Skip to content
Back to search results

Individual Patient Data Meta-Analysis

Introduced by
Cochrane Collaboration, Pioneered by Stewart & Clarke
Year
1990s
Type
Method
Family
Advanced Meta-Analysis
Font size

What does it do?

Individual patient data meta-analysis (IPD-MA) is a systematic synthesis method where researchers obtain and analyze raw data at the patient level from multiple randomized controlled trials, rather than relying on published summary statistics (aggregate data).

Pioneered by the Cochrane Collaboration and formalized by Stewart, Clarke, and Riley, IPD-MA is considered the gold standard for evidence synthesis because it enables consistent outcome definition across trials, robust subgroup analysis, and detection of treatment-covariate interactions. Though time-intensive and resource-demanding, IPD-MA provides the most reliable estimates of intervention effects and is preferred for critical clinical decisions, particularly for identifying which patients benefit most from treatment.

The intuition behind the method

Standard meta-analysis combines summary results (odds ratios, mean differences) reported in published papers. However, published papers report only the outcomes and subgroups investigators chose to highlight, and different trials may define outcomes or subgroups inconsistently. IPD-MA solves this by obtaining the actual patient-level data from trial databases: each patient's age, baseline characteristics, treatment assigned, and outcomes. With raw data, researchers can (1) standardize outcome definitions across trials, (2) perform subgroup analyses fairly (comparing patients with Characteristic X across all trials, not just trials that published this subgroup), (3) investigate treatment-covariate interactions (does the treatment work better in older vs younger patients?), and (4) detect differential treatment effects that summary-level meta-analysis may miss. IPD-MA is the benchmark for precision medicine and treatment personalization.

History

Individual patient data meta-analysis emerged in the 1990s as the Cochrane Collaboration recognized that IPD could address limitations of aggregate meta-analysis. Pioneer studies, particularly in cancer and cardiovascular disease, demonstrated that IPD-MA could identify subgroup effects and provide more nuanced treatment guidance. Stewart and Clarke's methodological work in the 2000s standardized IPD-MA protocols. Riley and colleagues advanced statistical methods for handling IPD in mixed-model frameworks. Major clinical consortia (CTG, EBCTCG) now use IPD-MA as standard for meta-analyses addressing precision medicine questions. IPD repositories (Vivli, YODA Project) have grown to facilitate data sharing for meta-analysis and reduce the burden on individual trial authors.

When to use it

Conduct IPD-MA when (1) a precise, subgroup-specific understanding of treatment effects is critical for clinical decision-making (e.g., identifying which cancer patients benefit from a new drug), (2) published summary-level meta-analysis shows heterogeneity or conflicting results and IPD analysis might clarify mechanisms or identify moderators, (3) important subgroup effects are suspected but not consistently reported across trials, (4) you can secure reasonable data response from trial authors (typically 50-80% of trials), or (5) you have resources (time, funding, data management expertise) to obtain, harmonize, and analyze patient-level data. Common applications include cancer trials, cardiovascular interventions, and chronic disease management, where precision medicine questions are important. Avoid IPD-MA if data access is severely restricted or if no meaningful subgroup questions exist.

Limitations and considerations

  • Substantially more time-consuming and resource-intensive than aggregate meta-analysis. Obtaining data from trial authors, curating, and harmonizing patient datasets requires 12-24 months.
  • Dependent on data availability: trials may refuse to share data due to privacy concerns, lack of resources, or loss of data. Non-response can bias results if negative trials are less likely to share.
  • Data quality may vary across trials; incomplete or inconsistently measured variables limit analyses. Missing data patterns differ by trial and require careful handling.
  • Statistical complexity increases; fitting hierarchical regression models with multiple interactions requires statistical expertise and can be prone to overfitting if not carefully planned.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Enables standardized outcome definition and consistent subgroup analysis across heterogeneous trials, increasing transparency and reducing bias from selective reporting.
  • Allows investigation of treatment-by-covariate interactions, identifying which patient subgroups benefit most—essential for precision medicine and personalized treatment recommendations.
  • Typically yields more precise effect estimates than aggregate meta-analysis by using all available data and correcting for trial-level clustering.
  • Detects differential treatment effects that summary-level meta-analysis may miss, particularly when subgroup effects are small or inconsistently reported.
  • Enables investigation of trial-level heterogeneity: trial design features, implementation fidelity, or patient population characteristics that explain outcome variation.

Applications

  • Cancer drug trials: IPD-MA of multiple RCTs to identify which tumor subtypes, molecular markers, or patient ages benefit from a new oncology therapy.
  • Cardiovascular intervention: synthesizing individual patient data from multiple heart disease trials to determine which patients (by age, baseline risk, comorbidities) benefit from angioplasty vs medical therapy.
  • Psychiatric treatment: IPD-MA of psychotherapy RCTs to identify which patients (by symptom profile, age, comorbidity) respond best to cognitive-behavioral vs interpersonal therapy.
  • Diagnostic test accuracy: IPD-MA of diagnostic imaging or biomarker studies to assess test accuracy across different patient risk strata, refining when to use the test.
  • Precision medicine trials: IPD-MA combining multiple studies to identify patient-treatment interactions, supporting development of treatment selection algorithms for personalized medicine.

How it works

How it works

1

Conduct a comprehensive systematic review following standard systematic review protocols: register your protocol (PROSPERO), define PICO criteria, search multiple databases, screen titles and abstracts, and identify eligible randomized controlled trials (typically RCTs; observational studies can be included but must be clearly distinguished). Contact all trial authors and data custodians requesting individual patient data for inclusion. Specify the data variables you need: baseline characteristics (age, sex, disease severity, comorbidities), treatment assignment, protocol adherence, and outcome measurements (timing must be standardized across trials). Provide data transfer agreements, secure data repositories (e.g., university servers with encryption), and timelines. Expect that not all trials will respond or provide data; document response rates and reasons for non-response. Harmonize data across trials: variables may be coded differently (age ranges in Trial A vs exact ages in Trial B); create a master dataset with consistent variable definitions. Assess completeness: identify missing data patterns and decide on imputation strategies.

2

Perform a one-stage meta-analysis by fitting a regression model (most commonly a linear mixed model for continuous outcomes or a logistic mixed model for binary outcomes) that includes patient characteristics, trial identity, treatment assignment, and interactions of interest. This approach correctly accounts for clustering of patients within trials. Alternatively, use two-stage methods where trial-specific analyses are conducted first, then trial results are combined (similar to conventional meta-analysis, but with standardized analyses across trials). Examine treatment effects overall and by pre-specified subgroups (age groups, sex, disease severity): does the treatment work equally well for all patients, or are there differential effects? Test for treatment-by-covariate interactions. Assess heterogeneity across trials: Is the treatment effect consistent, or does it vary significantly by trial context? If heterogeneity is high, explore sources (trial design features, patient populations, intervention implementation). Perform sensitivity analyses: exclude small trials, exclude trials with high attrition, or exclude outlier trials to test robustness.

3

Report results with clear stratification by patient subgroups: Effect size (odds ratio, hazard ratio, or mean difference) with 95% credible or confidence intervals, overall and for each pre-specified subgroup. Present interaction p-values and plots showing how treatment effect varies across covariate ranges. If treatment benefit is similar across all subgroups, conclude that the intervention is broadly effective. If significant interactions exist (e.g., treatment works in older patients but not younger, or in severe disease but not mild), emphasize these differential effects—this information is actionable for clinical practice. Compare results with prior summary-level meta-analysis if available: IPD-MA often yields more precise estimates and may reveal subgroup effects missed in aggregate analysis. Document which trials contributed IPD and any trials excluded due to non-response; assess whether missing trials could bias results (e.g., if negative trials are less likely to share data). Discuss implications for practice: Recommend treatment for which subgroups? Highlight populations needing further research. Use GRADE approach to rate certainty of evidence for overall effect and for key subgroup effects.

Common pitfalls

  • Conducting exploratory subgroup analyses without pre-specification. Many subgroup interactions tested (e.g., treatment × age, treatment × sex, treatment × baseline severity) increase false-positive rate. Pre-register analyses and adjust p-values for multiple testing.
  • Assuming missing data are missing completely at random (MCAR) without testing. If sicker patients are more likely to have missing outcome data, analysis may overestimate treatment benefit. Conduct sensitivity analyses under different missing data assumptions.
  • Over-interpretation of subgroup effects in small subgroups. If only 50 patients exist in a subgroup, credible intervals around the subgroup effect estimate will be wide; avoid strong clinical recommendations based on imprecise estimates.
  • Ignoring trial-level clustering. Analyzing all patient data as if independent (rather than using mixed models accounting for trial clustering) underestimates variance and inflates Type I error rate.

Frequently asked questions

What is the difference between IPD meta-analysis and aggregate (summary-level) meta-analysis?

Aggregate meta-analysis combines published effect estimates (e.g., odds ratios) from multiple trials. IPD meta-analysis obtains raw patient-level data and re-analyzes it centrally with standardized methods. IPD allows consistent subgroup analyses and exploration of patient-treatment interactions; aggregate meta-analysis relies on subgroup results reported in papers, which are often inconsistent or missing. IPD is more rigorous but much more resource-intensive.

How much missing data is acceptable in IPD meta-analysis?

Missing data at the patient level (e.g., some patients missing an outcome measurement) is common and can be handled with imputation methods. Missing data at the trial level (entire trial not providing IPD) is more problematic. Aim for IPD from 60-80%+ of eligible trials. Document response rates and assess whether missing trials differ systematically (e.g., are negative trials less likely to respond?). If response is <50%, bias risk increases; sensitivity analyses become essential.

How do I test for treatment-by-covariate interactions in IPD meta-analysis?

Include interaction terms in your mixed-effect regression model (e.g., treatment × age, treatment × baseline severity). Test the interaction term's statistical significance. Estimate the treatment effect separately for different covariate strata (e.g., age <65 vs ≥65). Plot the treatment effect across covariate ranges to visualize interactions. Pre-specify which interactions you will test before analyzing data to avoid false positives from multiple testing.

Can I combine RCTs and observational studies in IPD meta-analysis?

Yes, but clearly distinguish them and account for study design in analysis (e.g., use a mixed model including a study design indicator). RCTs provide strongest causal evidence; observational studies add breadth but carry confounding risk. Pre-specify whether observational studies will be analyzed separately, together with RCTs in a sensitivity analysis, or excluded entirely. Document any differences in findings by study design.

Related methods

Sources

  1. Stewart, L. A., Clarke, M. J., & Cochrane IPD Meta-analysis Methods Group. (2015). Practical methodology of meta-analyses (including IPD) of randomised trials reporting time to event data. Cochrane Database of Systematic Reviews, 2015(10), MR000027.

    Open publication reference
    Copied.
  2. Riley, R. D., Lambert, P. C., & Abo-Zaid, G. (2010). Meta-analysis of individual participant data: rationale, conduct, and reporting. BMJ, 340, c221.

    Open DOI reference
    Copied.
  3. Higgins, P. T., & Green, S. (Eds.). (2011). Cochrane Handbook for Systematic Reviews of Interventions (Version 5.1.0). The Cochrane Collaboration.

    Open publication reference
    Copied.