Can Contemporary Large Language Models Provide the Domain Knowledge Needed for Causal Inference? Evaluating Automated...

Area of Science:

• Epidemiology
• Public Health
• Causal Inference
• Health Informatics

Background:

• Directed acyclic graphs (DAGs) are essential for epidemiological study design and bias reduction.
• Developing accurate DAGs for causal inference demands significant domain expertise.
• Large language models (LLMs) offer potential for automating DAG generation due to extensive training data.

Purpose of the Study:

• To evaluate the effectiveness of prompt engineering strategies for LLMs in generating DAGs for population health research.
• To assess the performance of OpenAI's GPT-4o and GPT-o1 in creating DAGs depicting causal relationships.

Main Methods:

• Four prompt engineering strategies were tested: zero-shot, one-shot, instruction-based, and chain of thought (CoT).
• A hypothetical study on statins for cardiovascular disease prevention was used as a case example.
• Generated DAGs were evaluated for consistency, acyclicity, source accuracy, completeness (ASCVD criteria), and prompt adherence.

Main Results:

• All generated DAGs were acyclic, except for one instruction-based prompt instance.
• Over half of the DAGs met 6/7 ASCVD criteria, but race was consistently omitted.
• Chain of Thought prompts yielded the most complete DAGs; one-shot prompts offered the highest consistency and adherence.
• Zero-shot prompts performed better on GPT-o1, providing justifications and sources.

Conclusions:

• LLMs demonstrate a foundational ability to generate DAGs aligning with basic epidemiological standards.
• Limitations include a lack of justification, systematic omission of race, and frequent source hallucination, necessitating human expert review.
• Current LLMs are valuable as brainstorming or pre-analysis tools for DAG development, not replacements for expert judgment.