Healthcare decision makers and reimbursement agencies are continuously evaluating new healthcare technologies, by assessing clinical and cost-effectiveness data. These organizations, such as the National Institute for Health and Care Excellence (NICE) in the United Kingdom, make recommendations on the funding of new medical technologies on the basis of this evidence. Decision analytic models form the basis of the economic evaluation of these technologies as they are able to synthesize evidence on health consequences and costs from many different sources, and link intermediate outcomes from trial data to long term survival.
The NICE multiple technology appraisal (MTA) process considers evidence from an independent assessment group, together with submissions from the manufacturers of the health technology. In order for these models to be helpful in the decision-making process, it is necessary for them to have credibility and validity (Reference Turner, Raftery and Cooper1). To ensure methodological quality of the submitted models, they are assessed against the requirements of a checklist for methodological quality and generalizability to the United Kingdom (Reference Philips, Ginnelly and Sculpher2). However, even in cases where models have adhered to these requirements for methodological quality, there are often still large differences in the results and conclusions between the models. Differences between models can be due to a combination of differences in parameter values, methods and structures and this can make between-model comparison problematic (Reference Turner, Raftery and Cooper1).
Although many guidelines exist for the development of cost-effectiveness models, and these are continually developed and refined, relatively little guidance has been developed for the assessment of model structure and assumptions, except that these should be described clearly and justified (Reference Philips, Ginnelly and Sculpher2). In this article, we demonstrate the issues surrounding the choice of model structures and assumptions through the use of a comparative study of models developed for the evaluation of first-line treatment of multiple myeloma. These models were used to inform the guidance developed by NICE for these treatments (3). We compare the results from the independent assessment group model (the authors of this article) with those developed by the manufacturers and evaluate any differences between the models’ structure, parameters, assumptions and results.
TREATMENTS FOR MULTIPLE MYELOMA
Multiple myeloma (MM) is the second most common hematological cancer in the UK, characterized by unregulated plasma cell proliferation. Myeloma is not curable, but can be treated with a combination of supportive measures and chemotherapy. The aim is to extend the duration and quality of survival by alleviating symptoms and achieving disease control while minimizing the adverse effects of the treatment. Survival of patients from diagnosis can vary from months to over a decade.
In the United Kingdom, the choice of first-line treatment depends on a combination of factors. The majority of patients are not able to withstand intensive treatment, such as high-dose chemotherapy (HDT) with autologous stem-cell transplantation (SCT), because of age or poor performance status. These patients are therefore offered single agent or combination chemotherapy which is less intensive. This study concerns the use of more recent combination therapies that incorporate drugs such as thalidomide and bortezomib (Reference Messori, Maratea, Nozzoli and Bosi4). The three cost-effectiveness models were developed to compare the cost-effectiveness estimates of bortezomib in combination with melphalan and prednisolone (BMP) and thalidomide in combination with melphalan and prednisolone (MPT) versus melphalan and prednisolone (MP) for the first-line treatment of MM.
METHODS
The models were developed by us (the independent assessment Group, SHTAC) and the manufacturers of the two drugs under consideration (Celgene and Janssen-Cilag). The models are similar in structure, with each being a lifetime model with 6-week cycles and health states that include pre- and postprogression, and death. Health-related quality of life (HRQoL) is incorporated into the models for each of the health states and the models estimate lifetime quality-adjusted life-years (QALYs). The perspective of the analyses was that of the National Health Service (NHS) and Personal Social Services (PSS) in the United Kingdom. The models estimated the lifetime costs and benefits of treatment with discount rates of 3.5 percent, and a base price year for the costs of 2009. We describe the model developed by the authors of this study below (SHTAC model) and then outline the main differences in model structure, and assumptions used in the manufacturers’ models. The results from each of the models are compared and analyzed with respect to the differences in model structure and assumptions. Results have been converted from UK Pounds to Euros for this article (with exchange rate GBP1 = EUR1.2).
DESCRIPTION OF SHTAC MODEL
The SHTAC model was used to compare the cost-effectiveness of BMP and MPT versus MP for the first-line treatment of MM (Reference Picot, Cooper, Bryant and Clegg5). The model used a survival analysis approach to estimate the overall survival (OS) and progression-free survival (PFS) for each of the interventions for a patient with newly diagnosed MM. The parameter values used in the model are shown in Table 1. The model was for the duration of trial follow-up and an exponential distribution was used to extrapolate beyond the length of the trial, that is, after 36 months. Second-line treatment costs were included.
Table 1. Parameters and Distributions for the SHTAC Cost-Effectiveness Evaluation
a Cumulative hazard rates and ratios at 36 months for OS and 24 months for PFS. Calculated using duration in cycles.
bShows the proportion of those receiving second-line treatment, who receive bortezomib, according to their first-line treatment. For the MP and MPT groups on second line therapy that is not BMP, 15% receive CDTa or HDD; for the BMP group 70% receive CTDa and 15% HDD.
B = bortezomib; M = melphalan; P = prednisolone; T = thalidomide; OS, overall survival; PFS, progression free survival; OP, outpatient; BMP, bortezomib in combination melphalan and prednisolone; MPT, thalidomide in combination with melphalan and prednisolone; MP, melphalan and prednisolone.
Two survival curves were constructed for OS and PFS, based on the derived probability of death and progression respectively in each model cycle. The mean survival time for OS and PFS was calculated from the survival curves for OS and PFS, using the area-under-the-curve method. The difference between the two curves provides a direct estimate of the mean time alive following disease progression until death (Figure 1). Survival was classified into three health states: Treatment is the time patients are treated with first-line therapy, posttreatment is the mean time from end of first-line treatment therapy until disease progression and postprogression is the mean time from disease progression until death.
Figure 1. Schematic of the SHTAC cost-effectiveness model.
The clinical parameters for the models were derived from a systematic review of the clinical effectiveness of bortezomib and thalidomide (full details given elsewhere) (Reference Picot, Cooper, Bryant and Clegg5). Three studies were identified that compared MPT with MP (Reference Facon, Mary and Hulin10–Reference Palumbo, Bringhen and Caravita12) and one study was identified that compared BMP with MP (Reference San Miguel, Schlag and Khuageva14). The data from the MP arms of the randomized controlled trials (RCTs), identified by the systematic review, were pooled to form baseline MP OS and PFS curves. The OS and PFS curves were derived from the studies included in the systematic review of clinical effectiveness (Reference Picot, Cooper, Bryant and Clegg5).
Each health state is associated with a health related quality of life (HRQoL) utility estimate which is multiplied by the length of time spent in that state. The total QALYs over the life time of a patient is calculated by aggregating the estimated QALYs from each health state. Values for HRQoL were estimated for the treatment and posttreatment period and for those with complete response by mapping quality of life values from the European Organization for Research and Treatment of Cancer QoL questionnaire C-30 (EORTC QLQ-C30) to the EQ-5D for a cohort of MM patients receiving MP (Reference Gulbrandsen, Hjermstad, Wisloff and Nordic8).
The costs in the model comprise drug treatment, consultation, and monitoring costs and costs for treating adverse events (AEs). Patients remain on drug treatment for the full treatment course unless their disease progresses or they die. Drug unit costs and doses were based on the British National Formulary 2009 (15). The duration of treatment was based on recommendations from the relevant Summary of Product Characteristics (6;7), expert clinical opinion and the published trials. The duration of treatment varies between eight cycles for MPT to nine cycles for BMP. Bortezomib is administered several times per cycle as an intravenous injection, made from a single 3.5-mg vial (7). All patients who remain alive receive second-line therapy and this is assumed to start at the mean time of disease progression for the cohort. Second-line treatment consists of either bortezomib and high dose dexamethasone (HDD); cyclophosphamide, dexamethasone, and thalidomide (CTDa); or HDD. Third-line therapy was not included as it was assumed that most patients would receive lenalidomide, irrespective of the initial treatment, as per NICE guidance (TA 171) (16). Based on clinical advice, we assumed patients attend one hospital consultation every month during their treatment period and one consultation every 3 months thereafter. The monitoring tests used for the management of MM, based on those used for the MMIX RCT, were full blood count, biochemistry, protein electrophoresis, immunoglobin, and urinary light chain excretion. For each comparator, the incidence of AEs was estimated using evidence from the RCTs included in our systematic review of clinical effectiveness (Reference Picot, Cooper, Bryant and Clegg5). AEs included in the model were treatment-related serious (grade 3 and grade 4) AEs. The unit costs of treating AEs were estimated based on those used in a previous NICE technology appraisal for MM (16).
In each cycle, the total costs and QALYs are calculated by multiplying the individual costs and HRQoL by the number of people in the cohort still alive for each of the treatments. The total lifetime costs and QALYs are calculated by aggregating the costs and QALYs for all cycles. The total discounted QALY gain and cost of treatments are then calculated.
DESCRIPTION OF CELGENE MODEL
A Markov model was developed by Celgene, the manufacturer of thalidomide, to compare the difference in the progression of MM and of the costs of treatment when managed with the three different treatment options of MPT, BMP, or MP through a series of different health states
The model has four health states that are defined by the stage of disease progression or the occurrence of AEs. The four states are pre-progression without AEs, pre-progression with AEs, post-progression and an absorbing state of death. All patients start in the pre-progression without AEs health state and move to other states if their condition worsens or they incur an AE. Death can only occur at or after progression. The model does not include overall survival; rather it estimates the survival time before and after progression and then applies different utility values to these health states. Postprogression survival was modeled to be the same across different treatment strategies.
The model has a maximum of twelve treatment cycles for MPT and MP and nine treatment cycles for BMP. Treatment effects were calculated from a random-effects Bayesian mixed treatment comparison of data originating from three RCTs (Reference Facon, Mary and Hulin10;Reference Hulin, Facon and Rodon11;Reference San Miguel, Schlag and Khuageva14). The model does not include treatment costs for second or third-line therapies.
DESCRIPTION OF JANSSEN-CILAG MODEL
A survival cost-utility model was developed by Janssen-Cilag, the manufacturer of bortezomib, to compare the costs and benefits for BMP with those of MPT and MP in people with previously untreated MM who are not eligible for HDT with SCT (19). The model estimates OS and PFS curves for each of the comparators. Survival is partitioned into four different health states: before response to treatment; response but no progression; postprogression and death.
OS and PFS were estimated for MP from a meta-analysis of the MP arms of RCTs for thalidomide and bortezomib. The times to response or death were estimated from life tables constructed directly from the VISTA trial patient level data (Reference San Miguel, Schlag and Khuageva14). For the comparator treatments, relative hazard ratios were taken from a random effects meta-analysis that used OS and PFS summary data. For estimation of the OS hazard for thalidomide, data from five RCTs were used (Reference Facon, Mary and Hulin10;Reference Hulin, Facon and Rodon11;Reference Gulbrandsen, Waage and Gimsing20–Reference Wijermans, Zweegman and van Marwijk22), which included RCTs that had included thalidomide maintenance therapy.
The model includes the costs of second and third-line therapy where second-line treatment consisted of bortezomib + HDD, CTDa, or HDD. All patients receive lenalidomide plus dexamethasone as third-line treatment. HRQoL utility values are assigned to each of the states: before response to treatment, response to treatment without progression, and postprogression, based on a study evaluating chemotherapy followed by SCT in people with MM (Reference van Agthoven, Segeren and Buijt23).
The duration of treatment with MP is seven cycles as per the VISTA trial (Reference San Miguel, Schlag and Khuageva14). For bortezomib, 31.5 vials were used per patient based on usage in the VISTA trial, which is lower than the full treatment course of 52 vials. For thalidomide, the model used an average duration of treatment of 315 days based on the duration reported in the MPT RCTs.
RESULTS
Comparison of Economic Evaluation Results
The results for the manufacturers’ and SHTAC's economic analyses are shown in Table 2. The different assumptions and methodology described above result in a range of estimates for the cost and benefits of the treatment options.
Table 2. SHTAC and the Manufacturers’ Baseline Cost-Effectiveness Results versus MP
B, bortezomib; T, thalidomide; MP, melphalan and prednisolone; ICER, incremental cost effectiveness ratio; EUR, euro; QALY, quality-adjusted life-year; BMP, bortezomib in combination melphalan and prednisolone; MPT, thalidomide in combination with melphalan and prednisolone; MP, melphalan and prednisolone; SHTAC, Southampton Health Technology Assessments Centre.
The incremental cost-effectiveness ratio (ICER) for MPT versus MP varies between EUR10,694 (Janssen-Cilag) and EUR28,057 (Celgene) per QALY gained. The ICER for BMP versus MP varies between EUR12,598 (Janssen-Cilag) and EUR54,029 (Celgene) per QALY gained.
The results of the analyses comparing BMP and MPT vary considerably. For BMP versus MPT, the ICER was estimated as EUR14,288 (Janssen-Cilag), and EUR364,614 (Celgene) per QALY gained. For the SHTAC economic analysis, MPT dominated BMP, that is, MPT is cheaper and more effective, for the base case analysis. Thus the conclusions differ between the analyses, with Janssen-Cilag concluding that BMP is cost-effective compared with MPT, whereas the other two analyses disagree.
The costs vary substantially between the analyses, for example the cost of MP varies between EUR1,638 for the Celgene submission and EUR65,321 for the Janssen-Cilag submission. The costs from the Celgene analysis were lower as they had not included any subsequent treatment costs, whereas the SHTAC analysis included costs for second-line treatment and the Janssen-Cilag included costs for second- and third-line treatment.
The incremental costs for MPT versus MP vary between EUR5,866 (Janssen-Cilag) and EUR23,722 (Celgene). The Celgene submission uses higher dosages of thalidomide (238 mg/day) for longer periods (eleven cycles) than the other two analyses. The incremental costs for BMP versus MP vary between EUR14,690 (Janssen-Cilag) and EUR49,501 (Celgene). These differences are largely due to the assumptions around the number of vials of bortezomib used, with Janssen-Cilag assuming a mean of 31.5 vials used per person, whereas the mean number of vials used is over forty in the SHTAC and Celgene economic evaluations.
The total QALY estimates between the studies are reasonably similar with estimates for each treatment: MP (range, 2.42–2.86), MPT (range, 3.28–3.64), and BMP (range, 3.35–4.03). The incremental QALY estimates for MPT versus MP vary widely and these differences are due to the estimates chosen for the hazard ratio for OS compared with MP, range from 0.55 (Janssen-Cilag) to 1.22 (SHTAC).
Sensitivity Analyses of the SHTAC Model
One-way deterministic sensitivity analyses were previously performed on all the parameters in the SHTAC model, described in detail elsewhere (Reference Picot, Cooper, Bryant and Clegg5), and the model results were found to be most sensitive to the hazard ratio for OS, cost and dosage of the treatment, and the overall baseline survival curve used for MP. Sensitivity analyses were then performed on the SHTAC model to assess the effect of the different assumptions used between the models for the estimates of hazard ratio for OS for MPT and cost for bortezomib.
The estimate for the effectiveness of OS for MPT versus MP varies according to the sources of data chosen. The SHTAC model based its estimate on a systematic review (Reference Picot, Cooper, Bryant and Clegg5), (hazard ratio 0.62). This review excluded trials in which participants had received maintenance therapy with thalidomide. In contrast, the Janssen-Cilag analysis used an OS estimate based on a meta-analysis that included trials in which patients had maintenance therapy. The inclusion of studies with maintenance therapy resulted in lower improvement in OS for MPT versus MP than the meta-analysis without studies of maintenance therapy.
The cost of BMP varies substantially between the analyses. The SHTAC model based its estimate on the number of cycles and doses specified by the VISTA trial and the Summary of Product Characteristics (7;Reference San Miguel, Schlag and Khuageva14). In the model, this amounted to approximately forty-eight vials of treatment. In contrast, the Janssen-Cilag analysis uses a lower cost, due to a lower number of treatment doses, based on the actual number of treatment vials used in the VISTA trial, that is, 31.5 vials. The reduced number of vials may be due to fewer treatment cycles due to early discontinuation of treatment.
In the sensitivity analyses the hazard ratio for OS for MPT versus MP was varied from 0.62 (base case) to 0.8. The cost of bortezomib was varied according to the proportion of early discontinuation from treatment. In each cycle, the proportion who discontinued treatment was varied between 0 (baseline) and 10 percent. The results of the sensitivity analyses are shown in Table 3.
Table 3. Deterministic Sensitivity Analyses from the SHTAC Model
SHTAC, Southampton Health Technology Assessments Centre; BMP, bortezomib in combination melphalan and prednisolone; MPT, thalidomide in combination with melphalan and prednisolone; MP, melphalan and prednisolone; OS, overall survival.
The sensitivity analyses show that these parameters have a large effect on the model results, when comparing BMP to MPT, while the comparison between MPT vs. MP and BMP versus MP are more robust. In the base case analysis, MPT dominates BMP, as MPT is both cheaper and more effective than BMP. With the alternative hazard ratio for overall survival, the ICER is EUR44,294 per QALY gained for BMP versus MPT. For both the alternative hazard ratio for OS and a reduced cost for BMP, the ICER is EUR24,001 per QALY gained. These results are consistent with the results estimated by each of the models with the assumptions they used.
This uncertainty within the three models was explored further by conducting a probabilistic sensitivity analysis (PSA). All parameters were sampled probabilistically using the ranges and values shown in Table 1. However, in this case the hazard ratio OS for MPT versus MP was varied between the two estimates (0.62–0.8) using a beta distribution, and the treatment discontinuation rate for BMP was varied between the two estimates (0–10 percent) using a uniform distribution to represent the uncertainty between the parameter inputs in the three models. One thousand simulations were run. The cost-acceptability curve is shown in Appendix 2 and indicates that at lower willingness to pay thresholds of between EUR24,000 and EUR60,000 per QALY gained (GBP20,000 and GBP50,000 per QALY gained) MPT has the highest probability of being cost-effective. For a willingness to pay threshold higher than EUR60,000 per QALY gained, BMP is the treatment with the highest probability of being cost-effective.
DISCUSSION
In the United Kingdom, NICE has provided guidance for the use of bortezomib and thalidomide for first-line treatment of MM, based upon the clinical and cost-effectiveness evidence (3). It recommended thalidomide, in combination with an alkylating agent and a corticosteroid, as a cost-effective option for the first-line treatment of MM, with bortezomib recommended for those unable to tolerate or who have contra-indications to thalidomide. NICE assessed evidence from the manufacturers of thalidomide and bortezomib, from our clinical and economic evaluation, and the opinion of clinical experts and service users. For the estimation of the clinical effectiveness of MPT, the NICE appraisal committee decided that it was appropriate to exclude trials in which participants received maintenance therapy with thalidomide. For the estimation of the costs of bortezomib, the NICE appraisal committee accepted the number of bortezomib vials during treatment would be 31.5 (as proposed by Janssen-Cilag).
The approach taken in this article to compare cost-effectiveness models has shown that it is possible to compare and evaluate cost-effectiveness models, by identifying the most influential differences with respect to model structure, parameter and assumptions, and making a judgment on these differences. Although this approach is necessary within a decision-making context for national regulatory bodies, such as NICE, it is not common in the medical literature. Turner et al. (Reference Turner, Raftery and Cooper1) investigated the feasibility of between-model comparison by comparing four UK models developed for coronary heart disease. They concluded that, while checking between model consistency requires a potentially large investment in terms of researcher time and effort, there were situations where it would be useful for decision makers, for example where there were large differences in model results, or where results from different models cross decision-maker's thresholds. In this case, there may be considerable uncertainty to the implications of results to decision making. They noted that often detailed information on the models was restricted by word count limitations and recommended that modeling articles should include detailed Web appendices to aid replication and checks of between-model consistency.
In the current article, comparison was aided as each of the models involved had been constructed to conform to NICE guidelines for technology appraisals (24), and full details of the economic models were available and the authors had access to electronic versions of the models. Although there were many differences between the model results, it was possible to identify those parameters that primarily caused the differences and examine them for validity. Through analyzing these differences, we concluded that each modeling approach and structure was appropriate and that many of the differences between the models in terms of parameter inputs and assumptions have negligible effect on the model results.
Haji Ali Afzali and Karnon (Reference Haji Ali Afzali and Karnon25) propose the concept of reference models for specific disease areas, which could be made available to sponsors submitting health technologies for assessment by reimbursement bodies. These resulting models would be more likely to represent a comprehensive, unbiased representation of the disease. They argue that there is a diversity of model structures within disease areas which increases the complexity of comparing evaluations of alternative technologies for the same conditions. Furthermore, they consider that the consequences of inconsistencies in the choice of model structure within a specific disease for submissions to a national regulatory body can lead to inconsistent reimbursement decisions because changes in model structure and analysis can produce substantially different results. The comparison of existing models is a useful method to reach consensus between modelers. For example, the Mount Hood Challenge (26) was a forum for computer modelers of diabetes to discuss and compare models and identify key areas of future development to advance the field. By performing systematic comparisons and validation exercises enabled the identification of key differences among the models, as well as their possible causes and directions for improvement in the future.
CONCLUSIONS
A comparison of alternative cost-effectiveness models is not straightforward, and often differences in model structure, parameters, and assumptions are hard to identify and lead to large differences in results and conclusions. By comparing models developed for assessing treatments for multiple myeloma, we demonstrated that it was feasible to compare models, which then aided decision makers in making reimbursement decisions.
CONTACT INFORMATION
Keith Cooper, PhD (k.cooper@soton.ac.uk), Senior Research Fellow, Southampton Health Technology Assessments Centre (SHTAC), University of Southampton, UK
Joanna Picot, PhD, Senior Research Fellow, SHTAC, University of Southampton, UK
Jackie Bryant, MSc, Principal Research Fellow, SHTAC, University of Southampton, UK
Andrew Clegg, PhD, Professor of Health Services Research, and Director of SHTAC, University of Southampton, UK
CONFLICTS OF INTEREST
All authors report they have received a grant to their institution from National Institute for Health Research to undertake research.
APPENDIX 1
Cost-acceptability curve from the probabilistic sensitivity analysis from the SHTAC model (Appendix Figure 1) available online at http://dx.doi.org/10.1017/S0266462313000615.
