparison (MAIC) and simulated treatment comparison (STC) techniques that incorporated AS patients from the etanercept 314-EU trial. These analyses are referred to collectively as ‘analyses adjusting for differences in study baseline characteristics’. A comparison of the results from the MAIC and STC approaches show that while the results of the two analyses are similar, when considering comparisons between etanercept and both adalimumab and certolizumab, the MAIC analysis provides a lower overall comparative estimate of the benefit of etanercept, and is therefore considered overall to be the more conservative of the two approaches. To maintain consistency in the analysis utilised in the economic section, the MAIC was used throughout as the adjusted comparative efficacy measure between etanercept versus adalimumab and etanercept versus certolizumab. For the analysis comparing etanercept against certolizumab pegol, it was possible to address the issue of patient reclassification and differences in baseline characteristics by utilising the RAPID-axSpA trial results that were also available at the level of AxSpA patients, an approach not possible in the comparison of etanercept versus adalimumab. We note that although not explicitly detailed within the scope, the AxSpA population encompasses both nr-AxSpA and AS patients, thus making it a relevant comparison to the decision problem outlined in the scope.

Manufacturer’s submission, pp. 226–736

The manufacturer argued that the use of DES conferred potential advantages in relation to modelling non-linearity because of heterogeneous patient characteristics and in relation to modelling time dependency. The latter was also argued as an advantage to considering the impact of sequential therapy which was argued to be complex within a more conventional Markov type structure. Pfizer’s model was the only model which explicitly explored issues of treatment sequences. However, in the base case the use of second-line TNF-α inhibitor treatment was restricted to those patients who withdrew because of AEs and was assumed equal efficacy to first line usage. The schematic of the model provided by the manufacturer is replicated in Figure 6.

FIGURE 6.

Pfizer DES model schematic. NMA, network meta-analysis. 36

Etanercept RCT data were used to predict an initial 12-week response (in terms of reduction in BASDAI and BASFI) for etanercept for both nr-AxSpA and AS populations. Separate multivariate regressions were used to account for correlation between BASDAI and BASFI. A range of variables were initially included in the regression models based on potential predictors of response identified from their review of economic studies. The statistical significance and direction of effect were evaluated before final models were specified. The 12-week models of BASDAI and BASFI for the nr-AxSpA had R² values of (AiC information has been removed) respectively. For the AS population, the equivalent R² values were (AiC information has been removed). The regressions were used to estimate mean change in BASDAI and BASFI which through the patient level simulation were used to assign patients into BASDAI 50 responder/non-responder categories and to assess the associated magnitude of change at 12 weeks for these categories.

Relative effects from the MTC (or MAIC in the analyses adjusting for differences in study baseline characteristics), in terms of mean differences in BASDAI and BASFI, were applied in order to predict equivalent response and change scores for the other anti-TNF agents and CC at 12 weeks.

From week 12, the BASFI scores for CC were assumed to increase at a rate of 0.7 units per annum (0–100 scale). The modelling of change in BASDAI and BASFI at week 48 for responders to etanercept was conducted using the same approach used for the week 12 treatment response. However, change in BASDAI and BASFI from baseline at week 12 were included as additional covariates within the resulting models in order to ensure that an individual’s response at week 48 was dependent on their response at week 12. The 48-week models of BASDAI and BASFI for the nr-AxSpA had R² values of (AiC information has been removed), respectively. For the AS population, the equivalent R² values were (AiC information has been removed). In the absence of relative effect estimates at week 48 for other therapies, it was assumed that patients who remained on TNF-α inhibitor treatment beyond week 12 (i.e. responders) would converge at the BASDAI and BASFI levels predicted for etanercept by week 48. Constant BASDAI and BASFI scores for TNF-α inhibitor responders were assumed at the level observed at week 48 for subsequent periods.

Treatment discontinuation was modelled by fitting separate parametric survival curves to long-term open-label study data from etanercept for the AS and nr-AxSpA populations. In order to predict treatment cessation in the population that was likely to continue treatment after 12 weeks, parametric curves were fitted only to subjects who achieved a BASDAI 50 response at week 12. Only patients who were randomised to etanercept at baseline were retained within these survival analyses and patients who began etanercept during open-label phases of studies were excluded. The distributions that provided the best fit were exponential (AiC information has been removed) and log-normal (AiC information has been removed), based on the minimisation of the Akaike information criterion and the Bayesian information criterion. The exponential model was chosen based, in part, on the goodness of fit but also because the use of hazard ratios, which were applied to estimate the effect of other anti-TNFs on the rate of discontinuation, required the use of a proportional hazard survival model [to avoid making further assumptions when applying the hazard ratio to the log-normal (accelerated failure time) model].

The same risk of discontinuation was applied to all individuals in the model. The models of discontinuation translate into annual probabilities of discontinuation for etanercept, for patients who achieve a BASDAI 50 response, of 5% and 11% for nr-AxSpA and AS populations, respectively. Based on data from the DANBIO registry, it was assumed that other anti-TNFs have an increased risk of discontinuation compared with etanercept: a hazard ratio of 1.3 is applied for infliximab and 1.12 for adalimumab. In the absence of evidence for golimumab and certolizumab, it was assumed that the relative effect is the same as for adalimumab on the basis that these have common molecular structure and belong to monoclonal antibodies.

After discontinuation of the first treatment, an alternative TNF-α inhibitor was modelled as second-line treatment for patients who discontinued due to AEs [(CiC information has been removed)% for AS and (CiC information has been removed)% for nr-AxSpA]. The same efficacy as applied for first-line treatments was assumed for second-line treatments for patients switching because of AEs. For patients who discontinued because of loss of efficacy, no further TNF-α treatment was modelled. These assumptions were considered by the manufacturer to be consistent with current NICE guidance. For the base-case model, it was assumed that following discontinuation from anti-TNFs, patients would rebound back to their baseline BASDAI and BASFI scores and that the rebound takes 6 months based on the approach used within the TA23333 submission to NICE.

In the absence of previously published studies reporting the relationship between BASDAI/BASFI and EQ-5D utility scores in the nr-AxSpA population, a de novo relationship was estimated from the 1031 study;166 variables included age, sex, baseline BASDAI and BASFI. Ordinary least squares regression models were used, with SEs clustered around each subject to account for repeated observations. For consistency, a similar relationship was estimated for the AS population using the 314-EU study. 167 Alternative linear and non-linear relationships were evaluated and final model selection based on Akaike information criterion statistics. In the nr-AxSpA population, the final model included squared terms for BASDAI and BASFI and an interaction between BASDAI and BASFI, while in the AS population, the covariates for the interaction term, age and male were not included. Scenario analyses considered using alternative model specifications for mapping. The manufacturer reported that according to visual inspection, the estimated models were very similar between populations and reported a high degree of similarity between the results of the de novo estimated models and those published previously.

Figures 7 and 8 replicate the relationships reported by the manufacturer between EQ-5D, BASDAI and BASFI in the nr-AxSpA and AS populations, respectively. Additional figures were also presented by the manufacturer for predicted versus observed EQ-5D in each of the populations. The manufacturer concluded that the models overpredicted EQ-5D at low observed EQ-5D and underpredicted at higher observed EQ-5D values. The manufacturer argued that this was a common feature of mapping algorithms and argued that the approach would be conservative towards the use of anti-TNFs.

The manufacturer included the acquisition, administration and pre-treatment monitoring costs of TNF-α inhibitors. Subsequent monitoring costs were not included in order to avoid potential double counting of the costs which were estimated as a function of BASDAI and BASFI. In the base-case analysis the manufacturer used data from Rafia et al. 168 based on BASDAI scores only. A categorical approach was applied to BASDAI scores based on the following annual costs: BASDAI score of < 40 = £151.96; 40 ≤ BASDAI score < 60 = £311.08; and BASDAI score of ≥ 60 = £1039.16. The manufacturer justified the use of this source as it provides the most recent UK specific data reported and permitted separation of particular cost items. The costs and HRQoL of AEs (serious infections only) were included in the base-case analysis (none observed in the nr-AxSpA trial 1031 study;166 however, serious infections were observed in the AS trial 314-EU study167). A separate sensitivity analysis included the costs of serious infections.

A SMR of 1 for the nr-AxSpA population and 1.5 for the AS population were applied to general population life-tables.

Results of Pfizer (etanercept) model

The main base-case ICER results from the manufacturer36 are summarised in Table 50 for the AS population, together with a fully incremental comparison of ICERs in Table 51. The ICERs versus CC ranged from £19,586 per QALY (certolizumab) and £37,741 per QALY (infliximab). In common with the UCB model, it should be noted that the costs of certolizumab assumed within Pfizer’s model were also based on the PAS for certolizumab which has been proposed but is not yet formally agreed with the Department of Health and NICE. Hence the ICER for certolizumab versus CC without the PAS will be higher than the estimates reported here.

In the fully incremental analysis, adalimumab was extendedly dominated. Of the remaining non-dominated treatments, the ICERs of the next most costly interventions compared with the previous non-dominated alternative were £19,586 (certolizumab vs. CC), £28,834 (etanercept vs. certolizumab), £30,376 (golimumab vs. etanercept) and £1,131,181 (infliximab vs. golimumab).

The main base-case ICER results from the manufacturer36 are summarised in Table 52 for the nr-AxSpA population together with a fully incremental comparison of ICERs (Table 53). The ICERs versus CC ranged from £23,195 (etanercept) and £23,575 (certolizumab). In contrast to the UCB analysis, the ICERs for the nr-AxSpA population were marginally less favourable than the results for the AS population. There was also less of a marked difference between the ICERs for each of the anti-TNFs and CC compared with the UCB results, although a large difference was evident relating to the magnitude of the incremental QALY estimates for certolizumab vis-à-vis the other anti-TNFs. Table 54 reports the results of the fully incremental analysis. None of the anti-TNFs was ruled out via dominance or extended dominance, and the ICER of each comparison remained below £30,000 per QALY for each successively more expensive and effective treatment.

To address the concerns noted by Pfizer relating to the heterogeneity across the different trials in the nr-AxSpA population, a separate matched indirect comparison was presented for etanercept versus adalimumab. A separate comparison was also presented versus certolizumab for the combined axSpA population in the manufacturer’s submission. Using the MAIC approach, the ICER for etanercept versus adalimumab was £23,195 per QALY. Total cost and QALYs estimates were reversed in the MAIC approach compared with the base-case analysis (adalimumab generated greater QALYs at increased cost), demonstrating the potential impact of trying to minimise observable sources of possible confounding.

Overview of Merck Sharp & Dohme (golimumab, infliximab) model

The economic models submitted by Merck Sharp & Dohme37 compared the cost-effectiveness of golimumab and infliximab with conventional therapy and other licensed anti-TNFs for AS. Although the manufacturer made separate submissions for golimumab and infliximab, the model structures and data sources used to inform the economic models are identical across the submissions. Hence this review focuses on the specific submission for golimumab but also considers key data sources and assumptions specific to infliximab. The model base case is based on a lifetime time horizon (approximately 60 years), and costs and benefits are discounted at an annual rate of 3.5%. A NHS and PSS perspective is used for costs. The reference year for costs was reported to be 2012/13.

The economic model submitted by the manufacturer for golimumab37 is based on the same model structure submitted as part of NICE TA23333 and summarised previously in Data extraction (Armstrong et al. 2013)163 Hence, a description of the structure of the model is not repeated in this section. In summary, the manufacturer’s cost-effectiveness model was based on a short-term decision tree (based on an assessment of BASDAI 50 response at 12 weeks in the base case) and a longer-term Markov model.

The proportion of patients achieving BASDAI 50 at week 12 (± 2 weeks) for each TNF-α inhibitor was obtained from a systematic review and MTC undertaken by the manufacturer. The results are summarised in Table 55.

Data from the GO-RAISE90 trial and the open-label extension period (up to week 108) were used to develop predictive equations of mean change from baseline in BASDAI and BASFI scores over time. Two separate equations were developed based on the 24-week data (0–24 weeks) for all patients and post-24-week (week 24–108) data from GO-RAISE90 for patients who remained on treatment. The variables applied in each equation are summarised in Tables 56 and 57.

The treatment coefficient (and interaction term) in the short-term regression equation is used to estimate separate BASDAI/BASFI scores for anti-TNFs and CC. Hence, up to week 24, the same estimate of BASDAI/BASFI appears to be applied to all TNFs (i.e. regardless of the differential response rates assumed). Beyond week 24, the same BASDAI/BASFI score is applied to a responder to any of the TNFs, although a different response rate for each TNF-α inhibitor is assumed based on the MTC. The BASDAI/BASFI regressions are applied to responders who continue on TNF-α inhibitor therapy up to week 108 for BASDAI and up to week 108 for BASFI.

The BASDAI and BASFI scores beyond week 108 for responders who continue to receive anti-TNFs beyond this period in the model are assumed to remain constant (at the week 108 value). The BASFI scores beyond week 256 for responders who continue to receive anti-TNFs beyond this period in the model are assumed to remain constant (at the week 108 value) but are also subject to an annual progression rate of BASFI at this point which is set to half the rate of CC in the baseline (0.035 units per annum, 0–10 scale). The justification for this is not explicitly made by the manufacturer. For the base-case model, BASFI scores for CC patients on conventional therapy are assumed to progress at a rate of 0.07 units per year after week 24.

An annual discontinuation rate of 6.1% is applied for the entire time horizon after week 12 in the base-case analysis. This estimate is derived from data reported between weeks 24 and 256 in the 50 mg arm of golimumab from the GO-RAISE90 extension period. The manufacturer does not formally state whether or not this is specific to those patients who were identified to be responders at 12 weeks. However, it appears to be based on all patients who continued to receive golimumab beyond 24 weeks regardless of their response status. The same discontinuation rate is applied to all TNF-α inhibitors. Following discontinuation from anti-TNFs, the BASFI and BASDAI scores are assumed to deteriorate/rebound over a 24-week period back to their baseline BASFI and BASDAI scores (i.e. rebound equal to gain). Therefore, in common with other models which apply this rebound assumption, patients are assumed to achieve a lifetime benefit from treatment with anti-TNFs for BASFI.

Utilities were derived from a NICE TA (TA14317) and incorporated age, sex, BASFI score and BASDAI score. Costs included in the model comprised drug acquisition, short-term (12-week) costs, longer-term disease costs and AEs. Longer-term disease costs were based on BASFI scores from the GO-RAISE90 trial using the same regression equation used for NICE TA143. 17

The proportion of males and females recruited in the GO-RAISE90 trial is used to estimate a weighted average mortality risk by sex. The sex-specific SMR for AS from a study by Bakland et al. 16 is applied to the mortality rates from the general population to calculate adjusted mortality rates for AS patients in the model. The study by Bakland reported a SMR of 1.63 (95% CI 1.29 to 1.97) for males and 1.38 (95% CI 0.48 to 2.28) for females.

Results of the Merck Sharp & Dohme (golimumab, infliximab) model

The main base-case ICER results from the manufacturer37 are summarised in Table 58 for the AS population, together with a fully incremental comparison of ICERs in Table 59. The ICERs versus CC ranged from £19,070 (golimumab) to £42,532 (infliximab). In the fully incremental analysis, golimumab and certolizumab were the non-dominated anti-TNFs. The ICER for golimumab versus CC was £19,070 and for certolizumab versus golimumab was £21,441 per QALY.

Summary and critique of de novo cost-effectiveness submissions

In general, the manufacturer models34–37 appeared to be constructed to a high standard and it is evident that significant work had been undertaken by each to identify and use previously published studies and to exploit existing individual patient data from their own RCTs and open-label extension periods to generate estimates that were appropriate for the requirements of the model.

Despite the different model structures and assumptions applied across the various manufacturer’s submissions,34–37 the ICERs reported for the anti-TNFs versus CC were remarkably consistent in the AS population. Table 60 presents a summary of the ICER reported by each manufacturer for each of the anti-TNFs versus CC. It is perhaps expected that the majority of manufacturer’s reported the lowest ICER versus CC for their own products. The only exception to this is the Pfizer model, which estimated the lowest ICER versus CC for certolizumab in this population (etanercept was the next lowest), although it should be noted that Pfizer included the proposed PAS costs for certolizumab which was not universally applied across the different manufacturer’s submissions. Hence, although differences between the ICER versus CC were quite similar, the variation in approaches used by each manufacturer appears partially driven by maximising any potential comparative advantage considered vis-à-vis other manufacturer products (i.e. in terms of assumptions made about similarities and differences for response rates, magnitude of changes in BASDAI and BASFI and withdrawal rates). However, it should be noted that no manufacturer makes a strong claim regarding differential efficacy between the alternative anti-TNFs which is borne out in the relatively small differentials reported between the different products in each of the submissions.

TABLE 60 - Comparison of manufacturer ICER estimates vs. CC (AS population)

The figures reported in bold indicate which specific TNF-α inhibitor has the lowest ICER vs. CC in each of the manufacturer’s submissions.

Technology	AbbVie34 (adalimumab), ICER (£)	UCB35 (certolizumab), ICER (£)	Pfizer36 (etanercept), ICER (£)	Merck Sharp & Dohme37 (golimumab, infliximab), ICER (£)
CC	–	–	–	–
Adalimumab	16,391	19,932	20,909	19,275
Certolizumab	17,067	16,647	19,586	19,401
Etanercept	16,897	19,272	20,938	21,972
Golimumab	16,535	19,049	21,288	19,070
Infliximab	44,448	42,671	37,741	42,532

Table 61 presents a summary of the ICERs reported by each manufacturer for each of the anti-TNFs versus CC for the nr-AxSpA population. There appears to be much more heterogeneity across the manufacturer’s submissions34–36 compared with the AS population. There appears an almost twofold difference in the ICERs reported across the submissions for each of the anti-TNFs. Importantly, there also appears variation across the populations with more favourable ICERs reported than CC for the nr-AxSpA population vis-à-vis the estimates by AbbVie (both adalimumab and certolizumab) and UCB (certolizumab only). Hence, the differences in structural and parameter assumptions appear more evident in the results for the nr-AxSpA population compared with results for the AS population.

TABLE 61 - Comparison of manufacturer ICER estimates vs. CC (nr-AxSpA population)

The figures reported in bold indicate which specific TNF-α inhibitor has the lowest ICER vs. CC in each of the manufacturer’s submissions.

Technology	AbbVie34 (adalimumab), ICER (£)	UCB35 (certolizumab), ICER (£)	Pfizer36 (etanercept), ICER (£)
CC	–	–	–
Adalimumab	13,228	30,370	23,242
Certolizumab	12,866	15,615	23,575
Etanercept	Not assessed	50,692	23,195

To assist in identifying possible reasons for the differences between populations, summaries of the key structural assumptions used by each manufacturer are provided in Tables 62 and 63. A more micro-level analysis of comparison of specific parameter estimates is reported separately in Appendix 11.

TABLE 62 - Model structure and key structural assumptions: AS population

Parameter	Merck Sharp & Dohme economic model37 (infliximab, golimumab)	AbbVie34 economic model (adalimumab)	UCB35 economic model (certolizumab)	Pfizer36 economic model (etanercept)
Model type	Decision tree followed by Markov model	Markov model	Markov model	Patient-level simulation model (DES)
Time horizon	Lifetime	40 years	Lifetime	Lifetime
Response criterion	BASDAI 50 response at week 12	ASAS 20 response at week 12	ASAS 20 response at week 24	BASDAI 50 response at week 12
Response criterion justification	Efficacy outcome in GO-RAISE90 study; recommended by the ASAS Working Group (Keat 2005)169	Primary end point of ATLAS61 study	ASAS 20 is the primary end point of RAPID-axSpA64 study	Based on the current NICE definition of treatment response (TA14317)
Progression assumption BASDAI
Anti-TNFs responders	Constant after week 108	Constant after week 260	Constant after week 24	Constant after week 48
Anti-TNFs non-responders	Constant	Constant	Constant	Constant
CC	Constant after week 24	Constant	Constant	Constant after week 12
Progression assumption BASFI
Anti-TNFs responders	Constant after week 108; 0.035 after week 256	Constant after week 260	Constant after week 24	Constant after week 48
Anti-TNFs non-responders	0.07	0.056	0.07	0.07
CC	0.07 after week 24	0.056	0.07	0.07 after week 12
Rebound assumption	Rebound to baseline	Rebound to baseline	Rebound to conventional therapy	Rebound to baseline
Rebound assumption duration	Over a 6-month period	Immediately	Over a 6-month period	Over a 6-month period
Placebo response	14.5% at week 12; loss or maintenance of placebo response not reported	BASDAI and BASFI scores return to baseline at week 12	No placebo response	BASDAI and BASFI scores return to baseline at 12 weeks

TABLE 63 - Model structure and key structural assumptions: nr-AxSpA population

Parameter	AbbVie34 economic model (adalimumab)	UCB35 economic model (certolizumab)	Pfizer36 economic model (etanercept)
Model type	Markov model	Markov model	Patient-level simulation model (DES)
Time horizon	40 years	Lifetime	Lifetime
Response criterion	ASAS 40 response at week 12	ASAS 20 response at week 12	BASDAI 50 response at week 12
Response criterion justification	Primary end point of ABILITY-158 study	Primary end point of RAPID-axSpA64 study	Based on the current NICE definition of treatment response (TA14317)
Progression assumption BASDAI
Anti-TNFs responders	Constant after week 140	Constant after week 12	Constant after week 48
Anti-TNFs non-responders	Constant	Constant	Constant
CC	Constant	Constant	Constant after week 12
Progression assumption BASFI
Anti-TNFs responders	Constant after week 140	Constant after week 12	Constant after week 48
Anti-TNFs non-responders	0.084	0.07	Constant/0.07
CC	0.084	0.07	0.07 after week 12
Rebound assumption	Rebound to baseline	Rebound to conventional therapy	Rebound to baseline
Rebound assumption duration	Immediately	Over a 6-month period	Over a 6-month period
Placebo response	BASDAI and BASFI scores return to baseline at week 12	No placebo response	BASDAI and BASFI scores return to baseline at 12 weeks

In general, it is difficult to identify the specific factors which can easily explain differences within and between the two populations across the manufacturer’s submissions. In general, similar model structures were applied by each manufacturer across the separate populations. However, it is evident that there are important differences based on a number of key structural issues: (1) the response criteria and timings applied; (2) the magnitude of change scores and particularly the assumption concerning the time at which these were assumed to ‘level off’ (generally longer in the AS populations because of the longer open-label extension periods); (3) the underlying rate of progression of BASFI with CC and the impact of anti-TNFs on this rate; and (4) the rebound assumption and timing of this.

Given the complex inter-relationship between these structural assumptions and subsequent parameter estimates, it is difficult to identify single specific reasons for differences. However, the structural differences clearly lead to marked differences in the BASDAI and BASFI scores estimated over time by each manufacturer for each population. Figures 9–11 provide a graphical summary of the cohort BASDAI and BASFI scores, for the AS population, from three of the manufacturers. These highlight the significant differences in subsequent parameter estimates applied at a cohort level. Equivalent estimates are not presented for the Pfizer model because of the complexities of generating this data from the DES model. The BASDAI and BASFI scores are presented here only for the case made by each manufacturer for their own product.

FIGURE 9.

Comparison of cohort BASDAI/BASFI scores for AS population from AbbVie model (adalimumab). 34

FIGURE 10.

Comparison of cohort BASDAI/BASFI scores for AS population from UCB model (certolizumab). 35

FIGURE 11.

Comparison of cohort BASDAI/BASFI scores for AS population from Merck Sharp & Dohme models (golimumab/infliximab). 37 MSD, Merck Sharp & Dohme.

Tables 64 and 65 summarise the mean difference in BASDAI and BASFI scores applied to responders to anti-TNFs and those applied to CC at various time points in each model. The tables clearly highlight the range of different values applied across the separate manufacturers. This further emphasises the variation in approaches, sources and assumptions.

TABLE 64 - The BASDAI score difference for treatment responders vs. CC: AS population

Time	Adalimumab vs. CC	Certolizumab vs. CC	Infliximab/golimumab vs. CC
12 weeks	–2.98	AiC information has been removed	–2.01
24 weeks	–4.42	AiC information has been removed	–2.05
1 year	–4.9	AiC information has been removed	–2.77
3 years	–5.23	AiC information has been removed	–2.83
5 years	–5.31	AiC information has been removed	–2.83
10 years	–5.31	AiC information has been removed	–2.83
20 years	–5.31	AiC information has been removed	–2.83
40 years	–5.31	CiC information has been removed	–2.83

TABLE 65 - The BASFI score difference for responders vs. CC: AS population

Time	Adalimumab vs. CC	Certolizumab vs. CC	Infliximab/golimumab vs. CC
12 weeks	–2.03	AiC information has been removed	–1.68
24 weeks	–3.28	AiC information has been removed	–1.74
1 year	–3.71	AiC information has been removed	–2.49
3 years	–4.25	AiC information has been removed	–2.59
5 years	–4.25	AiC information has been removed	–2.66
10 years	–4.53	AiC information has been removed	–2.85
20 years	–5.09	AiC information has been removed	–3.18
40 years	–6.21	AiC information has been removed	–3.75

The equivalent figures and tables are reported for the nr-AxSpA population (Figures 12 and 13 and Tables 66 and 67).

FIGURE 12.

Comparison of cohort BASDAI/BASFI scores for nr-AxSpA population from AbbVie model (adalimumab). 34

FIGURE 13.

Comparison of cohort BASDAI/BASFI scores for nr-AxSpA population from UCB model (certolizumab). 35

TABLE 66 - The BASDAI score difference for responders vs. CC: nr-AxSpA population

Certolizumab patients who remain nr-AxSpA and do not transition to AS.

Time	Adalimumab vs. CC	Certolizumaba vs. CC
12 weeks	–3.89	AiC information has been removed
24 weeks	–5.54	AiC information has been removed
1 year	–5.42	AiC information has been removed
3 years	–5.99	AiC information has been removed
5 years	–5.99	AiC information has been removed
10 years	–5.99	AiC information has been removed
20 years	–5.99	AiC information has been removed
40 years	–5.99	AiC information has been removed

TABLE 67 - The BASFI score difference for responders vs. CC: nr-AxSpA population

Certolizumab patients who remain nr-AxSpA and do not transition to AS.

Time	Adalimumab vs. CC	Certolizumaba vs. CC
12 weeks	–2.95	AiC information has been removed
24 weeks	–4.11	AiC information has been removed
1 year	–4.12	AiC information has been removed
3 years	–4.55	AiC information has been removed
5 years	–4.72	AiC information has been removed
10 years	–5.14	AiC information has been removed
20 years	–5.98	AiC information has been removed
40 years	–7.66	AiC information has been removed

The differences across manufacturers and between the populations are further illustrated by the summary of key parameter inputs reported in Appendix 11. As well as reporting the main parameter inputs, Appendix 11 also explores differences in approaches at a parameter level for key inputs (e.g. withdrawal, costs, etc.).

It is evident from these comparisons that there are significant differences across the manufacturers in terms of key structural and parameter estimates. While it might appear reassuring that these differences do not appear to lead to significant differences across the ICER estimates reported for the AS population, the greater heterogeneity reported in the ICER estimates for the nr-AxSpA is clearly an issue. However, even within the AS population, any reassurance that one might have regarding the robustness and appropriateness that these estimates have for informing NHS practice needs to be carefully considered in relation to the key conceptual issues and concerns highlighted in Summary and critique of published cost-effectiveness studies for previously published cost-effectiveness studies. Despite significant work undertaken by each manufacturer in support of existing and new indications for their products, it is particularly concerning that many of the key conceptual issues and concerns appear to have not been fully addressed. Indeed, many of these issues seem to have not been addressed at all, such that many models still seem reliant on the use of open-label extension data (and even more so with the extended follow-up reported in the AS population) without any formal consideration of the potential issues with selection that the use of these studies inevitably are subject to. Consequently, the benefits of anti-TNFs are being projected over significant periods of time without any evidence on the counterfactual (i.e. what happens to patients who do not enter into the open-label extension periods? What happens to patients who subsequently withdraw from anti-TNFs? And what would have happened to patients over a longer-time horizon who did not receive anti-TNFs?).

It appears that much of the case being made concerning the cost-effectiveness of the anti-TNFs rests on comparison of single-arm studies (the subject of open-label data) and retrospective comparisons against historical cohorts (as the counterfactual, for patients not on treatment, is unknown). While such a comparison may be necessitated by the short-term nature of the double-blind periods, the lack of a more detailed consideration of the appropriateness of the comparisons being made in relation to sources of natural history data (and subsequent assumptions made concerning the BASDAI/BASFI trajectories of the different patient categories) is concerning and, hence, current ICER estimates reported by the manufacturers must be considered to be both speculative and highly uncertain.

Many of these problems can be associated with whether or not BASDAI and BASFI scores provide an appropriate conceptual basis for modelling the chronic and progressive nature of AS and nr-AxSpA. Hence current models appear largely driven by data availability (i.e. the extensive evidence which has been generated and continues to be generated investigating the relationship between BASDAI/BASFI and costs/utilities) rather than trying to develop a clearer underlying biological or clinical process which may better characterise the disease and subsequent progression across the separate populations.

Until such time that sufficient data-linking costs and utilities to other measures are reported, it seems inevitable that models will continue to be driven largely by BASDAI and BASFI scores over time together with assumptions concerning the longer-term effect of anti-TNFs. However, given the nature of existing models and the reliance on uncontrolled longer-term follow-up of anti-TNFs, and comparison with historical ‘controls’ (particularly in relation to BASFI progression over time and the assumptions being made concerning the potential disease modification properties of anti-TNFs in both AS and nr-AxSpA populations), it is surprising that greater efforts have not been made by the manufacturers to try to more formally link to the increasing evidence base being generated in relation to radiographic progression in the AS population.

It is also surprising that more thought has also not been given to characterising the potential difference in BASFI progression across the separate populations and how generalisable assumptions may be between these. The result is that many of the key assumptions concerning whether the anti-TNFs are primarily symptom control treatments or whether they are also potential disease modifiers remains implicitly dealt with within existing submissions. The result is that several manufacturers use identical assumptions across populations with respect to BASFI progression and the effect of the anti-TNFs. Interestingly, only one manufacturer appears to use differential rates of BASFI progression across the populations (AbbVie), although the same structural assumption concerning the effect of anti-TNFs is still made. Interestingly, this manufacturer applies a higher rate of change in BASFI for patients receiving CC in the nr-AxSpA population vis-à-vis the AS population. However, while such a difference is interesting, the basis of and implication for this differential is not fully explained or justified by the manufacturer.

The issue of intermittent and sequential use of anti-TNFs remain important clinical questions but the existing models do not provide a robust basis for informing these decisions. The cost-effectiveness of intermittent therapy versus continuous therapy was not formally considered in any model identified. However, it could be argued that such a comparison might be deemed outside the scope of a NICE appraisal. Although one manufacturer (Pfizer36) explored the potential cost-effectiveness of sequential therapy, much of this has been done via assumptions (e.g. assuming equal efficacy second line in patients who discontinue first line as a result of an AE) or via adjustments applied to first-line efficacy estimates based on ‘real-world’ evidence reported from large-scale registries (which typically show anti-TNFs to be clinically effective but with lower response rates than reported in naive patients). Consequently, existing attempts to model sequential therapy are largely based on applying adjustments to first-line efficacy data using observational evidence which are clearly subject to potential confounding. In large part, the limitations of existing cost-effectiveness models for informing these clinical questions appears less a function of the models themselves but rather that robust clinical data to date has not been generated to inform unbiased estimates of relative efficacy of alternative strategies for using the anti-TNFs.

The following sections report the development of a de novo model to address some of the key issues and uncertainties which have been identified in this review. Chapter 5 reports the results of an extended synthesis which has been developed to provide a more generalisable framework for synthesising clinical efficacy data ensuring that appropriate estimates are generated for the model which make use of all relevant and available evidence. This is followed in Chapter 6 by a description of the de novo model (York model) which attempts to link this framework to a more coherent conceptual model of the chronic and progressive nature of AS and nr-AxSpA.