Warning message: In cov2cor(w.cov) : diag(.) had 0 or NA entries; non-finite result is doubtful

[franz...@gmail.com]

Hello,

I tried to create the pooled within groups covariance matrix for a CFA using this function:
ml.s.change <- mcfa.input("Vkst", s.change)

Unfortunately, r shows this "Warning message: In cov2cor(w.cov) :   diag(.) had 0 or NA entries; non-finite result is doubtful" afterwards.
However, it is possible to fit a model. So I am a bit confused: what does the message mean? Can I trust the fit measures r displays when I fit the pw-matrix with my CFA?
And most important: how can I get rid of the message?

Thanks in advance!!!

All the best,
Franzi

[Terrence Jorgensen]

I tried to create the pooled within groups covariance matrix for a CFA using this function:
ml.s.change <- mcfa.input("Vkst", s.change)

I had to use Google to figure out that you were using a function (which is not part of the lavaan package) from this article:

http://faculty.missouri.edu/huangf/data/mcfa/MCFA%20in%20R%20HUANG.pdf

Did you know that the development version of lavaan can handle some simple cases of multilevel SEM?  Only 2 levels, continuous data, random intercepts (sufficient for multilevel CFA).  No random slopes, categorical data, or missing data.  Maybe lavaan provides what you need now, so you don't need this article's ad-hoc solution?

http://lavaan.ugent.be/development.html

http://users.ugent.be/~yrosseel/lavaan/tubingen2017/multilevelSEM_part1.pdf

Unfortunately, r shows this "Warning message: In cov2cor(w.cov) :   diag(.) had 0 or NA entries; non-finite result is doubtful" afterwards.
However, it is possible to fit a model. So I am a bit confused: what does the message mean?

A correlation is a standardized covariance.  You standardize the covariance between X and Y by dividing it by the SD of X and the SD of Y.  If the diagonal of the covariance matrix (i.e., the variances) includes a 0 or NA, then you cannot get a correlation because you cannot divide by zero (or NA, which is R's missing-data code).

 

Can I trust the fit measures r displays when I fit the pw-matrix with my CFA?

I don't know anything about Huang's ad-hoc solution.  Perhaps you can email him directly, since he is the author of that software.

And most important: how can I get rid of the message?

 

Instead of loading the source code from the article, trying actually stepping through it to see why the error occurs:

http://faculty.missouri.edu/huangf/data/mcfa/mcfa.R

Terrence D. Jorgensen

Postdoctoral Researcher, Methods and Statistics

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

UvA web page: http://www.uva.nl/profile/t.d.jorgensen