AMR-Climate Atlas — Causal Attribution of Climate to Antimicrobial Resistance

Project 2 of a three-project climate-health portfolio (see ../harvard-dataverse-climate-health-projects.md). Scoping: ../amr-climate-causal-attribution-scoping.md · Build plan: ../amr-climate-causal-attribution-phases.md · Full write-up: docs/paper.md.

Status: all four phases complete. EU/EEA panel (29 countries, 2000–2023, 4 pathogen-antibiotic combinations), a properly-tested causal model, a genomic cross-check, a live Streamlit app, an MCP server exposing the live data sources, and an automated eval-gated validation pipeline.

The finding

A two-way (country + year) fixed-effects panel regression, with a mandatory 200+ permutation placebo test on every result before it's allowed to be called significant:

Combo	n	Coef	Parametric p	Permutation p
E. coli / Fluoroquinolones	642	+0.11	0.77	0.62
E. coli / 3rd-gen cephalosporins	642	+0.21	0.54	0.31
K. pneumoniae / Carbapenems (2010-2022, consumption-adjusted)	355	−1.73	0.008	0.000
S. aureus / MRSA	667	−0.13	0.89	0.69

3 of 4 combinations show no weather effect anywhere. The one exception — K. pneumoniae carbapenem resistance, in the ESAC-Net-consumption-adjusted 2010–2022 subsample — is real (leave-one-country-out stable, not COVID-driven) and directionally confirmed by independent genomic sequencing data (NCBI Pathogen Detection/MicroBIGG-E, same negative sign, r=+0.38 correlation with the phenotypic rate), but does not appear in the full 2000–2023 panel, and antibiotic consumption itself isn't significant in the model that's supposed to control for it. Read as: a real, narrow, sample-window-specific pattern, not a confirmed causal climate→AMR link.

This result survived a real QA process — an independent testing pass caught a missing permutation test on the consumption-adjusted model, and a later robustness check caught a confounder column-selection bug that meant antibiotic consumption was never actually in the model as first reported. Both are disclosed plainly in docs/paper.md rather than smoothed over.

What's in this repo

app.py                         Streamlit atlas — combo explorer, causal-test caveat, rankings
mcp_server/                    live-data MCP server (NASA POWER + BigQuery, not the frozen CSV)
data/
  eu_eea_pilot_panel.csv       joined country-year-combo panel (join_panel.py's output)
  validation_baseline.json     accepted "known good" snapshot for the eval-gate pipeline
  raw/                         gitignored raw pulls (ECDC, NASA POWER, ESAC-Net, EPI)
validation_runs/               timestamped JSON logs from validate_pipeline.py runs
docs/
  phase1_access_check.md       access-verification memo
  phase3_scoping.md            causal-design scoping, grounded in the real pilot data
  phase4_scoping.md            scale-up fork (WHO GLASS vs. BigQuery vs. deepen EU/EEA) + genomic cross-check
  paper.md                     full write-up: methods, bugs found and fixed, results
src/amr_climate_atlas/
  config.py                    country lists, centroids, DOIs, ECDC/NASA POWER/BigQuery constants
  fetch_ears_net.py            EARS-Net resistance rates via ECDC Atlas (fully scriptable, see below)
  fetch_esac_net.py            antibiotic-consumption confounder (static XLSX annexes, 2010-2022)
  fetch_weather_anomaly.py     daily weather per EU/EEA country centroid (NASA POWER)
  fetch_confounders.py         EPI 2024 tabular confounder from Harvard Dataverse
  join_panel.py                joins resistance + weather + EPI into the country-year panel
  stage3_identification.py     two-way FE model + mandatory permutation test + lag check, per combo
  compare_genomic_phenotypic.py genomic (BigQuery) vs. phenotypic (EARS-Net) cross-check
  validate_pipeline.py         automated eval-gate: rerun + placebo-sanity-check + baseline drift check
.github/workflows/validate.yml ready-to-run CI (fires once this repo is pushed to GitHub)

Two findings worth flagging up front

1. ECDC's Surveillance Atlas export is fully scriptable, not manual-UI-only. The export fires a plain parameterized GET, and its downloadTokenId param is not actually validated server-side — any value works once a session cookie is bootstrapped. fetch_ears_net.py pulls the full country×year grid for a given pathogen-antibiotic combo in two HTTP requests.
2. Dataverse's API 403s the default python-requests User-Agent — undocumented anywhere, a browser-like UA is required (see config.DATAVERSE_REQUEST_HEADERS).

Quickstart

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

PYTHONPATH=src python -m amr_climate_atlas.fetch_ears_net        # ECDC EARS-Net resistance rates
PYTHONPATH=src python -m amr_climate_atlas.fetch_esac_net        # antibiotic consumption confounder
PYTHONPATH=src python -m amr_climate_atlas.fetch_weather_anomaly # NASA POWER, EU/EEA centroids
PYTHONPATH=src python -m amr_climate_atlas.fetch_confounders     # EPI 2024 from Dataverse
PYTHONPATH=src python -m amr_climate_atlas.join_panel            # join panel
PYTHONPATH=src python -m amr_climate_atlas.stage3_identification # causal model + permutation test

streamlit run app.py                                            # the live atlas

MCP server — live data, not just the frozen CSV

mcp_server/ exposes two of this project's data sources over the Model Context Protocol, so any MCP-compatible client (Claude Desktop, a custom agent) can query live weather or live genomic AMR data on demand, rather than only reading eu_eea_pilot_panel.csv. Built on Anthropic's official mcp Python SDK (not the fabricated "MCP 2.0"/"Agent Plugins" terminology an earlier draft of this scope used — see mcp_server/README.md for the full terminology correction and a real API-drift finding: the FastMCP class most tutorials use has been removed from the current SDK). Tested with a genuine external-process client (mcp_server/test_external_client.py), not just an in-process call — real weather data for Paris, real carbapenem-gene prevalence (16.83%) for French E. coli isolates. See mcp_server/README.md for setup and a documented naming quirk (NCBI groups E. coli under "E.coli and Shigella" in its own schema, not "Escherichia coli").

Automated eval-gated validation

validate_pipeline.py automates stage3_identification.py's fit + permutation test into a rerunnable gate: it fails if the permutation test's null distribution looks degenerate (a real check for the shuffle logic silently breaking) or if results drift meaningfully from a stored baseline. This is designed to have caught the confounder bug above automatically, not just when a human happened to look closely. A GitHub Actions workflow (.github/workflows/validate.yml) is ready to run this on every push, once the repo is on GitHub.

Deferred / not fetched by default

GRDI v1.10 (raster GeoTIFFs, 932MB) and EnvClim (2.6GB geopackage) — gridded/sub-national products disproportionate to a country-year panel. fetch_confounders.py prints their file listings so a future subnational build can decide if the GIS-tooling investment is worth it.