A Primer on Individual Participant Data Meta-Analysis and Its Strengths and Limitations.

Abstract

In conventional (aggregate data) meta-analysis, the results of many similar studies are statistically combined to yield a single pooled result. Conventional meta-analyses have many limitations. They cannot examine research questions that were not examined in the source studies, interactions between variables cannot be studied, granular analyses cannot be performed, and systematic biases in the source study data will be retained in the pooled results. Individual participant data meta-analysis (IPD-MA) differs from conventional meta-analyses in that, instead of pooling the results of already completed analyses from source studies, the statistical team obtains and processes individual participant data from the source studies. This allows the specification of a new study protocol that can be uniformly applied, across source studies, to the individual participant data. Matters that can thus be harmonized across the source studies include participant eligibility criteria, choice of exposures and outcomes, operational definitions of exposures and outcomes, time points for data examination, and the method of data analysis. IPD-MA can be performed as a 1-stage or 2-stage procedure; the latter is simpler. Whereas IPD-MA overcomes some of the limitations of conventional meta-analysis, it has its own limitations. Obtaining individual participant data can be difficult and time-consuming, reprocessing and reanalyzing source study data requires time and effort, and new biases may be introduced. The new biases arise from lack of availability of individual participant data from all source studies, limitation of the generalizability of findings when harmonization of the study protocol excludes subjects from analysis, loss of randomization structure when participant eligibility restrictions are applied in IPD-MAs of randomized controlled trials, and failure to adequately adjust for necessary covariates. Readers need to be aware of these biases, and authors of IPD-MAs need to report on the potential impact of these biases on their results.