---

0.6 series

Version 0.6-6

• Released on CRAN: 13 May 2020
• New features and user-visible changes:
  • lavPredict() + se = TRUE now works correctly in the presence of missing data and missing = “fiml”
  • lavResiduals() + summary = TRUE now provides confidence intervals and better labelling
  • lavResiduals() + summary = TRUE now works for categorical data, but only in models without exogenous covariates (for now)
  • lavCor() gains cor.smooth= argument to force the correlation matrix to be positive definite
  • sum(sampling.weights) no longer needs to equal the total sample size
  • sampling.weights can now be used in combination with categorical data and estimator WLSMV and friends
  • information/h1.information/observed.information arguments now accept a vector of two elements: the first one is for the standard errors, the second for the test statistic
  • the data (provided by the data= argument) is now always converted to a data.frame (in case it is a tibble, or another class that inherits from data.frame)
  • experimental feature: when the sample size is (very) small, adding the argument bounds = TRUE may stabilize estimation, by automatically choosing upper and lower bounds for several sets of model parameters
• Known issues:
  • same as 0.6-1
• Bugs/glitches discovered after the release:
  • (just released)

Version 0.6-5

• Released on CRAN: 28 Aug 2019
• New features and user-visible changes:
  • a new option `effect.coding’ can be used to scale the latent variables; a typical use is effect.coding = “loadings”, or effect.coding = c(“loadings”, “intercepts”); if “loadings” is included, equality constraints are used so that the average of the factor loadings (per latent variable) equals 1; if “intercepts” is included, equality constraints are used so that the sum of the intercepts (belonging to the indicators of a single latent variable) equals zero; as a shortcut, you can also set effect.coding = TRUE, which implies effect.coding = c(“loadings”, “intercepts”); note that his may not work well with bifactor models, or any other type of model where items depend on multiple factors
  • summary(fit, nd = 4) now shows 4 digits (after the decimal point) for all (non-integer) numbers, including the header; the header has been re-arranged slightly: the estimator is now shown on the top; the test statistic(s) are now in a section ‘Model Test User Model:’
  • the test= argument now accepts a vector of multiple test statistics, for example test = c(“satorra.bentler”, “mean.var.adjusted”)
  • parameterEstimates(), standardizedSolution() and fitMeasures() have a new output= argument which can be set to “text” for pretty printing
• Known issues:
  • same as 0.6-1
• Bugs/glitches discovered after the release:
  • in the categorical setting (estimator U/WLSMV) + missing = “pairwise” + the number of thresholds is larger than one: some entries in the weight matrix are not correct, resulting in (small) bias for some parameters
  • optim.method=”none” is ignored in nlminb.constr() settings; as a side-effect, this produces wrong results in lavTestLRT() if method=”satorra.bentler.2010”, and explicit inequality constraints are used in the model
  • wrong labels for lavInspect(,”lv.mean”) and lavInspect(,”cov.all”) if both formative and reflexive latent variables are used
  • the “<~” operator did not work well in the categorical setting in combination with conditional.x = TRUE
  • parameterEstimates + fmi = TRUE often gives crazy results
  • lavTestLRT + MLR + df=0 for one of the models results in an error
  • estimator = “WLS” + continuous data + missing = “pairwise” does not work
  • loglikelood value if fixed.x = TRUE and only sample statistics are provided is still reflecting the joint loglikelihood

Version 0.6-4

• Released on CRAN: 3 July 2019
• New features and user-visible changes:
  • this release includes mostly changes under the hood
  • an extra check is done for convergence (even if the optimizer claims convergence): we check if all the elements of the (unscaled) gradient are smaller than 0.0001 (in absolute value). If not, we set converged = FALSE. The tolerance value 0.0001 can be changed by the option optim.dx.tol. To switch off this check, use optim.dx.tol = +Inf
  • lavInspect() gains new options: “vcov.def.joint” (and standardized variants) to give the (joint) variance-covariance matrix of both the free and `defined’ (using the := operator) variables; “loglik.casewise” to extract casewise (log)likelihood values (only if ML has been used); “h1.information” to extract the information matrix of the unrestricted h1 model
  • lavTestScore() + uni table and modindices() output (in particular epc and epv values) are now more alike
  • lavPredict() + Bartlett now uses FIML to compute factor scores (in the presence of missing data); new append.data= argument cbinds factor scores and raw data; if assemble = TRUE, factor scores of multiple groups will be joined in a single data.frame (with a group column); accept newdata= argument, even if sample statistics were used when fitting the original model
  • using ordered = TRUE implies all variables should be treated as ordinal
  • new logical arguments orthogonal.y and orthogonal.x can be used to specify zero covariances between endogenous and exogenous latent variables (only) respectively
  • new option sampling.weights.normalization (set to “total” by default)
  • group.equal now also accepts “composite.loadings” when the “<~” operator is used
  • parameterEstimates() gains a remove.nonfree= argument
  • initial (but still experimental and undocumented) support for rotation, EFA, and ESEM (contact me if you wish to test this)
  • initial (but still experimental and undocumented) support for the `Structural After Measurement’ (SAM) approach (see the new lavaan:::sam() function; contact me if you wish to test this)
• Known issues:
  • same as 0.6-1 (although many long standing github issues have been resolved)
• Bugs/glitches discovered after the release:
  • twolevel models: standard errors are off for fixed effects (only) if the model contains a regression of the form y ~ x, both variables are observed, and x is not centered
  • twolevel models: the chi-square test statistic for the baseline model is wrong (too low): the chi-square test statistic for the full model was substracted; therefore, if the user model fitted really bad, the baseline test statistic was too low; if the user model fitted reasonably well, there was (almost) no impact on the CFI/TLI fit measures
  • lavPredict(): labels were wrong in the multigroup twolevel case

Version 0.6-3

• Released on CRAN: 22 September 2018
• New features and user-visible changes:
  • new function lavResiduals() provides raw and standardized residuals in addition to summary statistics (rmr, srmr and crmr, including their standard error, a test for exact fit, and an unbiased version); for continuous data only (single-level, conditional.x = FALSE)
  • new option missing = “ml.x” or missing = “fiml.x” will not delete cases with missing values on exogenous covariates, even if fixed.x = TRUE (this restores the behavior of < 0.6); this can be useful for models with a large number of exogenous covariates, which you can treat as stochastic, but where fixed.x = TRUE is just more convenient
  • lavInspect/lavTech + “sampstat” now always returns the data summaries from the h1 slot; this implies that if data is missing, and missing = “(fi)ml)”, you get the ‘EM’ estimated covariance matrix and mean vector (instead of the pairwise deleted covariance matrix from the samplestats slot)
  • lavInspect/lavTech + “sampstat” and “implied” no longer return the mean vector if meanstructure = FALSE
  • srmr_bollen is renamed crmr in the output of fitMeasures(), and does no longer count the (zero) diagonal elements when taking the average
  • if conditional.x = TRUE, the baseline model will now set the slopes free (instead of fixing them to zero); this will usually result in a much better fit for the baseline model, and a less optimistic value for CFI/TLI and friends; to get back the old behaviour, set the option ‘baseline.conditional.x.free.slopes’ to FALSE
• Known issues:
  • same as 0.6-1
• Bugs/glitches discovered after the release:
  • mimic = “Mplus” was no longer setting meanstructure = TRUE
  • if a partable was provided as input (instead of model syntax), the `y’ variables were identified by using lavNames(fit, “ov.y”) instead of lavNames(fit, “ov.nox”); therefore, we ‘missed’ some variables in lav_dataframe_check_ordered(), not (automatically) treating them as ordered
  • mimic = “EQS” was not setting all the correct settings any longer
  • lavPredict() + Bartlett + missing data did not produce ‘true’ FIML factor scores (although they were very close)
  • lav_partable_merge() did not set the group/level column right, affecting lavTestLRT() when method = “SatorraBentler2010” (producing a warning that the optimizer could not find a solution, and no test statistic)
  • when a second-order factor was included in the model, the summary output did not print a dot (‘.’, to indicate endogenous) in front of the (residual) variances of the first-order latent variables
  • lavTestLRT() + Satorra2000 with two models having the same df resulted in a strange error
  • lavTestScore() + uni resulted in epc values with the wrong sign; as a result, the epv values were in the wrong direction
  • lavaanList() produced “nlev too low for some vars” error, when ordered variables were involved that included missing values (NA) (the NA was counted as an additional category)

Version 0.6-2

• Released on CRAN: 16 July 2018
• New features and user-visible changes:
  • cluster-robust standard errors are reported when the cluster= argument is used in combination with a standard single-level syntax (without the “level:” keywords); this can be used in combination with missing = “fiml” (to handle missing values) and sampling weights
  • added lower() and upper() modifiers in the model syntax; it works in a similar way as the start() modifier to provide starting values; for example, to force a variance (say, y1 ~~ y1) to be larger than 0.001, you can now write y ~~ lower(0.001)*y1
  • even if the optimizer claims to have found a solution, we check if the gradient is (almost) zero, and warn if not; this check can be turned on/off by the option check.gradient = TRUE/FALSE
  • more effort is done to check if the model is identified in the presence of equality constraints; a warning is given if vcov is (nearly) singular; this check can be turned on/off by the option check.vcov = TRUE/FALSE
  • header of the summary() function now shows the optimizer, the number of free parameters, and the number of constraints (if any)
  • fabin3 starting values for factor loadings are now also computed if std.lv = TRUE
• Known issues:
  • same as 0.6-1
• Bugs/glitches discovered after the release:
  • ifi.scaled fit measure was not scaled
  • lavTables() did not work if the first argument was a data.frame
  • two-level + optim.method = “em” resulted in an error (about the check argument)
  • modindices() did not work if conditional.x = TRUE
  • lavInspect(, “data”) did not return covariates if conditional.x = TRUE

Version 0.6-1

• Released on CRAN: 22 May 2018
• New features and user-visible changes:
  • initial support for two-level SEM with random intercepts; see example(Demo.twolevel)
  • initial support for sampling weights (for non-clustered data only for now)
  • the code to compute (robust) standard errors and (robust) test statistics has been rewritten; the code is now much more consistent, but this may result in small changes in the values of some robust test statistics as printed in the output of summary(); a meanstructure is no longer added automatically when switching to a robust estimator
  • conditional.x = FALSE now works in the categorical case (both with estimators DWLS and PML)
  • if fixed.x = TRUE, we compute the loglikelihood without the exogenous covariates, and we do not allow for missing values in the exogenous covariates when missing = “ml” (and fixed.x = TRUE)
  • estimator = “PML” now works for continuous-only data, or mixed continuous/ordinal data
  • lavTestLRT(, method = “satorra2000”) which is the default when comparing models that were fitted with estimator MLMV or WLSMV now uses the so-called ‘delta’ method by default; this allows for comparing models that are nested in the covariance sense (and not just in the parameter sense); in addition, a scaled.shifted test statistic will be used by default
  • the warning about empty cells in bivariate cells (when categorical data is used) is no longer shown
  • many, many small changes under the hood
• Known issues:
  • two-level SEM: no random slopes, no missing data, no categorical data
  • the marginal ML infrastructure (estimator = “MML”) has not been updated yet
  • several github issues have not been resolved yet
• Bugs/glitches discovered after the release:
  • robust CFI/TLI and robust RMSEA values are printed as NA when estimator = “MLR”
  • lavPredict(, level = 2) does not handle phantom latent variables properly

---