Comparing factor scores

[Reza Norouzian]

When using predict for a fitted model in package lavaan, we can obtain the factor scores (fscores). However, by default these fscores are all made to have a mean of 0 (i.e., centered).

However, I want to compare fscores from two factors to see their difference. How can I compare the fscores of lavaan package?

Here is reproducible data and R code: 

 set.seed(0)

  D <- mapply(sample, 6, rep(18, 6), T) # DATA
  colnames(D) <- paste0("v", 1:6)

 library(lavaan)                     # MODEL
 m1 <- " f1 = ~v1+v2+v3              
         f2 = ~v4+v5+v6 "

fit1 <- cfa(m1, data = D)            # RUN lavaan

fscore <- data.frame(predict(fit1))  # Get fscores

round(c(f1.mean = mean(fscore$f1), f1.sd = sd(fscore$f1), f2.mean = mean(fscore$f2), f2.sd = sd(fscore$f2)), 3)

#> f1.mean   f1.sd   f2.mean   f2.sd   ## Notice both f1 and f2 have a mean of ZERO. 

    0.000    0.511    0.000    2.141  # How to compare f1 with f2

[Balal Ezanloo]

Hi

I think you have to run measurement invarince (MI) to compare your groups' means. for example look at https://groups.google.com/forum/#!searchin/lavaan/invariance$20test|sort:date/lavaan/O_dRFM2XtwM/WypNClyfBgAJ

[Balal Ezanloo]

Invarince can be done in different forms that are:

1- Configural Invariance

2- Weak Invariance

3- Strong Invariance

4- Strict Invariance

5- Invariance of the Factor Variance-Covariance Matrixes

6- Latent Mean Invariance                                                 # this is what you want

but i think you have to read more about its condition for example read appendix in this article https://www.tandfonline.com/doi/abs/10.1080/00220973.2013.876231

HTH

 

[Yves Rosseel]

The means of factor scores correspond to the (estimated) means of the
latent variables in the model.

In a single group analysis (which is what you seem to be doing), the
means for all latent variables are (by default) fixed to zero. As a
result, the factor scores also have zero means.

There are two things that puzzle me in your question:

1) if your goal is to compare latent means, you don't need factor scores
at all; you can estimate the latent means in your model directly

2) in a single-group, single-timepoint setting, the means of latent
variables are arbitrary, so there seems no point trying to compare them.

But perhaps we misunderstood your goal. Can you give us more background
on what you wish to accomplish?

Yves.