Assessment of cost-effectiveness

The AG conducted a systematic review of the economic literature to identify the existing evidence assessing the cost-effectiveness of treatment with lenvatinib and sorafenib (vs. each other and vs. BSC) for people with progressive, locally advanced or metastatic RR-DTC. The review focused on the decision problem outlined in the final scope issued by NICE.53 The economic evaluations presented in the submissions by Eisai Ltd8 and Bayer HealthCare7 are discussed and critiqued separately in Summary of the key features of the companies’ economic models.

Search strategy

The AG identified cost-effectiveness studies by searching EMBASE, MEDLINE, the NHS Economic Evaluation Database via The Cochrane Library and EconLit from 1999 onwards. The starting date for all of the searches was 1999 and all databases were searched on 10 January 2017. Based on the fact that the FDA approved sorafenib for its first indication in 2005, and lenvatinib in 2015, the AG considered that this date span would allow all relevant economic evidence to be identified. The reference lists of included publications, in addition to the NICE, SMC and CADTH websites, were hand-searched. The results of the searches were entered into an EndNote X7.4 library and de-duplicated.

Study selection and inclusion criteria

Publications were selected for inclusion in the review based on their relevance to the decision problem and the specific economic criteria presented in Table 12. In addition to costs, quality-adjusted life-years (QALYs), cost–benefit and cost-effectiveness outcomes, such as cost per PFS year, were also extracted from relevant publications.

Two reviewers (RH and NF) independently screened the titles and abstracts of all publications identified by the searches. The same two reviewers then independently retrieved and assessed (for inclusion) the full texts of the publications that had been identified as being potentially relevant to the review. Disagreements about inclusion in the review were resolved through discussion and, in all cases, a consensus was reached; it was, therefore, not necessary to consult a third reviewer during the screening and selection process.

Quantity of evidence

The searches for economic evidence identified 19 citations in total: 14 were obtained from the database searches and five were identified from other sources. Once duplicates were removed, 18 publications remained and, after assessment of the titles and abstracts, 10 publications5,38,48,158–164 were retrieved and a detailed assessment of their eligibility was undertaken.

Nine of these 10 publications were included in the review. The AG included four publications158–160,163 that clearly met the inclusion criteria. The AG considered that the economic evidence for lenvatinib and sorafenib that had been submitted to the SMC38,48 and CADTH5,162 was also relevant to this review and so these four records,5,38,48,162 one for each drug’s individual submission to each regulatory agency, were included in the review. One further relevant publication161 was identified during the citation search of the included publications; this publication became available online only after the AG’s database searches had been completed.

One publication164 was a budget impact analysis and was, therefore, excluded from the review.

A Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram showing the process of study selection is shown in Figure 8.

FIGURE 8 The PRISMA flow diagram: the AG economic literature review. (Figure not reproduced here; see the original article.)

A summary of the characteristics of the nine included publications5,38,48,158–163 is presented in Table 13.

Quality of the included evidence

The quality of the included evidence was assessed using the NICE reference case checklist169 and the Drummond checklist.170 Summary tables of the AG’s quality assessments are presented in Appendices 11 (see Table 72) and 12 (see Table 73). Full details of the completed checklists are presented in Appendices 13 (see Tables 74–81) and 14 (see Tables 82–89). The publications by Huang et al.158,159 have been evaluated together as the same economic model was used to generate results for both publications.

Only the Wilson et al.161 publication was available as a full-text paper published in a peer-reviewed journal. Three of the included publications158,159,163 were available only as abstracts, and one publication160 was available as a poster. The submissions to the regulatory bodies in Scotland38,48 and Canada5,162 were available only as summary reports. As a result, only limited information was available from most of the included publications and this hindered the quality assessment of some of the methodologies described in the publications.

The authors of all of the included publications produced incremental cost-effectiveness estimates enabling a single metric [an incremental cost-effectiveness ratio (ICER) per QALY gained] to be used for comparative purposes. All of the publications included a discussion of the certainty associated with study results; however, full details of the sensitivity analyses and parameter values were not always available in the text.

Generally, the text describing the assumptions and data sources used to generate resource use, costs and HRQoL estimates within the economic models was not clear. In addition, it was unclear whether or not the costs and benefits described in the publications were discounted appropriately. Results from analyses of the cost-effectiveness of all the relevant comparators (lenvatinib, sorafenib and BSC) were available from only four of the reviewed publications.158–161

None of the publications considered the decision problem from the perspective of the NHS in England. However, as the Scottish NHS provides a sufficiently similar environment to the NHS in England, the AG considered that, for the purposes of this appraisal, the results from the SMC submissions38,48 are broadly generalisable to patients in England. The characteristics of the health-care systems, in terms of the way treatments are procured and used in the USA,158,159,161 Canada5,162 and Turkey,163 make the results from analyses based on these perspectives less useful when considering treatment options for patients in the NHS in England. However, including these studies5,158,159,161–163 in this review allows a broad range of cost-effectiveness estimates to be considered and provides some indication of the effect of varying assumptions, such as the model time frame and estimates of HRQoL.

Assessment Group economic review: overview of included publications

The AG identified nine relevant publications5,38,48,158–163 describing the cost-effectiveness of treatment with lenvatinib and sorafenib in a population of patients with RR-DTC. When necessary, authors were contacted and asked to provide further information on methodological aspects that lacked clarity in the publications; only one lead author160 replied and provided the discount rate used in the model.

One publication163 considered the cost-effectiveness of treatment with sorafenib compared with usual care in the Turkish setting. Four publications158–161 compared treatment with lenvatinib with treatment with sorafenib from a US perspective. The SMC submissions38,48 considered resource use in the Scottish NHS, and the CADTH submissions5,162 included analyses that were undertaken from the perspective of the Canadian health-care system. The results reported in the publications5,38,48,158–162 comparing the cost-effectiveness of lenvatinib with the cost-effectiveness of sorafenib are based on the results of indirect comparisons. This means that the authors considered that the trial and patient characteristics of SELECT and DECISION were sufficiently comparable for their data to be compared using this methodology. The AG discusses the limitations of using data from SELECT and DECISION in an indirect comparison in Chapter 4, Indirect comparison feasibility assessment.

The costs, benefits and incremental results from each of the publications are presented in Table 14. All costs from 2014 have been inflated to 2015/16 prices using the hospital and community health services index.171 Analyses conducted using 2015 and 2016 prices have not been inflated as the 2016/17 inflation indices were not available. When the year that costs used within the model is not reported, the year of publication is used as a proxy. When necessary, all cost data have been converted to Great British pounds using the Bank of England exchange rate as of 25 May 2017.172

Erdal et al.163

The authors described a partition survival model that used clinical evidence from DECISION, supplemented with Turkey-specific resource use and cost information, to generate estimates of the cost-effectiveness of treatment with sorafenib versus BSC in a population of people with locally advanced or metastatic RR-DTC. Deterministic results were presented and the ICER per QALY gained for the comparison of treatment with sorafenib with treatment with BSC was £23,859. The authors concluded that the results of the one-way deterministic analyses and probabilistic sensitivity analysis (PSA) were similar to the main set of deterministic results. However, as few details of the parameters and values that were used to estimate the level of uncertainty around results were reported in the publication, the AG was not able to ascertain the reliability of results generated by the sensitivity analyses. Despite not reporting a willingness-to-pay threshold, the authors considered sorafenib to be a cost-effective treatment compared with BSC.

Huang et al.158

The Markov model described by the authors used effectiveness evidence from the Phase III trials SELECT and DECISION. Results from one-way sensitivity analyses showed that the base-case model results were sensitive to changes to the costs of lenvatinib and sorafenib and the utility benefit of continuing with lenvatinib. The AG notes that the value and duration of the utility benefits were not reported. The base-case ICER for the comparison of treatment with lenvatinib with treatment with sorafenib was £61,109 per QALY gained.

Huang et al.159

The authors reported the methods and results of an expected value of perfect information (EVPI) analysis using the same model described in the abstract by Huang et al.158 An ICER of £73,913 per QALY gain was reported, indicating that treatment with lenvatinib offers an increase in benefit over sorafenib, but at an additional cost. At a willingness-to-pay threshold of approximately £77,000 per QALY gained, the probabilities of lenvatinib and sorafenib being cost-effective were low (37% and 33%, respectively). Owing to uncertainty around the reliability of model results, the authors were not certain that treatment with lenvatinib was cost-effective when compared with sorafenib and placebo.

Tremblay et al.160

The poster included results from a cost-effectiveness analysis from a partition survival model designed to compare treatment with lenvatinib and treatment with sorafenib using clinical evidence from the Phase III SELECT and DECISION. The base-case ICER for the comparison of treatment with lenvatinib with treatment with sorafenib was £81,338 per QALY gained when a 10-year time horizon was modelled, and £96,671 per QALY gained when a 5-year time horizon was modelled.

Costs per PFS year (£58,833 with a 5-year time horizon and £62,318 with a 10-year time horizon), costs per responder (£77,372 with a 5-year time horizon and £84,841 with a 10-year time horizon) and life-year saved were also reported in the publication. The authors did not set a willingness-to-pay threshold to determine at what level the cost per responder, for example, would offer good value for money. The authors refer to PSA in the publication but do not report the methods or the results of the analysis.

Wilson et al.161

The same set of authors who produced the abstracts by Huang et al.158,159 authored a full-text paper comparing the cost-effectiveness of treatment with lenvatinib with that of sorafenib, in which they described a Markov model that used effectiveness data from the Phase III trials SELECT and DECISION. ITCs to compare the effectiveness of lenvatinib with the effectiveness of sorafenib were made following adjustments to the placebo arms of the trials as the authors considered that the placebo arm of SELECT included patients who appeared to be healthier than those in the comparator arm of DECISION. However, the AG does not consider that the adjustments are sufficient to generate reliable estimates of the comparative effectiveness of lenvatinib and sorafenib. In addition, as discussed in Chapter 4, Indirect comparison feasibility assessment, the AG does not consider that it is appropriate to undertake an indirect comparison of the effectiveness of lenvatinib and the effectiveness of sorafenib using data from SELECT and DECISION.

The results of the authors’ cost–utility analysis differ from those reported in the abstracts.158,159 In the base-case analysis, treatment with lenvatinib generated more benefits (+ 1.34 QALYs) than treatment with sorafenib (+ 0.96 QALYs), as well as more benefits than placebo (+ 0.71 QALYs), but at an increased cost of £7368 versus sorafenib and £19,921 versus placebo. The base-case ICER for the comparison of treatment with lenvatinib with treatment with sorafenib was £19,522 per QALY gained. The base-case ICERs for the comparison of treatment with lenvatinib with placebo and treatment with sorafenib with placebo were £31,566 and £49,484 per QALY gained, respectively.

Sorafenib Scottish Medicines Consortium submission48

For the comparison of treatment with sorafenib with BSC, the ICER was £32,083 per QALY gained; the Scottish PAS price of sorafenib was used in the analysis.48 These results were sensitive to the time horizon of the model and the approach used to estimate OS, with the ICER increasing with a shortened time horizon and with a change to the OS extrapolation method employed.

Lenvatinib Scottish Medicines Consortium submission38

For the comparison of treatment with lenvatinib with treatment with sorafenib, the base-case ICER was £49,525 per QALY gained; this analysis used the Scottish PAS price for lenvatinib and Eisai Ltd’s estimate of the Scottish PAS discount currently in place for sorafenib.38 The ICERs per QALY gained were sensitive to the estimates of OS for lenvatinib (ranged from £29,000 to £96,000 per QALY gained with PAS prices) and to changing the utility rates used in the model by 20% (ranged from £41,000 to £62,000 per QALY gained with PAS prices).

Sorafenib Canadian Agency for Drugs and Technologies in Health submission5

The company’s base-case cost-effectiveness estimate was that treatment with sorafenib versus BSC resulted in an ICER of £82,080 per QALY gained. Several other ICERs per QALY gained were also presented as a result of reanalyses suggested by the Economic Guidance Panel. The reanalyses included amendments to the time horizon, the duration of treatment and estimates of OS. The results from the reanalyses ranged from £108,974 to £118,913 per QALY gained.

Lenvatinib Canadian Agency for Drugs and Technologies in Health submission162

The base-case analysis for the comparison of lenvatinib with BSC, submitted by the company, generated an ICER of £72,536 per QALY gained. This increased to £101,293 per QALY gained when the amendments suggested by the Economic Guidance Panel were implemented. The reanalysis included amendments to OS estimates, time horizon, use of the intervention drug in terms of both wastage and the appropriate pack size to reach the required dosage, and the utility values used within the model.

Although the company submitted results from additional analyses comparing the cost-effectiveness of lenvatinib with the cost-effectiveness of sorafenib to CADTH, these results were not presented in the available CADTH guidance report.162

The AG notes that the SMC38,48 and CADTH5,162 reports highlight concerns about the clinical effectiveness data derived from SELECT and DECISION. Key issues of concern related to median OS not being reached and the high rates of treatment crossover from the placebo (BSC) arms to the intervention arms (lenvatinib or sorafenib) that took place during the trials.

The Assessment Group’s review of economic evidence: summary and conclusions

The published economic evidence163 shows that the ICER of £23,859 per QALY gained for the comparison of sorafenib with BSC (after conversion from Turkish lira) is within the willingness-to-pay threshold that is considered to reflect a cost-effective use of NHS resources. However, without further details of the economic model inputs, in particular the resource use and costs, the relevance of this finding to the NHS setting is unclear.

In the US setting, when compared with placebo, both treatment with lenvatinib and treatment with sorafenib appear to provide additional health benefits while either saving resources158 or yielding ICERs per QALY gained of < £50,000 after conversion from US$ (£31,566 per QALY gained161 for lenvatinib versus placebo and £49,484 per QALY gained161 for sorafenib vs. placebo). When treatment with lenvatinib is compared with sorafenib in the US setting, lenvatinib offers a health benefit over sorafenib but at an increased cost. Cost-effectiveness results ranged from £19,522 per QALY gained161 (lenvatinib vs. sorafenib) to £96,671 per QALY gained160 (lenvatinib vs. sorafenib), at current UK prices. Again, it is unclear whether or not these results are relevant to the NHS setting.

In 2015, sorafenib became the standard of care for patients in Scotland with locally advanced or metastatic RR-DTC, provided that the company supplied the drug to the NHS at the Scottish PAS price agreed by the company with NHS Scotland.48 The SMC sorafenib report48 states that sorafenib generated more benefit than BSC but at an increased cost. The ICER for this comparison was £32,083 per QALY gained. In 2016, an appraisal of treatment with lenvatinib38 versus sorafenib was submitted to the SMC; lenvatinib was considered by the SMC to be both an orphan drug and an end-of-life treatment. For the comparison of treatment with lenvatinib with treatment with sorafenib, based on survival outcome results generated using indirect comparison methods, and using the Scottish PAS price for lenvatinib, the ICER per QALY gained was estimated to be £49,525 and lenvatinib was accepted for use in NHS Scotland.

The AG notes that any discount to the list prices of the drugs agreed with the NHS in Scotland does not equate to an equivalent agreement with the NHS in England. All PAS prices are confidential and thus the applicability of the results presented within the Scottish submissions to the appraisal of lenvatinib and sorafenib for use in the NHS in England is unclear as it is not known whether or not the discounts agreed with the NHS in Scotland are the same as those agreed with the NHS in England.

In 2015, sorafenib was appraised by CADTH5 and, after reanalyses suggested by the Economic Guidance Panel, estimates of the most plausible ICERs for the cost-effectiveness of treatment with sorafenib versus BSC ranged from £108,974 to £118,913 per QALY gained (after conversion from CA$). Lenvatinib was considered for use by the Canadian health-care system in 2016. Estimates of the cost-effectiveness of treatment with lenvatinib versus both BSC and sorafenib were generated but only the comparisons with BSC are reported in the CADTH report.162 After the Economic Guidance Panel’s suggested amendments were carried out, the best estimate for the comparison of treatment with lenvatinib versus BSC was £101,293 per QALY gained. Both lenvatinib and sorafenib have been recommended for use in Canada. The relevance of these results to patients in the NHS is unknown.

What is lacking from the current evidence base are any cost-effectiveness analyses of direct relevance to the NHS in England. The SMC submissions38,48 provide an insight into the costs and consequences associated with treatment with lenvatinib, sorafenib and BSC, and these are likely to be similar for patients treated in England. However, the PAS prices agreed with the NHS in Scotland are confidential and this prevents the reported cost-effectiveness estimates being directly applicable to the NHS in England.

Head-to-head comparisons of the effectiveness of treatment with lenvatinib with the effectiveness of treatment with sorafenib depend on results from indirect comparisons, whether conducted in a formal statistical framework5,38,48,160,162 or with adjustments made to the placebo arms of the Phase III trials,161 which provide estimates based on the pooling of the comparator arms within the SELECT and DECISION Phase III trials. The AG considers that, because of the issues discussed in Chapter 4, Indirect comparison feasibility assessment, it is not appropriate to employ indirect comparisons of the effectiveness of lenvatinib and the effectiveness of sorafenib using data from SELECT and DECISION.

Summary of the companies’ systematic reviews of economic evidence

Both of the companies carried out systematic reviews to identify published cost-effectiveness studies that included lenvatinib and/or sorafenib. Both companies concluded that there are no cost-effectiveness studies conducted in the UK from the perspective of the NHS that were relevant to decision-making in England. Therefore, both companies produced their own de novo economic evaluations.

Summary of the key features of the companies’ economic models

This section includes summary details of the key features of the economic models submitted to NICE, from Eisai Ltd and Bayer HealthCare, as part of the MTA process. All of the company data presented in this section are drawn from the company submissions7,8 and models.

Population

Both companies state that their economic evaluations focus on patients with progressive RR-DTC. However, in the submission from Eisai Ltd,8 it is highlighted that the SELECT definition of progressive RR-DTC was locally advanced or metastatic DTC confirmed by radiographic evidence of disease progression within the prior 13 months and that some patients participating in this trial had received prior vascular endothelial growth factor (VEGF) therapy. Eisai Ltd8 points out that, in contrast, no patients recruited to DECISION had received prior VEGF therapy and that, to be eligible for recruitment, evidence of disease progression within the 14 months prior to commencing the trial was required. The AG describes other differences in the two trial populations in Chapter 4, Trial characteristics, Participant characteristics and Indirect comparison feasibility assessment.

Model structure

Key elements of the structure of the economic models submitted by Eisai Ltd and Bayer HealthCare are included in Table 15. The structure of the two company models is similar and is in line with the structure of models that have previously been submitted to NICE to inform appraisals of interventions used to treat patients with cancer. The structure of both models conforms to specifications detailed in the final scope issued by NICE.53

Therapies

Details about the intervention and comparators included in the company models are provided in Table 16. Both models included the therapies listed in the final scope issued by NICE.53 The AG highlights that the lenvatinib and sorafenib doses in the models are based on average levels of use in SELECT and DECISION and are lower (approximately 17 mg for lenvatinib and 651 mg for sorafenib) than the licensed doses (24 mg for lenvatinib and 800 mg for sorafenib). Possible reasons include dose interruptions/reductions as a result of AEs; in some cases, intolerance may lead to a treatment being stopped.

Survival modelling

Summary details of the general approach the companies used to model patient survival (OS and PFS) are provided in Tables 17 and 18, respectively.

Measurement and valuation of health effects

Sources of utility values

The base-case utility values used in the Eisai Ltd model were stated to be taken from EQ-5D values for sorafenib from DECISION. Disutilities were then applied as a weighted proportion, based on values obtained from a vignette study carried out by Fordham et al.168 The AG notes that only the utility values used in the progressive state were the same as the utility values derived from DECISION.

The source of the utility values used in the Bayer HealthCare model7 was the EQ-5D data collected during DECISION. No additional utility decrements associated with AEs were included in the model.

The use of utility values derived from EQ-5D data collected during clinical trials is in line with the approach set out in the NICE Guide to the Methods of Technology Appraisal 2013.169

Utility values

The utility values used in the companies’ models are provided in Table 19.

Health-care costs

Levels of resource use

Eisai Ltd obtained estimates of the level of health-care utilisation inputs for the pre-progression and progressive disease states from physician surveys conducted in Europe; these estimates were then validated by four practising clinical experts employed by NHS England. Mortality-related costs were obtained from the Nuffield Trust173 and adjusted for inflation to 2016 values based on PSSRU171 inflation rates for 2016.

Expert advice from oncologists was the basis for Bayer HealthCare’s resource use estimates. Unit costs were obtained from the NHS Reference Costs 2015 to 2016174 and the Unit Costs of Health and Social Care 2016.171 In the model, it is assumed that resource use associated with treatment with lenvatinib is the same as the resource use associated with treatment with sorafenib.

The monthly routine care costs used in both company models are provided in Appendix 10 (see Table 70).

Eisai Ltd’s routine costs included physician visits and disease-associated hospitalisation days. Bayer HealthCare’s routine costs included inpatient stay, outpatient appointments and pharmaceutical costs.

Eisai Ltd’s end-of-life costs (£7450) included secondary care, local-authority-funded social care, district nursing and GP contacts.

Adverse event costs

The Eisai Ltd model includes the following AEs:

• lenvatinib – grade 3 and 4 treatment-emergent AEs and AEs that required hospitalisation in SELECT
• sorafenib – grade 3 and 4 treatment-emergent AEs in DECISION and AEs that required hospitalisation based on proportions from SELECT.

The Bayer HealthCare model7 only includes grade 3 and 4 AEs occurring in > 5% of patients in the lenvatinib arm of SELECT or in the sorafenib arm of DECISION.

Bayer HealthCare also included AE management costs (per 28 days), see table 29 in the company submission for details.7

Frequencies/rates and costs associated with AEs included in the company models are presented in Appendix 10 (see Table 71). Eisai Ltd’s cost sources are a mix of NHS Reference Costs 2015 to 2016174 and Unit Costs of Health and Social Care 2016.171 Bayer HealthCare’s cost sources are a mix of NHS Reference Costs 2014 to 2015,175 Unit Costs of Health and Social Care 2015176 and British National Formulary (BNF) costs.47

Cost-effectiveness results

Base-case cost-effectiveness results

The base-case cost-effectiveness results from the Eisai Ltd8 and Bayer HealthCare7 submitted economic models are shown in Table 20.

Bayer HealthCare also carried out cost-effectiveness analyses using the adjusted MAIC HRs. The effect on the company’s ICERs was small. The resultant base-case ICERs per QALY gained for the comparison of treatment with sorafenib with BSC and the comparison of treatment with lenvatinib with BSC are commercial in confidence and cannot be reported.

Probability of being the most cost-effective

For the Eisai Ltd model, the PSA results suggest that, at a willingness-to-pay threshold of £50,000 per QALY gained, the probability of lenvatinib being more cost-effective than sorafenib or BSC is 60%.

For the Bayer HealthCare model,7 the PSA results suggest that, at a willingness-to-pay threshold of £30,000 per QALY gained, the probability of sorafenib being cost-effective is 30%, the probability of BSC being cost-effective is 54% and the probability of lenvatinib being cost-effective is 16%.

The PSA results from the Eisai Ltd and Bayer HealthCare submitted economic models are shown in Table 21.

Sensitivity and scenario analyses

Both companies carried out a range of deterministic sensitivity analyses and scenario analyses.

In the Eisai Ltd model, for the comparison of lenvatinib and sorafenib, the two most influential parameters in the deterministic sensitivity analysis were OS HR versus sorafenib (lenvatinib dominates) and PFS HR versus sorafenib (£5000 to £35,000 per QALY gained). In the scenario analyses, the most influential parameters were the treatment duration for lenvatinib (treatment to progression rather than clinical trial duration; £71,978 per QALY gained) and the cut-off point for OS and PFS extrapolation (20 weeks for OS and PFS; £29,874 per QALY gained).

In the Bayer HealthCare model, for the comparison of sorafenib and lenvatinib, the largest deviations from the base-case ICER per QALY gained were attributable to variation in the OS HR for lenvatinib and lower lenvatinib progression-free utility. The scenario analyses that had the biggest effects on the companies’ cost-effectiveness results were the time horizon (reduction to 10 years) and lower lenvatinib progression-free utility. The ICERs per QALY gained for these analyses are commercial in confidence and cannot be reported here.

The Assessment Group’s independent cost-effectiveness assessment

Model design

In common with the two companies, the AG has used a standard partitioned survival model structure, applied to the patient population specified in the final scope issued by NICE,53 to consider the cost-effectiveness of treatment with lenvatinib and sorafenib compared with BSC (as represented by data from the placebo arms of SELECT and DECISION).

Two particular differences should be noted:

1. The AG has not included a separate health state for patients who respond to treatment. On clinical advice, the AG considers that there is little merit in this addition to the standard three-state structure (in which patients begin in the progression-free health state and, following assessed disease progression, transfer to the postprogression state in which they receive only BSC prior to death). For responding patients, who are mostly symptom-free, response alone is unlikely to have a measurable effect on patient-perceived quality of life/utility and has no effect on resource use.
2. The AG has designed a model that allows each intervention (lenvatinib and sorafenib) to be represented in its natural time metric: 30-day cycles for lenvatinib and 28-day cycles for sorafenib. This involved creating two parallel models using the same assumptions and model parameters, but each with its own placebo arm calibrated from its respective clinical trial data. Although not ideal, the AG has provided an illustrative structural sensitivity analysis (Figure 9) based on applying data from the counterfactual placebo arm of both trials to illustrate the extent of uncertainty involved in comparisons between the active treatments with the currently available clinical evidence. The reason for this unusual approach is to demonstrate non-equivalence of the placebo arms of the two clinical trials, which renders indirect comparison of the two treatments via a common comparator invalid (as discussed in Chapter 4, Indirect comparison feasibility assessment, and illustrated graphically in Figure 9).

FIGURE 9 Model structure featuring two simple trial-based comparisons, with additional cross-trial comparisons as a structural sensitivity analysis to illustrate the uncertainty associated with choice of comparator. (Figure not reproduced here; see the original article.)

Resource use estimation, the sources for unit costs and selection of health-related utility values used in the AG’s model are presented in this section. Standard discount rates of 3.5% per annum are used for discounting both costs and benefits (measured as QALYs), but not for life-years (survival). The AG model is structured with a maximum time horizon of 40 years.

Effectiveness data

Modelling long-term outcomes from trial data

Both companies have followed a conventional approach to the general problem of identifying an appropriate method by which to extrapolate time-limited follow-up trial data for PFS, OS and time to treatment discontinuation. This involves attempting to fit a range of prespecified statistical functions to the available evidence, and selecting one that appears to be optimal according to particular ‘measures of fit’ (principally the Akaike information criterion and the Bayesian information criterion).

This paradigm is wholly dependent on the limited data available and the restricted armoury of ‘standard’ models. In particular, it fails to take into account a wider evidence base specifically related to the natural history of the disease, and the influence of particular characteristics of both the recruited patient group and of the trial design.

The AG has investigated long-term survival trends in patients diagnosed with Stage 3 or 4 (locally advanced or metastatic) thyroid cancer in the USA and recorded on the Surveillance Epidemiology and End Results (SEER) database.177 A total of 32,818 patients (male and female) followed for 15 years yielded a persistent trend from 18 months after diagnosis. Figure 10 demonstrates the very close match between these data and a simple linear model, indicating that the risk of death remained unchanged throughout this period, which is indicative of a simple exponential survival process.

FIGURE 10 Cumulative hazard data from follow-up of patients diagnosed with stage III/IV thyroid cancer for 15 years. (Figure not reproduced here; see the original article.)

This evidence is sufficiently compelling to give the AG confidence to employ exponential extrapolation as the default method of modelling incomplete trial data in this appraisal. The nature of clinical trials (selecting patients who have suffered a recent disease progression, and administering a novel treatment that takes time to reach full effectiveness) means that the initial period post randomisation will give rise to temporary distortions to the underlying disease process. However, thereafter, it is likely that the natural history of the condition will be re-established, so that a long-term exponential function will reappear. The mean time since diagnosis of patients randomised in DECISION was 7.24 years, suggesting that the trial cohort lies in the middle of the follow-up range shown in Figure 10. The AG is therefore confident that outcome data extrapolation should be focused on fitting exponential models to estimate lifetime survival expectation.

Data issues

Following the initial stakeholders’ meeting for this appraisal (17 February 2017), the AG submitted identical requests to the two companies, asking for a set of detailed analyses of the latest data available from the two clinical trials, based on common analytical methods to allow comparative analyses to be carried out by the AG; thus, minimising the risk of methodological bias. Eisai Ltd provided the requested data relating to SELECT as an appendix to the submission.8 Unfortunately, Bayer HealthCare chose not to address the AG’s request. As a consequence, the AG was unable to conduct some comparative analyses based on common assumptions, and the potential for bias and uncertainty in the data available to the AG remains.

The two clinical trials that provide the effectiveness evidence for this appraisal share common features, which result in interpretive complexity and uncertainty. In particular, in both trials patients were permitted to cross over from the placebo control to the active treatment (lenvatinib or sorafenib) following disease progression. As a consequence, randomisation was broken in both trials and some outcome variables may not be mutually compatible, even after attempts to adjust for crossover effects.

Both companies assume that, in addition to the active treatments, a third comparator (BSC) may be represented by the placebo arms of the two trials. Moreover, it is implicitly assumed that the randomised patients are drawn from similar populations with reference to their risk profile for the various time-to-event outcomes measured [PFS, OS, postprogression survival (PPS) and time-to-treatment discontinuation]. In Chapter 4, Indirect comparison feasibility assessment, the non-equivalence of PFS data from the placebo arms of the two clinical trials has been clearly demonstrated. This is of crucial importance to attempts to employ relative effectiveness measures reliant on the PHs assumption in relation to PFS, which is the only standard outcome variable reported in these trials that is free from any contamination by crossover effects (both trial protocols required confirmation of disease progression before patients were allowed to enter the open-label phase in which patients in the placebo arm were offered crossover treatment).

The problem of devising a credible approach to indirect comparison between lenvatinib and sorafenib for PFS cannot be resolved by appeal to technical argument alone. The pattern of hazard over time for disease progression in the two active arms is sufficiently similar to justify a simple HR approach. However, the placebo arms exhibit unexpectedly inconsistent patterns of temporal change, not compatible with the assumption of similarity between the patient groups not receiving active treatment. The AG, therefore, considers that the patients enrolled in the two trials cannot be considered to derive from a common population. This degree of difference precludes the use of either placebo arm as being representative of untreated patients across both trials.

The data for both placebo arms exhibit an unexpected improvement in long-term survival (reducing progression hazard) for which there is no obvious explanation. The effect of this phenomenon is to produce a varying differential in performance when comparing survival components across the two trials without any clear confirmatory evidence. Therefore, the AG is unable to support the use of a conventional ITC in this appraisal. The AG considers that it is preferable to model the relative effectiveness and cost-effectiveness of each active treatment against its own placebo comparator, and then generate results for each drug relative to the placebo of the other clinical trial as a sensitivity analysis, in order to allow assessment of the uncertainty associated with the choice of comparator.

Progression-free survival

The AG chose to use data for locally assessed PFS rather than centrally assessed PFS, as local assessment is generally more closely related to normal clinical practice.

Lifetime mean PFS for patients in DECISION who received placebo may be readily estimated from trial data (for the period available) and a simple exponential curve that conforms closely to the reported trial data (Figure 11). The AG estimated lifetime mean PFS from the area under the K–M data to 16.5 months of elapsed time followed by the area under the exponential function thereafter, giving a lifetime mean PFS estimate of 7.56 months. The sorafenib PFS arm of DECISION exhibits a simple constant hazard (exponential) relationship (see Figure 11), allowing the lifetime mean PFS to be estimated in a similar fashion, using the area under the curve (AUC) of the K–M data until 25 months, and the exponential extrapolation thereafter. This shows a lifetime mean PFS estimate of 47.18 months for patients receiving sorafenib, and a mean gain in PFS of 39.62 months compared with receiving placebo.

FIGURE 11 Progression-free survival K–M data from DECISION modelled by an exponential function. (Figure not reproduced here; see the original article.)

The SELECT data for PFS exhibit a more complex pattern in each arm. The cumulative hazard plots (Figure 12) reveal two distinct phases, both of which follow a constant hazard. Patients in the placebo arm who remain progression-free after 312 days experience a reduction in hazard of about 53%, which is sustained thereafter. Similarly, patients in the lenvatinib arm experience a reduction of progression hazard of about 47% at 529 days. As before, the estimated mean lifetime PFS for these patient groups was estimated as the sum of the AUC in each trial arm, followed by lifetime extrapolation using the long-term exponential hazard of progression or death. This approach yields estimates of mean lifetime PFS of 41.00 months for patients receiving lenvatinib and 6.92 months for patients in the placebo arm of SELECT. Thus, the estimated net lifetime gain in PFS for patients receiving lenvatinib is estimated to be 34.08 months.

FIGURE 12 Cumulative hazard for disease progression for SELECT, with two-phase fitted exponential models. (Figure not reproduced here; see the original article.)

Time to treatment discontinuation

As illustrated in Figure 13, the SELECT data are virtually complete for the cycles of lenvatinib dispensed during the trial. The AG estimates mean usage of lenvatinib as 12.61 30-day cycles per patient.

FIGURE 13 The 30-day cycles of lenvatinib dispensed in SELECT. (Figure not reproduced here; see the original article.)

The DECISION trial data are also complete for the cycles of sorafenib dispensed during the trial, as illustrated in Figure 14. The AG estimates mean usage of sorafenib as 14.36 28-day cycles per patient.

FIGURE 14 The 28-day cycles of sorafenib dispensed in DECISION. (Figure not reproduced here; see the original article.)

Overall survival

Data provided by the company for lenvatinib-treated patients in SELECT (Figure 15) show a simple long-term exponential trend indicating a constant mortality risk throughout the trial period (19.6% per year). This allows the mean lifetime OS for patients treated with lenvatinib to be estimated using the AUC of the trial K–M curve until 34.7 months plus a simple exponential extrapolation thereafter, giving a total mean OS of 55.1 months.

FIGURE 15 Overall survival: lenvatinib-treated patients in SELECT with a fitted exponential model, and RPSFTM adjusted for placebo patient crossover with a long-term exponential-fitted model. (Figure not reproduced here; see the original article.)

Both companies have employed RPSFTM adjustments to data from the placebo arms of their clinical trials to correct for patients crossing over to the active treatment following disease progression. Adjusted OS placebo arm data from SELECT are also displayed in Figure 15 and indicate that, after RPSFTM adjustment, a similar long-term exponential (constant risk) trend also applies to the placebo arm beyond 6 months. Using the AUC of the adjusted K–M curve until 19.1 months plus the exponential extrapolation thereafter yields a lifetime estimated mean OS for the corrected placebo arm of 29.9 months and a net estimated OS gain attributable to treatment with lenvatinib of 25.3 months.

An examination of the OS data from DECISION (Figure 16) indicates that both patients in both treatment arms were subject to a period of relatively low mortality hazard, followed by transition to a higher constant risk of death. This transition took place after 11.2 months for sorafenib patients and 6.4 months for placebo patients.

FIGURE 16 Cumulative mortality hazard for sorafenib-treated patients in DECISION with a fitted two-phase exponential model, and for RPSFTM-adjusted placebo patients with a fitted two-phase exponential model. (Figure not reproduced here; see the original article.)

Using the AUC of the RPSFTM-adjusted K–M curve for the placebo arm until 6.4 months plus the AUC of the exponential extrapolation thereafter yields a lifetime estimated mean OS for the placebo arm of 47.18 months. Similarly, combining the AUC of the sorafenib arm up to 11.96 months with the exponential trend thereafter yields an estimated lifetime mean OS of 56.66 months. Thus, the net mean OS gain attributable to sorafenib is 9.48 months.

Postprogression survival

Assessment of PPS may be carried out at an aggregate level by calculating the difference between model estimates of OS and PFS. However, it can also be informative to consider this outcome at the level of individual patients, at which it may provide useful insight into possible post-treatment long-term effects of treatments even after active treatment has ceased. The AG asked both companies to provide PPS data from their respective primary clinical trials. Unfortunately, only data from SELECT have been received. As with OS, it is important to allow for the effects of crossover on PPS by using RPSFTM-adjusted data.

In Figure 17, the beneficial effect of crossover to lenvatinib for patients initially randomised to the placebo arm is clearly apparent. Both trial arms exhibit a similar early pattern, albeit at different absolute levels of survival, and thereafter show similar long-term exponential trends after 15 to 18 months from the time of disease progression. When the RPSFTM adjustment is applied, the corrected placebo arm very closely follows the trajectory of the lenvatinib arm (although the effect of RPSFTM revised censoring does not allow direct comparison beyond 16 months). Nonetheless, these data suggest that, after crossover adjustment, there is probably no additional benefit to individual patients crossing from placebo to lenvatinib beyond what would have been gained by treatment prior to disease progression.

FIGURE 17 Postprogression survival: lenvatinib in SELECT with a fitted exponential model, and RPSFTM adjusted for placebo patient crossover with a long-term exponential-fitted model. (Figure not reproduced here; see the original article.)

Summary of time-to-event outcome data analysis

Estimates of PFS, OS and PPS and mean cycles of active treatment received in the two clinical trials are displayed in Table 22. The main difference occurs in the PFS results in which lenvatinib provides substantially greater benefit than sorafenib (34.1 additional months before progression compared with only 6.3 months, respectively). However, the estimated OS results are very similar (55.1 for lenvatinib vs. 56.8 months for sorafenib), and, consequently, estimated PPS is reduced with lenvatinib treatment but increased for sorafenib treatment). Thus, it appears that lenvatinib shows effect more strongly in initially delaying progression, but does not offer additional benefit over sorafenib in terms of long-term survival. The duration of active treatment in the two trials is very similar when measured in days rather than cycles, with a difference of < 7%.

Health-related utility data

The AG has carefully considered the opposing approaches used by the two companies to estimate appropriate health-related utility values to assign to health states and to AEs. The Eisai Ltd model relies heavily on the Fordham et al.168 vignette study (which it sponsored), whereas the Bayer HealthCare model draws on EQ-5D-3L data collected during DECISION.

On theoretical grounds, directly collected evidence from patients with the condition (as used in the Bayer HealthCare model7) should always be preferred to the results of an artificial study without recourse to the views of patients either in design or calibration (as used in the Eisai Ltd model). Of particular concern is the serious overestimation of baseline utility values in the Fordham et al.168 study when compared with UK general population values for people of a similar age. The contrary position argues that DECISION data include the disutility of AEs in estimates of health-state utilities, and, therefore, are biased without any objective means of adjusting the health-state estimates.

On balance, the AG considers that the data from DECISION should be used in the base case (Table 23) with a sensitivity analysis using the Eisai Ltd model values.

Resource use and cost data used in the Assessment Group’s model

Active treatments (lenvatinib and sorafenib)

The lenvatinib full acquisition cost is £4311.00 per 30-day treatment (NHS Indicative Price, BNF June 2017).47 This is reduced by the SELECT dose intensity factor (72.5%) so the true cost per cycle is £3089.55.

The sorafenib full acquisition cost is £3576.56 per 28-day treatment (NHS Indicative Price, BNF June 2017).47 This is reduced by the DECISION dose intensity factor (81.40%) so the true cost per cycle is £2911.32.

There is no administration cost associated with either drug, both of which can be safely taken unsupervised. The NHS Reference Costs174,175 figures quoted by both companies for administration of oral treatment relate to particular drugs that may cause serious rapid-onset reactions, and so the patient must be monitored following administration. Thus, it is not appropriate to use these costs when estimating the cost of either sorafenib or lenvatinib.

Routine care costs

Table 24 summarises the schedule of itemised routine care tests, treatments and specialist visits identified by the AG’s clinical advisor, in terms of use per quarter (3 months), per 28-day cycle and per 30-day cycle. These items are considered to be applicable to all patients.

Adverse events

Three common AEs feature in the two company models for which treatment types and resource use were estimated by the AG’s clinical advisor. The cost estimates shown in Table 25 are for only a single cycle (28 days or 30 days) and take no account of AE episodes that do not resolve within that time or that subsequently recur.

End-of-life care

Health-care costs during the last 90 days of life were estimated using the results presented in Table 9 of the paper by Georghiou and Bardsley;173 costs were uplifted from 2010/11 to 2015/16 using the Hospital and Community Heath Services inflation index178 as shown in Table 26.

Cost-effectiveness results

Deterministic cost–utility results from the AG model using public list prices are compared with submitted results from the two companies in Tables 27 (vs. the Eisai Ltd model) and 28 (vs. the Bayer HealthCare model). Overall, the estimates of incremental costs from the three models are not very different, but estimates of outcomes (life-years and QALYs) show larger discrepancies across the models, reflecting the different assumptions and estimation methods employed. The ICERs per QALY gained reported from the AG model are substantially greater than those obtained from the Bayer HealthCare model, but the Eisai Ltd’s model results show a much larger ICER per QALY gained for sorafenib versus BSC than that obtained from either of the other models.

Inevitably, the relative economic performance of the treatments in all three models will change significantly when final discounted acquisition prices are applied.

Structural sensitivity analysis

The AG cross-trial ICERs per QALY gained can be readily calculated by interchanging the results shown in the two AG BSC columns of Tables 27 and 28. For sorafenib, this results in an incremental cost per patient of £47,993 and incremental QALYs per patient of 1.150, leading to an exploratory ICER of £41,716 per QALY gained. However, for lenvatinib, the incremental cost per patient is £77,148 and the incremental QALYs per patient are 0.591, leading to an amended ICER of £130,592 per QALY gained.

These very large changes (an increase of 105% in the lenvatinib ICER per QALY gained, and a decrease of 54% in the sorafenib ICER per QALY gained) serve to illustrate that the choice of BSC comparator is of major importance in this appraisal, and that the absence of credible indirect comparison results precludes any simple resolution of this difficulty.

Deterministic sensitivity analyses

Sensitivity analyses have been conducted on the cost-effectiveness results obtained using the AG model and the results from these analyses are shown in Tables 29–31.

The AG identified five modelling issues, which do not involve stochastic uncertainty, and the implications, in terms of changes to the size of the estimated ICER per QALY gained in the AG model, that result from changes to these parameter values are shown in Table 29. Assuming that a change in the estimated ICER per QALY gained of < £5000 is not considered substantial, all but one of the five issues generated important changes in the ICER per QALY gained estimates for either sorafenib or lenvatinib (the exception being the discount rate applied to costs).

The AG identified 18 parameter values for which stochastic uncertainty could be quantified in the AG model, and the findings from adjusting these values are summarised in Tables 30 and 31. Only three parameters (the utility values for the PFS and PPS health states estimated from EQ-5D-3L patient data in DECISION, and the sorafenib OS AG extrapolation hazard) were found to lead to substantial effects on the size of the estimated ICER per QALY gained when varied between the lower and upper 95% confidence limits. In particular, the AG considers that uncertainty in specific unit costs (other than drug acquisition costs) is not an important factor when generating uncertainty in ICER per QALY gained estimates.

Probabilistic sensitivity analyses

The AG carried out PSA varying model parameters subject to quantifiable stochastic sampling uncertainty:

• nine routine care cost variables
• seven AE incidence rates
• seven health-related utility values
• seven end-of-life health and social care costs.

In most cases, probabilistic values were drawn from normal distributions around the standard error of the mean, except for incidence rates, for which beta distributions were employed.

Using list prices, the in-trial comparisons of lenvatinib and BSC (Figure 18) and of sorafenib and BSC (Figure 19) yielded similar deterministic and probabilistic ICERs per QALY gained.

FIGURE 18 Probabilistic sensitivity analysis: lenvatinib vs. BSC in SELECT. (Figure not reproduced here; see the original article.)

FIGURE 19 Probabilistic sensitivity analysis: sorafenib vs. BSC in DECISION. (Figure not reproduced here; see the original article.)

Unfortunately, information relating to the key outcome variables (PFS, OS and time to treatment discontinuation) provided to the AG by one of the companies was not in the form requested, and information on uncertainty in the estimated treatment dose intensity was not included by the other company in its submission or their model. Without these key data items, it was not possible to incorporate these important components of the normal PSA on this occasion. Therefore, the results presented below should be treated with caution.

For the comparison of lenvatinib with BSC, the deterministic ICER is £65,872 per QALY gained and the probabilistic ICER is £66,038 per QALY gained.

For the comparison of sorafenib with BSC, the deterministic ICER is £85,644 per QALY gained and the probabilistic ICER is £83,547 per QALY gained.

The variation in additional cost per patient is much smaller relative to the uncertainty in outcomes (QALYs) gained because of the dominance of drug acquisition costs, which constitute 85–90% of the incremental cost per patient when full list prices are assumed to apply.

Clearly, both treatments exhibit estimated ICERs well above £50,000 per QALY gained if list prices are applied. This is confirmed by the cost-effectiveness acceptability curves (CEACs) presented in Figures 20 and 21. An examination of the CEACs shows that, compared with BSC, the probability of sorafenib being cost-effective at a threshold of £50,000 per QALY gained is < 0.05% and the probability of lenvatinib being cost-effective is 5.4%.

FIGURE 20 Cost-effectiveness acceptability curves for sorafenib vs. BSC (DECISION). (Figure not reproduced here; see the original article.)

FIGURE 21 Cost-effectiveness acceptability curves for lenvatinib vs. BSC (SELECT). (Figure not reproduced here; see the original article.)

Discussion and summary of cost-effectiveness results

The comparison of data from the placebo arms of SELECT and DECISION indicated that the experience of patients differed markedly for PFS, the principal outcome of both trials, to the extent that the PHs assumption is violated. This invalidates the derivation and application of HRs in order to model an indirect comparison to compare the effectiveness of lenvatinib with that of sorafenib. As a consequence, the AG was only able to carry out separate economic assessments of each active treatment against its trial comparator, using common methods and shared parameter values.

In order to assess the importance of the available placebo data (used to represent long-term BSC), a structural sensitivity analysis was carried out substituting the placebo arm data from each trial as the comparator for the intervention treatment. These analyses resulted in very large changes to the AG’s estimated base-case ICERs per QALY gained, and confirmed the suspicion that the two trial populations are not equivalent.

Using published list prices47 in the AG model, neither treatment was found to be cost-effective at a willingness-to-pay threshold of £50,000 per QALY gained. Moreover, neither treatment meets the NICE end-of-life criteria for special consideration [the AG analyses show that both are indicated to have a lifetime mean estimated OS of 55 to 57 months, and a survival gain versus standard of care (BSC/placebo) of > 9 months].

A comparison of the patterns of clinical effectiveness of the two treatments suggests that the proportion of the average gain in PFS, which is subsequently translated to a gain in OS, is very different between the treatments (73% for lenvatinib vs. 24% for sorafenib). This suggests quite different modes of action, which may have important consequences for patients’ long-term prognoses.

The estimated mean time spent in the PFS and OS health states in the AG model show little difference between the two active treatments, so that apparently different net outcome gains are mainly attributable to large differences in the experience of patients in the comparator arms of the two trials. This consistency of outcomes for the active treatments, and the apparently different modes of action, may suggest that these treatments could be used sequentially to generate additional long-term benefit.

Assessment of factors relevant to the NHS and other parties

Lenvatinib and sorafenib are both MKIs and have been approved for use for treating RR-DTC in NHS Scotland (contingent on the continuing availability of PAS prices). Sorafenib is currently available in NHS England via the CDF. Therefore, it is not anticipated that, if recommended by NICE, the use of lenvatinib and sorafenib would have major implications for NHS service provision, particularly as the administration and AEs from both therapies are broadly in line with those of other TKIs already used to treat patients with cancer in the NHS.