A Tool for Distributed Collaborative Causal Discovery

The development of accurate causal models is crucial for achieving and explaining desired outcomes that require interventions. Building these models efficiently requires combining available data with expert causal knowledge. Often experts have unique data and model insights, but sharing them is challenging due to privacy or security concerns. Federated machine learning addresses similar issues by allowing multiple sites to collaborate on a common model without sharing private datasets. This paper introduces CCaT, a distributed causal discovery tool enabling collaborative development of a shared causal model while preserving local models and data privacy. CCaT allows each site to evaluate and refine the shared model using its private dataset, sharing only summary statistics or suggested new causal relations. The tool supports maintaining distinct local causal models, as analysts can choose to adopt or change parts of the shared model. CCaT enhances the accuracy of causal models by leveraging diverse expertise and data, achieving a generality and accuracy unattainable by individual sites. We present several common scenarios with the CCaT to demonstrate its effectiveness.