Test 16 — Warming Can Add Snow
Supplied working report; original verdicts not validated.
Review status: Needs review. This is an archival source note, not a passed proof.
README.md
# Test 16 — Warming Can Add Snow
## Verdict: supported
All three pre-specified criteria hold on GHCN-Daily US/Canada stations for winters 1980/81–2022/23: the cold stations' median snowfall trend is positive with a 95% bootstrap CI entirely above zero, the cold stations are significantly more positive than the warmest bin, and a significant majority of cold stations show positive trends.
## Headline numbers
| Quantity | Value |
|---|---|
| Stations analysed (≥ 30 valid winters, US/CA) | **1 127** (of 1 500 candidates; 373 excluded with < 30 valid winters) |
| Cold stations (mean DJF TMAX < −5 °C) | **68** |
| Cold median Theil–Sen snowfall trend | **+0.77 in/decade**, 95% bootstrap CI **[+0.33, +1.94]** |
| Cold vs warmest bin (mean DJF TMAX > 5 °C, n = 480) | difference **+0.95 in/decade**, permutation **p = 0.00025** |
| Cold stations with positive trends | **64.7%** (44/68); exact binomial vs 25%: p < 0.0001 |
| Bin medians (in/decade) | < −10: +3.37 (n = 3) · −10…−5: +0.77 (n = 65) · −5…0: +1.37 (n = 252) · 0…5: +0.70 (n = 327) · > 5: 0.00 (n = 480) |
## What this shows and cannot show
- Shows: cold-region DJF snowfall has not declined over 1980–2023 in the compiled station set; the median cold station gained about three quarters of an inch of snow per decade, and nearly two thirds of cold stations trend upward. The coldest stations are significantly more positive than the warmest bin.
- Shows (with a caveat): the "warmer stations decline" half of the posted prediction is only weakly visible. The warmest bin's median is exactly zero largely because most warm stations now record no measurable snowfall in most winters (a floor effect — e.g. Nashville: 12 of 43 winters entirely snowless, maximum 14.3 in); the 0–5 °C bin, where snowfall is still common, has a *positive* median (+0.70 in/decade), not a declining one. The cold-vs-warm contrast is therefore driven substantially by snowless winters at the warm end, not by declining snow among snowy warm stations.
- Cannot show: a causal attribution to warming and moisture. The test measures observed trends and their gradient, not the physical mechanism (moisture supply versus snowline retreat); a warmer and wetter North Atlantic/Arctic storm track is one plausible explanation for the cold-bin rises, but so are station moves, gauge changes and siting artefacts.
## Limitations
1. **Mechanism untested.** The prediction is causal ("warming raises snowfall because moisture rises faster than the snowline retreats"); the test only measures trend patterns. Rising cold-region snowfall is consistent with the mechanism but does not identify it.
2. **The floor effect at the warm end.** The warmest bin's zero median reflects winters with no snowfall at all, so the cold-vs-warm contrast partly measures "snow exists" versus "snow rarely exists", not competing snowfall trends. The 0–5 °C bin's positive median shows the warmer-bin decline predicted by the claim is not cleanly observed.
3. **Station-network artefacts.** GHCN stations move, change instruments and change observing practices over 43 years; snowfall is especially sensitive to gauge undercatch and wind. No homogenisation was applied here (GHCN-Daily is not homogenised), and a rival explanation for some large cold-station trends (e.g. +6–8 in/decade at a few Maine/Quebec stations) is measurement change rather than climate.
4. **Selection and threshold dependence.** Results apply to US/Canada stations with ≥ 30 valid winters at ≥ 80% daily completeness, binned by mean DJF TMAX. The coldest bin (< −10 °C) holds only three stations, so the combined cold-bin result carries the criterion 1 verdict.
## Structure
- `hypothesis.md` — claim and pre-specified criteria (written before analysis)
- `run.py` — parses cached NOAA CSVs, Theil–Sen with permutation p-values, bins, criteria; regenerates `results.json`, `data/*.csv`, `figures/*.png`
- `results.json` — per-prediction verdicts and decisive numbers
- `SOURCE.md` — endpoints, hashes, exclusions
hypothesis.md
# Test 16 — Warming Can Add Snow (warming-more-snow) ## Posted claim (Vol XXXII, warming-more-snow), verbatim "In cold regions, warming raises snowfall because moisture rises faster than the snowline retreats." ## Posted prediction Stations with cold mean winter temperatures (< −5°C) show rising or stable DJF snowfall over 1980–2023, while warmer stations decline. ## Posted refutation Universal snowfall decline. ## Source links - Daily data: NOAA NCEI Climate Data Online access services, `https://www.ncei.noaa.gov/access/services/data/v1?dataset=daily-summaries&stations=<ID>&startDate=1980-01-01&endDate=2023-12-31&dataTypes=SNOW,TMAX,TMIN&format=csv&units=standard` (route verified by a prior probe). - Station inventory: `https://www.ncei.noaa.gov/pub/data/ghcn/daily/ghcnd-stations.txt`; element coverage prescreen from `https://www.ncei.noaa.gov/pub/data/ghcn/daily/ghcnd-inventory.txt`. - Menne et al. 2012 (GHCN-Daily) for the dataset itself. ## Pre-specified criteria for this test (written before analysis) Station selection and processing (fixed before analysis): - US and Canada stations only (country codes US, CA), with ≥ 30 valid winters in 1980–2023, capped at ~1500 stations by earliest-complete inventory order; every exclusion documented in `data/`. - A winter = December of year Y−1 + January–February of year Y (labelled by Y; winters 1980/81–2022/23, i.e. 43 winters: Dec 1980 … Feb 2023 — see note below). A winter is valid only if ≥ 80% of its 90/91 daily SNOW values are present; DJF TMAX mean uses all present TMAX days (winter valid for temperature if ≥ 80% present). - DJF SNOW total per winter in inches (as served, units=standard); station DJF mean TMAX in °C (converted from °F). - Per-station snowfall trend: Theil–Sen slope (median of pairwise slopes, mm/inch per year... reported in inches per decade) with a permutation p-value (1999 permutations, seeded). Trend computed only on stations with ≥ 30 valid winters. - Temperature bins (by station mean DJF TMAX): < −10, −10…−5, −5…0, 0…5, > 5 °C. "Cold stations" = the two coldest bins (< −5 °C), matching the posted < −5 °C threshold. Note on window: the posted span "1980–2023" contains 43 complete DJF winters bounded by Dec 1980 and Feb 2023 (winter 1980/81–2022/23). This is the pre-specified test window; stations contribute whichever of these 43 winters are valid. Verdict criteria: 1. **Cold-bin stability** — supported if the coldest bin's (or combined cold-bins') median trend is ≥ 0 with its 95% bootstrap CI above 0 or straddling 0; not supported if the cold-bin median is negative with a CI entirely below 0. 2. **Cold-vs-warm contrast** — supported if the cold (< −5 °C) station slopes are significantly more positive than the warmest bin (0–5 °C or > 5 °C, whichever is larger) by a permutation test on the difference of means (p < 0.05, two-sided, 19999 permutations, seeded); partly supported if the sign of the difference is positive but p ≥ 0.05; not supported if the difference is not positive. 3. **Positive minority** — supported if ≥ 25% of cold stations have positive trends and the exact two-sided binomial test against 25% gives p < 0.05 (i.e. the share is significantly above the quarter the claim needs); not supported otherwise. 4. **Overall verdict** — supported if (1), (2) and (3) all hold; partly supported if the cold-bin trend is flat (criterion 1 holds) while warm bins decline but (2) or (3) fail; not supported as stated if all bins decline (refutation); not testable only if the endpoint fails after retries and no fallback works. All randomness (permutations, bootstrap) seeded. Missing years handled as specified above; exclusions documented.