A public map of the 2026 Bundibugyo virus outbreak
An interactive record of how Bundibugyo virus disease spread in 2026, limited to claims a person has approved.
Cases, contacts, and sources on one map
Research software
DAGpedia, a living repository of causal DAGs for epidemiology
DAGpedia is Diaphana’s public repository of causal directed acyclic graphs for epidemiology, with an evidence rating on every edge.
- Study field
- Epidemiology
- Discipline
- Research software
- Year
- 2026
- Duration
- Ongoing
The challenge
Directed acyclic graphs (DAGs) are how epidemiologists write down their causal assumptions before an analysis. Most are drawn once, used once, and discarded. The next team working on the same question starts again.
Reynolds (2026, American Journal of Epidemiology) argued that DAGs should work as shared infrastructure instead: annotated with the strength of the evidence behind each edge, revised as knowledge accumulates, and reused across studies rather than rebuilt from scratch.
What we did
DAGpedia is a public website at https://dagpedia.org that keeps DAGs in that form. The graph is written in dagitty syntax. Every edge carries one of six evidence ratings: strong, moderate, weak, conflicting, expert opinion, or unknown. Nodes come from a shared vocabulary that is checked against MeSH.
Each entry also states how its variables are measured, why each edge is drawn, why other edges are left out, and the population, place, and period it is meant for. The rationale for edges that are not obvious cites the literature by DOI or PMID. A DAG for a different context is kept as a separate entry and linked as an alternative.
The site draws the graph interactively and lists the edges with their ratings. Adjustment sets and conditional independencies are computed from the graph with dagitty.js, not typed in by hand. Every change is validated automatically and reviewed before it is merged, and the history of each DAG is kept.
DAGpedia follows the method described in Reynolds RJ, “Living DAGs: the future of DAGs in epidemiology”, American Journal of Epidemiology 2026; 195(5):1365–1367, doi:10.1093/aje/kwag029 .
Two further steps are planned for version 0.2. Each DAG is to receive its own DOI through Zenodo, so that contributors can cite their work. A language model is to read each submission first and report possible problems, such as a missing common cause or a collider in the adjustment set. A person confirms or dismisses each point, and the decision to accept a DAG stays with that person.
Keeping edges current means reading new studies as they appear, and checking every claim by hand does not scale. Diaphana is studying whether System One models can carry part of that load. These are typed models that return a calibrated probability over a fixed set of options in a single pass, without generating text, at a fraction of the cost of a generative model. The study tests them on causal claims in cohort study abstracts: how a claim is worded, what the study design and adjustment were, and which way the effect runs. It compares their accuracy, calibration, and cost per claim with generative models, against labels from two human annotators. Whatever a model extracts is still reviewed by a person before it becomes evidence for an edge. A preprint is planned.
Browse the DAGs at dagpedia.org .
Results
- Public site at dagpedia.org, currently in beta
- An evidence rating on every edge, including conflicting evidence
- Adjustment sets computed from the graph
- Changes validated and reviewed before they are merged
- Version history kept for every DAG
Contact
Write to Diaphana
For questions about our work, or about working with us, write to [email protected].