Chapter 4Cost-effectiveness review and health economic modelling

Search strategy

A comprehensive search of the literature for published economic evaluations (including any existing models), cost studies and QoL (utility) studies was performed. The systematic search included searching the following electronic databases during December 2014 (from 12 to 17 December 2014):

• MEDLINE (via Ovid) (1946 to Week 3 November 2014)
• MEDLINE In-Process Citations and Daily Update (via Ovid) (11 December 2014)
• EMBASE (via Ovid) (1947 to 15 December 2014)
• NHS Economic Evaluation Database (The Cochrane Library)
• Science Citation Index (Web of Knowledge) (1970–present)
• Cost-effectiveness Analysis Registry
• EconPapers (Research Papers in Economics)
• School of Health and Related Research Health Utilities Database.

The search included terms for CD, anti-TNF-α drugs and the different assay kits, combined with economic and QoL terms. The search was limited to studies published in the English language. The search strategy developed was based on the clinical effectiveness review, with input from a health economist. Details of the full search strategies are provided in Appendix 3.

Inclusion criteria

Only studies meeting the following inclusion criteria were included in the review:

• study type – fully published economic evaluations (including economic models)
• population – people with CD
• intervention – anti-TNF-α drugs (ADA and IFX) and antibody drug testing (LISA-TRACKER ELISA kits, TNF-α-Blocker ELISA kits and Promonitor ELISA kits) for any dosage or treatment regimen
• comparator – standard care treatment: anti-TNF-α drugs (ADA and IFX) for any dosage or treatment regimen
• outcomes – cost-effectiveness or cost–utility studies reporting outcomes as clinical effectiveness measures or utility measures [utility, EQ-5D, Short Form questionnaire-6 Dimensions score or quality-adjusted life-years (QALYs)].

Exclusion criteria

Studies meeting the following exclusion criteria were excluded from the review:

• non-English-language publications
• studies in the health areas where these anti-TNF-α drugs have also been used, such as UC, rheumatoid arthritis, psoriasis and tuberculosis.

Assessment of eligibility and data extraction

All retrieved records (citations and abstracts) were collected in a specialist database (EndNote) and duplicate records were identified and removed. Two reviewers independently reviewed titles and abstracts to identify potentially relevant papers for inclusion. Any discrepancies were resolved by discussion. See Appendix 13 for the table of full-text studies excluded with reasons.

Data extraction was carried out in two stages by one reviewer using standardised data extraction sheets (see Appendix 14) and was then checked by a second reviewer. Stage 1 considered all eligible studies (fully published economic evaluations including any economic models) and stage 2 considered studies assessed for usefulness for populating the economic model. Data extracted during stage 1 included the following:

• study details – author names, source of publication, language and publication type
• baseline characteristics – population, intervention, comparators, outcomes and type of economic evaluation
• methods – target population and subgroups, setting and location, study perspective, time horizon, discount rate, measurement of effectiveness, measurement and valuation preference-based outcomes, resource use and costs, currency, price date and conversion, model type, assumptions and analytical methods
• results – study parameters, incremental costs and outcomes and characterising uncertainty
• discussion – study findings, limitations, generalisability and conclusions
• other – sources of funding, conflicts of interest and comments.

Quality assessment

The quality of full economic evaluation studies that were identified was assessed using the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) checklist (see Appendix 15) by one reviewer and cross-checked by a second reviewer. The CHEERS checklist comprises six dimensions: title and abstract, introduction, methods, results, discussion and other. Under these dimensions, a series of questions check whether or not the criteria have been clearly reported. Any studies containing an economic model were further assessed using the framework for the quality assessment of decision-analytic modelling by Philips et al.¹⁴⁹ (see Appendix 15). The Philips checklist contains two main dimensions: structure of the model and data used to parameterise the model. Under these dimensions several questions assess whether or not the criteria has been clearly reported.

Data synthesis

Information extracted from the included studies were summarised and tabulated. Findings from individual studies were compared narratively.

Summary of the Health Technology Assessment report by Dretzke et al.⁵

The main aim of this HTA report was to assess the cost-effectiveness of anti-TNF-α drugs in the management of adult patients with moderate to severe CD in the UK NHS. The authors described induction therapy as the use of anti-TNF-α therapy with the aim of achieving remission (a repeated reinduction treatment was considered, rather than a one-off induction therapy) and maintenance therapy as the use of anti-TNF-α therapy to maintain remission in patients who have responded (and continue to respond) to anti-TNF-α therapy when in relapse. Response by the authors was defined as remission within 8 weeks.

The authors developed a Markov model from a NHS and Personal Social Services (PSS) perspective to estimate the incremental cost per QALY gained for both ADA and IFX (anti-TNF-α therapy) compared with standard care. Mortality was not included in the model, as the authors found no difference in the mortality rates that were reported in the clinical trials reviewed and, therefore, felt that a lifetime horizon would not improve the precision of the cost-effectiveness estimate. Instead, the time horizon for the model was 1 year and the cycle duration was 4 weeks. The model for both induction and maintenance therapy started with a cohort of patients in the standard care refractory relapse health state. The model had four main health states and at any time, and on any given treatment, a patient was in remission, in relapse, undergoing surgery or in post-surgery remission.

Transition probabilities for the standard care health states were based on Silverstein et al.¹⁵⁰ Transition probabilities for both the induction and maintenance model were assigned a treatment effect by using relapse to remission probabilities from RCT evidence; however, for the maintenance model there was a lower remission to relapse rate.

The majority of utility values for the model were based on the study by Gregor et al.,²⁰ which used the time-trade off measure to estimate the health-related QoL in CD. A utility value for surgery was not available in the published literature; therefore, it was assumed that the average utility value for surgery would be equivalent to EQ-5D health state 22222, with a utility weight of 0.516.

The direct costs to the NHS were the sum of the anti-TNF-α costs and type-specific health-state costs. The costs of anti-TNF-α therapy, both induction and maintenance, were derived from the BNF (2007/8),^151,152 and administration costs were also included for IFX. Type-specific health-state costs included costs for surgery, which were modelled as the cost of inpatient IBD interventions, and post-surgery remission costs, which were based on outpatient surgical gastrointestinal follow-up. Moderate and severe relapse costs were modelled as the cost of IBD outpatient major and intermediate interventions. Relapse costs were based on a gastrointestinal admission to hospital. Remission costs were modelled using literature. The majority of health-state unit costs were obtained from the NHS reference cost database (NHS Reference Costs 2005 to 2006¹⁵³).

Incremental cost-effectiveness ratios (ICERs) and cost-effectiveness acceptability curves were presented. One-way sensitivity analyses and probabilistic sensitivity analyses (PSAs) using 10,000 simulations were conducted to characterise uncertainty in the model.

For induction therapy for severe CD, both ADA and IFX dominated standard care (i.e. they were cheaper and more effective). For maintenance therapy for severe CD, neither drug was cost-effective (well above NICE thresholds). For moderate CD, for maintenance therapy for both drugs and induction therapy for IFX, these were not cost-effective (well above NICE thresholds); however, for induction therapy, ADA dominated standard care.

Sensitivity analysis showed that, in patients with severe disease, IFX induction treatment was cost-effective relative to maintenance treatment and standard care in > 99% of cases at all points up to £100,000 per QALY. Likewise, ADA induction treatment was found to be cost-effective relative to maintenance treatment and standard care for thresholds up to £100,000 per QALY.

The key limitations of this model was a short time frame (1-year time horizon); the exclusion of death from the model; no randomised controlled data available for maintenance therapy; and the use of Silverstein et al.¹⁵⁰ data for transition probabilities, which inherently had their own problem (i.e. surgery rates were higher and relapse rates much lower than in routine practice).

Search results for objective D

The literature search identified 2466 records through electronic database searches and other sources. After removing duplicates, 1527 records were screened for inclusion. On the basis of a title and abstract sift only, 1518 records were excluded. The remaining nine records were subjected to full-text screening. A further five articles^{5,26,154–156} were excluded at the full-text stage, as these studies did not use assay kits to measure levels of TNF-α inhibitors and anti-drug antibodies. The literature search identified four studies^{73,123,124,157} of the cost-effectiveness of different assay kits for measuring levels of TNF-α inhibitors and of anti-drug antibodies (Figure 27).

FIGURE 27

The PRISMA flow diagram of cost-effectiveness studies.

Overview of included studies

The literature search identified four studies^{73,123,124,157} that met our inclusion criteria (studies looking at the cost-effectiveness of different assay kits for measuring levels of TNF-α inhibitors and of anti-drug antibodies) and were reviewed. In the following sections we present an overview of the included studies by population (responders and those showing LOR) of interest.

Comparison of the included studies

All four studies included in this review have been summarised in Table 27. Three studies were based on RCTs^73,123,124 and only one study¹⁵⁷ presented an economic model. Of the RCTs, two^123,124 were conducted in Denmark and one⁷³ was conducted in Belgium. All four studies^{73,123,124,157} conducted cost-effectiveness analyses: Vande Casteele et al.⁷³ compared concentration-based with clinician-based dosing; Steenholdt et al.^123,124 compared IFX treatment failure using a treatment algorithm compared with IFX dose increasing; and Velayos et al.¹⁵⁷ compared a testing-based strategy with an empiric dose escalation strategy. All studies^{73,123,124,157} clearly stated the type of assay used to analyse serum levels and antibodies to anti-TNF-αs. Two studies^123,124 used RIA in the base case, one study⁷³ used an assay developed in house and the remaining study¹⁵⁷ used a PROMETHEUS ELISA.

TABLE 27

Summary characteristics of the economic studies comparing ELISA kits

The patient populations for three studies^73,123,124 included eligible patients with moderate to severe CD, whereas the study by Vande Casteele et al.⁷³ included patients with UC. The study perspective was not reported in two studies,^123,124 whereas the other two studies^73,157 conducted the analysis from a third-party payer perspective. The time horizon varied from 12 weeks to 1 year. Steenholdt et al.¹²³ based their analysis on a 12-week horizon, whereas the other three studies^73,124,157 used a 1-year time horizon to estimate the cost-effectiveness of the different strategies.

In two studies,^73,157 outcomes were reported as cost per QALYs gained. Vande Casteele et al.⁷³ used the EQ-5D measure to estimate QALYs, whereas Velayos et al.¹⁵⁷ did not explicitly report how the QALYs were estimated, except to say that they were obtained from a secondary source.²⁰ The two studies by Steenholdt et al.^123,124 reported outcomes in terms of cost per ITT and cost PP population.

Three studies^123,124,157 provided quite a comprehensive breakdown of resource use and costs, whereas the study by Vande Casteele et al.⁷³ did not elaborate on resource use, apart from the drug costs. Three studies^73,123,124 reported costs in 2012 prices, whereas Velayos et al.¹⁵⁷ did not report the price year explicitly; however, we assumed that costs are most likely to be in 2012 prices, as the study was published in 2013.

No studies conducted discounting for either costs or benefits as the time horizon for these studies was ≤ 1 year.

The results and conclusions reported differed between studies, Vande Casteele et al.⁷³ demonstrated that concentration-based dosing was slightly less effective and less costly than clinically based dosing, but overall differences were small. Steenholdt et al.¹²³ showed that the intervention based on the algorithm achieved similar clinical and life quality outcomes to dose intensification, but at a lower cost at 12 weeks. These results were maintained at both 20 weeks and 1 year.¹²⁴ Velayos et al.¹⁵⁷ showed that the testing strategy was cheaper and more effective than the empiric strategy.

All four studies^{73,123,124,157} conducted sensitivity analyses to deal with uncertainty around key parameters. The sensitivity analyses ranged from the most simplistic one-way sensitivity analyses^123,124 to the more sophisticated probabilistic analyses.¹⁵⁷

Quality assessment

We present, in Appendix 15, a summary of the reporting quality of the studies included in the current review against the CHEERS checklist.¹⁵⁸ Using a 25-point CHEERS checklist, one article⁷³ did not identify the study as an economic evaluation in the title. All studies provided background information to the study and clearly outlined the objectives of the study. Two studies^73,157 reported the viewpoint of the economic analysis. All studies described the comparators fully and reported the time horizon. However, because of the short time horizon, no studies conducted discounting of costs and benefits. In addition, the choice of health outcomes was well reported by all four studies;^{73,123,124,157} however, only one study⁷³ reported how these health states were valued. Resource use and costs were well reported in three studies^123,124,157 apart from that by Vande Casteele et al.,⁷³ who described only the drug costs. The majority of the studies^73,123,124 conducted an economic analysis alongside a RCT, whereas one study¹⁵⁷ developed an economic model. In terms of analytical methods, study parameters, incremental costs and outcomes and uncertainty were well reported by all four studies. Limitations were provided by all four studies and generalisability was only partially reported by three studies.^123,124,157

From the studies identified, one¹⁵⁷ conducted a model-based economic analysis to determine whether or not a testing-based strategy was more cost-effective than an empiric dose escalation strategy. We present, in Appendix 15, a summary of the reporting quality of this study against Philips’s checklist.¹⁴⁹ In general, Velayos et al.¹⁵⁷ conformed to best practice for reporting model-based economic evaluations in terms of clearly stating the decision problem, adequately outlining the objectives, clearly stating the viewpoint of the analysis and describing the model structure, which represented the clinical pathway that patients with CD may follow. Time horizon and cycle length were stated and justified. In terms of the data required to populate the model, Velayos et al.¹⁵⁷ adequately provided references, but they were unclear on the choices made between data sources and the quality of information used in the model. In addition, it was unclear whether or not any expert opinion had been used when choosing baseline information for the model. The other limitations identified were the lack of explanation of pre-model analysis (e.g. calculation of transition probabilities, and methods and assumptions used to extrapolate short-term results into final outcomes) and the omission of half-cycle correction.

Concurrent testing

In the concurrent testing strategy, patients undergo tests for serum anti-TNF-α levels and antibodies to anti-TNF-α simultaneously, and once the test results are available follow the proposed algorithm. Patients are classified, on the basis of their test results, into one of four groups: drug absent and antibodies present, drug absent and antibodies absent, drug present and antibodies present, or drug present and antibodies absent. Alternatively, patients may be categorised according to levels of drug regardless of antibody levels (e.g. as in the TAXIT trial⁷³). Details of test results and proposed algorithms from Steenholdt et al.¹²³ for patients with LOR and from Vande Casteele et al.⁷³ for responders are presented in Chapter 3, Objective B: description of algorithms prescribing patient management following test outcomes for drug and/or anti-drug antibody levels.

Reflex testing

In the reflex testing strategy, patients would receive a test to analyse serum anti-TNF-α levels. As a result of testing, two test outcomes are likely: drug absent or drug present. Based on the drug result, patients would undergo further testing for the presence or absence of antibodies. In this section we outline the health states and the pathways for patients undergoing reflex testing for both responder and LOR models. No study was identified that tested an algorithm for reflex testing. The algorithm followed in the model was therefore based on that of the TAXIT⁷³ trial for responders and the Steenholdt et al.¹²³ algorithm for patients with LOR using concurrent testing. Further details of test results and proposed algorithms are presented in Chapter 3, Objective B: description of algorithms prescribing patient management following test outcomes for drug and/or anti-drug antibody levels.

Proportions

The proportions of patients required to populate various model decision tree branches were obtained from secondary sources [e.g. management studies described in Chapter 3, Objective C1: clinical studies evaluating drug monitoring for the management of Crohn’s disease patients (management studies)] and, when such data were lacking, from clinical input. Proportions that were estimated included partitioning of patients by presence or absence of IFX and of antibodies to IFX in responders and in those with LOR; partitioning of responders according to defined IFX trough levels; and partitioning by treatment options following surgery.

Table 29 summarises the partitioning of IFX responders based on the study of Imaeda et al.,⁹⁹ discussed in Chapter 3, Analysis of correlation studies of tumour necrosis factor alpha/anti-drug antibodies level and response, that used concurrent monitoring for the absence or presence of IFX and antibodies to IFX.

TABLE 29

Proportions derived based on concurrent testing of patients responding to IFX

The proportions of IFX responders with various trough levels of IFX were based on Vande Casteele et al.⁷³ (discussed in Chapter 3, Vande Casteele et al.: Trough level Adapted infliXImab Treatment study⁷³). These authors screened a cohort of patients with IBD who were receiving maintenance IFX treatment, and further categorised patients by drug concentration based on test result. Drug levels < 3 µg/ml were considered below the target range, levels between 3 and 7 µg/ml were considered within range and those > 7 µg/ml were above target range. Table 30 shows the proportions of responders with different trough drug levels and the proportions of responders derived from this study.

TABLE 30

Proportions according to IFX trough levels of patients responding to IFX

The partitioning of IFX patients with LOR according to concurrent test monitoring of IFX and antibodies to IFX was based on information obtained from Steenholdt et al.¹²³ Table 31 summarises these proportions.

TABLE 31

Proportions based on concurrent testing of patients with LOR to IFX

Patients who have undergone surgery may receive post-operative treatment to maintain remission. These options include an anti-TNF-α, an immunosuppressant, a combination of an anti-TNF-α and an immunosuppressant or no treatment. Table 32 shows these proportions based on the study of van der Have et al.¹⁶⁰

TABLE 32

Treatment following surgery

Table 33 summarises the proportions of IFX responders based on the study by Imaeda et al.,⁹⁹ in which reflex testing was used to test for the absence or presence of IFX. In patients in whom IFX was present, we used the proportions according to IFX trough levels based on the Vande Casteele et al.⁷³ study, as shown in Table 30.

TABLE 33

Proportions derived based on reflex testing of patients responding to IFX

The partitioning of IFX patients with LOR according to reflex test monitoring of IFX was based on information obtained from Steenholdt et al.¹²³ Table 34 summarises these proportions.

TABLE 34

Proportions based on reflex testing of patients with for LOR to IFX

Time-to-event transition probabilities

Table 35 summarises the transition probabilities for time-to-event outcomes used in the models.

TABLE 35

Summary of parametric models used for estimating transition probabilities for time-to-event outcomes

Resource use and costs

The resource use and costs included were those directly incurred by the NHS. The costs of reagents for monitoring trough concentration of anti-TNFs and of antibody-measuring kits, treatment for CD and laparoscopic ileocolic resection were all included in the analysis. Resource use and costs associated with occupying all health states except dead were also included. Unit costs are presented in Table 36. The majority of the cost information used in the analyses was obtained from secondary sources.

TABLE 36

Resource use and costs and utilities used in the models

The costs of monitoring kits for IFX and for antibodies to IFX were obtained from Theradiag/Alpha Laboratories. In Appendix 18, we present a breakdown of the resource use and costs associated with monitoring kits for IFX and antibodies to IFX. In the models, we used a cost of £39.58 per person for concurrent testing for IFX and antibodies to IFX. In the case of reflex testing, we used a cost of £43.48 for patients in whom testing for IFX was followed by testing for antibodies because the results of the former were negative. For patients in whom a test for IFX was positive, no subsequent antibodies monitoring test was undertaken and, hence, we used a cost of £21.74.

The costs of maintenance treatment were obtained from the BNF (2013/14).¹⁶⁶ The costs of treatment associated with the induction phase (weeks 0–6) were not included. IFX treatment costs comprised its acquisition and administration costs. In the base case, we assumed that patients receiving maintenance therapy have received infusions of IFX 5 mg/kg every 8 weeks and that patients weighed, on average, 70 kg. For IFX maintenance, we derived a cost of £1966.41 (assuming four 100-mg vials at £419.62 plus administration costs of £287.93 per infusion) every 8 weeks. For patients switching to ADA, we derived a cost of £704.28 (2 × £352.14, assuming 40 mg of ADA is required every 2 weeks) per 4-week cycle. We assumed that patients would self-administer ADA; hence, no administration costs were included.

The estimated costs of management (outpatient visits to consultants and further investigations) associated with occupying all health states except the dead state were obtained from NHS Reference Costs 2013 to 2014¹⁶⁷ and in consultation with a clinical expert. These health-state costs include outpatient visits, colonoscopy and magnetic resonance imaging. In Table 36, we present the unit costs per year associated with each health state.

Costs obtained from published sources were adjusted to 2013/14 prices using the Hospital and Community Health Service Pay and Price Index¹⁷⁰ and future costs were discounted at a rate of 3.5% per annum, as recommended by NICE.

Sensitivity analysis

In addition to our base-case analysis, we have undertaken a number of sensitivity analyses. These analyses are summarised below:

1. undertake concurrent testing and reflex testing every 12 months in the responder and LOR models
2. estimate the mean costs and effects associated with each strategy using a 1-year time horizon with 4-week cycle lengths
3. in the responders model – three possible modes of one-off testing:
   1. one-off testing at 3 months followed by yearly retesting
   2. one-off testing at 3 months and one retest for those who regained response
   3. one-off testing at 3 months and no retesting for responders/regained response.
4. in the LOR model – 3-monthly testing for patients with LOR; no testing for patients who have regained response
5. no regain of response following best supportive care (responders)
6. no regain of response following best supportive care (LOR).

Probabilistic sensitivity analyses

Probabilistic sensitivity analyses were undertaken to determine the joint uncertainty in key model input parameters of test results and expected QALYs. The PSA was undertaken based on the outcome of cost per QALY only. In PSA, each model parameter is assigned a distribution reflecting the amount and pattern of its variation, and cost-effectiveness results are calculated by simultaneously selecting random values from each distribution. The distributions used in the PSA are presented in Table 36. We have calculated probabilities that each strategy is the most cost-effective, at a willingness to pay of £20,000/QALY.