I came across this magnificent post from Rose Hartman (http://www.understandingdata.net/2017/03/22/cfa-in-lavaan/#writeup) that suggests standardising latent factors (constraining them to have a mean of 0 and a variance of 1) in order to allow free estimation of all factor loadings when performing CFA. I did it with the lavaan package in R by setting std.lv=TRUE when calling cfa().

My question is: am I standardising unique factors (measurement errors) too, by doing so?


No, the latent variable standardization protocol with CFA only standardizes the latent variables. If, before you ran the analysis, you standardized your manifest measures, then the factor loadings can be interpreted exactly as correlations (between latent factor and measured item). However, even then, your error variances will be constrained by the factor loading (correlation). In particular, if the manifest item has $\lambda$ factor loading, than the variance of the standardized manifest item with be estimated as $1-\lambda^2$. Thus, the measurement errors will only be standardized if all $\lambda$ are zero...which defeats the purpose of CFA (i.e., dimension reduction), as it suggests each item is its own “dimension”.

  • $\begingroup$ Thank you @GreggH!! I have two follow up (probably silly) questions: I didn't standardise my manifest measures, but based the CFA on polychoric correlations (and standardised the latent factor) - therefore my factor loadings can also be interpreted as correlations, isn't it? The second question is: considering what you said about measurement errors, I understand they are simply a logic complement of factor loadings - is it necessary then to report them on a paper (for instance, in the diagram representing the model tested) after running a CFA in which the latent variable is standardised? $\endgroup$
    – mireiabo
    Mar 27 '18 at 16:11
  • $\begingroup$ Question 1: standardized factor loadings (meaning the output is standardized), yes...just using standardized latent factors (without standardized manifest measures), no, the loadings won't be the correlations. Question 2: I wouldn't report them (that doesn't mean a reviewer less familiar with the mathematics behind CFA/SEM won't ask for them). $\endgroup$
    – Gregg H
    Mar 27 '18 at 18:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.