Skip to contents

Changelog

Source: NEWS.md

slca 1.4.3

  • Added the reference to the accompanying Journal of Statistical Software article (Kim and Chung, 2026, doi:10.18637/jss.v117.i08) in CITATION, DESCRIPTION, and the package documentation.

slca 1.4.2

CRAN release: 2026-10-27

  • Filled missing defaults when reusing historical control objects, and ignored saturated-table convergence metadata when comparing otherwise identical datasets.

  • Prevented undefined subtraction in posterior calculations when a latent-state subtree has zero likelihood, retaining zero posterior mass for that path.

  • Restored the 1.4.0 internal response convention (0 for missing, 1,...,K for categories), retaining missing-score and range-check fixes. Continued estimation rebuilds saved response caches, and model comparison ignores derived encoding caches. Public labels and fitted numerical results are unchanged.

  • Retained existing zero constraints when re-estimating fitted objects unless fix2zero is supplied explicitly, and rejected unknown, unnamed or duplicate control options before estimation.

  • Added warnings for numerically rank-deficient information matrices without changing existing covariance calculations. Skipped bootstrap histograms when a zero likelihood ratio bypasses simulation.

  • Fixed out-of-bounds reads when reordering or summarizing fitted objects saved with the older response encoding, including the bundled NLSY fits. Native likelihood calculations now reject out-of-range response codes.

  • Regenerated the bundled NLSY objects with current preprocessing and calculation code at exactly the original CRAN 1.4.0 parameters, without refitting. Original convergence history and numerical results are retained; missing control defaults and current internal metadata are populated.

  • Increased the default absolute-bootstrap iteration limit in gof() from 100 to 5000; explicit maxiter values and compare() defaults are unchanged.

slca 1.4.1

CRAN release: 2026-10-23

  • Aligned formula-environment regression covariates by their observation IDs before transformations, including factors and matrices. Unnamed environment covariates now require explicit positional alignment; ambiguous original-length inputs receive actionable errors.

  • Reported saturated-table EM convergence and withheld absolute fit statistics for unfinished iterations.

  • Aligned regression covariates by row identity even with equal row counts; added explicit positional alignment.

  • Clarified categorical level requirements for constrained indicators and added regression coverage.

  • Required final-stage convergence for hybrid bootstrap refits.

  • Declared testthat >= 3.1.7 for the mocked-binding test helpers.

  • Corrected missing-item scores used for standard errors and covariance matrices.

  • Included compatible unobserved cells in large saturated response tables; oversized tables now warn and return unavailable absolute fit statistics.

  • Preserved missingness and rejected nonconverged fits in absolute bootstrap tests, with finite-sample p-values; relative tests reject negative observed likelihood ratios.

  • Corrected reference logits and Hessian transformations after class reordering.

  • Fixed confidence interval subsets and regression odds-ratio intervals, multiple starts with zero constraints, and simulated log-likelihoods for multiple measurement blocks.

  • Assigned modal ties to the first class consistently, rejected cyclic model structures, and excluded all-indicator-missing observations as documented.

  • Fixed saturated-frequency handling for partially missing data when missing patterns map to candidate observed cells of equal length.

  • Updated internal response encoding to use -1 for missing values and zero-based category indices for C++ calculations.

  • Kept simulated manifest responses as plain factor data frames without internal encoding attributes.

  • Fixed predict.slcafit() when newdata is omitted.

  • Added validation for predict() inputs, simulation levels, and initial parameter lengths.

  • Added validation for model formulas, regression latent outcomes, bootstrap counts, and simulation counts.

  • Fixed simulated factor responses when some generated categories are absent in a sample.

  • Fixed Hessian-based covariance extraction and regression confidence interval returns.

  • Fixed regression standard errors and Wald p-values in summary.reg.slca().

  • Hardened gof() list inputs, do.call() calls, and bias-adjusted regression internals.

  • Recorded bootstrap replicate failures in model-fit diagnostics instead of returning error objects.

  • Strengthened the relative bootstrap likelihood-ratio test with convergence checks, multiple starting values, missing-pattern preservation, and finite-sample p-values.

  • Clarified the documented structure of convergence, predict(), and simulated responses.