weatherxbiodiversity-substrate-sensitivity

Initial release

Methodological follow-up to two single-substrate replications of Soroye et al. 2020 (Climate change contributes to widespread declines among bumble bees across continents, 10.1126/science.aax8591) for Iberian Bombus under DestinE Climate DT SSP3-7.0:

• HEALPix nside=64 (~92 km): weatherxbiodiversity-projection
• HEALPix nside=128 (~46 km): weatherxbiodiversity-projection-nside128

Both single-substrate Outcomes confirm Soroye's TEI mechanism is substrate-robust at fit time (sc_TEI_delta within ±30% across CEA, nside=64, nside=128). When the same fitted GLMMs are used to project to future climate, per-species rankings diverge across substrates by 1–9 logits.

This repo asks: why does that happen, and how should it be done properly?

Headline finding

The substrate-coupling at projection time is not caused by per-species random-effect refit (substrate-stable, ΔRE ≤ 0.6 logits) or by per-species niche-limit refit (modest, ΔT_range = 0–3°C). It is caused by two mechanisms acting together:

1. Per-species sample size at projection time. Below ~10 occupied + active cells per substrate, per-cell extrapolation noise dominates the species-mean η regardless of which projection variant is used.
2. The GLMM interaction term sc_TEI_delta:sc_PEI_delta — the largest single contributor to projection η at SSP3-7.0 — compounds substrate-specific predictor standardisation quadratically when future predictors extrapolate 2–4σ outside the training distribution.

Recommended reporting protocol

For any future TEI-based extirpation projection that compares across substrates:

1. Report only on species with at least 10 occupied + active cells per substrate.
2. Drop the GLMM interaction terms at projection time — keep them in the fit, but use main-effects-only η to extrapolate. At n≥10 this lifts cross-substrate Spearman ρ from +0.59 to +0.97 (mid-term horizon, both substrates).
3. Cross-check against a substrate-invariant physical metric — mean future TEI, or fraction of cells where future TEI > 0.5. Both hit ρ ≥ 0.66 across the entire species set including small-N species.

Empirical evidence

Cross-substrate Spearman ρ at the SSP3-7.0 mid-term horizon (2030–2039), across five projection variants and four cell-count filters:

| Variant | n≥1 | n≥5 | n≥10 | n≥20 | |---|---:|---:|---:|---:| | (a) Full GLMM η | +0.27 | +0.51 | +0.59 | +0.77 | | (b) Main-effects-only η | +0.40 | +0.52 | +0.97 | +0.98 | | (c) Shared CEA reference η | +0.27 | +0.49 | +0.52 | +0.55 | | (d) Mean future TEI | +0.66 | +0.69 | +0.90 | +0.82 | | (d2) Frac TEI_future>0.5 | +0.66 | +0.71 | +0.88 | +0.83 |

(Same qualitative pattern at the near-term horizon 2020–2029.)

Refuted hypotheses

The diagnostic also empirically refutes three intuitive but incorrect explanations of substrate-coupling: (i) per-species random-intercept refit; (ii) per-species niche-limit refit; (iii) shared-reference standardisation alone (without refitting the GLMM β). See results/SUBSTRATE_SENSITIVITY_FINDINGS.md for the full analysis.

Reproducibility

git clone https://github.com/annefou/weatherxbiodiversity-substrate-sensitivity.git
cd weatherxbiodiversity-substrate-sensitivity
mamba env create -f environment.yml
mamba activate weatherxbiodiversity-substrate-sensitivity
# Until upstream Zenodo URLs are wired into 01_inputs_fetch:
INPUTS_FETCH_MODE=local snakemake --cores 1

Companions

• weatherxbiodiversity-projection — canonical Iberian Bombus replication at HEALPix nside=64.
• weatherxbiodiversity-projection-nside128 — substrate-extension at HEALPix nside=128.

The Jupyter Book is at https://annefou.github.io/weatherxbiodiversity-substrate-sensitivity/.