omegasCFA() no longer reports a zero-width confidence interval when the standard
errors of the omega coefficients could not be computed. lavaan returns the point
estimate as both interval bounds when the information matrix cannot be inverted,
which reads as a perfectly precise estimate. The interval is now NA and a warning
is issued. This affects in particular the second-order model with two group factors,
where the general factor is not identified: the general factor has one loading per
group factor, but the group factors supply only a single correlation. The point
estimates are still returned.
omegasCFA() no longer fails entirely when fit.measures = TRUE and the fit measures
cannot be computed, which lavaan refuses for a model that did not converge. The
coefficients are returned and the fit measures are omitted.
omegasCFA() now warns when the factor model did not converge, and when the solution
is inadmissible, for instance because of a negative variance estimate. Previously a
non-converged model returned plausible-looking coefficients without any indication.
omegasCFA() gained diagnostics, a list with the entries
converged, admissible, and se.available, and lavaan.fit, holding the fitted
lavaan object. print() reports the diagnostics when any of them fail.bomegas() now calls the function supplied to callback during the Gibbs
sampling loops. Previously the argument was accepted but never invoked, so
external progress reporting never advanced.
bomegas() no longer errors for model.type = "bi-factor" in combination
with param.out = TRUE. The array holding the posterior samples of the
factor variances was allocated without room for the general factor.