Individual Patient Data Meta-Analysis

Individual Patient Data Meta-Analysis: Concept Architecture

Orientation and learning roadmap

The page follows participant records from eligible studies through checking, harmonisation, and synthesis. It compares two-stage and one-stage analyses, works a simple pooling calculation, and shows how participant-level modifiers can inform a health-economic model. The last sections examine missing study data, trial-level clustering, and interpretation limits.

Assemble the eligible studies before requesting records

An IPD meta-analysis begins with a systematic question and an explicit search for all eligible studies, not just investigators who agree to share records. The team seeks participant-level observations such as treatment allocation, baseline covariates, outcomes, follow-up, and censoring from each study. Published study summaries remain relevant for assessing studies whose individual data cannot be obtained.

Data sharing requires agreements, privacy protection, and a reproducible analytic plan. Reanalysis may resolve inconsistencies in reported methods, but access to rawer data does not remove original trial bias or measurement limitations. The analysis should document changes from the protocol.

Choose a two-stage or one-stage analysis

In a two-stage analysis, each study's participant records are analysed separately to obtain a compatible effect estimate and uncertainty measure. Those study-specific estimates are then synthesised using a prespecified meta-analysis model. This preserves the study as the unit of evidence in the second stage and makes study-level heterogeneity visible.

A one-stage analysis fits a joint model to participant records from all studies while explicitly accounting for participants clustered within studies. Study-specific baselines, treatment effects, and relevant random effects or stratification should reflect the outcome and estimand. Pooling all records into one ordinary analysis while ignoring study membership can produce misleading precision or confound study differences with patient effects.

Either approach can estimate an overall effect when correctly specified; neither is inherently superior for every outcome or sparse-data setting. One-stage and two-stage estimates can differ because their models, weighting, treatment of sparse studies, or assumptions differ. State the estimand and assumptions before comparing them.

Check a small two-stage calculation

Suppose two hypothetical randomised studies provide harmonised participant-level estimates of an additive treatment effect on a common outcome: $\widehat\theta_1=0.10$ with standard error 0.05, and $\widehat\theta_2=0.20$ with standard error 0.10. Under a common-effect inverse-variance illustration, their weights are $1/0.05^2=400$ and $1/0.10^2=100$. The pooled effect is therefore weighted more heavily toward the more precise study.

$$\widehat\theta=\frac{\sum_j w_j\widehat\theta_j}{\sum_j w_j} =\frac{400(0.10)+100(0.20)}{500}=0.12, \qquad \operatorname{SE}(\widehat\theta)=\sqrt{1/500}\approx0.0447.$$

This arithmetic assumes the study estimates are independent, comparable on the same additive scale, and treated as sharing an underlying effect for this example. A random-effects model or a time-to-event outcome would require appropriate study-specific effect measures and synthesis assumptions. The example demonstrates the second stage; obtaining IPD is what enables the consistent first-stage reanalysis and patient-level investigations.

Estimate participant-level effect modification

Participant records can examine whether treatment effects differ by a baseline characteristic, such as disease severity, age, or a validated biomarker. The relevant evidence for a patient-level interaction is the comparison of treatment groups across characteristic levels within each study. Comparing studies with older average participants against studies with younger average participants can produce an ecological association that does not represent individual-level modification.

In a two-stage design, estimate a treatment-by-characteristic interaction within each trial and then pool those interaction estimates. In a one-stage design, separate the within-study characteristic deviation from its study mean when forming interactions, or use an equivalent parameterisation that keeps within- and between-study information distinct. Prespecify clinically plausible modifiers, their scale, missing-data handling, and multiplicity safeguards; a subgroup pattern from a small sample is uncertain.

Translate estimates into health-economic inputs

An economic model may use pooled comparative treatment effects, baseline-risk relationships, subgroup effects, and uncertainty estimates from an IPD synthesis. Keeping participant records can support consistent time-to-event analysis and exploration of covariate distributions in the target population. The model still needs absolute baseline risks, utility values, resource use, time horizon, and external validity assumptions rather than a pooled relative effect alone.

If patient-level costs and health outcomes are jointly analysed, preserve their dependence and the study structure where possible. Avoid treating participant records as if everyone came from one homogeneous trial or copying a subgroup interaction into an unrelated population without checking transportability. Differences in eligibility, clinical practice, follow-up, and comparator availability can be decision-critical.

Audit missing data, bias, and reporting

IPD may be unavailable for eligible studies, while participants or outcomes may be missing within studies that provide files. These are distinct gaps and can require different sensitivity analyses. Compare characteristics and published results for contributing and non-contributing studies, evaluate risk of bias, and report the flow of studies and participants as PRISMA-IPD recommends.

• Prespecify the review question, eligibility rules, outcomes, effect scale, and analysis methods.
• Reconcile received records with trial reports and explain any discrepancies in sample size or events.
• Keep study membership in the analysis and inspect between-study heterogeneity.
• Derive patient-level interactions from within-study comparisons, not study-average proxies.
• Explore the impact of studies without IPD and participants with missing outcomes or covariates.
• Report uncertainty and limits on applying results to the health-economic target population.

Sources and technical basis

Cochrane's IPD chapter describes the study search, data checks, one-stage and two-stage synthesis, and within-trial interactions. PRISMA-IPD provides reporting requirements for requested and received data, while methodological research explains ecological bias in participant-level interactions. The two-study pooling numbers above are original teaching values, not empirical treatment effects.

• Cochrane Handbook, Chapter 26: Individual participant data.
• PRISMA-IPD Statement.
• Burke and colleagues, One-stage and two-stage IPD meta-analysis approaches.
• Hua and colleagues, Avoiding ecological bias in one-stage treatment-covariate interactions.