Tools
ferx-core fits one model to one dataset. Everything that calls fit() more than once — resampling, screening, searching a model space — lives in ferx-tools and is exposed as a ferx <tool> subcommand. Each tool carries a maturity tag; see Feature Maturity for what the tags mean.
Uncertainty
- Bootstrap — non-parametric case bootstrap: resample subjects with replacement, refit, and read bias, standard errors and confidence intervals off the spread of the estimates. PsN
bootstrapfeature parity.
Covariate screening
- GAM covariate screening — rank covariate–ETA pairs by ΔAIC on the EBEs before paying for a stepwise search. The ferx equivalent of Xpose4’s
xpose.gam().
Model-space searches
All searches read the same .ferxsearch configuration file, whose search space is written in Pharmpy’s Model Feature Language (MFL).
- Search configuration — the
.ferxsearchTOML file, the MFL grammar ferx accepts, and the coverage table of what is and is not supported. - Covariate search — stepwise covariate modelling: PsN
scmforward / forward-then-backward and Pharmpycovsearch, plus allometric scaling. - Structural model search — Pharmpy
modelsearchover absorption route, peripheral compartments, transit compartments and lag time. - Residual-error search — Pharmpy
ruvsearchover the[error_model]block, tested by likelihood ratio. - Variability-structure search — Pharmpy
iivsearch: which parameters carry an η and how the omegas are blocked. - Inter-occasion variability search — Pharmpy
iovsearch: which parameters carry an occasion-level κ.
End to end
- Automatic model development — Pharmpy
amd: the searches above run in order from one.ferxsearchfile, each seeded from the model the last one selected, reported as one object.
Global search
- Global model search — pyDarwin’s genetic algorithm or exhaustive enumeration over one grid of structural and covariate choices, ranked on pyDarwin’s penalized fitness; for the model spaces whose axes interact, where a stepwise search can lock in a local optimum.