Pipeline guide

End-to-end walk-through of the five stages. All scripts run from the repo root.

Stage 1 — Curated event panels

Two parallel panels live under data/raw/:

  • uprisings_temperature.csv — 112 violent uprisings, revolutions, massacres, and prison revolts (1750–2024). Hand-curated from Wikipedia, Britannica, USHMM, BlackPast, the Encyclopedia of Arkansas, the National Security Archive, and the historiographic monographs cited row-by-row in event_data_source.

  • peaceful_gatherings.csv — 41 peaceful mass gatherings (carnivals, papal visits, royal weddings, large religious pilgrimages) used as the Field-1992 outdoor-opportunity falsification panel.

Coordinates live in event_geo.csv and peaceful_geo.csv. Editing rule: changes to the event lists go in raw/; resolved temperatures and analysis frames are derived artefacts and live elsewhere.

See data/README.md for the column schema.

Stage 2 — Tiered weather-archive cascade

The cascade is implemented in thermostrife.lookup.resolve_event_anomaly and dispatched through thermostrife.sources.*:

Tier

Source

Coverage

Adapter

1

GHCN-Daily / ECA&D / DWD via meteostat 2.x

~1880+ where stations exist

sources.meteostat_src

2

HadCET (Hadley Centre Central England Temperature)

1772+ (mean), 1878+ (max), British Isles only

sources.hadcet_src

3

ECMWF ERA5 reanalysis (0.25° grid) via cdsapi

1981+ global

sources.era5_src

4

NOAA 20CRv3 reanalysis (1° grid, ensemble mean) via PSL THREDDS

1806–1980 global

sources.twentycr_src

The resolver tries each tier in order and accepts the first tier that returns both an event-day Tmax and a same-source baseline window with ≥ 20 days of data. The baseline is built from the same station / grid cell that resolved the event day, so the anomaly event − mean(baseline) is internally consistent. Each resolved row carries a temp_provenance flag (tier1_ghcn, tier2_hadcet_max, tier2_hadcet_mean, tier3_era5, tier4_20crv3) so the analysis can audit which events came from which tier.

Stage 3 — Anomaly construction

Built inside the cascade rather than as a separate stage. For each resolved event:

anomaly_C = event_day_Tmax − mean(baseline_window_Tmax)

where baseline_window is the ±5-year same-station, same-calendar-month daily Tmax record from the resolving tier, excluding the event window ± 7 days. This anchors anomaly_C against the local climate of the event year rather than today’s climatology and removes the warming-trend bias that would otherwise distort pre-1900 anomalies.

Stage 4 — Case-crossover inference

python scripts/run_inference.py --panel violent
python scripts/run_inference.py --panel peaceful

The primary test (H2) is a time-stratified case-crossover fitted with statsmodels.discrete.conditional_models.ConditionalLogit. For each event i, the control set is every day at the same station, same calendar month, in years event_year ± 5, excluding the event window ± 7 days. The stratification absorbs all time-invariant station, regional, and era confounds by construction. A daylight-hours covariate is included as a direct rebuttal to the Field (1992) outdoor-opportunity critique.

Robustness battery (same script):

  • H1 descriptive — Wilcoxon signed-rank, sign test, bootstrap mean CI on the per-event anomalies.

  • Stratified permutation within each event’s matched set (B = 10 000) as a non-parametric backup for H2.

  • H3 within-event temporal contrast — paired test against the “hot week happened to contain an event” alternative.

  • Burke-2015 σ rescaling — per-event z-score so the effect translates into the +2.4 % per 1 σ Burke meta-analytic currency.

  • Benjamini–Hochberg FDR across the auxiliary battery (q = 0.05), with Bonferroni columns alongside.

Stage 5 — Reporting

python scripts/compare_panels.py
python scripts/run_stratifications.py

Outputs in reports/:

  • inference_results.{md,json} — violent panel.

  • inference_results_peaceful.{md,json} — peaceful control.

  • peaceful_vs_violent.md — side-by-side comparison.

  • stratification.md — era / hemisphere / duration / event-type sensitivity analysis with per-stratum forest plot.

  • cascade_validation.md — diagnostic audit of the tier cascade.

  • figs/ — raincloud, null density, anomaly-by-year, warming-stripes timeline, superposed-epoch composite, forest plot. Every figure is a triple SVG + PNG + CSV.