A power prior is a Bayesian method that uses historical data as prior information for a current study. The likelihood of the historical study is raised to a weight between zero and one before it enters the prior distribution. This weight determines how strongly the earlier evidence influences the current estimate and decision; it can therefore selectively discount historical information.
How historical evidence is weighted
A weight of zero excludes the historical data from prior construction, while a weight of one carries over its likelihood without discounting. Intermediate values reduce its influence. The prior then contains both a baseline prior and the down-weighted historical information, which is combined with the data from the new study to form the posterior distribution. The amount of prior information can be expressed as an effective historical sample size: it describes how many additional observations the chosen weight adds to the current study in computational terms, making the extent of borrowing tangible for assessment.
The method is suitable even when only one relevant prior study is available. It does not require a hierarchical model across multiple historical studies to form a distribution. This simplicity is precisely why the choice of weight is central: it is not a minor technical setting, but determines the strength of borrowing from the single historical source.
Data comparability and conflict
Historical and current data are not interchangeable simply because they share the same indication. Eligibility criteria, endpoint definition, background therapy, geographic standard of care, follow-up duration, and handling of intercurrent events can limit the relevance of the prior information. The FDA highlights that differences in estimands or estimators between an external source and a prospective study can affect the suitability of the data for borrowing.
A high degree of borrowing may appear to improve precision, but can increase bias if the historical study systematically reflects different patients or treatment conditions. The rationale must therefore disclose data provenance, selection criteria, the comparability assessment, and how potential conflicts are handled. Sensitivity analyses with alternative weights show whether the conclusion materially depends on the historical source.
The weight can be fixed in advance or determined in a data-dependent manner. With a data-dependent choice, the particular challenge is that the current study is used both to assess historical agreement and to estimate the treatment effect. The properties of this procedure must be examined for realistic conflict scenarios. Presenting the posterior effect alone is not sufficient; it is also important how the degree of borrowing changes when historical and current data diverge.
The power prior can accommodate different data types, but the parameterization and weight must fit the respective endpoint model. A weight transferred from a different endpoint would not constitute an empirical validation of the historical information.
Distinction from other priors
The power prior discounts historical data directly via the exponentiated historical likelihood. It is not the same as a meta-analytic-predictive (MAP) prior, which combines multiple historical studies in a hierarchical model and derives a prediction for the current study. In contrast, a power prior generally requires only a single prior study.
The commensurate prior also follows a different modeling idea: it models similarity between the historical and current parameter via a distributional structure. The power prior sets the extent of information borrowing through its weight. The umbrella term Bayesian statistics includes both approaches, but does not describe the specific mechanism of discounting.
Relevance for clinical trials
For planned use, the sponsor and statistics team must define in advance which studies or datasets will enter the prior, how their weight will be determined, and which robustness analyses will accompany the regulatory interpretation. The current study must provide sufficient independent evidence so that a non-matching historical population does not covertly determine the treatment effect. Reproducibility of the historical analysis is particularly important if individual patient data are to be re-analysed according to the current estimand.
Full-service CROs such as Mediconomics support the identification and data qualification of suitable prior studies, create comparability matrices for population, endpoint, and therapy context, specify weighting and sensitivity analyses in the statistical analysis plan, and prepare a transparent presentation of historical borrowing for the clinical study report.
Frequently Asked Questions (FAQ)
Does a power prior always use exactly one historical study?
No. The concept can pool historical information, but unlike the MAP prior it does not depend on hierarchically modeling multiple studies. It can therefore also be used with a single prior study.
Is a high weight proof of good comparability?
No. The weight is a modeling decision. Comparability must be justified independently based on the data sources, study conditions, and alignment with the estimand.
Can a power prior replace a randomized control group?
That depends on the clinical question and the overall evidence plan. Historical borrowing does not eliminate the risks of systematic differences between the old and current data source.
Regulatory References
- FDA, Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products – addresses informative priors and external data in drug development.
- EMA/CHMP/1813/2026, Concept Paper on Bayesian Methods in Clinical Development – highlights comparability, estimands, and error control for informative priors.
- ICH E9(R1), Addendum on Estimands and Sensitivity Analysis – requires analysis aligned with the estimand and sensitivity assessment.
