lavCor() ubiquitously sets correlation = categorical

[TDJorgensen]

This is not a problem for the default behavior of lavCor(..., output = "cor"), but when lavCor(..., output = "fit") and any variables are ordered= , the saturated model generated by lav_partable_unrestricted() will fix ALL variances to 1, not just the latent-response variances for ordered= variables. This is not a problem when ALL variables are ordered=, but with a mix of categorical and continuous variables, an inappropriately restrictive correlation=TRUE model is fitted:

HS3 <- HolzingerSwineford1939[,c("x1","x2","x3","sex")]
parTable(lavCor(HS3, ordered = "sex", output = "fit"))

Granted, the model isn't actually fitted, sbut it does not seem necessary to set correlation=categorical here.

• None of the output depends on all the variables having a variance fixed to 1 during estimation, since lav_cor_output() applies cov2cor() or standardizedSolution() anyway.
• The lav.data object informs lav_partable_unrestricted() which variables are ordered=, so those particular variances are fixed to 1 anyway.

So I think correlation=categorical can be deleted without affecting anything lavCor() is utilized for, except that the correct model is specified when output = "fit". That is something lavaan.mi::poolSat() utilizes, which is how some noticed this when trying to fit an efa() to multiply imputed data.