### INTRODUCTION

- Top of page
- Abstract
- INTRODUCTION
- METHODS
- RESULTS
- DISCUSSION
- Acknowledgements
- REFERENCES
- Supporting Information

Rheumatoid arthritis (RA) is a chronic, inflammatory systemic disease. Its most prominent feature is persistent synovitis in the peripheral joints that can lead to serious morbidity (1, 2). Because there is no definite cure for the disease, the goal of pharmacotherapy is to achieve and maintain low disease activity and to prevent disease progression and the resulting joint destruction and functional disability.

There are several strategies for treating RA (3–5). A usual approach in the Netherlands with newly diagnosed patients is to start with sulfasalazine and, if there is insufficient effect or toxicity, to prescribe methotrexate. If this is also ineffective or leads to major toxicity, a range of other classic disease-modifying antirheumatic drugs (DMARDs) may be used, though they are often only moderately effective and may have considerable toxicity (4). Fortunately, new drugs have recently been introduced, including agents against tumor necrosis factor (TNF) α. TNF-blocking therapy has been effective in patients for whom methotrexate treatment failed (6–8). Another relatively new drug is leflunomide, which has been shown to be at least as effective as methotrexate in the treatment of RA (9).

In the Netherlands, the indication for TNF-blocking treatment is active RA (Disease Activity Score including 28 joints [DAS28] > 3.2) in patients who have been adequately treated with at least 2 DMARDs of which 1 was methotrexate (25 mg/week if tolerated) with insufficient effect. Within 8–12 weeks, an adequate response (an improvement in the DAS28 of ≥1.2) should be present for the therapy to be continued. This is in accordance with a consensus statement of an international group of rheumatologists (10). For leflunomide, a similar indication might apply (9).

The major drawback of these new drugs (especially the TNF-blocking agents) is their high price. Therefore, studies of the cost-effectiveness of treatment strategies including TNF-blocking treatment and leflunomide for the above indication are relevant. Because TNF-blocking treatment and leflunomide have only recently been introduced, no long-term data on costs and effects on quality of life are available. Modeling can be used to overcome the problem of incomplete data and to extrapolate short-term endpoints to long-term outcomes. Modeling involves mathematically combining (empirical) data from different sources (11, 12).

Because RA is a chronic disease characterized by periods of high disease activity alternating with periods of low disease activity, simple decision analysis, which considers outcomes only at 1 time point (instead of outcomes over time), does not suffice. A Markov model considers outcomes over time and might therefore be an appropriate model for RA (13).

In a Markov model, a hypothetical cohort of patients is simulated. The patients are classified in a finite number of health states (Markov states) defined by the severity of the disease. These Markov states can be valued in terms of costs and effects. Development of the disease and the effectiveness of treatment are represented as transitions from one Markov state to another in a defined time span (cycle length). These transitions occur with a certain probability (transition probabilities). According to these transition probabilities, the simulation cohort is distributed over the Markov states, inducing expected costs and expected effects over a certain period (time horizon of the analysis) (11–13).

The aim of this study was to compare the expected effects on disease activity, quality of life, and costs over 5 years of different treatment strategies (including TNF-blocking treatment and leflunomide treatment) in patients who satisfy the indication for TNF-blocking agents in the Netherlands. For this cost-effectiveness study, a societal perspective and a third-party payer perspective were used (11).

### METHODS

- Top of page
- Abstract
- INTRODUCTION
- METHODS
- RESULTS
- DISCUSSION
- Acknowledgements
- REFERENCES
- Supporting Information

A Markov model consisting of Markov states defined by the DAS and a cycle length of 3 months was used (14). Markov states for remission (DAS < 1.6), low disease activity (1.6 < DAS < 2.4), moderate disease activity (2.4 < DAS < 3.7), and high disease activity (DAS28 > 3.7) were defined (15). A time limit of 5 years (20 cycles) was used for the analysis; a longer time horizon implied too many assumptions (Figure 1). Cost and quality of life values were assigned to the Markov states using data from a 48-week multicenter trial with methotrexate that included 411 patients (see Appendix A, available at http://www.interscience.wiley.com/jpages/0004-3591:1/suppmat/index.html). (16). In this trial, medical (consultations physicians, surgery, etc.) and nonmedical (absence from paid labor, travel expenses, etc.) costs were measured using a questionnaire and a patient diary. Costs for drug treatment were excluded because they were partly due to trial medications and were estimated separately (see below). For quality of life valuation, a utility value (derived using the Euroqol questionnaire) was assigned to the Markov states. Utilities refer to the preferences that individuals or society have for a particular health outcome, as measured on a scale between 0 (death) and 1 (perfect health) (11, 12) and can be used to calculate quality-adjusted life years (QALYs). The simulation cohort starts in the Markov state for high disease activity because we considered patients satisfying the current indication for TNF-blocking treatment in the Netherlands, which implies that patients have active disease (Figure 1). In this evaluation, the following treatment strategies were considered: 1) usual treatment (U); 2) treatment with leflunomide, in the case of nonresponse after 3 months, switch to usual treatment (Lef); 3) TNF-blocking treatment, in the case of nonresponse after 3 months, switch to usual treatment (TNFb); 4) treatment with leflunomide, in the case of nonresponse after 3 months, switch to TNF-blocking treatment, in the case of nonresponse to TNF-blocking treatment after 3 months, switch to usual treatment (Lef-TNFb); 5) TNF-blocking treatment, in the case of nonresponse after 3 months, switch to leflunomide treatment, in the case of nonresponse to leflunomide after 3 months, switch to usual treatment (TNFb-Lef). Assessment of response after 3 months of treatment is in accordance with the consensus statement that TNF-blocking treatment should be stopped when a patient does not respond well at 8–16 weeks (10).

#### Costs of drug treatment.

The costs of TNF-blocking treatment and leflunomide in the Netherlands were derived from the College Voor Zorgverzekeringen (Health Care Insurance Board), and were €12,706 per year or €3,177 per cycle for TNF-blocking treatment and €908 per year or €227 per cycle for leflunomide. The cost of usual care for patients who satisfied the indication was calculated from an open longitudinal study of early RA (disease duration <1 year with no prior use of DMARDs) that has been underway since 1985 at the department of Rheumatology of the University Medical Center Nijmegen in the Netherlands. From this study, patients were selected who stopped treatment with sufasalazine and methotrexate due to insufficient effect or toxicity (as noted in the medical record) and had a high disease activity (DAS > 3.7). The followup data from then onward were used to calculate medication use by multiplying the times for which patients used different DMARDs and corticosteriods by the usual dosages. Costs for these treatments were derived using the prices of the drugs. These costs were found to be €150 per year or €38 per cycle.

#### Effectiveness of treatment strategies.

In a Markov model, the effectiveness of treatment is expressed as transition probabilities, the probability that a transition from one state to another in a defined time span will occur for all transitions. The data needed to calculate these transition probabilities for the different treatments were derived from the following 3 sources.

1) The followup data of patients from the open study who stopped treatment with sulfasalazine and methotrexate due to insufficient effect or toxicity and who had high disease activity were used to calculate the transition probabilities for usual treatment.

2) A dataset made available by Wyeth Pharmaceuticals (Madison, NJ) from clinical trials of monotherapy with etanercept in patients who failed DMARD treatment (1–4 DMARDs) and of combination therapy with methotrexate in patients with insufficient response to methotrexate alone was used (6, 7). From these trials, patients with high disease activity at baseline and a good or moderate response to etanercept (according to the European League Against Rheumatism [EULAR] criteria) after 3 months were selected (15). For the calculation of transition probabilities, data from these 2 trials were combined because they did not differ in efficacy or patient characteristics.

3) Because no detailed data concerning treatment with leflunomide was available, transition probabilities could not be calculated so information from a published clinical trial was used (17). The American College of Rheumatology (ACR) 20%, 50%, and 70% response criteria after 1 and 2 years of treatment, as reported in the article, were considered to represent the Markov states for moderate disease activity, low disease activity, and remission, respectively. It was assumed in the model that the distribution remained constant after 2 years. The patients in the published clinical trial did not satisfy the indication. It seems likely that leflunomide would be less effective when used in patients satisfying the indication. In the analyses, it was therefore arbitrarily assumed that leflunomide treatment is 25% less effective than was reported in the published clinical trial.

#### Economic evaluation.

For each of the treatment strategies, a specific Markov model was used with the same structure (Figure 1) and the same cost and utility values of the Markov states, but the models used specific transition probabilities and costs for the drug treatments. Using these models, the expected costs and expected effects (outcomes of the model) were compared between the alternative treatment strategies.

##### Baseline analysis.

The decision model was calculated with the simulation cohort using the point estimates for the transition probabilities and the cost and utility values of the Markov states. Expected patient-years spent in the Markov states (disease activity), cost (medical as well as total), and QALYs over 5 years were compared between the treatment strategies. Cost-effectiveness ratios and incremental cost-effectiveness ratios (ICERs) were calculated. To calculate a cost-effectiveness ratio, the costs of a treatment are divided by the effect of this treatment. To calculate an ICER, the *additional* costs of a treatment as compared with an alternative are divided by the *additional* effect as compared with that of the alternative (11, 18). The primary outcome measures of the economic evaluation were the ICERs for the comparisons of the treatment strategies using all costs (medical as well as nonmedical) and QALYs as effectiveness measure. Discounting with 4% per year for the cost as well as the effects was used.

##### Uncertainty analyses.

Model uncertainty was explored using probabilistic sensitivity analysis. In a probabilistic sensitivity analysis, uncertainties in all input parameters (i.e., transition probabilities, cost, and utility values of the Markov states) of the model are considered simultaneously (19). Instead of a point estimate, a distribution (representing the uncertainty) is defined for all input parameters. When the decision model is calculated, the simulation cohort is simulated through the model repeatedly, and every time the cohort is simulated, a value from the distributions is used (with the chance depending on the distribution) for all input parameters. In this way, the expected outcomes of the model are also represented by distributions, representing the uncertainty of the outcomes.

Distributions were specified for the transition probabilities, the cost, and utility values of the Markov states and for the response of etanercept (EULAR good/moderate) and leflunomide treatment (ACR 20%) after 3 months, and were based on the data. The effectiveness of leflunomide was varied from 50% less effective to equally effective compared with the extrapolation from the effectiveness reported in the published clinical trial, using a uniform distribution. The simulation cohort was simulated 1,000 times. To determine the relative importance of the uncertainty in the different input parameters for the primary outcome measure, correlations between the input parameters and the outcome were calculated in this analysis. Important model parameters as defined by this correlation were varied in a one-way sensitivity analysis.

The Markov models were built using the spreadsheet program Microsoft Excel (Redmond, WA); the program Crystal Ball (version 4.0, Decisioneering, Denver, CO) was used for the probabilistic, sensitivity analysis.

### DISCUSSION

- Top of page
- Abstract
- INTRODUCTION
- METHODS
- RESULTS
- DISCUSSION
- Acknowledgements
- REFERENCES
- Supporting Information

In this modeling study it was found that, over a period of 5 years, the expected effect of treatment strategies that include etanercept on disease activity and QALYs is larger than the expected effect of usual treatment or leflunomide treatment in RA patients satisfying the indication for TNF-blocking agents in the Netherlands. This reduces medical and nonmedical costs by about 16% and 33%, respectively, as compared with usual treatment; this does not, however, counter-balance the extra costs of the drug treatment for the treatment strategies starting with etanercept. When these costs are included, the extra costs for the strategies starting with etanercept are considerable, leading to large (incremental) cost-effectiveness ratios.

The strategy of starting with leflunomide and, in case of nonresponse, starting etanercept seems to have an effect similar to that of treatment strategies starting with etanercept, but the total costs are lower. This results in more acceptable ICERs for the comparison with usual treatment or leflunomide treatment alone. The ICERs for the different comparisons were sensitive to the effectiveness of the treatments, and the utility values of the Markov states, but the strategies using etanercept were almost always more effective and the strategies starting with etanercept were always more expensive.

This study has some deficiencies. The first of these relates to the populations used to derive efficacy data for the treatments. The indication for TNF-blocking treatment in the Netherlands is patients with a DAS28 > 3.2 who have been adequately treated with at least 2 DMARDs. This relatively low level of disease activity was chosen to present rheumatologists with the opportunity to prescribe TNF-blocking treatment in individual patients with a lower DAS. However in clinical practice and in clinical trials, patients generally have a DAS > 3.7. Therefore for this study, a DAS28 > 3.7 was chosen as the definition of active disease. Although all patients had active disease, the requirement of DMARD treatment may not have been satisfied for all patients because some patients had not been treated with a dosage of methotrexate that is currently considered adequate before stopping the treatment (for the open study and the clinical trial with etanercept), and some patients may not have taken at least 2 DMARDs that resulted in failure (in the clinical trials with etanercept). However, we selected patients such that most of them would meet the inclusion criteria for TNF-blocking treatment in the Netherlands. Although patients from the leflunomide trial did not satisfy the indication and no selection was possible because we did not have detailed data, we assumed leflunomide to be 25% less effective with this indication. Varying this decrease in effect, in a separate analysis, showed that leflunomide would have to be more than 90% less effective for the ICER to be below €100,000/QALY for the comparison of TNFb-Lef versus Lef-TNFb, which is not probable. For the effectiveness of leflunomide, an extrapolation of the ACR 20%, 50%, and 70% response criteria to the Markov states for moderate disease activity, low disease activity, and remission, respectively, was used. This probably leads to an underestimation of effectiveness and is thus a conservative estimate (21).

A second concern relates to the fact that this study is not a direct comparison between treatment strategies but a modeling study. Before using the model for this cost-effectiveness study, it was validated according to a method proposed by McCabe and Dixon (22) and found to be valid for use in economic evaluations in RA (14). Disease activity was found to be the most important factor determining the cost and utility values for the Markov states, and the toxicity of the methotrexate treatment did not influence these values, using data from the methotrexate trial (14, 22). No clear relationship of costs with disease duration was found. This indicated that the data from the trial with methotrexate could be used to calculate the cost and utility values for the Markov states in our context. Furthermore, it was found that transition probabilities calculated from data from the open inception cohort could be extrapolated to 5 years. The expected distribution of the patients over the Markov states (results of the model) resembled the actual data fairly well (14). In modeling, assumptions are unavoidable, but they should be made realistically, or even conservatively, for the results to be credible. Extensive sensitivity analyses are necessary to judge whether the results and conclusions can be regarded as robust. This is one of the main advantages of modeling (12). An assumption in our model was that the costs for drug treatment were equal for all Markov states. This assumption is valid because patients do not usually stop taking or change their medication when they go into remission or low disease activity. Use of nonsteroidal antiinflammatory drugs (NSAIDs) was not included in the costs of drug treatment. However, large differences in NSAID use are not expected between the different treatment strategies. A further assumption in this model was that when a patient responds well to treatment with TNF-blocking treatment or leflunomide treatment, this treatment would be continued for the rest of the 5 years. The percentage of patients with a response after 3 months was varied in the sensitivity analyses and did not have a large influence.

The time horizon for our analyses was 5 years. It is possible that there would be additional savings from the use of TNF-blocking treatment (such as prevention of surgical procedures, admission to nursing homes, etc.) outside the 5-year time horizon of the analyses. These savings are uncertain and rely on the long-term effects of TNF-blocking treatment on disease activity and toxicity. Recently, reference case recommendations for economic evaluations in rheumatoid arthritis were published by the Outcome Measures in Rheumatology Clinical Trials group (23). Our study is in accordance with almost all of the recommendations. Only mortality was not accounted for in this study, however it is not expected that over the study horizon of 5 years, a difference in mortality is apparent between the treatment strategies.

Notwithstanding the limitations of the study we think the results of this analysis are valid because the extensive sensitivity analyses did not change the conclusions of the study.

In this study, the cost effectiveness of treatment strategies using etanercept was investigated. Etanercept was used merely as an example of TNF-blocking treatment. It is not clear that the drugs differ in effectiveness regarding RA and Infliximab, for example, is probably even more expensive due to the mode of administration. We therefore consider the conclusions of this study to be representative for other TNF-blocking agents. Extrapolating these findings to other countries or comparing them with results for other countries should be done with caution because the analyses and the cost values were specific to the situation in the Netherlands.

Few studies have investigated the cost effectiveness of treatment strategies for a patient population as was done in this study. Choi et al (24) compared treatment options for methotrexate-resistant RA. In this study, a reference analysis of methotrexate therapy for methotrexate-naive RA compared with no treatment, costing $1,100/ACR 20% response, was used to compare cost-effectiveness ratios. The authors concluded that triple therapy (hydroxychloroquine, sulfasalazine, and methotrexate) is probably cost effective in this patient group because this costs only 1.3 times more per patient ACR 20% outcome, and etanercept in combination with methotrexate had a high cost-effectiveness ratio of $42,600/ACR 20% response (38 times more compared with the reference case). The other treatment options (continuing methotrexate, etanercept monotherapy, cyclosporine plus methotrexate, and no second-line agent) were dominated by methotrexate therapy. This study had a time horizon of only 6 months and did not use non–disease-specific outcome measures, such as utility, for comparison with cost-effectiveness studies in other diseases. Leflunomide treatment was not considered in this cost-effectiveness study. In our study, triple therapy was not one of the comparators because this kind of therapy is not often used in the Netherlands.

Wong et al (25) estimated the cost-effectiveness of 54 weeks of infliximab as compared with methotrexate treatment in patients that had active disease despite treatment with methotrexate. They projected the 54-week results from the Anti-Tumor Necrosis Factor Trial in Rheumatoid Arthritis with Concomitant Therapy Study Group (ATTRACT) into lifetime economic and clinical outcomes using a Markov model with Markov states defined by the health assessment questionnaire (HAQ). Because the effectiveness of infliximab beyond the 54 weeks was not available, they assumed that infliximab would be discontinued and the patients would start methotrexate treatment. This assumption is probably not valid and results in an underestimation of the ICER. Furthermore, infliximab treatment was compared with methotrexate treatment in a group of patients that did not respond well to treatment; this is probably not the best available alternative or the usual treatment for these patients. The authors concluded that 54 weeks of infliximab treatment should be cost effective.

Recently, Brennan et al compared a DMARD sequence with etanercept as a third line to a DMARD sequence excluding etanercept in RA patients that did not get improvement from 2 DMARDs (26). They calculated lifetime economic and quality of life outcomes using a decision model that focused on the progression of HAQ disability and simulates individual patient histories according to the chance of responding to treatment (ACR 20%), adverse event, and death. Although they found a low ICER (£16,330 per QALY) as compared with our study, the results are difficult to compare because of the different modeling approach.

We found 2 studies (27, 28) that investigated the costs of RA care in patients who started treatment with leflunomide as compared with etanercept (and infliximab) (28). These studies concluded that RA care costs in the leflunomide group were lower. The treatments were not compared on effectiveness, however, and one of the studies was retrospective (27). The costs were based on databases containing patient-level medical and pharmaceutical claims. It is unclear whether the patient populations of these 2 studies are comparable with the patient population used in our study.

One study compared a DMARD sequence including leflunomide treatment with a DMARD sequence without leflunomide in patients with RA severe enough to require treatment with methotrexate (29). They used a decision model in which patients could respond to therapy or not (ACR 20%), continue therapy, stop therapy, and experience adverse events. The time horizon of the analyses was 5 years. It was concluded that adding leflunomide as a new option to a conventional sequence of DMARDs was reasonably cost effective, with a ICER of $54,229 or $71,988 per QALY gained, depending on the method of calculating QALYs. The population of this study was not comparable to the patients in our study.

We conclude that treatment strategies including TNF-blocking treatment are probably the most effective for patients for whom 2 DMARDs have failed, of which 1 is methotrexate. From these strategies, a strategy starting with leflunomide, and in the case of nonresponse, switching to TNF-blocking treatment, probably results in the most favorable ratio between the extra costs and effects.

We would like to recommend that in addition to the current indication for TNF-blocking treatment in the Netherlands, unless it is contraindicated, leflunomide should also be used for at least 3 months, with insufficient effect, before starting TNF-blocking treatment.