Written and maintained by CASRAI Editorial Board
Last updated
Individual participant data (IPD) meta-analysis pools the raw, participant-level records from every included study instead of the summary numbers each study published — and that difference changes what you can ask the data, not just how precisely you can answer it. Cochrane’s own guidance treats IPD synthesis as the gold-standard variant of meta-analysis specifically because it lets a review team apply consistent outcome definitions, consistent analysis methods, and consistent handling of missing data across studies that originally analyzed and reported their results differently. The trade-off is that obtaining, harmonising and analyzing raw data from multiple independent trial teams is a materially bigger undertaking than extracting numbers from published tables, and the review has two genuinely different ways to model the pooled data once it has it: one-stage or two-stage.
Individual participant data vs. aggregate data: what actually changes
A conventional (“aggregate data” or AD) meta-analysis works from whatever each study already published — a mean difference, an odds ratio, a hazard ratio, each with its own confidence interval. A meta-analysis built this way is constrained by how the original authors chose to define subgroups, handle missing outcomes, and report covariates; if a trial published results only for the full sample, a review can’t re-derive the effect in a subgroup it never reported.
IPD meta-analysis instead collects the underlying dataset — one row per participant — from each contributing study’s investigators, then re-analyzes it centrally under one pre-specified protocol. That makes several things possible that aggregate data cannot support well: consistent covariate-adjusted analyses across studies, reliable subgroup and interaction analyses at the participant level rather than the ecologically-biased study level, standardised handling of missing data and dropout, and outcome definitions that are re-derived the same way in every study rather than trusted as reported. The Cochrane Handbook covers this as its own chapter (Chapter 26, Individual participant data) precisely because the methods, and the practical project management involved, diverge enough from standard aggregate-data synthesis to need separate treatment.
Getting the data: sharing agreements, harmonisation, and partial IPD
None of this is available until the review team actually has the data in hand, and that step is usually the slowest part of an IPD project, not the statistics. In practice it means:
- Requesting raw data from each trial’s investigators or sponsor, typically under a formal data sharing agreement or data transfer agreement — sometimes a data use agreement where a US-based data holder is involved — that specifies what is transferred, how it may be used, retention and destruction terms, and any restrictions on re-identification risk.
- Harmonising variables across studies before any pooled model runs: recoding categorical variables to a shared scheme, aligning units and cut-points, and reconciling how each study originally defined the outcome so the “same” variable actually means the same thing everywhere it appears in the pooled dataset.
- Deciding how to handle non-participating studies. It is common for some eligible trials to be unable or unwilling to share IPD. A review can proceed as a hybrid design — IPD from the studies that provide it, aggregate data (from publications) for the rest — but this needs to be pre-specified in the protocol, not improvised after the fact, since it changes what statistical model is appropriate.
This is also where an IPD review’s timeline diverges most sharply from a standard systematic review: data requests, agreement negotiation, and cleaning/harmonisation routinely take longer than the eventual analysis.
One-stage vs. two-stage: the actual analytical choice
Once the pooled IPD dataset exists, there are two established ways to analyze it.
| Dimension | Two-stage | One-stage |
|---|---|---|
| What happens first | Each study is analyzed separately to produce a study-level effect estimate and its variance — the same output an aggregate-data review would use. | All participants from all studies are modelled together in a single statistical model. |
| Second step | The study-level estimates are pooled with a conventional meta-analysis model (fixed-effect or random-effects), the same way an AD meta-analysis pools published estimates. | There is no second pooling step — between-study heterogeneity is a random-effect term inside the same model, typically a hierarchical or mixed-effects regression. |
| Familiarity | Uses the same well-documented pooling methods described in the Cochrane Handbook’s core meta-analysis chapter — easier to explain to a review team already comfortable with aggregate-data synthesis. | Less standardised across software and review teams; more prone to model-convergence issues with complex random-effects structures. |
| Where it tends to do better | Larger studies with reasonably common events, where a normal-approximation summary estimate per study is a safe simplification. | Rare outcomes, small trials, or few events per study — because it uses an exact likelihood rather than approximating each study’s result as a normally-distributed summary statistic. |
| Clustering by study | Handled automatically, since each study is analyzed on its own before any pooling happens. | Must be explicitly specified in the model (e.g. stratifying the baseline term by study); getting this wrong is a documented source of bias unique to the one-stage approach. |
Why the two approaches can give different answers
It’s tempting to assume one-stage is simply the more sophisticated, more correct option. The methodological literature doesn’t support that as a blanket rule. Burke, Ensor and Riley’s widely-cited comparison (Statistics in Medicine, 2017;36(5):855–875, doi:10.1002/sim.7141) found that most of the numerical differences between one-stage and two-stage results trace back to differing modelling assumptions — which weighting scheme is used, whether treatment effects are fixed or random, how the heterogeneity variance is estimated, whether an exact or approximate likelihood is used for rare outcomes — rather than to the one-stage/two-stage choice itself. When the same assumptions are applied consistently in both approaches, the results converge. Their practical conclusion: the choice matters most for rare events, small trials, or where within-study clustering and correlated parameters need careful handling — not as a default preference for one method over the other.
PRISMA-IPD: reporting an IPD review is not the same as reporting a standard one
Standard PRISMA 2020 reporting items assume the review pooled published, aggregate results. An IPD review has additional things a reader needs to know that PRISMA 2020 doesn’t ask for: which eligible studies were approached for data and which actually provided it, what data-sharing arrangement governed the transfer, how variables were harmonised across datasets, how missing participant-level data were handled, and whether the analysis used a one-stage or two-stage model and why.
PRISMA-IPD (Stewart LA, Clarke M, Rovers M, Riley RD, Simmonds M, Stewart G, Tierney JF — JAMA, 2015;313(16):1657–1665) is the dedicated reporting extension that covers exactly this gap, with its own checklist and flow diagram distinct from the standard PRISMA 2020 documents. A protocol or manuscript for an IPD meta-analysis should be checked against PRISMA-IPD specifically, not just PRISMA 2020, since a reviewer or editor working in this space will expect the IPD-specific items (data provenance, sharing terms, harmonisation, one-stage/two-stage justification) to be addressed explicitly rather than folded into a generic methods paragraph.
The misconception this term invites
“IPD meta-analysis” sounds like it describes one specific statistical technique. It describes a data-acquisition strategy — using participant-level rather than published-summary data — that is compatible with either analytical approach above, and it is not automatically superior on every dimension. IPD synthesis is resource- and time-intensive, depends entirely on how many eligible studies are willing and able to share data (a review that only obtains IPD from a minority of eligible trials introduces its own selection question), and a well-conducted two-stage IPD analysis can be entirely appropriate and defensible — it is not a downgrade from one-stage, just a different, more familiar modelling route to the same pooled question.
Frequently asked questions
Is IPD meta-analysis always better than aggregate-data meta-analysis?
Not automatically. IPD gives a review team more analytical control — consistent outcome definitions, participant-level subgroup analysis, standardised missing-data handling — but it requires far more time, cooperation from every contributing study’s investigators, and data-governance work. When aggregate data already answers the question adequately and no participant-level analysis is needed, the added cost of an IPD approach may not be justified.
Do I have to choose one-stage or two-stage before starting?
The choice, and the reasoning behind it, should be pre-specified in the review protocol, the same way a pooling model choice is pre-specified in an aggregate-data meta-analysis. Deciding after seeing preliminary results risks selecting whichever method produces a more favourable answer.
What if only some of the eligible trials will share individual participant data?
This is common, not exceptional. A hybrid design that combines IPD from participating studies with aggregate data (from publications) for the rest is an established approach, but the protocol needs to specify how the two data types are combined statistically — this isn’t something to improvise once data collection is underway.
Is PRISMA-IPD a replacement for PRISMA 2020?
No. PRISMA-IPD is an extension used alongside standard PRISMA 2020 reporting, adding the items specific to individual participant data acquisition, sharing terms, harmonisation, and the one-stage/two-stage analytical choice that PRISMA 2020 doesn’t cover.
Does a one-stage model always produce a different answer from a two-stage model on the same data?
Not if the same modelling assumptions are used in both. Documented divergence between the two mostly traces back to different assumptions about weighting, fixed vs. random treatment effects, and heterogeneity estimation — not to an inherent property of the one-stage or two-stage structure itself.
Related CASRAI resources
- Meta-analysis — operational definition and worked examples for standard aggregate-data pooling.
- Systematic Review vs. Meta-Analysis — where pooling (of either kind) fits into the broader review process.
- Fixed-Effect vs. Random-Effects Meta-Analysis — the pooling-model choice that also applies inside a two-stage IPD analysis.
- Heterogeneity in Meta-Analysis — the between-study variance that both one-stage and two-stage models have to account for.
- Meta-Regression in Meta-Analysis — a study-level alternative to IPD subgroup analysis when only aggregate covariates are available.
- How to Peer Review a Systematic Review or Meta-Analysis — includes checking whether an IPD review’s data provenance and analytical choices were reported adequately.
- Data Sharing Agreement — the governance instrument that typically covers an IPD data transfer between trial teams.
- PRISMA 2020 — the base reporting standard that PRISMA-IPD extends.
