semTests 2.0.0

Breaking changes

pvalues_nested() has a new default restriction matrix, A.method = "weighted", which does not depend on the units of the observed variables or on the factor identification. lavaan’s construction remains available as A.method = "delta", which reproduces lavTestLRT(method = "satorra.2000") but changes with the units.
semTests requires lavaan (>= 0.7-3). Its asymptotic covariance for mixed continuous and ordinal data is now robust for the continuous variables.

semtest() computes score and criterion-difference tests of overall fit and of restrictions in nested models. The restriction tests remain valid when the larger model is misspecified. Overall fit covers complete-data ML and FIML; restriction tests also cover continuous least squares and complete, all-ordered DWLS, WLS, and ULS fits.
pvalues() accepts semtest() results, so the usual calibrations (ALL, SB, SS, PEBA<j>, …) apply to them.
inference_control() collects the test conventions (curvature, score covariance, fitting-weight effect) and numerical settings.
New vignette: vignette("score-criterion", package = "semTests").

fiml.convention is deprecated. Its "lavaan" option depended on the fit’s information and h1.information options; use lavaan::lavTest() or lavaan::lavTestLRT() to reproduce lavaan’s tests.

pvalues_nested(A.method = "exact") was ignored for complete-data fits.
Nested comparisons with nonlinear equality constraints evaluated the constraint Jacobian at the starting values instead of the estimates. This affected categorical data.
Nested FIML tests lost accuracy when the variables had very different scales.

Requests for several tests share their statistics and eigenvalues, and complete-data ML no longer builds the full Gamma matrix.
The vignettes are reorganized into four: an introduction, ordinal indicators, missing data, and score and criterion tests.

Earlier versions marked several estimators as experimental. This release promotes them to stable.

GLS, ULS, categorical DWLS and ULS, mixed continuous and ordinal indicators, and FIML, each with one or several groups, now work for both single models and nested comparisons.
A fiml.convention argument selects the FIML information convention. The default disagrees with lavaan but has better statistical properties; set fiml.convention = "lavaan" to reproduce lavaan’s numbers exactly.
New vignettes cover measurement invariance with ordinal indicators and latent growth models under FIML.

The reweighted least-squares statistic was read as NULL under lavaan 0.7-2 and is now correct. semTests requires lavaan (>= 0.7-2).