Methods, made visible.
Four ways to turn biological and public data into a decision you can inspect.
- Built in
- Grapevine Red Blotch Virus gene data, UC Davis
- Re-tested on
- World crop yield data, 1961-2023 (FAOSTAT)
Variance breakdown
A Bad Year, or a Bad Place?
One model gave the wrong picture. Looking at each crop separately showed that place usually matters much more than year.
- Linked to country
- 69.1%
- Linked to year
- 5.6%
- Data points
- 305,676
- Crops
- 63
- Built in
- GBIF occurrence data for Xylella fastidiosa
- Re-tested on
- Country-level outbreak surveillance
Effort normalization
When More Records Do Not Mean More Disease
Normalize pathogen observations by reporting effort so the map reflects surveillance intensity instead of simply rewarding countries that publish more data.
- Records
- 1,170
- Mapped
- 309
- United States
- 7th
- Portugal
- 0
- Built in
- OpenAlex AI plant-disease publications
- Re-tested on
- FAOSTAT world crop production, 2023
Exposure normalization
Count Research Relative to the Crop
Compare AI plant-disease research with crop production so publication volume is interpreted against agricultural scale rather than as an isolated count.
- Records
- 170,605
- Tomato papers
- 2,848
- Attention gap
- 11.5×
- AI share
- 1 in 7
- Built in
- Grapevine Red Blotch Virus transcriptomics, UC Davis
- Re-tested on
- FAOSTAT world crop production, 1961-2023
Causal mediation
Is It the Heat, or the Virus?
Split one biological signal into direct and virus-mediated routes, then test the same model on 63 years of crop data.
- Via virus
- 61.8%
- Genes
- 39
- Via land
- 97.2%
- Bootstraps
- 2,000