mescla.emma.diagnostics

Goodness-of-fit diagnostics for a mixing model (Hooper, 2003).

The question these answer is not “how much variance did I explain” but “does this data set actually behave like a mixture of this many sources”. A species the model cannot reproduce is telling you something specific: it is non-conservative, or the rank is too low, or the end-members are not constant in time.

The summary statistics rank the suspects; the residual structure convicts them. Always plot residuals against observed concentration.

Functions

cross_site_projection	Project another site's chemistry into this site's mixing subspace.
hooper_diagnostics	Per-species relative bias and relative RMSE of a mixing model.
residual_structure_test	Test each species' residuals for structure against observed concentration.