Is treatment with ICS and LABA cost-effective for COPD? Multinational economic analysis of the TORCH study

Chronic obstructive pulmonary disease (COPD) is an incurable, debilitating and progressive disease that can be fatal. Patients typically suffer a prolonged respiratory and systemic functional decline punctuated by acute exacerbations often requiring hospital treatment 1. The incidence of COPD is increasing world wide, making it one of the greatest disease burdens in many countries 2. Pharmacological treatments that can extend life and improve the quality of life of COPD patients have the potential to offset this burden. Nevertheless, if they are to become routinely available, health systems around the world need evidence of not only the effectiveness but also the cost-effectiveness of such interventions.

Despite the importance of COPD as a major disease burden, there is a paucity of high-quality cost-effectiveness data concerning potential treatments. The TOwards a Revolution in COPD Health (TORCH) study has recently reported its clinical findings 3. This multinational 3-yr study was the first trial of pharmacotherapy to adopt mortality as a primary end-point in COPD, and included a rigorous tracking of deaths in order to ensure complete follow-up. Furthermore, the TORCH study prospectively collected data on medical resource use and health-related quality of life for patients in the trial. This provides an opportunity to assess the impact of study treatments on health economic outcomes in order to judge whether or not they are worthwhile from the perspective of the healthcare system.

The aim of the present article is to report the analysis of the economic data from the TORCH study in order to inform decision-makers of the potential cost-effectiveness of alternative treatments for COPD. Special consideration is given to the need to estimate cost-effectiveness results that are relevant to decision-makers in particular jurisdictions covered by this multinational trial.

METHODS

Analysis

Figure 1⇓ provides an overview of the methods used to estimate the components of the cost-effectiveness calculation. In the first stage, multivariate regression models were fitted separately for survival, study medication cost, other medical cost and preference data. In the second stage, estimates of these four components were combined to give total cost, QALY number and cost-per-QALY estimates. This approach is described in more detail below, with particular emphasis on the methods for handling missing data due to attrition, which affected the secondary end-points of medical resource use and health status, and the adjustment of the estimates to the regions of the study.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1—

Schematic of analyses. In the first stage, multivariate regression models were fitted to the TOwards a Revolution in COPD Health (TORCH) data, with treatment and region as covariates (alongside other prognostic factors). In the first model, a parametric (Weibull) time-to-event model was fitted directly to the mortality data since these data were complete. Owing to attrition, there were missing data in the preference (European quality of life five-dimension (EQ-5D)) and cost data; therefore, the second model estimated a set of weights based on the estimated propensity for observations to be missing via logistic regression. These weights were then used in weighted regression analyses of the preference and cost data in models III, IV and V. In the second stage, the fitted models were used to estimate the components of the cost-effectiveness calculation. The health-related quality of life (HRQoL) preference estimates were combined with the number of life years to estimate the number of quality-adjusted life years (QALYs), and the study medication (Rx) and other medical service use costs were combined to estimate total cost.

Missing data

There were no missing survival data, except for 1 yr of follow-up in one subject 3. Withdrawal from the study or missed follow-up visits resulted in 22.6% of the medical service use data and 26.1% of the EQ-5D data being missing. This missingness was addressed by use of inverse probability weighting 13, 14. Weights were defined as the reciprocal of the estimated probability of being observed (generated from logistic regression with missingness status as the dependent variable) and then utilised in weighted regression models to estimate study medication cost, other medical cost and preference scores adjusted for the missing values. Further details relating to the modelling of each component of the cost-effectiveness calculation are given below.

Survival

Mortality was estimated as a function of explanatory prognostic factors and treatment allocation by use of a parametric Weibull survival model. The number of life years associated with each treatment arm (overall and as a function of other important prognostic factors, including geographical region) were estimated as the area under the Weibull curve.

Cost

Separate estimates of the cost of study medication and other costs were made, recognising that the treatment cost was part of the trial protocol, whereas other costs related to the health system. Generalised linear models were used to estimate study medication cost and other costs, utilising separate link functions and families guided by the fit to the data 15.

Preferences

Ordinary least squares regression was employed to model the EQ-5D scores. The number of QALYs in each treatment arm were calculated by weighting the per-period estimates of preference scores by the probability of survival at the corresponding visit.

Regional estimates

The TORCH study included participants from five geographic/economic regions of the world: the USA, Western Europe, Eastern Europe, Asia/Pacific and all other countries. Since no validated EQ-5D instrument was available for the languages in the Asian/Pacific region, the economic analysis presented here relates only to the remaining four regions. Regional estimates of costs and number of QALYs for each of the four arms of the TORCH study were generated by including regional main effect terms alongside the treatment indicators and other explanatory variables in each of the four regression models described above. Treatment-by-region interaction terms were evaluated for each regression, but these terms were only included in the final regression model if the joint test of significance gave a p-value of <0.05. Treatment, region and treatment-by-region interaction terms were estimated taking into account the baseline characteristics of subjects in that region. It should be noted that the economic results are aggregated and presented at the regional rather than the country level since the number of patients recruited by individual countries was low, with the exception of the USA, which is considered as a region in its own right.

Cost-effectiveness

Trial-wide cost-effectiveness ratios are reported, as are region-specific ratios, which were calculated by use of the region-specific estimates of costs and number of QALYs described above. The method of recycled predictions 15 was employed to estimate costs and number of QALYs, and these were discounted at a rate of 3% per year. Cost and QALY number results are presented on the cost-effectiveness plane 16 for the US and Western European regions. The US region was chosen since this is the region with the highest healthcare costs, and the Western European region was chosen as being the region in which the majority of formal reimbursement authorities who make use of cost-effectiveness data reside. Incremental cost-effectiveness ratios are presented in detail for the primary comparison of SFC versus P (the regulatory comparison) and for SFC versus Salm (since Salm is likely to be the standard of care in most countries).

Uncertainty

sem were estimated by use of a nonparametric bootstrap, and 95% confidence intervals (CIs) for incremental cost-effectiveness ratios were based on the bootstrap percentiles 17. Cost-effectiveness acceptability curves 18 were employed to represent the joint uncertainty across all four treatment options included in the TORCH trial for the US and Western European regions. In common with the majority of cost-effectiveness studies reported alongside clinical trials, the prices in the present analysis were treated as fixed weights. As a result, statistical measures of uncertainty in the cost of each treatment option relate to the sampling variability associated with the observed resource use in the study.

RESULTS

Cost-effectiveness

The trial-wide estimate of cost per QALY gained for the comparisons of SFC versus P and of SFC versus Salm is shown in table 3⇓, together with cost-effectiveness estimates by region. SFC had a trial-wide point estimate of 43,600 USD compared with P (95% CI 21,400–123,500 USD). The trial-wide point estimates for Salm versus P (197,000 USD) and FP versus P (78,000 USD; data not shown) were less favourable, whereas the trial-wide cost-per-QALY estimates for SFC versus Salm (26,500 USD) and SFC versus FP (27,000 USD; data not shown) were necessarily more favourable. The cost-effectiveness estimates for all therapies in the USA were greater than those in other regions, reflecting the higher cost structure in the USA. In the USA, the point estimate for SFC versus P was 77,100 USD, which, again, was less than the US point estimates for Salm versus P (291,000 USD) and FP versus P (124,000 USD).

View this table:

• View inline
• View popup

Table 3—

Regional cost^# and number of quality-adjusted life years (QALYs) for salmeterol (Salm)/fluticasone propionate combination (SFC) versus placebo (P) and SFC versus Salm

Ratios from the comparison of SFC versus Salm were lower than those from the comparison of SFC versus P because Salm produces fewer additional QALYs than does SFC, at a higher cost per QALY, and the same was true for the SFC versus FP comparison (results not shown, but available from the authors on request). This relationship is illustrated in figure 2⇓, which plots the incremental cost and number of QALYs gained for the three active treatments compared to P in the Western European region and in the USA. In these plots, the gradient of the line from the origin to the plotted point represents the point estimate for the cost-effectiveness ratio for the therapy compared to P. The difference in the gradient of the line between the different active therapies represents the incremental cost-effectiveness ratio for the relevant therapy comparison. In both regions, Salm and FP produce fewer QALYs than does SFC, at a higher cost-per QALY ratio, compared to P, and this same pattern was apparent in the Eastern European and other regions (data not shown, but available from the authors on request).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2—

Cost-effectiveness plane for all three treatments (salmeterol (Salm), fluticasone propionate (FP) and Salm/FP combination (SFC)) compared to placebo for: a) the Western European region; and b) the US region. The gradient of the line connecting the SFC point estimate to the origin (placebo) represents the incremental cost-effectiveness of SFC. The Salm and FP point estimates lie above and to the left of this line, indicating that, although these options are of lower cost than SFC, the proportional increase in the number of quality-adjusted life years (QALYs) per additional US dollar (USD) spent was less than that of SFC. They are both, therefore, less efficient than SFC in producing QALYs given their cost.

Figure 3⇓ shows the cost-effectiveness acceptability curves for the simultaneous comparison of uncertainty between the four arms of the TORCH trial in the Western European and US regions. These show the likelihood that each treatment arm is cost-effective for a given willingness to pay for an additional QALY. At a threshold willingness to pay for a QALY of 50,000 USD, the (one-tailed) likelihood for each of SFC, P, Salm and FP being cost-effective are ∼0.93, ∼0.04, ∼0.01 and ∼0.03, respectively, in the Western European region. Assuming a threshold willingness to pay in the USA of 100,000 USD per QALY gained, the corresponding likelihoods are 0.65, 0.21, 0.04 and 0.10. The results for the Eastern European and other regions are similar to those of the Western European region (data not shown, but available from the authors on request).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3—

Cost-effectiveness acceptability curves for all treatment arms (–––: placebo; –– –– ––: salmeterol (Salm); – – – –: fluticasone propionate (FP); --------: Salm/FP combination) for: a) the Western European region; and b) the US region. Each curve represents the proportion of times in 1,000 bootstrap replications of the data that a treatment is preferred as a function of the threshold cost-effectiveness (ceiling cost-per-QALY ratio). The graph should be read vertically, with a given threshold cost-effectiveness being represented as a vertical line (·······; lines shown represent possible decision thresholds of 50 thousand US dollars (TUSD) in Western Europe and 100 TUSD in the USA). For each bootstrap replication at a particular threshold, one treatment is preferred. The proportion of times (across 1,000 replications) that each treatment is preferred is given by the cost-effectiveness likelihood at the intersection of the vertical line and the therapy curve. The sum of the four intersection likelihoods is 1.0.

DISCUSSION

The present article reports on the cost-effectiveness of ICSs and LABAs, alone or in combination, for the treatment of COPD, based on an economic analysis of the recently reported TORCH study 3. The trial-wide point estimate for the cost-per-QALY ratio was 43,600 USD for combination therapy versus P (95% CI 21,400–123,500 USD), but there was clear evidence of heterogeneity in study medication costs across regions, which limits the applicability of this overall result to the individual regions and countries that participated in the trial. When cost-effectiveness ratios were estimated for each region, it was found that, in the USA, the incremental cost-per-QALY ratios for Salm, FP and SFC were all higher than those observed in other regions, reflecting the higher cost structure in the USA.

There is currently no single agreed methodology for the estimation of regional cost-effectiveness 19. A recent suggestion 20 has been to use a formal heterogeneity test of cost-effectiveness in order to choose between use of a single pooled estimate across the trial versus splitting the data and calculating region-specific cost-effectiveness ratios. Comparison of the pooled estimate reported here with four region-specific estimates based on splitting the data failed to reject the null hypothesis of homogeneity across regions (results not shown, but available from the authors on request). However, such an approach suffers from an acknowledged lack of power 20, which makes such negative test results difficult to interpret. The method reported here represents a variation on the approach suggested by Cook et al. 20, with heterogeneity tests conducted on individual components of the cost-effectiveness calculation. Although the potential problem of lack of power remains, our preferred approach has the advantage that known differences in prices of medication between regions are not masked by high levels of uncertainty in other components of cost-effectiveness.

One of the challenges of interpretation of the economic evaluation of the TORCH trial is the simultaneous comparison of the four arms of the trial, as shown on the cost-effectiveness plane of figure 2⇑. Incremental cost-per-QALY ratios involve the comparison of two alternatives. Since the incremental cost per QALY of SFC versus P is lower than that for Salm versus P or FP versus P, the appropriate summary cost-effectiveness for SFC comes from the comparison with P (this is an example of the principle of extended dominance 21). Although this is straightforward for point estimates, the analysis of uncertainty in cost-effectiveness results is less straightforward. The policy decision is to choose between all four options: withholding treatment, or use of either FP or Salm alone or in combination. One approach to expressing the uncertainty in this decision is represented by the multiple cost-effectiveness acceptability curves of figure 3⇑. Such presentations encourage the reader to think about the appropriate level of evidence that is sufficient to recommend adoption of a particular treatment rather than relying on arbitrary threshold p-values 22.

Another challenge faced in this analysis was the high levels of missing data for medical service costs (22.6%) and preference scores (26.1%) due to attrition during the study. The approach taken here was to use the inverse proportional weighting technique in combination with a set of multivariate regression models for the estimation of the components of the cost-effectiveness calculation. This approach gives greater weight to the observed values of patients who have similar characteristics to those patients who have dropped out of the study. Since this approach is based on the observed data, it is unable to account for situations in which the propensity to be missing is directly related to the unobserved values (so-called non-ignorable missingness). Nevertheless, if it is assumed that attrition in the study is associated with poor outcomes and higher cost, and that the P group suffered the highest rates of attrition, the approach used in the present study is likely to be conservative with respect to estimating the value of treatment (versus P).

The scope of the resource use data collection in the TORCH study was limited by practical necessity. COPD-related costs were collected for medication and primary and secondary care resource categories. For non-COPD-related resource use, only secondary care (hospitalisations) was recorded. No rehabilitation data were collected, nor were any data collected on the costs borne by patients or carers, nor productivity losses to society. In practice, this limitation in scope is conservative with respect to the estimation of the cost-effectiveness of active treatment. Active treatments in the TORCH trial were shown to reduce COPD exacerbation rates compared to P, with SFC achieving the greatest reduction. The analysis of all costs other than those related to study medication showed nonsignificant trends in favour of active treatment, and broadening the scope of cost data collection (particularly patient costs and productivity losses, which might be expected to be strongly correlated with exacerbations of COPD) would most probably have reinforced the study results in favour of active therapy.

The TORCH study has two great strengths. First, it is the first randomised controlled trial to be powered to examine mortality as a primary end-point. Secondly, it collected economic data prospectively within the trial, which permits the prospective assessment of value in terms of cost per QALY. Previous analyses of cost-effectiveness alongside clinical trials in COPD have only been able to report results in terms of cost per avoided exacerbation 23, 24, which does not facilitate comparisons of results across disease areas (required for reimbursement decisions). QALY number results have previously been presented for COPD treatments, but based on economic modelling studies, which typically project treatment effects from short-term randomised controlled trials over the lifetime of patients 25, 26. The cost-per-QALY results reported in the present article relate to the observed 3-yr duration of follow-up of the TORCH study. Relatively simple models projecting this experience beyond the 3-yr results suggest that cost-effectiveness ratios for long-term therapy are likely to fall by ∼50% after 6 yrs of projection (i.e. after the ninth year of therapy), in line with other studies in which benefits have been reported first over the period of the trial and then extrapolated to patient lifetimes 27–30. Therefore, the within-trial results presented here are likely to represent a conservative estimate of the true cost-effectiveness of long-term COPD treatment.

Ultimately, it will be for individual countries to judge whether the combined ICS and LABA therapy represents a cost-effective use of resources within their own jurisdiction, given the point estimates of cost-effectiveness and associated uncertainty. In the UK, for example, it is commonly regarded that treatments costing <20,000 pounds sterling (GBP) per QALY are likely to be considered cost-effective, whereas those of >30,000 GBP are unlikely to be acceptable on cost-effectiveness grounds alone 31. In the USA, there is no single reimbursement authority; nevertheless, a threshold of 50,000 USD per QALY is commonly quoted, and, more recently, commentators have suggested that a contemporary threshold is likely to be ≥100,000 USD 32. In comparison to other disease areas, the cost-effectiveness of SFC therapy falls somewhere between primary prevention of cardiovascular disease with statins 33 and secondary prevention with implantable defibrillators 34, both of which have received positive recommendations from the National Institute for Health and Clinical Excellence in the UK.

To summarise, based on the analysis of data from the TORCH study, SFC is more effective and has a lower incremental cost-effectiveness ratio (compared with P) than either FP or Salm used alone, and, therefore, this combination therapy is preferred to the existing use of monotherapy on efficiency grounds. For patients who are treatment-naive, whether combination therapy offers good value for money is a judgment for the decision-makers in each jurisdiction. Nevertheless, the present economic analysis of the TORCH trial provides a strong empirical basis for making those decisions.

Support statement

Financial support was provided by GlaxoSmithKline.

Statement of interest

Statements of interest for all of the authors and the study itself can be found at www.erj.ersjournals.com/misc/statements.dtl