An open dataset of causal DAGs from epidemiology papers
Paste the DOI or PMCID of an open-access observational or clinical study. NVIDIA Nemotron models on Nebius Token Factory extract its causal diagram and flag likely missing confounders; Tavily finds supporting evidence; you review and add it to the dataset.
- Fetching open-access full text from Europe PMC
- Extracting the DAG with Nemotron Nano
- Critiquing confounding & bias with Nemotron Super
- Finding evidence with Tavily
Recently added
Edit the DAG
Nemotron Super made the accept/reject decisions on this paper. Check them, change anything you disagree with, and approve to mark it as human-reviewed.
Study
Bias risks AI-suggested
DAGitty
Paste into dagitty.net
or dagitty::dagitty() in R to find adjustment sets.
OpenDAG dataset
Causal diagrams (DAGs) reconstructed from open-access epidemiology papers by NVIDIA Nemotron, critiqued for missing confounders, and reviewed before entering the dataset: by a person, or where marked AI-reviewed, by Nemotron Super. Free to reuse under CC BY 4.0.
Ask the dataset Nemotron
Effect estimates across studies
Main reported estimate and 95% CI per paper, log scale. Dashed line = no effect (1).
Most often missed confounders AI-suggested
How many papers the reasoning model flagged for each unadjusted confounder (similar names grouped).
| Exposure | Outcome | Design | Paper | Year | Nodes | Status |
|---|
No DAGs match. Try another search, or add a paper to grow the dataset.
How the studies connect
Every variable in the dataset, merged across papers. Bigger nodes appear in more papers; arrows are the causal links papers assume. Drag to rotate, scroll to zoom, click a variable.
About OpenDAG
Causal diagrams (DAGs) make a study's assumptions explicit: what causes what, and which variables were adjusted for. Most papers never publish one. OpenDAG reconstructs them from open-access papers and builds an open, reusable dataset.
How it works
- Fetch: Europe PMC; only Creative Commons-licensed papers are accepted.
- Extract: (NVIDIA Nemotron Nano) reads the methods and returns exposure, outcome, adjusted covariates and edges as JSON.
- Critique: (larger Nemotron with reasoning) proposes unadjusted confounders and bias risks.
- Evidence: Tavily searches for literature supporting each suggested confounder.
- Validate: the graph is checked to be acyclic and schema-valid.
- Review: a person edits, accepts or rejects suggestions, and approves.
What we store
Bibliographic metadata, the DAG, AI critique, evidence links and short supporting phrases (≤20 words). Never the paper's full text.
Code: MIT. Dataset: CC BY 4.0.