Skip to contents

Changelog

Source: NEWS.md

mixedsubjectsirt 1.1.0

  • fit_2pl() now estimates the 2PL by marginal maximum likelihood with the package’s own quadrature and L-BFGS-B machinery, the same objective used by fit_mixed_subjects_mml() at lambda = 0, instead of calling mirt. Estimates agree with the previous mirt backend to about four decimal places on simulated data. This makes fit_2pl() the 2PL counterpart of the already native fit_1pl().
  • Breaking: the mirt-specific technical, verbose, and ... arguments of fit_2pl() are removed, replaced by n_quad, initial_pars, quadrature, slope_lower, slope_upper, and control (matching fit_1pl()). Code calling fit_2pl(resp, technical = list(NCYCLES = 500)) should now call fit_2pl(resp). $model is retained in the returned list and is NULL.
  • mirt moves from Imports to Suggests. It is now needed only to read coefficients from a fitted mirt model passed in as item parameters, and the package no longer pulls in the SimDesign/qs2/RcppParallel chain. This fixes installation-dependent failures on macOS, where loading qs2 (reached through mirt) could abort with a missing TBB symbol and take every call to fit_2pl() down with it.

mixedsubjectsirt 1.0.0

CRAN release: 2026-06-25

Initial CRAN release.

  • Mixed-subjects 2PL/1PL IRT calibration that augments human responses with LLM-generated responses through a PPI++ marginal-MML estimator (fit_mixed_subjects_mml() and relatives). The estimator is anchored to the human data and is asymptotically unbiased for the human item parameters at any tuning weight.
  • Power tuning by ability-score risk (tune_lambda_ability_risk()), which selects the tuning weight by direct 1-D optimization of propagated ability-recovery risk (pass method = "grid" to scan a grid instead). Also included: a theoretical PPI++ score diagnostic (tune_lambda_ppi_score()), cross-fitted tuning (tune_lambda_ability_risk_crossfit(), the recommended workflow for reported analyses), and experimental per-item tuning (tune_lambda_ability_risk_item()). All non-experimental tuners use the marginal-MML estimator by default; the frozen expected-count estimator remains available via fit_fn but is discouraged.
  • Louis-corrected marginal sandwich covariance through the vcov() S3 method (vcov_mixed_subjects_mml()), with ability scoring and item-parameter uncertainty propagation (score_theta(), ability_risk()).
  • Vignettes covering the recommended workflow, lambda tuning, the 1PL model, per-item tuning, scale linking, and a simulation-validation study; an R-CMD-check GitHub Actions workflow.
  • Currently predicted and generated data must be binary 0/1 responses in the high-level fitting and PPI-score functions; the low-level quadrature utilities accept fractional input.