NEWS
VCMoE 0.2.0 (2026-07-15)
Fitting engines
- Adds stable single-response joint-path EM through
vcmoe_fit(..., engine = "joint_path_em").
- Keeps
engine = "local_grid_em" as the default and preserves the 0.1.0
local-grid numerical path when engine is omitted.
- Adds dense-grid cost guardrails, nearest-grid observation assignment,
joint-path convergence traces, and optional progress logging.
Inference and selection
- Makes analytic confidence bands available through the common fitted-object
interface for both engines. Joint-path bands identify their JASA
local-likelihood sandwich plug-in scope and expose score-imbalance
diagnostics.
- Makes parametric bootstrap refits and bandwidth selection engine-aware.
- Adds sample-weighted projected joint-path constrained-null fitting for GLRT and
vcmoe_fit_reduced() while retaining the established local-grid BFGS null
optimizer. Joint-path GLRT uses sample-level nearest-grid likelihood
contributions and blocks failed or nonconverged null fits.
- Changes the GLRT calibration default to
"none". The statistic is available
without a reference distribution; analytic Epanechnikov calibration is an
explicit approximation, and parametric bootstrap calibration is
engine-preserving.
- Adds criterion, projection, constrained-MLE, calibration, and theory-status
metadata to GLRT results. Analytic confidence bands now state whether their
covariance target matches the fitted estimator and which shared-path and
label uncertainties are excluded.
Scope
- Version 0.2.0 remains a single-response package for Gaussian, Binomial, and
Negative-Binomial models with 2 through 10 components.