03 — Five-variant cross-substrate concordance - weatherxbiodiversity-substrate-sensitivity

This is the headline statistic of this repo. For each species and horizon, compute five candidate per-species rankings:

(a) Full GLMM η, within-substrate standardisation
(b) Main-effects-only η, within-substrate standardisation
(c) Full η, shared CEA reference standardisation
(d) Mean future TEI (substrate-invariant physical metric)
(d2) Fraction of cells with TEI_future > 0.5

Then compute Spearman ρ between nside=64 and nside=128 rankings at four minimum-cell-count filters: ≥1, ≥5, ≥10, ≥20.

The headline finding: at n≥10, (b) main-effects-only η lifts ρ from +0.59 to +0.97 (mid-term horizon). The substrate-invariant physical metric (d) hits ρ ≥ +0.66 at all sample sizes.

Outputs (per horizon):

results/variant_comparison_<horizon>.csv
results/variant_comparison_<horizon>.json

import subprocess
import sys
from pathlib import Path

REPO_ROOT = Path(__file__).resolve().parent.parent

for horizon in ("2020_2029", "2030_2039"):
    cmd = [sys.executable, str(REPO_ROOT / "scripts" / "compare_variants.py"),
           "--horizon", horizon, "--out-dir", str(REPO_ROOT / "results")]
    print(f"\n=== Horizon {horizon} ===")
    print("Running:", " ".join(cmd))
    subprocess.run(cmd, check=True)

What to look at¶

The Spearman heatmap (variant × n_cells filter) is the cleanest summary — see 04_figures.py for the visualisation.
The per-species CSV lets you inspect why a particular species’ ranking diverged: compare (a)_eta_64 vs (a)_eta_128 against the same species’ (b) main-effects-only η; if the gap shrinks, the interaction term was the cause.
Variant (c) — shared CEA reference — barely improves on (a). This was a refuted hypothesis: just rescaling future predictors at projection time does not decouple the substrate, because the GLMM β coefficients are themselves substrate-fit. To properly test (c) the GLMM would need to be refit with shared standardisation, which is out of scope for this repo.