New meta-analysis method maximises data use to reliably determine which participants benefit most from treatments
18 Aug 2022
A new article sets out the optimal way to use data from multiple trials to identify subgroups of patients who particularly benefit, or not, from an intervention. This information is important to ensure that patients receive the best treatment. The study has been recently published, open access, in the journal Research Synthesis Methods.
Traditional approaches for investigating patient subgroups in meta-analysis may be too simplistic. They may lead to “aggregation bias”, which can result in misleading conclusions. This is because these approaches assume that the association between the effect of the intervention and the subgroup measurement is the same at the trial level and the patient level. This phenomenon is known as Simpson’s Paradox.
As an example, suppose that the estimated intervention effects from one group of patients (say, males) from each trial are combined; and similarly for another group (say, females). The traditional approach would compare these two combined groups (see animation below). However, if the trials in the meta-analysis have different make-up of patients (for example, if one trial recruits 70% males whilst another trial recruits 70% females) then there is a risk that this variation across aggregated trials could mask, or “confound”, the true difference in effect between individual patients.
The traditional approach is illustrated in the animation below
To solve this problem, the authors propose the use of a “within-trial framework”. This approach focuses on estimating relevant quantities within trials first, one trial at a time (beginning with the difference between subgroups; see right-hand panel in the animation below).
In doing so, there is no opportunity for variation in the patient make-up of trials to cause confounding or aggregation bias. Furthermore, data from all subgroups (instead of, say, just the male patients) is used at each stage of estimation. Hence, this new method gives reliable evidence about whether groups of patients benefit differently from an intervention.
The new approach is illustrated in the animation below
This new article does not simply describe the proposed new approach. It also explains how to use it, highlighting that researchers do not need any more information than before. They simply need to analyse it differently. To help researchers use this method, the authors have also developed a Stata software package.
In fact, this new approach is already in use. It was first piloted in reviews of abiraterone treatment for advanced prostate cancer and of radiotherapy for locally-advanced prostate cancer. More recently, it was used within a large collaborative meta-analysis run by the World Health Organisation looking at whether Interleukin-6 antagonists were beneficial for patients hospitalised with COVID-19. Further improvements to the methodology are at the forefront of an Individual Participant Data meta-analysis funded by Prostate Cancer UK, looking at the effect of adding chemotherapy agents to hormone therapy for men with advanced prostate cancer.
These examples demonstrate the ability to reliably identify groups of patients that either did or did not benefit from the intervention of interest, using practical methodology, and to show the evidence clearly. Routine use of this method as the emerging standard for patient subgroups in meta-analysis will give researchers substantial added value from the data they already collect and bring targeted benefits to patients sooner.
Further information: