Generalized likelihood-ratio test for VCMoE coefficient variation

Usage

vcmoe_glrt(
  fit,
  data,
  test = c("coefficient", "constant_block", "constant_all"),
  coefficient_set = c("expert", "gating", "sigma", "theta"),
  component = NULL,
  term = NULL,
  calibration = c("none", "bootstrap", "analytic_epanechnikov", "both",
    "parametric_bootstrap"),
  B = 200L,
  seed = NULL,
  control = list(),
  refit_control = list(),
  verbose = FALSE
)

Arguments

fit: A vcmoe fit.
data: Original data frame used to fit fit.
test: Test type. "coefficient" tests one coefficient function; "constant_block" tests all expert or gating functions jointly; "constant_all" tests all fitted coefficient functions jointly.
coefficient_set: Coefficient block for coefficient-specific or block-constant tests.
component: Component label or index for coefficient-specific tests.
term: Term name for coefficient-specific tests.
calibration: Calibration method. The default "none" returns the statistic without attaching a reference distribution. "analytic_epanechnikov" uses the Epanechnikov modified chi-square calibration; "bootstrap" uses parametric bootstrap calibration; "both" reports both. The analytic calibration is retained as an explicitly requested approximation because the implemented statistic is not identical to the manuscript criterion.
B: Number of bootstrap calibration replicates.
seed: Optional random seed.
control: Controls for constrained null optimization and diagnostics.
refit_control: Controls overriding bootstrap full-model refits.
verbose: Whether to message bootstrap progress.

Value

A vcmoe_glrt object.

Details

Local-grid fits retain the 0.1.0 constrained BFGS null optimizer. Joint-path fits use a paper-inspired sample-weighted grid-projected null: after every M-step, each constrained coefficient path is replaced by its mean weighted by the number of observations assigned to each nearest grid point, and constrained local slopes are set to zero. Its statistic compares sample-level likelihood contributions evaluated at each observation's nearest grid point. The projected update is not a generic constrained optimizer and its diagnostic likelihood trace need not be monotone. Bootstrap calibration preserves both the full-fit engine and its matching null engine.