0.7 series

Version 0.7-3
  • Released on CRAN: 8 October 2026
  • New features and user-visible changes:
    • the transition to snake_case is now (almost) complete: the documentation and all messages use the snake_case option names; the elements of list-valued options (estimator_args, rotation_args, fm_args, …) are now also snake_case; the keywords used as selectors (the what= argument of lavInspect(), fit measure names, type=, method=, …) can be written with dots or underscores, in lower or upper case; the old (dotted) names still work
    • estimator = “ML”: Browne’s residual-based test (model-based, normal theory; “browne.residual.nt.model”) is now computed by default and shown in the summary, next to the standard chi-square test; in small samples, its type I error rates are much closer to the nominal level
    • categorical data + estimator WLSMV/ULSMV (and friends): the ‘bread’ of the sandwich used for the robust standard errors is now based on the observed information (instead of the expected information); for mixed (ordered + continuous) data, observed derivatives are now used for the continuous variables; as a result, the standard errors may differ slightly from 0.7-2 (use information_bread = “expected” to get the classic standard errors); point estimates and test statistics are not affected
    • new option information_bread= to set the type of information used for the bread of the sandwich, independent of the meat; new option information_meat_hc= for small-sample (HC1/HC2/HC3) corrections of the robust (sandwich) standard errors
    • correlation structures (correlation = TRUE): the default estimator is now ML (instead of GLS); support for conditional_x = TRUE, a mean structure, missing = “ml”, composites, and partial correlation structures; lavPredict() now works in the correct metric
    • new rescale_data option: if the variances of the observed variables are very different, the variables are internally rescaled (by default), and the solution is transformed back to the original metric
    • if the first indicator of a latent variable is a poor marker, lavaan now only produces a warning (and no longer switches to another marker automatically); to switch the marker, use bad_marker_switch = TRUE (github issue 628)
    • new function lavTestNET() for nesting and equivalence testing; lavTestLRT() now checks if the models are nested (check_nested = TRUE), and warns if the models were not fitted to the same data
    • new function lavH1() to compute the unrestricted (h1) sample statistics (optionally with standard errors); with output = “lavMoments”, the result can be used to refit a model from summary statistics only (including two-stage and least-squares approaches)
    • the noniterative estimators “JS”/“JSA” (James-Stein) and “MGM” (multiple group method) have been rewritten, and now offer (delta method) standard errors, Browne’s residual-based test, fit measures, multiple groups, equality constraints, and two-stage missing data (see ?estimator_js); “MGM” also handles cross-loadings and correlated residuals, and can be used for the measurement blocks in sam()
    • estimator = “IV”: new iv_method = “LIML” and “FULLER” (next to the default “2SLS”); much faster standard errors for multiple groups
    • new operator “:~” for data-defined (moderated) parameters, where a parameter is a function of observed variables (wrapped in data()); lavaan only handles the syntax and the housekeeping; the estimation is left to external backends
    • rotation: the gradient projection algorithm now uses a Barzilai-Borwein step size (faster); see the new gpa_algorithm element of rotation_args (use “legacy” for the old behavior)
    • the optimizer now detects ‘runaway’ solutions (where a parameter drifts to infinity) instead of reporting convergence
    • sam(): local sam + latent interactions now supports missing = “ml”; new local options (lambda1_floor, lambda2_floor) for a (debiased) truncation of the lambda correction in small samples/low reliability settings, with standard errors that take the truncation into account; the corrected (Hayakawa) adjusted test statistics are now available for the structural part of a local sam model
    • sam(): se = “local” now uses the observed information for the bread of the structural sandwich
    • lavPredict(): transform = TRUE now also works for categorical data, and for missing = “ml” (github issue 624)
    • missing = “two.stage”/“robust.two.stage” now supports sampling weights
    • lavInspect()/lavTech(): many new keywords to extract the internals of the data and sample statistics slots (e.g., “sampling_weights”, “patterns_stats”), the constraints (“constraints”, “con_jac”) and the results of the random starts (“rstarts”)
    • a warning is given if the sample_cov= matrix is not symmetric
    • continuous data + missing = “pairwise” + robust standard errors/test statistics now results in an informative error (for the ML and GLS estimators, if the data contain missing values)
    • mimic = “EQS” and mimic = “LISREL” now set fixed_x = FALSE
  • Bugs/glitches discovered after the release:
    • (just released)
Version 0.7-2
  • Released on CRAN: 16 July 2026
  • New features and user-visible changes:
    • the formal arguments of (almost) all user-facing functions have been renamed to snake_case (e.g. sampling.weights -> sampling_weights, sample.cov -> sample_cov, int.ov.free -> int_ov_free); the old (dotted) names still work, but are deprecated and may be removed in a future release; the same holds for the lavaan option names, which may now also be written in snake_case or with capital letters; in the next release of lavaan (0.7-3), we hope to finalize this transition to the snake_case pattern, thereby complying with modern coding guidelines
    • the ‘open’ syntax parser is now the default
    • new operator “|~” to specify (external) instruments
    • new estimator = “IV” for model-implied instrumental variables (MIIV-2SLS) estimation; supports multiple groups, categorical data, (simple) equality constraints, two-stage missing data, user-specified/external instruments, and Sargan/Hansen overidentification tests (see ?estimator_iv for more information and examples)
    • two-level (D)WLS estimation, including WLSMV for categorical data, and exogenous covariates (fixed_x/conditional_x)
    • random slopes for two-level models via the new rv() modifier; ML (nlminb and EM), robust (MLR) standard errors, empirical-Bayes predictions, and support for both complete and missing (missing = “ml”) data (continuous data only) (see the Multilevel section in the tutorial for an example)
    • optim.method = “em” is now the default for twolevel + missing = “ml”; EM acceleration (SQUAREM and quasi-Newton) and analytic (Louis) observed information for faster/more accurate standard errors
    • composites (using the “<~” operator) now use an analytic gradient (the optim.gradient = “numerical” workaround is no longer needed), and gain a robust mean structure, higher-order factors, multilevel support, and a new composites_cov option
    • initial support for the reduced-bias M-estimation (RBM) framework to improve finite-sample bias (see this paper)
    • new Foldnes-Moss-Gronneberg (FMG) test statistics; new Hayakawa test statistic: test = “mean.var.adjusted.corrected” and “scaled.shifted.corrected” (see ?lavTestLRT and ?lavTest for references)
    • sampling weights: new sampling.weights.type= argument (design vs frequency weighting), per-group normalization by default, and a properly weighted Gamma (NACOV) for the least-squares estimators
    • a new goodness-of-fit index (GFI, Maydeu-Olivares et al. 2024) is added to fitMeasures(); the old gfi/agfi are renamed gfi_lisrel/agfi_lisrel
    • lavResiduals() overhaul: output = “text” (with largest-residuals tables), per-element residual standard errors with z-tests, support for conditional.x and multilevel models, and a user-provided h1= saturated model
    • new function lavEffects() for total/indirect/direct effects, with delta-method, Monte Carlo and bootstrap standard errors
    • new function lavResidualsY() for residuals of the ‘y’ variables
    • sam(): cluster-robust standard errors for single-level clustered data, local standard errors and a corrected structural test for two-level models, and the Yuan-Chan rescaled test for sam.method = “global”
    • sam(): conditional.x = TRUE is now supported, both for continuous and categorical (ordinal indicators and ordered outcomes) data, including a corrected structural test and robust standard errors; new (experimental) se = “twostep.huber.white”; se = “local” now also works for estimator = “PML”
    • sam(): fitMeasures(), lavResiduals(), modindices() and friends now work on sam objects
    • new se.def = “mc” for Monte Carlo standard errors/confidence intervals of defined (“:=”) parameters; new se.delta.second.order option
    • new aux= argument for auxiliary variables under missing data
    • support for a partial correlation structure (a vector of variable names passed to correlation=), including fixed.x = TRUE
    • lavPredict(): standard errors for categorical factor scores; newdata= is now supported for two-level models
    • lavPredict()/lavInspect(): casewise Mahalanobis-distance diagnostics (mdist=) are now also available for categorical/ordered data
    • the estimator=, rotation= and bootstrap= arguments can now be given as a list (replacing the separate *.args arguments)
    • missing = “ml” (FIML) and missing = “two.stage”/“robust.two.stage” now support conditional.x = TRUE (single-level, continuous data), including the standard errors and a correct Gamma (NACOV)
    • new auto.cov.x option: freely estimate the covariances between latent and observed exogenous variables (currently set to FALSE by default)
    • multilevel models: level-specific fit measures (based on partially saturated models) are shown in the summary and can be requested via fitMeasures(fit, level = …); controlled by the new fit_by_level option
    • lavTestScore() and lavTestWald() now also report scaled, adjusted and robust versions of their test statistics when a robust estimator is used
    • optim.method = “gn” (Gauss-Newton) has been rewritten as a Levenberg-Marquardt damped Fisher-scoring optimizer (faster and more robust); its options now live in the gn_args= list
    • ESEM: a separate rotation target matrix can be specified per EFA block, via a named list in rotation_args$target
  • Bugs/glitches discovered after the release:
    • the parsed-syntax cache could (rarely) return the parsed version of a different model, if two model strings happened to share the same hash key; the wrong model was then silently fitted (github issue 636)
    • the automatic marker switch (bad_marker_crit) could result in a doubly-scaled factor (and negative latent variances) if the marker loading was labeled and involved in an equality constraint; more generally, switching the marker silently changed the metric of the latent variable (github issue 628); see the 0.7-3 changes
    • mixed categorical + continuous data fitted from sample statistics: the means/intercepts of the continuous variables had the wrong sign
    • categorical data (WLSMV) or correlation structures with an observed mediator: the analytic Jacobian ignored the dependence of the (derived) residual variance of the mediator on the regression coefficients, resulting in convergence problems and/or (slightly) wrong standard errors
    • composites + categorical data: the analytic model Jacobian was wrong
    • composites = FALSE crashed with the (new) open parser
    • passing rotation = list(…) to sem()/cfa()/lavaan() (instead of efa()) failed (github issue 618)
    • rotation_se = “bordered” used the wrong variances to standardize the rotated loadings when the rotation was oblique; rotation_se = “delta” and “bordered” now give identical results
    • oblique GPA rotation could crash on a (numerically) singular transformation matrix when factors collapse (e.g., oblimin with gamma > 0)
    • lavTest(fit) returned NA (instead of the available test statistic) for estimators without a standard chi-square test (e.g., estimator = “IV”)
    • estimator = “IV”: a label repeated within a single equation (e.g., y ~ ax1 + ax2) produced inconsistent estimates; the standard errors of the intercepts used the wrong divisor when only sample statistics were provided
    • the Browne residual-based NT test statistics were wrong in several settings: (near) zero when fixed_x = TRUE and the ‘y’ block was saturated, inflated when missing = “ml”, and “browne.residual.adf.model” ignored the model-based moments when sampling weights were used
    • optim_parscale = “standardized” (also used in the automatic retry attempts when the optimizer fails): the bounds were not rescaled, and general (in)equality constraints (e.g., a == 2*b) were imposed on the scaled parameters
    • a parameter with both a lower and an upper bound that was estimated at its lower bound was not treated as an active constraint when computing the standard errors
    • the baseline model for two-level categorical data could not be fitted (“non-conformable arrays”)
    • level-specific fit measures: the between-level RMSEA used the total sample size instead of the number of clusters
    • two-level (D)WLS: the observed information matrix was not scaled correctly
    • conditional_x = TRUE crashed when an observed variable only appeared in covariance ("~~") formulas
    • group_equal = “all” ignored an explicit meanstructure = FALSE (github issue 631)
    • lavPredict(transform = TRUE) ignored the missing patterns when missing = “ml” (github issue 624), and crashed when conditional_x = TRUE
    • lavPredict(): the EBM/ML factor scores for categorical data were wrong in the presence of a mean structure (intercepts/latent means were added twice)
    • simulateData(empirical = TRUE) failed when skewness/kurtosis values were provided (github issue 622); for multilevel data, a warning about ignored arguments was always shown
    • sam(se = “standard”) stopped with the error “object ‘i_22’ not found” (github issue 626)
    • local sam() + latent interactions + missing = “ml”: all standard errors were NA (github issue 621)
    • local sam() + latent interactions (or quadratic terms) + more than one measurement block: the step-1 contribution to the standard errors (se = “local”) used the wrong parameter order
    • re-applying sam() on a stored sam object failed when se = “local”
    • estimator = “PML” with ordered indicators, exogenous covariates and missing = “available.cases” failed as soon as a response was missing; do_fit = FALSE returned a wrong objective value (github issue 635)
    • post_check did not flag a non-positive-definite set of residual covariances among ordered endogenous variables (github issue 633)
    • sampling weights + conditional_x = TRUE + missing = “pairwise” failed (github issue 639)
    • cluster= + ordered data + missing = “pairwise”: the cluster-robust standard errors were NA (github issue 642)
    • estimator = “DWLS” + missing = “two.stage” + sampling weights crashed; two-stage (ML) standard errors were not available with sampling weights
    • correlation = TRUE + sample_cov=: the sample covariance matrix was standardized too late, resulting in an inconsistent objective function and gradient if the input was not a correlation matrix; user-specified variances of observed variables were not fixed
    • defined parameters (using the “:=” operator) could be assigned to the wrong rows if they were not specified in alphabetical order
    • model syntax using Windows line endings (CRLF) was not parsed correctly
    • lavParTable(auto = TRUE) ignored model_type = “cfa”
    • objects created with (very) old versions of lavaan (< 0.5-19) could no longer be read
Version 0.7-1
  • This version was never released on CRAN. Development versions use the 0.Y-y.xxxx versioning scheme, where xxxx denotes the Git commit number. Starting with the 0.7 series, the CRAN release will always have a higher version number than the corresponding development versions (e.g., 0.7-2 > 0.7-1.2959).