Individual participant data meta-analysis
What question does this help answer?
What is the overall effect when every study is analysed in the same way, and does the effect differ between kinds of participant, questions that published summaries often cannot answer?
When to use it
Use it when consistent definitions, outcomes, follow-up or analyses across studies matter, when investigating whether effects vary with participant characteristics, when time-to-event or unpublished outcomes are needed, or when published reports are incomplete. It usually requires collaboration with the original investigators and substantial time.
When not to use it
Do not assume it is always worth the cost: when published summaries are complete and consistent and the question concerns only the overall effect, an aggregate-data meta-analysis may answer it adequately. Nor is it a remedy for poorly conducted studies or for a set of studies whose data could only partly be obtained.
How it works in plain language
The team obtains, checks and harmonises each study's data, then analyses them in one of two ways. A two-stage approach analyses each study separately and combines the study estimates as an ordinary meta-analysis would. A one-stage approach fits a single model to all participants while keeping each study's data distinct. When the two give different answers, it is usually because they were set up with different assumptions rather than because one approach is inherently right.
What a reversed trend looks like
Key assumptions and design requirements
- A protocol and data-sharing agreements are in place before data are requested, and studies that do not provide data are reported with the reasons.
- Each dataset is checked for integrity, including randomisation patterns, baseline balance and missing data, before analysis.
- Participants remain clustered within their original study; analysing them as if they came from one large study is an error.
- Effect modification by participant characteristics is estimated from within-study information, because an association across study averages is subject to ecological bias.
- The possibility that studies providing data differ from those that did not is examined, for example by comparing with the published results of the missing studies.
Why Mallard does not run it
Mallard does not perform individual participant data meta-analyses or size systematic reviews, and it does not handle data transferred from other investigators. This page supports specifying and reading such a synthesis.
Clinical example
Trials of a rehabilitation programme reported function at different time points and did not report results by age. A collaborative group plans an IPD meta-analysis, harmonises the function scale and time points, checks each dataset, and prespecifies estimating whether the effect varies with age using within-trial interactions only. The example illustrates planning choices only.
What to watch out for
Having participant-level data does not remove the biases of the studies that produced it, and studies that could not or would not share data can differ systematically from those that did. Comparing effects between kinds of participant needs within-study information: an association across study averages is an ecological comparison, however the data were obtained.
Before using this in your workspace
Record the eligibility of studies, the data requested and the harmonisation rules, data-sharing and governance arrangements, integrity checks, the one-stage or two-stage approach and its assumptions, how clustering and effect modification will be modelled, and how studies without available data will be handled.
- Meta-analysis of individual participant data: rationale, conduct, and reporting (opens in a new tab) · BMJ · doi 10.1136/bmj.c221
- Meta-analysis using individual participant data: one-stage and two-stage approaches, and why they may differ (opens in a new tab) · Stat Med · doi 10.1002/sim.7141
- Preferred Reporting Items for Systematic Review and Meta-Analyses of individual participant data: the PRISMA-IPD Statement (opens in a new tab) · JAMA · doi 10.1001/jama.2015.3656