Can algorithms replace expert knowledge for causal inference? A case study on novice use of causal discovery

Area of Science:

• Epidemiology
• Causal Inference
• Machine Learning

Background:

• Growing interest in causal inference and machine learning among epidemiologists.
• Increasing discussion of causal discovery algorithms for guiding covariate selection.

Purpose of the Study:

• Present a case study of novice application of causal discovery tools.
• Validate causal discovery results against a well-established causal relationship.

Main Methods:

• Applied 4 causal discovery algorithms to the Coronary Drug Project (CDP) data set.
• Investigated the effect of adherence on mortality in the placebo arm.
• Varied model inputs and identified 15 adjustment sets from 17 parameterizations.

Main Results:

• Identified adjustment sets that produced effect estimates with similar bias as prior published results in baseline analyses.
• Observed greater residual bias with causal discovery methods compared to expert-selected sets when controlling for time-varying confounding.

Conclusions:

• Causal discovery algorithms can perform comparably to expert knowledge.
• Recommend expert support for novice users in algorithm selection, parameter tuning, assumption assessment, and variable finalization.