Power priors for replication studies

Area of Science:

• Statistics
• Scientific Methodology

Background:

• The replication crisis in science necessitates robust methods for evaluating replication studies.
• Assessing the compatibility between original and replication study findings is crucial for scientific integrity.

Purpose of the Study:

• To propose a novel Bayesian analysis approach using power priors for evaluating replication studies.
• To quantify the compatibility between original and replication data.
• To explore dynamic information borrowing from original studies in replication analyses.

Main Methods:

• A Bayesian power prior approach where the original study's likelihood is raised to a power $$ and used as the prior for replication data.
• Posterior distribution and Bayes factor hypothesis tests on the power parameter $$ to assess study compatibility.
• Generalization of the connection between normal power priors and normal hierarchical models.

Main Results:

• The power prior approach quantifies the degree of compatibility between original and replication studies.
• Inferences for parameters like effect sizes dynamically borrow information from the original study, adapting to data conflict.
• A generalized connection between power prior models and hierarchical models is established, clarifying their relationship.

Conclusions:

• The proposed Bayesian power prior method offers a flexible framework for analyzing replication studies.
• This approach provides a principled way to quantify evidence for or against replication success.
• Understanding the link to hierarchical models enhances the interpretability and application of power prior methods in scientific research.