Elsevier

The Lancet

Volume 365, Issue 9456, 22 January 2005, Pages 341-346
The Lancet

Series
Can meta-analysis help target interventions at individuals most likely to benefit?

https://doi.org/10.1016/S0140-6736(05)17790-3Get rights and content

Summary

Meta-analyses of randomised trials aim to summarise the effects of interventions across many patients, and can seem remote from the clinical issue of how individual patients should be treated and which patient groups will benefit the most from treatment. One method that attempts to address this point entails relating the overall effect in every trial to summaries of patient characteristics. This is called meta-regression. The interpretation of such analyses is not straightforward, however, because of a combination of confounding and other biases. Much more useful is to compare the outcomes for patient subgroups within trials and combine these results across trials. Unfortunately this method is rarely possible using published information, so analyses of individual patient data from trials are necessary. Also, although meta-analyses generally summarise an intervention's effect as a relative risk reduction, the groups of patients with the greatest absolute risk reduction have the most to gain.

Introduction

Systematic reviews of health-care interventions are an attempt to collate information from all relevant studies and, if deemed appropriate, combine their results using meta-analysis.1 This process inevitably brings together studies that are diverse in their designs (in terms of outcomes assessed and length of follow-up, for example), in the specific interventions used (method, intensity, and duration), and in the types of patients studied (demographic and clinical characteristics). Thus the results, based on such a broad range of evidence, can seem remote from the issue of how to treat individual patients, and even somewhat irrelevant to clinical practice.2 Nevertheless it is incontrovertible that treatment decisions should be based on evidence when it exists, and that good quality systematic reviews provide an essential mechanism in reviewing available evidence.3 The issue is how best to bridge the gap between evidence based on many patients and making decisions about treating individuals.

The larger randomised trials are, the less their results will be subject to chance. Many patients are needed to distinguish true treatment benefits that are clinically important, but moderate in size, from chance effects.4 Increasing numbers of patients by combining results across trials provides a principal rationale for meta-analysis.3 At the other extreme, n-of-1 trials attempt to isolate effective treatments for a particular individual;5 however, such trials can only be undertaken in specific clinical situations, for example, for treatments to relieve symptoms in chronic disorders, and do not provide evidence about medical policy that can be generalised to new patients. In between these extremes lies the aim of targeting interventions by identifying subgroups of patients most likely to benefit. Subgroup analyses within a clinical trial investigate the effects of an intervention for specific groups of patients—eg, defined by their clinical characteristics—in an attempt to refine how the treatment might best be used in practice.6 Such analyses are, however, inevitably plagued by chance effects—both wider confidence intervals due to the fewer patients involved, leading to more uncertain inferences, and false positive results arising from the multiplicity of subgroups typically investigated.7

Comparing patient subgroups within a meta-analysis might help to ameliorate the tension between decision making in clinical medicine and overall statements of evidence in systematic reviews. Researchers have suggested that meta-analysis should go beyond estimating one overall effect,8, 9 although this expansion has drawbacks.10 One aim of meta-analysis should be to estimate how treatment effectiveness varies according to patients' characteristics.11 In this article, we discuss the extent to which this aim is achievable, and investigate whether we can progress beyond the general statement that meta-analytic conclusions should be borne in mind in clinical decision making. In doing so, we need to distinguish the relative risk reductions usually summarised in meta-analyses from their implications for absolute risks, which describe how much patients benefit.

Section snippets

Conventional meta-analysis

To focus the discussion, we introduce a specific example. The effectiveness of platelet glycoprotein IIb/IIIa inhibitors (PGIs) in acute coronary syndromes (non-Q-wave infarction and unstable angina) has received much attention, being the subject of a Health Technology Assessment review,12 National Institute for Clinical Excellence guidance,13 and a Cochrane systematic review.14 Although PGIs reduce the risk of death and myocardial infarction in patients undergoing percutaneous coronary

Meta-regression

Meta-regression aims to relate the treatment effects recorded in different trials to the overall characteristics of those trials. We will consider the example of whether the effectiveness of PGIs is different between men and women. The basic characteristics of patients recruited into trials are usually reported fully in publications. For example, we can relate the odds ratio noted in every trial to the proportion of women in that study (figure 2). Meta-regression assesses the strength of the

Discussion

Identification of patient groups who benefit most from an intervention is never going to be easy, since it is a task for which enormous quantities of randomised evidence are necessary. Even in large trials, apparent subgroup differences can result merely from chance. Meta-analyses of large trials based on individual patient data allow subgroups to be contrasted within trials, and for these results to be combined across trials, producing more reliable evidence. Individual patient data also allow

References (48)

  • M Egger et al.

    Systematic reviews in health care: meta-analysis in context

    (2001)
  • RE Wittes

    Problems in the medical interpretation of overviews

    Stat Med

    (1987)
  • CD Mulrow

    Rationale for systematic reviews

    BMJ

    (1994)
  • R Peto

    Why do we need systematic overviews of randomised trials?

    Stat Med

    (1987)
  • G Guyatt et al.

    Determining optimal therapy: randomized trials in individual patients

    N Engl J Med

    (1986)
  • AD Oxman et al.

    A consumers' guide to subgroup analyses

    Ann Intern Med

    (1992)
  • SG Thompson

    Why sources of heterogeneity in meta-analysis should be investigated

    BMJ

    (1994)
  • JA Berlin

    Benefits of heterogeneity in meta-analysis of data from epidemiologic studies

    Am J Epidemiol

    (1995)
  • G Davey Smith et al.

    Meta-analysis: beyond the grand mean?

    BMJ

    (1997)
  • MS McDonagh et al.

    A rapid and systematic review of the clinical effectiveness and cost-effectiveness of glycoprotein IIb/IIIa antagonists in the medical management of unstable angina

    Health Technol Assess

    (2000)
  • Guidance on the use of glycoprotein IIb/IIIa inhibitors in the treatment of acute coronary syndromes (NICE Technology Appraisal Guidance No 12)

    (2000)
  • X Bosch et al.

    Platelet glycoprotein IIb/IIIa blockers for percutaneous coronary revascularization, and unstable angina and non-ST-segment elevation myocardial infarction (Cochrane Review)

    Cochrane Database Syst Rev

    (2001)
  • Platelet Receptor Inhibition in Ischemic Syndrome Management (PRISM) Study Investigators

    A comparison of aspirin plus tirofiban with aspirin plus heparin for unstable angina

    N Engl J Med

    (1998)
  • The Platelet Receptor Inhibition in Ischemic Syndrome Management in Patients Limited by Unstable Signs and Symptoms (PRISM-PLUS) Study Investigators

    Inhibition of the platelet glycoprotein IIb/IIIa receptor with tirofiban in unstable angina and non-Q-wave myocardial infarction

    N Engl J Med

    (1998)
  • Cited by (154)

    View all citing articles on Scopus
    View full text