Skip to contents

Changelog

Source: NEWS.md

mixqrgate 0.1.2

  • Documentation. Added an applied primer vignette (“A Primer on Location-Varying Gating for Quantile Mixtures”) that walks social-science researchers through the concepts, data, fitting, interpretation, diagnostics, visualisation, and practical guidance with worked ggplot examples; and a formal PDF reference manual. Both are linked from the package website navbar.
  • Louis observed-information gate SEs. variance = "louis" computes the gate-block covariance by Louis’s (1982) identity – complete-data information minus the missing information from the latent labels – giving analytic, classification-aware standard errors without the imputation cost of "stochEM". In a Monte-Carlo check (n = 300) the Louis SE matched the empirical SD (0.207 vs 0.211) and reached 0.95 coverage, where the conditional sandwich gave 0.80. This is the variance DESIGN section 2.4 specified; it is now the recommended choice for gate inference. summary() labels the SE method in use.

mixqrgate 0.1.1

Revision after two adversarial reviews. The estimator was verified correct; this release fixes the inference honesty and the software contract.

  • Classification-aware gate SEs. mixqrgate(variance = "stochEM") adds a multiple-imputation gate variance that propagates the uncertainty about class membership. The default sandwich SE is conditional on the fitted memberships and under-covers when components overlap; summary() now says so.
  • Honest location-varying demo. sim_gate2() gains loc_vary, which makes membership genuinely depend on the quantile rank. The vignette is rewritten around inference (does the gate vary, with uncertainty shown) rather than an eyeballed trend; the previous example overstated a noisy drift.
  • Missing data. Component and gate designs are now built on the same complete-case rows, so y, X, and Z stay aligned.
  • confint() returns gate-coefficient intervals (previously empty); AIC(k=) honors k.
  • Guards: warnings on collapsed components, ill-conditioned gates, and non-convergence.
  • CRAN hygiene: .Rbuildignore, Remotes: for the non-CRAN mixqr dependency.

Still deferred (documented): full Louis joint gate+component information, a cross-tau Wald test, KDE-path classification-aware gate SEs, cross-tau relabel coherence, and vary_gating = "smooth".

mixqrgate 0.1.0

First release. A companion to mixqr adding a concomitant, quantile-indexed mixing gate to finite mixtures of quantile regressions.

  • mixqrgate() fits a gated mixture: the mixing probabilities follow a multinomial-logit gate on concomitant covariates (gating = ~ z), optionally refit per quantile level (vary_gating = "discrete", the Furno 2025 location-varying mixing mode). ALD and Wu & Yao kernel-density paths.
  • The gate M-step is a weighted multinomial-logistic regression solved by ridge-penalised Newton/IRLS, with sandwich standard errors – the inference on how membership varies that Furno’s heuristic lacks.
  • Reduces exactly to a constant gate (and the plain mixqr fit) when gating = ~1.
  • Reuses the mixqr component and constrained-error-density machinery via its extension API (weighted_rq(), constrained_kde()).
  • S3 methods: print, summary (component + gate blocks), coef, vcov, predict, plot(which = "gating") (the gate-vs-tau picture), logLik, AIC, BIC, fitted, residuals.
  • sim_gate2() simulates a two-component design with a concomitant gate.