Combining treatment effects from mixed populations in meta-analysis: a review of methods

Meta-analysis is a useful method for combining evidence from multiple studies to detect treatment effects that could perhaps not be identified in a single study. While traditionally meta-analysis has assumed that populations of included studies are comparable, over recent years the development of precision medicine has led to identification of predictive genetic biomarkers which has resulted in trials conducted in mixed biomarker populations. For example, early trials may be conducted in patients with any biomarker status with no subgroup analysis, later trials may be conducted in patients with any biomarker status and subgroup analysis, and most recent trials may be conducted in biomarker-positive patients only. This poses a problem for traditional meta-analysis methods which rely on the assumption of somewhat comparable populations across studies. In this review, we provide a background to meta-analysis methods allowing for synthesis of data with mixed biomarker populations across trials.

For the methodological review, PubMed was searched to identify methodological papers on evidence synthesis for mixed populations. Several identified methods were applied to an illustrative example in metastatic colorectal cancer.

We identified eight methods for evidence synthesis of mixed populations where three methods are applicable to pairwise meta-analysis using aggregate data (AD), three methods are applicable to network meta-analysis using AD, and two methods are applicable to network meta-analysis using AD and individual participant data (IPD). The identified methods are described, including a discussion of the benefits and limitations of each method.

Methods for synthesis of data from mixed populations are split into methods which use (a) AD, (b) IPD, and (c) both AD and IPD. While methods which utilise IPD achieve superior statistical qualities, this is at the expense of ease of access to the data. Furthermore, it is important to consider the context of the decision problem in order to select the most appropriate modelling framework.