Skip to contents

Profile-likelihood CI for per-trait total variance

Source: R/profile-derived.R
profile_ci_total_variance.Rd

Penalty-based chi-square_1 profile approximations (via fix-and-refit on log V_t) for the per-trait total variance V_t = (Lambda Lambda')[t,t] + psi[t]^2, the diagonal of the requested covariance tier. This is the same functional extract_Sigma() returns on its diagonal, and the same one bootstrap_Sigma() resamples.

Usage

profile_ci_total_variance(
  fit,
  tier = c("unit", "unit_obs", "phy", "B", "W"),
  trait_idx = NULL,
  level = 0.95
)

Arguments

fit

A fit returned by gllvmTMB().

tier

Covariance tier: "unit", "unit_obs", "phy", or the soft-deprecated legacy aliases "B" / "W". Default "unit".

trait_idx

Integer indices of traits, or NULL for all.

level

Confidence level. Default 0.95. No level currently carries an exact-profile coverage certificate.

Value

A data frame with one row per trait and columns:

trait, tier

Trait name and the canonical tier name.

estimate

Point estimate of V_t.

lower, upper

Two-sided profile bounds, NA when the profile could not be computed.

method

Always "profile".

interval_status

Claim-boundary marker: "route-only" for a computed but uncertified interval and "none" for a point-only row with no interval.

What the coverage evidence does and does not cover

The 2026-07-29 Gaussian n_units = 150, d in {1, 2} campaign and the 2026-08-25 n_units = 400, d = 2 campaign both measured the implemented penalty-profile approximation. The numerical coverage gates passed in those cells, but the retained endpoints do not establish that every constrained refit converged and attained its requested log(V_t) exactly. The implementation can accept target mismatch up to its numerical tolerance and interpolate after failed interior refits. Those campaigns are therefore measured, not exact-profile certificates. Every computed row is labelled "route-only"; a row with no interval is labelled "none". This route must be repaired and recalibrated before any coverage certificate is restored.

See also

extract_Sigma() for the point estimate, bootstrap_Sigma() for the resampling route to the same estimand.