Historical / exploratory workflow sections

Species-derived grouping examples on this page are retained for historical reproduction. The primary analysis now uses only recorded researcher cover_category labels; see the canonical label policy.

How do the whole-spectrum tests work?

The biological analysis compares sampled group curves under conditional label exchangeability. Tests validate an explicit null and sampling design; they do not remove species, spatial or unrecorded acquisition confounding.

Method	Question / statistic	Interpretation limit
Maximum	Largest absolute standardized contrast across supported wavelengths	Sensitive to a localized excursion; null envelope is not an effect confidence band
Integrated	Weighted sum of squared standardized contrasts	Sensitive to distributed differences; atmospheric gaps have no area
NPC	Synchronized Tippett combination of maximum and integrated permutation probabilities	Calibrated against the joint ensemble, not an uncorrected minimum p
Rank sensitivity	Rank-based contrast comparison in the earlier framework	Different estimand; robust ordering is not automatically a mean effect
Energy sensitivity	Within-block multivariate distribution-distance statistic	Can detect distribution/dispersion changes without a mean shift
FPCA sensitivity	Test in a fixed label-independent spectral principal-component subspace	Retained rank/variance choices can affect power; not proof from a PCA picture

Permute the right unit

Labels move as whole polygon/year curves within exact flight/source-history blocks. This preserves wavelength covariance and block membership. It assumes exchangeability conditional on the block, not merely equal means under arbitrary covariance differences. Single-class histories do not supply within-block contrasts. Pixel duplication must not create extra independent plants; the atlas aggregates before inference. Repeated GlobalIDs across years are coupled in the cluster bootstrap rather than resampled as unrelated plants.

The primary family contains 21 eligible inferred-PFT contrasts in this dataset. Holm-adjusted scalar probabilities control that declared family. The plotted maximum-component localization uses a fixed Bonferroni family envelope, which can be more conservative than Holm. An NPC-supported curve need not cross the maximum envelope if support is distributed. Pointwise bootstrap intervals are uncertainty aids, not simultaneous evidence across all bands.

Why the audit matters

Strong repeated DBS–ECT and DBT–ECT contrasts survive several checks. Weak DBT–ECS and borderline DBS–DBT findings should not be rewritten as universal separability. The inferred PFTs are attached to species; the design concerns the sampled mixtures of taxa, not a species-independent functional mechanism. Biological replicate counts, missing support, exact histories and multiplicity families are retained in the numerical tables.

Full method review and simulations · Module contracts · Scientific testing philosophy