Struct RemlCandidate 

pub struct RemlCandidate {
    pub index: usize,
    pub name: String,
    pub score: f64,
    pub edf: Option<f64>,
    pub log_lik: Option<f64>,
    pub family: Option<String>,
}

One fitted model in a REML/LAML evidence comparison.

Fields

index: usize

name: String

score: f64

Minimised REML/LAML cost. Lower is better. This is the model’s reported evidence headline (Model.evidence), kept verbatim in the score table.

edf: Option<f64>

log_lik: Option<f64>

Log-likelihood at the converged mode, on the engine’s constants-omitted scale (same as gam_inference::model_comparison). Present when the fit carries it; None for legacy payloads.

family: Option<String>

Response-family tag (e.g. “gaussian”, “gamma”, “binomial”). Carried so compare_reml_fits can REFUSE to rank fits whose REML/LAML scores are on incomparable base measures (a cross-family comparison is meaningless; #1384). None for legacy payloads that did not record it — those are not guarded (back-compatible), but every current FFI candidate carries it.

Implementations

impl RemlCandidate

pub fn ranking_score(&self) -> f64

Cost used to RANK candidates and pick the winner.

The REML/LAML marginal-likelihood evidence headline (score) does NOT reliably Occam-penalise an added pure-noise smooth: on y ~ s(x) vs y ~ s(x) + s(z) with z ⟂ y, the augmented model’s evidence is lower (apparently better) by a few nats on essentially every dataset, because the Gaussian REML Occam pair ½(log|H| − log|S|₊) collapses toward zero for a finite-λ̂ null term while that term still spends a few effective degrees of freedom fitting noise (issue #1362).

The conditional AIC −2ℓ + 2·edf prices exactly those spent degrees of freedom and discriminates correctly: it penalises the noise smooth (Δ ≈ +15 nats) yet rewards a genuinely relevant smooth (Δ ≈ −650), preserving power. We therefore rank on the conditional AIC whenever both the log-likelihood and the effective degrees of freedom are available, and fall back to the raw evidence headline otherwise. The reported score_table still carries the unaltered evidence (reml_score), so Model.evidence / bayes_factor_vs stay consistent with the table.

Trait Implementations

impl Clone for RemlCandidate

fn clone(&self) -> RemlCandidate

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RemlCandidate

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.