new lavaan.mi package on CRAN

[Terrence Jorgensen]

A new R package is available on CRAN, which contains some familiar but greatly improved functionality that is now deprecated from semTools 0.5-7.  Among the improvements is a stronger consistency in the user interface between lavaan and lavaan.mi, including new *.mi() functions listed in the README description.  The NEWS file includes a list of changes that make this package differ from the semTools functionality.  I also attached my presentation slides from the recent PsychoCo workshop, which provide a good overview of the package's functionality.

The semTools package still supports lavaan.mi-class objects, but until the deprecated semTools functions are completely removed from a future version of semTools, you should always load the packages in the following order:

library(semTools)

library(lavaan.mi)

Note that either line automatically first loads lavaan because it is a dependency for both packages.  So a separate library(lavaan) call is not necessary.

Best,

Terrence D. Jorgensen  (he, him, his)

Assistant Professor, Methods and Statistics

Research Institute for Child Development and Education, the University of Amsterdam

http://www.uva.nl/profile/t.d.jorgensen

[Christian Arnold]

Hi Terrence,

thanks for this fantastic package! I noticed that the following syntax for an MGSEM can cause problems:

…

f ~ c(M.G1, M.G2) * 1
M.G1 == 0

One can easily work around this by using a little trick (which isn't very pretty though):

group: 1

…

group: 2

…

f ~ M.G2 * 1

It seems like the first approach sometimes (apparently not always) outputs NA for the M.G1 se and thus excludes the imputation from imps2use, even if the model fits perfectly fine. I can send you a rereducible example if needed.

Since the vcov is available, should it actually be possible to determine the CIs (se's and the p-values) using Monte Carlo simulation? At least with many imputations, this should have a speed advantage compared to parameterEstimates.mi and the differences between the CIs should actually be negligible? monteCarloCI (semTools) does not currently accept a lavaan.mi object, or am I mistaken here?

Best

Christian

[Terrence Jorgensen]

It seems like the first approach sometimes (apparently not always) outputs NA for the M.G1 se

That often happens in lavaan with complete data, too.  A better syntax would specify labels and fixed/free parameters (when that is what you want):

f ~ c(M.G1, M.G2)*1 + c(0, NA)*1 

Since the vcov is available, should it actually be possible to determine the CIs (se's and the p-values) using Monte Carlo simulation?

Yes.

 

At least with many imputations, this should have a speed advantage compared to parameterEstimates.mi

I doubt it.  The normal-theory CIs are already calculated from vcov() and don't require simulating estimates.

 

monteCarloCI (semTools) does not currently accept a lavaan.mi object, or am I mistaken here?

 It does:

library(semTools)

library(lavaan.mi)

data(HS20imps) # import a list of 20 imputed data sets

## specify CFA model from lavaan's ?cfa help page
HS.model <- '
  visual  =~ x1 + b1*x2 + x3
  textual =~ x4 + b2*x5 + x6
  speed   =~ x7 + b3*x8 + x9
b_sum := b1 + b2 + b3 '

fit <- cfa.mi(HS.model, data = HS20imps)
monteCarloCI(fit)