Individual Participant Data (IPD)

Individual participant data (IPD) refers to the raw, observation level data collected for each participant in a clinical study, as opposed to the aggregate summary statistics reported in published articles. In IPD meta analysis, original datasets from multiple studies are obtained and reanalyzed to produce pooled estimates that can address questions not answerable from published summaries alone.

IPD meta analysis is widely regarded as the gold standard of evidence synthesis because it offers advantages that aggregate data meta analysis cannot provide. With access to individual records, analysts can standardize outcome definitions across studies, perform consistent statistical adjustments, explore patient level effect modifiers (such as age, disease severity, or biomarker status), and handle missing data more effectively. IPD meta analyses are particularly valuable when published studies use different subgroup definitions, when treatment effect heterogeneity is suspected, or when regulatory agencies and guideline panels need the highest quality evidence to support clinical recommendations.

Conducting an IPD meta analysis involves several stages beyond those required for a standard aggregate data meta analysis. First, the research team conducts a systematic review to identify all relevant studies. Second, they contact the investigators of each identified study to request access to the raw datasets. This process often requires data sharing agreements, ethical approvals, and considerable time; response rates vary, and some studies' data may be unavailable due to data loss, investigator refusal, or institutional restrictions.

Once obtained, the datasets must be harmonized. Variables may be coded differently across studies (for example, one study may record blood pressure in mmHg while another records it as a categorical variable). The IPD meta analysis team standardizes variable definitions, recodes data as needed, and resolves discrepancies with original investigators. Data quality checks identify implausible values, duplicate records, and patterns of missingness.

The analysis can follow either a one stage or two stage approach. In the two stage approach, each study is analyzed separately using a common statistical model, and the study specific results are then combined using standard meta analytic methods. In the one stage approach, all individual records are pooled into a single dataset, and a hierarchical or mixed effects model is fitted that accounts for clustering of participants within studies. The one stage approach is generally more flexible, particularly when studies are small or when interactions between participant characteristics and treatment effects are of interest. Such interactions should be estimated from comparisons within trials rather than across trials, because across trial comparisons are prone to ecological (aggregation) bias.

Reporting of IPD meta analyses is guided by the PRISMA-IPD statement (2015), an extension of PRISMA that adds items on checking the integrity of the IPD, reporting any important issues found, and exploring variation in effects, such as whether some types of participant benefit more than others.

IPD meta analysis is most valuable when the research question involves patient level effect modification (for example, does the treatment work better in younger patients?), when published studies report outcomes inconsistently, when the goal is to develop or validate a clinical prediction model, or when the number of events is too small for reliable subgroup analysis within individual studies. It is less practical when the number of relevant studies is very large, when data sharing barriers are insurmountable, or when the research question can be adequately addressed using published summary data.

A common error is conducting an IPD meta analysis without obtaining data from all or nearly all identified studies, which introduces availability bias if the studies with available data differ systematically from those without. Another mistake is failing to harmonize variable definitions across studies, leading to measurement heterogeneity that undermines the analysis. Some researchers treat the pooled IPD dataset as a single large study without accounting for the clustering of participants within studies, which can bias the pooled estimates as well as produce incorrect standard errors. Finally, inadequate data quality checks can allow coding errors and implausible values from original datasets to propagate into the pooled analysis.

A collaborative group of trialists conducts an IPD meta analysis of 12 RCTs (total N = 8,400) examining anticoagulation therapy after venous thromboembolism. They obtain individual records from 10 of the 12 trials (92% of participants). Harmonizing the datasets reveals that three trials used different definitions of "major bleeding," which the team standardizes using ISTH criteria. The one stage mixed effects analysis, which estimates the age interaction from comparisons within trials, finds that the treatment benefit is significantly greater among patients aged 65 or older compared to younger patients (interaction p = 0.008), a finding that no individual trial had the power to detect.