Bundle review table

FINDINGS — T01: Coral bleaching tracks accumulated heat (DHW) better than peak temperature

Supplied working report; original verdicts not validated.

Review status: Needs review. This is an archival source note, not a passed proof.

FINDINGS.md

# FINDINGS — T01: Coral bleaching tracks accumulated heat (DHW) better than peak temperature

**Verdict: SUPPORTED** (accumulated heat beats pure peak/threshold; conditional on satellite product quality)

## Headline numbers (core sample, n = 286 observations with complete 12-week satellite history)

| Quantity | Value |
| --- | --- |
| Spearman rho, bleached vs **DHW** (max over prior 12 wk) | **0.405** (p = 9.7e-13) |
| Spearman rho, bleached vs **peak SSTA** | **0.319** (p = 3.5e-8) |
| Spearman rho, severity (% bleached) vs **DHW at obs date** | **0.621** (p < 1e-10) |
| Spearman rho, severity vs peak SSTA | 0.514 |
| Model A (peak SSTA only) AIC / CV-AUC | 325.8 / 0.698 |
| Model B (peak SSTA + DHW) AIC / CV-AUC | **303.8 / 0.739** |
| **ΔAIC (B vs A)** | **22.0** → support rule met (ΔAIC ≥ 10) |
| DHW coefficient in B (per °C-week) | **+0.41**, Wald p = **1.9e-5**, positive given peak |
| LRT B vs A | chi2 = 24.0, p = 9.7e-7 |
| Peak coefficient in B (given DHW) | −0.12, p = 0.78 (adds nothing once DHW is in) |
| Model C (peak SSTA + hot-days ≥1 °C) | ΔAIC = 6.7, hot-days coef p = 0.0043 → weak only |
| DHW-only model vs peak-only model | AIC 301.9 vs 325.8; CV-AUC 0.751 vs 0.698 |

## Support-rule comparison (rule fixed in hypothesis.md before analysis)

- **Rule:** Model B (peak + DHW) must improve on Model A (peak only) substantially: ΔAIC ≥ 10
  in B's favour, **or** the DHW coefficient significant at p < 0.001 with positive sign given peak.
- **Result:** ΔAIC = 22.0 ≥ 10 **and** DHW coefficient positive (+0.41 °C-week⁻¹) at p = 1.9e-5 < 0.001.
  Both arms of the rule fire → **SUPPORTED**.
- The rival (pure threshold response to the hottest day) predicts duration should add nothing
  beyond peak. Observed: with peak in the model, DHW adds massively (LRT p = 9.7e-7) and the
  peak coefficient itself collapses to ≈ 0/− (p = 0.78) — the signature of an accumulation
  process, not a peak-threshold process.

## Equal-peak / duration check (Vol XXX #13's separator)

Within each quartile of peak SSTA, sites that bleached had **higher accumulated DHW** than
sites that did not:

| Peak SSTA quartile (range °C) | n | mean DHW bleached | mean DHW not bleached |
| --- | --- | --- | --- |
| Q1 (0.29–1.16) | 72 | 0.31 | 0.10 |
| Q2 (1.17–1.53) | 72 | 1.90 | 0.43 |
| Q3 (1.54–1.95) | 71 | 3.74 | 2.89 |
| Q4 (1.96–5.01) | 71 | 7.27 | 2.58 |

At equal peak, longer/more accumulated events bleach more — the direction the hypothesis
predicts and the threshold rival denies. (Same pattern for hot-days in Q2–Q4.)

## Secondary checks

- Severity model (severe, >50 % bleached, vs rest): peak+DHW beats peak-only by ΔAIC = 16.2,
  DHW coef +0.30 per °C-week, p = 4.1e-5. Consistent with the binary result.
- Pure duration (hot-days above 1 °C) helps less than DHW (ΔAIC 6.7, p = 0.004) — the
  intensity×duration integral carries more information than duration alone.
- DHW measured **on the observation date** (MB2) does even better than the window max
  (ΔAIC = 31.6 vs peak-only; coef p = 1.8e-6).

## Limitations (honest)

1. **Reduced sample.** ERDDAP fetching was truncated by a coordinator time limit at 536 of
   1,100 planned single-variable requests (fetched 2026-10-30 18:22–18:35Z). Of the planned
   n = 600 stratified sample, **286 observations** have complete metrics (86 not-bleached,
   68 mild, 50 moderate, 82 severe). The stratification oversamples bleached sites relative
   to natural prevalence, so intercepts/prevalence are not population-representative;
   contrasts and model comparisons are still valid within the design.
2. **Observation quality.** Global Coral Bleaching Database records come from ~1,400 sources
   with heterogeneous methods; % bleached is often a visual estimate or a range; ~half of
   records lack a day-of-month, so dates default to the 15th (adds up to ±2 weeks of date
   noise in the 12-week window).
3. **Satellite vs point survey.** CRW 5-km v3.1 (CoralTemp) daily products were matched to the
   nearest grid pixel (~2.5 km average offset, up to ~3.5 km); a 5-km pixel can miss local
   refugia/hydrodynamics. Pre-2002 SST is OSTIA reanalysis (the dataset's stated basis).
4. **Circularity risk.** CRW built/calibrated the bleaching database used here while
   developing the DHW product suite (stated in the dataset description), so a DHW–bleaching
   association is partly built-in; this affects the *strength* of support more than the
   peak-vs-accumulation contrast, which the equal-peak check addresses directly.
5. **Collinearity.** Peak and DHW are correlated (a hotter event accumulates more), inflating
   SEs; the DHW coefficient nonetheless stays positive and highly significant.
6. Single test, one database, one satellite product family; no correction for site clustering
   or repeated sites.

## What would change the conclusion

- A **null or negative DHW coefficient given peak** in the full n = 600 sample (once the
  remaining ~46 % of site windows are fetched) would demote this to Conditional — but the
  direction is consistent in every subsample checked (quartiles, severity strata, CV folds).
- Independent bleaching data **not used by CRW for calibration** (e.g., AIMS/Great Barrier
  Reef surveys, Reef Check transects with exact dates) showing duration adds nothing beyond
  peak would favour the threshold rival.
- A pure peak-threshold result in **controlled/manipulative tank experiments** (identical
  peak, varied duration) would override this observational result.

## Data & code

- data/Global_Coral_Bleaching_Database.csv (NOAA NCEI accession 0228498, 33,244 records)
- data/erddap_raw/ — 536 single-pixel-window daily CSVs (CRW_SSTANOMALY, CRW_DHW) from
  PacIOOS ERDDAP dataset dhw_5km (NOAA CRW 5-km v3.1)
- code/01_sample.py, 02_fetch_erddap.py, 03_compute_metrics.py, 04_analyze.py, statutil.py
- results/results.csv (per-observation metrics), results/model_coefficients.csv,
  results/correlations.csv, results/equal_peak_table.csv, results/model_results.json

hypothesis.md

# Hypothesis — T01: Coral bleaching tracks accumulated heat (DHW) better than peak temperature

## Claim (two Incoprea entries, same claim)

1. **Vol XXIX #35 — "Bleaching Tracks Accumulated Heat"**
   Coral bleaching tracks degree-heating weeks (accumulated thermal stress) better than peak temperature.
   Source: https://incoprea.com/research-library/original-theories-vol29-20261030/theories/bleaching-degree-heating.html

2. **Vol XXX #13 — "Accumulated Heat or Peak?"**
   Bleaching tracks accumulated heat rather than peak temperature; rival explanation is a
   threshold response to the hottest day. Events with equal peaks and different durations
   separate the two — at equal peak, longer events should bleach more.
   Source: https://incoprea.com/research-library/original-theories-vol30-20261030/theories/dhw-vs-peak-temp.html

## Pre-written support rule (fixed before analysis)

- Fit logistic models of bleaching (bleached vs not) on:
  - **Model A** = peak SSTA only
  - **Model B** = peak SSTA + DHW (and separately peak SSTA + hot-days above 1 °C, a duration measure)
- **Support** = Model B improves on Model A substantially: ΔAIC ≥ 10 in B's favour, **or** the
  duration/DHW coefficient is significant at p < 0.001 with positive sign given peak.
  A duration variable significant beyond peak counts as support for the accumulation claim and
  against the pure-threshold rival.
- If neither variable separates bleached from unbleached at all, the test is **undecisive** —
  report that honestly.
- Additionally report univariate Spearman correlations of bleaching with DHW vs with peak SSTA.

Verdict scale: Supported / Not supported / Conditional / Undecisive.

## Metrics computed per observation

For each bleaching observation, from daily NOAA Coral Reef Watch 5-km data (SSTA, DHW) in the
12 weeks before the observation date:
- `peak_ssta` — maximum daily sea-surface-temperature anomaly (°C) over the prior 12 weeks
  (the "hottest day" / threshold-rival variable)
- `dhw` — degree heating weeks at the observation date (CRW DHW is the rolling integral of
  hot-spot anomalies ≥ 1 °C over the prior 12 weeks — the accumulated-heat variable)
- `hot_days` — number of days with SSTA ≥ 1 °C in the prior 12 weeks (pure duration measure)

Included code and results