Errors in causal diagrams elicited from experts can lead to the omission of important confounding variables from adjustment sets and render causal inferences invalid. In this report a novel method is presented that repairs a misspecified causal diagram through the addition of edges. These edges are determined using a data-driven approach designed to provide improved statistical efficiency relative to de novo structure learning methods. Our main assumption is that the expert is "directionally informed", meaning that "false" edges provided by the expert would not create cycles if added to the "true" causal diagram. The overall procedure is cast as a pre-processing technique that is agnostic to subsequent causal inferences. Results based on simulated data and data derived from an observational cohort illustrate the potential for data-assisted elicitation in epidemiologic applications.
- causal diagrams
- causal inference
- directed acyclic graphs
- expert opinion
- structure learning algorithms