CITATION,
DESCRIPTION, and the package documentation.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.
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.