Existence of LLMs or prompts that can effectively select covariates for confounder identification

Determine whether any large language model or prompt design can effectively select covariates for confounder identification in real-world causal analysis contexts such as the Coronary Drug Project.

Background

Although current models show inconsistent and often inadequate performance for confounder selection, the authors acknowledge that their results do not rule out the possibility that future LLMs or improved prompting strategies could succeed.

They propose that the CDP setting and code can serve as a repeatable benchmark to test future models for improvements in covariate selection capabilities.