Individual Participant Data Meta-Analysis: What the Raw Records Add

An individual participant data meta-analysis, or IPD meta-analysis, requests the original participant-level records from eligible studies, checks and harmonizes them, and performs a new synthesis. It can answer questions that published summary tables cannot, especially about time-to-event outcomes and treatment-effect differences across participant characteristics, but its credibility still begins with a complete systematic review and depends on which investigators share usable data.

The systematic review comes first#

An IPD project should begin like any other systematic review: a protocol, a focused question, comprehensive study search, explicit eligibility criteria, duplicate-study resolution, risk-of-bias assessment, and a prespecified synthesis plan. Investigators should not begin with the teams most willing to share data and call that collection the evidence base.

The eligible-study set and the obtained-IPD set are separate. Reviewers may identify twenty trials but receive participant records from fifteen. Published aggregate results from the other five still matter, and the review should give which studies supplied IPD, which did not, why data were unavailable, and whether obtained and unavailable studies differ.

Data sharing can uncover unpublished trials or clarify whether a report is eligible. It can also introduce a new availability filter. Older studies, small teams, unfavorable results, and trials with weak data systems may be less likely to supply files; a high percentage of participants obtained is reassuring only when the search for all eligible studies was sound.

What arrives is not analysis-ready#

Trial teams may send different file formats, variable names, coding conventions, units, visit windows, and definitions. One trial records age in years and another in date fields; one calls an event “recurrence” while another separates clinical from imaging recurrence; and missing values may be blank, coded as 999, or buried in a status field. Nothing lines up by itself.

The IPD team builds a data dictionary and maps each study into a harmonized structure; this work should preserve the original values and create a traceable record of every recode or derived variable. Harmonization is a scientific decision. Combining similar labels that represent different clinical events can create artificial consistency.

Some outcomes cannot be harmonized without losing meaning. The appropriate response may be separate analyses, a broader but defensible definition, or exclusion from that outcome, with the limitation stated; uniform columns are not evidence that the underlying measurements are equivalent.

Data checking is a major advantage#

Participant records allow checks that are impossible from a journal table. Reviewers can inspect ranges, dates, duplicates, internal consistency, missingness, and treatment coding. In randomized trials, they can compare randomization lists where available, assess baseline balance, and check whether all randomized participants are represented.

Time-to-event data can be reconstructed using actual follow-up rather than a single published hazard ratio or event count. Longer follow-up may be available than at the original publication. Outcomes measured but omitted from the article may be recovered, reducing outcome-reporting bias within contributing studies.

Queries should go back to the original team. A surprising date sequence may be a formatting error, a protocol deviation, or a real clinical event. The IPD reviewers should not silently “clean” it according to expectation. Audit trails and agreed corrections preserve accountability.

These checks improve the available data, but they do not erase flaws in design or conduct: broken allocation concealment, biased outcome assessment, substantial loss to follow-up, or selective enrollment remain source-study limitations.

Preserve each trial's randomization#

The most tempting error is to stack every participant into one file and compare all treated people with all controls as though the data came from one enormous trial. They did not: participants belong to studies with different randomization schemes, baseline risks, follow-up schedules, and treatment contrasts, and ignoring study can produce confounding and confidence intervals that are too narrow.

Valid models stratify or otherwise account for study. In a two-stage approach, each trial is analyzed first to produce its treatment-effect estimate. Those estimates are then combined using conventional meta-analysis methods. This keeps trial-specific comparisons visible and produces familiar forest plots.

In a one-stage approach, participant records are analyzed in one hierarchical or stratified model that accounts for study, and this can handle complex outcomes and interactions efficiently, but assumptions about baseline risk, heterogeneity, and correlation must be explicit.

One-stage and two-stage analyses often agree when they target the same estimand and use compatible assumptions. Differences may reflect model choices rather than a fundamental superiority of one format. The protocol should choose the approach from the question, and include sensitivity analyses when a reasonable alternative would change what you conclude.

The strongest use case is effect modification#

A common question is whether treatment works differently by age, sex, baseline severity, biomarker, or another participant characteristic. Aggregate meta-analysis can be misleading here. Trials with older average participants may differ from trials with younger averages in setting, dose, era, or risk, and a relationship across trial averages does not establish that age modifies treatment response within a trial.

With IPD, reviewers can estimate a treatment-by-characteristic interaction within each randomized trial and then pool those interaction estimates. This preserves the protected comparison created by randomization. A one-stage model can do the same if it separates within-study from between-study information correctly.

Subgroup averages alone are not enough. Showing benefit in one subgroup and a nonsignificant result in another does not prove the effects differ, and the relevant test is the interaction, with its confidence interval and clinical magnitude.

IPD also makes it easy to search many thresholds and characteristics. That flexibility raises the risk of chance findings. Effect modifiers should be limited, biologically or clinically justified, prespecified, and assessed for consistency. Exploratory results should be labeled as such and not converted directly into treatment rules.

Outcome definitions and missing data become more transparent#

Published studies may report different time points or dichotomize a continuous scale at different thresholds. IPD can apply one definition and one follow-up window where the source data support it. It can analyze continuous values without unnecessary categorization and use actual survival times.

Participant-level missing data remain a problem. IPD allows more thoughtful modeling and sensitivity analysis, but no method can recover information without assumptions. The reason and pattern of missingness should be assessed within each study. A common imputation model should not erase differences in trial context.

Study-level missingness is separate. If an entire eligible trial does not provide IPD, sophisticated imputation within received files does nothing to address that missing study. Reviewers may combine IPD with available aggregate results, conduct sensitivity analyses, or present the two evidence sets side by side.

Privacy and governance are part of rigor#

Participant-level files require secure transfer, controlled access, data-use agreements, and compliance with applicable consent, ethics, and privacy requirements. The analysis team should request only variables needed for the protocol and document who can access them, how long they are retained, and how results avoid re-identification.

De-identification does not reduce scientific responsibility. Rare conditions, small sites, and detailed event dates can make records identifiable when combined. Governance should be planned with the same care as the statistical model.

Collaboration with original investigators can improve interpretation and resolve data questions. It can also create conflicts if contributors influence eligibility, analysis, or reporting. Roles, decision rights, funding, and conflicts of interest should be transparent.

IPD cannot solve every meta-analysis problem#

Publication bias persists if unreported studies are never identified. Data-availability bias arises if contributing trials differ from noncontributing trials. Clinical and methodological heterogeneity remain. A precisely harmonized outcome can still be the wrong outcome for your decision.

IPD also takes substantial time, funding, statistical skill, data-management work, and collaboration. Aggregate data may answer some questions adequately, and the added effort is most justified when published summaries cannot support a reliable synthesis, when consistent reanalysis materially improves the outcome, or when credible participant-level effect modification is central.

Calling IPD a “gold standard” can therefore stop you appraising it. Quality depends on the review question, search, completeness, source trials, harmonization, model, governance, and reporting. Raw records create opportunities for stronger analysis, not an automatic ranking.

How PRISMA-IPD improves transparency#

PRISMA-IPD extends systematic-review reporting for projects that collect participant data. The flow diagram should distinguish eligible studies from those supplying IPD and show participants included at each stage. Reports should describe data checking, integrity concerns, harmonization, unavailable data, risk of bias, and analysis methods.

The protocol and report should list requested variables and outcomes, define one-stage or two-stage methods, explain interaction analyses, and state how aggregate results from non-IPD studies were used. You need enough information to see where the review gained detail and where missing data still constrain it.

An appraisal sequence#

First confirm that the eligible studies came from a systematic search. Calculate the proportion of eligible studies and participants represented by IPD, and compare contributors with noncontributors. Review the harmonization rules and data checks. Check that models preserve study structure and randomization.

For subgroup claims, find a within-study interaction rather than separate significance tests. Inspect prespecification, multiplicity, missing-data assumptions, and heterogeneity. Finally, ask whether access to participant records changed the evidence you would act on, or mainly reproduced a conclusion the aggregate data already supported.