Simultaneous comparison of multiple treatments: combining direct and indirect evidence
BMJ 2005; 331 doi: http://dx.doi.org/10.1136/bmj.331.7521.897 (Published 13 October 2005) Cite this as: BMJ 2005;331:897 Deborah M Caldwell (d.m.caldwell{at}bristol.ac.uk), research associate1,
 A E Ades, professor of public health science1,
 J P T Higgins, senior research associate2
 ^{1} Medical Research Council Health Services Research Collaboration, Department of Social Medicine, University of Bristol, Bristol BS8 2PR
 ^{2} Medical Research Council Biostatistics Unit, Institute of Public Health, Cambridge CB2 2SR
 Correspondence to: D M Caldwell
 Accepted 29 June 2005
How can policy makers decide which of five treatments is the best? Standard metaanalysis provides little help but evidence based decisions are possible
Introduction
Several possible treatments are often available to treat patients with the same condition. Decisions about optimal care, and the clinical practice guidelines that inform these decisions, rely on evidence based evaluation of the different treatment options.1 2 Systematic reviews and metaanalyses of randomised controlled trials are the main sources of evidence. However, most systematic reviews focus on pairwise, direct comparisons of treatments (often with the comparator being a placebo or control group), which can make it difficult to determine the best treatment. In the absence of a collection of large, high quality, randomised trials comparing all eligible treatments (which is invariably the situation), we have to rely on indirect comparisons of multiple treatments. For example, an indirect estimate of the benefit of A over B can be obtained by comparing trials of A v C with trials of B v C,3^{–}5 even though indirect comparisons produce relatively imprecise estimates.6 We describe comparisons of three or more treatments, based on pairwise or multiarm comparative studies, as a multiple treatment comparison evidence structure.
The need to combine direct and indirect evidence
Concerns have been expressed over the use of indirect comparisons of treatments.4 5 The Cochrane Collaboration's guidance to authors states that indirect comparisons are not randomised, but are “observational studies across trials, and may suffer the biases of observational studies, for example confounding.”7 Some investigators believe that indirect comparisons may systematically overestimate the effects of treatments.3 When both indirect and direct comparisons are available, it has been recommended that the two approaches be considered separately and that direct comparisons should take precedence as a basis for forming conclusions.5 7
Difficulties arise, however, if the direct evidence is inconclusive but the indirect evidence, either alone or in combination with the direct evidence, is not. Furthermore, this approach becomes increasingly impractical as the number of treatments increases. If five treatments have been compared with each other, there are 10 possible direct pairwise comparisons and 70 indirect comparisons. Keeping this information separate makes little sense, particularly when the entire body of evidence could be summarised in terms of four relative treatment effects.
Suitable statistical methods for comparing multiple treatments that fully respect randomisation have been available for some time.8^{–}11 They have not been widely used, although applications of these methods have begun to appear in US medical journals12 and in medical decision making.13 14 Below, we provide a worked example of these methods and show their advantages over restricting attention to direct pairwise comparisons. We also examine some concerns about bias and randomisation.
Thrombolysis and angioplasty after myocardial infarction
Two recent overviews of treatment for acute myocardial infarction show the difficulties of relying on standard pairwise metaanalysis. Boland and colleagues reviewed 14 randomised controlled trials making two or three way comparisons of six thrombolytic treatments (table 1).15 Their findings are limited to summary statements on each of the pairwise comparisons for which direct evidence was available (streptokinase is as effective as nonaccelerated alteplase, tenecteplase is as effective as accelerated alteplase, reteplase is at least as effective as streptokinase, etc). Presenting results in this way makes it difficult to draw an overall conclusion about which treatment is best on the chosen outcome, or even to form an internally consistent summary of their relative effects.
At the other extreme, Keeley and colleagues looked at 22 randomised controlled trials that compared primary percutaneous transluminal coronary angioplasty with thrombolytic treatment (streptokinase, alteplase, or accelerated alteplase).16 They found that primary percutaneous transluminal coronary angioplasty was better, but only by “lumping” the three thrombolytic drugs to form a single comparator. This approach was criticised because the relevant comparison is with the best thrombolytic drug not the average one.17 18 The full set of evidence from both reviews is available on bmj.com.
To deal with such multiple treatment evidence structures we need a single statistical analysis providing estimates for all the 21 possible pairwise comparisons between seven treatments, and, more importantly, an assessment of which treatment is most likely to have the lowest mortality. The first objective can be achieved by a variety of traditional (frequentist)6 8^{–}10 and bayesian statistical methods,8 11 all of which are based on a logistic regression model.
Calculation of the probability that each treatment is best on some chosen outcome requires simulationbased methods. We have used an existing bayesian Markov chain Monte Carlo method,8 adapted to apply to any connected network set of treatment comparisons. Vague prior distributions are used for comparisons of treatments, so the findings should be close to those obtained with frequentist methods.19 Further technical details of the method, including computer programs that are applicable to a wide range of evidence structures, are available on bmj.com.
Multiple comparisons versus pairwise metaanalysis
2 compares results from a single analysis comparing multiple treatments with those from standard pairwise metaanalyses of direct comparisons. The multiple treatment analysis provides a full set of odds ratios for all the 21 comparisons, effectively combining all the direct and indirect evidence for each comparison. Direct data exist for only 10 pairwise comparisons, and for four of these comparisons only one randomised controlled trial is available. When direct evidence is available, it agrees with the results obtained by combining all the available evidence, and the multiple treatment analysis tends to produce narrower confidence intervals.
In 3 we present estimates of absolute risk of mortality for each treatment, along with the estimated probability that each treatment is best. The results decisively confirm Keeley and colleagues'conclusion that primary percutaneous transluminal coronary angioplasty is the most effective treatment but by using a method that retains the identity of each treatment.16 Note that, although on the direct evidence primary percutaneous transluminal coronary angioplasty is not significantly better than accelerated alteplase (fixed effect odds ratio = 0.81, 95% confidence interval 0.64 to1.02),17 in the multiple treatment comparison analysis it clearly is (0.74, 0.61 to 0.89). The random effects analysis produces an even greater narrowing of the confidence interval. The combination of all available data in this way allows relevant treatments to be compared and leads to more precise conclusions.
This statistical analysis can be embedded in a decision analysis by including data on costs and total life years gained for each treatment. The same methods can then be extended to calculate the probability that each treatment is the most cost effective.13 14
Bias, randomisation, and generalisability
What assumptions are being made in this analysis? One way to conceptualise this is to imagine that all 36 trials had examined all seven treatments but that in each trial results for all but two or three treatments had been lost at random. The key assumption for the fixed effect analysis is that the relative effect of one treatment compared with another is the same across the entire set of trials.3 4 This means that the true odds ratio comparing A with B in trials of A v B is exactly the same as the true odds ratio for A vBinthe A v C, B v C, and indeed E v F trials, even though A and B were not included in those studies. In a random effects model, where it is assumed that the odds ratios in each trial are different but from a single common distribution, the assumption is that this common distribution is the same across all sets of trials.
These assumptions are remarkably similar to those that underlie a standard pairwise metaanalysis. The only additional assumption is that the similarity of the relative effects of treatment holds across the entire set of trials, irrespective of which treatments were actually evaluated. Indeed, one way to question this assumption is to imagine all trials had compared the same two treatments and to judge whether they are sufficiently similar to be combined in a metaanalysis.
It may be helpful to consider how the target population relates to the patient groups in the studies. If we are attempting to find the single best treatment with a view to recommending it to all the various patient groups represented in previous trials, a single combined analysis is clearly indicated. If our target is a specific subgroup of patients who were represented in studies around a specific set of treatment comparisons, it may be more appropriate to consider restricting the analysis to a subset of the trials.
The assumptions behind comparisons of multiple treatments are unlikely to be statistically verifiable, and it seems reasonable to rely on expert clinical and epidemiological judgment, both for multiple treatment comparison and for standard pairwise metaanalysis. This is, after all, the criticism that clinicians often make of metaanalyses—that like is not being compared with like. Multiple treatment metaanalysis is no different from pairwise metaanalysis in requiring such a judgment. Poor judgments in both cases may induce heterogeneity of effects. Poor judgments in metaanalyses of multiple treatment comparisons may, in common with subgroup analyses and metaregression of pairwise comparisons, lead to confounding (if, for example, trials of A v B and A v C are systematically different from trials of B v C). However, bias would not be expected generally to operate in any particular direction.
A further assumption relates to scale of measurement. In common with studies based on indirect comparisons,3^{–}5 20 multiple comparison models are based on the assumption that treatment effects add together so that the relative effect of A v C can be predicted from the effects of A v B and B v C. This assumes that the appropriate measure of effect (log odds ratios, relative risk, or risk difference) has been chosen.21
Summary points
Healthcare decisions often involve choosing from a selection of treatment options
Most systematic reviews and metaanalyses focus on pairwise comparisons, forcing reliance on indirect comparisons
Statistical methods for comparing multiple treatments that combine direct and indirect evidence in a single analysis are available
These methods make the similar assumptions to standard pairwise metaanalyses but require that they hold over the entire set of trials
Multiple treatment comparisons should be more frequently used to inform healthcare decisions
Both of the pairwise metaanalyses described earlier were built on the above assumption.15 16 In our multiple analysis we made the further assumption that the two sets of studies are homogeneous regarding the relative effects of all the treatments, or (in the random effects analysis) that the relative effects are from the same common distribution. This was generally supported by the similarity between the direct evidence and the estimates obtained by combining direct and indirect evidence. This observation has been made in previous analyses.4
Another common feature of our proposed multiple comparison methods and standard pairwise metaanalyses is a lack of any assumptions about baseline risks across studies. Thus, our analyses are based only on randomised comparisons. Methods that compare patient outcomes in one treatment arm of one trial directly with those in a treatment arm in another trial break randomisation and have been rightly criticised.6
Can these methods be used routinely?
The precepts of systematic review enjoin us to put all the available evidence together, to avoid selection bias, increase precision, and examine generalisability.22 A unified, coherent analysis can be achieved only by analysing the entire collection of relevant randomised controlled trials while respecting randomisation. Rather than asking whether analyses comparing multiple treatments should be used routinely, it is more appropriate to ask whether they can be avoided. Methods for comparing multiple treatments will have increasing scope as new treatments proliferate. Most manufacturers' trials, aimed at obtaining licenses for new products, tend to use current or older treatments as comparators. Thus the head to head comparisons of new treatments that are most useful to clinical practice are not available. The need for integrated analyses to inform technology appraisals and clinical guidelines in these circumstances is a recurring theme at the National Institute for Clinical Excellence.1 Metaanalyses comparing multiple treatments are feasible and should be considered as the bedrock for decisions when several treatments are available.
Common sense is obviously required when applying these methods. Investigators must consider whether it is meaningful to generalise over the entire set of studies. Particular caution should be exercised in combining contemporary trials with historical trials, since earlier trials may include more severely ill participants. If efficacy depends on baseline risk,23 a multiple treatment comparison that does not take this into account will be biased, just as a pairwise metaanalysis would be.
No statistical method is a panacea. Other models making weaker assumptions remain to be explored. More advanced methods that allow for random errors in the generalisability assumption may also prove valuable.10 12 This is an area of active research, and we look forward to a period of extensive development and, most importantly, increasing application of these methods.
Further details of the method are on bmj.com
Footnotes

Contributors and sources DMC is working on the role of multiple treatment comparisons in decision analysis. AEA is leading a programme on methods for multiparameter evidence synthesis. JPTH developed the statistical method on which the current analyses are based. DMC wrote the first draft and carried out the statistical analyses. AEA supervised the analyses and is the guarantor. AEA, JPTH, and DMC contributed to the final draft.

Competing interests None declared.