topic Re: structural equation modeling (confirmatory factor analysis) problem in Discussions

structural equation modeling (confirmatory factor analysis) problem

lujc07 — Fri, 09 Jun 2023 00:40:10 GMT

Hi! I did a confirmatory factor analysis with three latent variables. However, there are some estimate problems. Some standardized coefficients are missing (eg. latent2 > DO/Temp; latent1 <-> latent2). One standardized coefficient greater than 1 (latent3 > SoftSed, 1.012). All variables are transformed and standardized, but some are still skew. I am guessing if this is the reason. I attach some analyses results. Could anyone give me the reason why this happen? Thanks.

Re: structural equation modeling (confirmatory factor analysis) problem

lujc07 — Sat, 16 Oct 2021 13:13:26 GMT

Also for variables DO and Temp, they have missing data, but each of them only have two observations missing compared with totoal 75 observations.

Re: structural equation modeling (confirmatory factor analysis) problem

LauraCS — Mon, 18 Oct 2021 17:21:59 GMT

Hi @lujc07,

The reason for the missing values on the diagram is because the variance of the latent variable is negative --variances shouldn't be negative so we don't try to give you a standardized estimate because it won't make sense. Also, it's important to check the convergence status of your model prior to examining the results; if the model fails to converge, the results are not interpretable. In your image, I see convergence is a problem:

Convergence failure can happen for a number of reasons but in this model I have two concerns:

1) It appears some of your latent variables might be nearly orthogonal (see, for example, the correlations between the Latent3 indicators and all other variables from the multivariate output), so the 2-indicator latent variables might not be estimable and this can lead to problems.

2) The negative correlation of DO and Temp could be making things harder.

My suggestions to address these concerns are:

1) You could change the specification of the 2-indicator latent variables, such that you fix their variance to 1 and then set the two loadings to equal:

For consistency, you could free the first loading of Latent3 and also fix its variance to 1.

2) You could try "reverse-coding" one of those two variables so that the correlation is positive. For example, for Temp, you could go to the column in the data table, right-click on it, and select New Formula Column > Transform > Negation. This will create a new data table column that's reverse coded and you can use it in place of Temp.

I hope this helps your model converge to a proper solution.

Best,

~Laura

Re: structural equation modeling (confirmatory factor analysis) problem

lujc07 — Fri, 22 Oct 2021 18:40:16 GMT

Hi Laura,

Thanks for your answer. This really solved my problems.