Chronic obstructive pulmonary disease (COPD) has become a leading cause of death worldwide and its management requires an enormous amount of human and financial resources [1, 2]. Reduction of exacerbation rates is one of the main treatment goals in COPD management since exacerbations bear heavily on the patient's health-related quality of life and prognosis as well as on COPD-related costs [3]. Several conventional meta-analyses provided evidence that long-acting beta-agonists, long-acting anticholinergics and inhaled corticosteroids reduce exacerbations in patients with COPD when compared with placebo [4–6]. The conventional meta-analyses are, however, less informative about the comparative effectiveness of long-acting beta-agonists and anticholinergics or about the additional value of inhaled corticosteroids when added to long-acting bronchodilators. The comparative effectiveness is of great interest to physicians because the predominant question in clinical practice is to choose between treatments rather than deciding whether to treat or not to treat [7, 8].

Opinions differ about whether a long-acting bronchodilator alone is sufficient or if an inhaled corticosteroid provides additional benefits at least for some patients [3, 9, 10]. Any added benefit of corticosteroids, if present, should outweigh the associated risk for adverse effects and their additional costs. Conventional meta-analyses [4–6, 11, 12] do not provide enough support to solve this debate because evidence from randomized head-to-head comparisons is often unavailable. Also, conventional meta-analyses cannot provide effect estimates for comparisons of more than two treatments at the same time, so that a ranking of competing treatments is not available. Finally, conventional meta-analyses cannot assess subgroup effects reliably [13] although such information is very valuable for clinicians. Theoretically, a single very large trial would overcome these three limitations of conventional meta-analysis. However, the sample size would need to be a multiple of that of the recently published TORCH trial [14] (more than 6000 patients) if subgroup effects were to be investigated.

Since such a trial is very unlikely to become available, a network meta-analysis or individual patient data meta-analysis can be very informative [15]. Such pooled analyses unify evidence from all randomized trials while fully preserving randomization [16–18]. Therefore, our aim was to assess the relative effectiveness of competing inhaled drug regimens for the prevention of exacerbations in patients with stable COPD in a pooled analysis of randomized comparisons. In addition, we assessed whether the effectiveness depend on the severity of COPD, treatment duration, or the definition of an exacerbation (event based or symptom based).

Based on 35 trials, our analysis showed that all inhaled drug regimens significantly reduced exacerbations but there were no significant differences between them. Thus, it appears that inhaled corticosteroids provide no additional value in reducing exacerbations when used concurrently with long-acting beta-agonists. However, combination treatment appeared to be more effective than beta-agonists alone in patients with low FEV₁.

According to the recently published BOLD study, the prevalence of COPD with GOLD stage II to IV in populations over 40 years of age is around 10%, and approximately 80% of these patients have an FEV₁ ≥ 50% predicted (GOLD stage II) [57]. For these patients, our study suggests that single treatment with long-acting beta-agonists or long-acting anticholinergics is not only the treatment of choice for symptom control but also for preventing exacerbations. Although common practice is to prescribe inhaled corticosteroids also [58, 59], our findings suggest that these patients do not need the additional medical therapy to reduce exacerbations.

In contrast, in patients with low FEV₁ (below 40% predicted), long-acting anticholinergics appear to be the most attractive choice. They are as effective as combination treatment in reducing exacerbations but offer advantages in terms of costs and adverse effects [53]. The only head-to-head comparison of a long-acting anticholinergic and combination treatment, the recently published INSPIRE study, identified outcomes in agreement with these findings [53]. The INSPIRE study also showed no clinically relevant differences in health-related quality of life (< 4 points on St. Georges Respiratory Questionnaire). A significant reduction in 2-year mortality was observed with combination treatment (3%) compared with the long-acting anticholinergic (6%). However, this mortality analysis should be interpreted with caution because a closer look at how death was ascertained in the INSPIRE trials reveals that death was only recorded for patients who did not withdraw from treatment (65.5% with combined treatment and 58.3% with long-acting anticholinergic). Thus, it appears that one third of patients did not enter the mortality analysis. Unfortunately, little evidence is available about the combination of long-acting anticholinergics and inhaled corticosteroids; however, a recent Canadian trial indicated that this combination could be the most effective treatment for reducing exacerbations in patients with low FEV₁ [20].

Exacerbation is not the only outcome that should inform the decision for or against adding inhaled corticosteroids to long-acting bronchodilators. Other outcomes should be considered as well. A recent systematic review found no risk reduction in terms of mortality if an inhaled steroid was added to a long-acting bronchodilator [11]. Health-related quality of life was statistically significantly better after combined treatment (difference of -1.64 units on St. Georges Respiratory Questionnaire; 95% CI -2.28 to -1) but the effects were well under the threshold representing a minimal important difference (4 units) [11]. Arguments for treating COPD solely with long-acting beta-agonists are the substantially lower costs and lower risk for adverse effects such as pneumonia, oral candidiasis, or loss of bone density compared with combination treatment [11, 14, 53, 60]. Integrating and presenting this complex information about benefits and downsides of inhaled drug treatments is challenging. One approach is to use decision aids that are particularly valuable for value-sensitive decisions where the balance of benefits and downsides is not straightforward. As the decision is, in the case of inhaled drug treatment for COPD, only between two treatment options (bronchodilator(s) versus combined treatment), information about the comparisons with placebo could now be excluded for simplicity, since offering no treatment is not in the best interest of the patient. However, whether patients benefit from such informed decision-making requires testing in additional trials.

Unlike previous meta-analyses based on comparisons of inhaled drugs with placebo or, in some instances, with another inhaled drug [4–6, 11, 12], we argue that the comparative effectiveness of available treatments are of greater interest to physicians [7, 8]. To provide estimates for this comparative effectiveness, we used a new analytical approach to pool evidence from all available randomized trials. Such analyses may be, in many instances, the only way to estimate comparative treatment effectiveness; head-to-head comparisons are unavailable, and the sample sizes required to detect small differences are unfeasibly large. For comparison of inhaled COPD drugs, for example, thousands of patients are required just for the main comparison, not even accounting for subgroup effects.

There are other approaches for indirect comparisons [16–18]. An early and important contribution emphasized the importance of preserving randomization [16]. However, that proposed approach only allows for a single head-to-head comparison and not for simultaneous comparison of all treatments under consideration. Two other approaches [17, 18] are available, but the complexity of their statistical methods may represent a barrier for their application and interpretation [61]. Our approach is particularly attractive because it is transparent and easily reproducible. The multivariable logistic regression analysis allows for an analysis like that used in any randomized trial while keeping intact the randomization of each trial. The comparisons with placebo showed the same results as those of the meta-analyses of the Clinical Practice Guideline of the American College of Physicians [6]. When the odds ratios are transformed to relative risks [62], the effect estimates are identical to those of a conventional meta-analysis.

A limitation of our study is that we focused on exacerbations and did not consider additional outcomes such as health-related quality of life or mortality. When we planned the current analyses the body of evidence on mortality was too small to provide precise effect estimates, even if pooled. The TORCH trial [14] published in 2007 was the first trial that was powered to assess the effects of inhaled drugs on mortality, but even this very large trial turned out to be too small. Therefore, we decided to postpone analyses on mortality to a point in time where more data would become available. Another limitation is that we performed only a few stratified analyses. For physicians, knowing more about profile-specific effect estimates would be very useful. For example, more detailed information about risk factors for exacerbations, such as previous exacerbations, would enhance understanding about who benefits from the addition of inhaled corticosteroids. Single trials are unlikely to provide such information because of the very large sample sizes they would require for sufficient power to detect effect modification. Individual patient data meta-analysis and pooled analyses may solve this problem to some extent. However, reporting of such patient characteristics is highly variable, and retrieval of individual patient data from primary studies is challenging [63].

A limitation of any approach to pool data on exacerbations is that (relative) treatment effects can only be estimated adequately based on the proportion of patients with at least one exacerbation. Analysis of exacerbation rates expressed as mean exacerbation rates per person-year would offer a more comprehensive use of the data and would be less dependent on the length of follow-up. However, such data cannot be pooled adequately without having access to adequate and fully reported analyses that took within- and between-patient variability in exacerbations into account. But as Suissa points out, most trials are not analyzed and reported adequately[19]. Ideally, an individual patient data meta-analysis would be conducted but it is very challenging to convince investigators of all relevant trials to share their data [63]. The advantage of using the proportion of patients with at least one exacerbation is the limited influence of few patients with many exacerbations on the results. Also, physicians may be more familiar with this format and respective estimates for treatment effects such as odds ratios or relative risks than with mean differences in exacerbation rates. However, there is little evidence on how effect estimates should be presented in order to facilitate the transfer from research into practice.

A common problem of systematic reviews in COPD is poor reporting of clinical and spirometric characteristics of patients enrolled in included trials. It is sometimes difficult to judge whether all patients had COPD, or if some patients had other lung disease such as asthma. Also, there is often no separate reporting for moderate and severe exacerbations, although it would be informative to estimate the effects of inhaled steroids stratified for the severity of exacerbations. COPD is a heterogeneous disease and it is important that future studies provide more information about their patients, including clinical characteristics, lung function data and also information on co-treatments.

The results of our study may support physicians in selecting inhaled drug treatments for patients with COPD. In general, long-acting beta-agonists or anticholinergics appear to be the treatment of choice to reduce exacerbation rates. Adding an inhaled corticosteroid does not provide additional protection from exacerbations. In patients with low FEV₁, combination treatment and long-acting anticholinergics should be favored because they reduce the risk for exacerbations more than single treatment with a long-acting beta-agonist. Our analyses, together with the existing meta-analyses that considered additional outcomes, inform patients and physicians to balance the benefits and downsides of different inhaled drug treatments for COPD.