OBJECTIVES: Meta-analysis on aggregate data (AD) is prone to bias, mainly due to differences in patient’s characteristics and study design. Meta-regression is used to adjust for these differences, but it is known to lack statistical power. Combining individual patient data (IPD) and AD in a meta-analysis is the gold standard to account for heterogeneity between studies and adjust treatment effects on patient’s characteristics. The objective of the study was to identify and compare existing methods for combining IPD and AD in a meta-analysis from the literature.
METHODS: We conducted a comprehensive search in Pubmed and Google scholar, until April 2016. All methodological papers describing a method for combining IPD and AD in a meta-analysis were selected.
RESULTS: Out of 21 papers considered as relevant, 9 suggested a new method or extended an existing method, 11 re-described or applied existing methods and 1 paper performed a review of existing methods. Four general methods were identified: 1- the two-stage method, 2- the analysis of partially reconstructed IPD, 3- the multi-level model, and 4- the Bayesian hierarchical related regression model (BHRR). The two-stage method is mostly used due to its simplicity of application. The analysis of partially reconstructed data is also simple to apply but restricted to binary and ordinal data. However they ignore patient-level information and are likely to be inadequate when specific patient-level characteristics are of interest. In contrast, the multi-level and the BHRR models distinguish between IPD and AD studies and model patient information from IPD and overall study-level effects from IPD and AD.
CONCLUSIONS: Further research is needed to inform methodological choices when combining IPD and AD in meta-analysis such as comparative studies based on real data or simulated data.