A meta-analytic predictive prior, or MAP prior for short, is a Bayesian prior distribution derived from multiple historical studies using a hierarchical model. The model estimates both the average historical evidence and the between-study heterogeneity, and uses these to generate a prediction for the effect in the current study. The method therefore requires multiple prior studies and a justified assumption that their effects are exchangeable with those of the new study in a relevant sense.
From the study pool to prediction
The MAP prior is not a simple summary of published effect sizes. In the hierarchical model, each historical study is assigned its own study-specific effect; the variation in these effects contributes to the uncertainty of the prediction for the new study. With heterogeneous historical results, the prior typically becomes wider and transfers less precise information to the new study parameter.
The predictive component focuses on the upcoming study, not merely on a common historical mean. Therefore, the selected studies must meaningfully match the new investigation with respect to population, comparator treatment, endpoint, follow-up, and care setting. A formally calculated prior is not a justification for treating different study contexts as equivalent.
Exchangeability and robust borrowing
Exchangeability does not mean identity. It states that the historical and current effects can arise from a common higher-level distribution, while allowing variation between them. This assumption becomes weak if standard of care, diagnostics, inclusion criteria, or the definition of the target endpoint have changed. In that case, hierarchical heterogeneity may capture the differences only inadequately.
Robust MAP approaches often supplement the informative prior with a weakly informative component. This allows the current study to rely more strongly on its own data when there is conflict with the historical evidence. Which robustness variant is used and which criteria apply for study selection must be defined before reviewing the new results; retrospectively shifting the historical selection would undermine the benefit of transparent borrowing.
The number of studies alone does not guarantee a reliable MAP prior. Several small studies with the same systematic limitations can produce an apparently narrow distribution without improving transferability to the current study. In addition, the choice of heterogeneity model influences how strongly rare outlying study results widen the prediction. The model report should therefore present study weights, estimated heterogeneity, and the effect of a robust mixture in a way that makes the contribution of historical evidence to the current conclusion traceable.
The current study remains the source of new evidence. The MAP prior changes how historical experience is incorporated into its evaluation, not the requirements for a clearly defined current study question.
Distinction from meta-analysis and power prior
A meta-analysis combines studies to synthesize evidence on effects that have already been investigated. A MAP prior uses a hierarchical synthesis as input for a future Bayesian analysis; its function is predictive prior information for the current study. The two methods may use similar models, but they do not necessarily address the same decision question.
In contrast to the power prior, historical influence is not controlled by directly raising a historical likelihood to a power with a single weight. The MAP prior models heterogeneity across multiple studies and is therefore not designed for only one historical study. A commensurate prior can also work with a single prior study and is therefore not an alternative name for the MAP approach.
Relevance for clinical trials
Implementation begins with a transparent search and selection strategy for historical studies. For each source, data quality, comparator treatment, observation window, and alignment of estimands must be assessed; published summaries are often insufficient to identify all differences in endpoint definitions. Development planning should also simulate how heterogeneous or conflicting historical data affect decision probabilities and erroneous decisions.
Full-service CROs such as Mediconomics support the structured identification and assessment of historical evidence, organize the harmonization of endpoint and variable definitions, develop the hierarchical model with biostatistics including a robust component, and document selection, assumptions, simulations, and results for the study protocol, analysis plan, and regulatory submissions.
Frequently Asked Questions (FAQ)
Can a MAP prior be constructed from a single historical study?
No. Without multiple studies, the between-study heterogeneity that characterizes the MAP prior cannot be estimated from a historical study pool.
Is a MAP prior itself a meta-analysis?
It uses a meta-analytic hierarchical structure, but it is applied as a prior for predicting and analyzing a new study. Its role therefore goes beyond the retrospective summary of existing studies.
Why is the same indication not sufficient for exchangeability?
The same indication does not rule out differences in patient selection, standard of care, endpoint measurement, or observation duration. These differences in particular can bias the transferred effect.
Regulatory References
- FDA, Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products – addresses the assessment of external information and informative priors.
- EMA/CHMP/1813/2026, Concept Paper on Bayesian Methods in Clinical Development – highlights the comparability of historical data and alignment with the estimand.
- ICH E9(R1), Addendum on Estimands and Sensitivity Analysis – is relevant for comparability of the target variable across data sources.
