ESEM Rotation PreviousNext
Mplus Discussion > Exploratory Factor Analysis >
Message/Author
 Madison Smith posted on Friday, May 08, 2020 - 8:31 am
I am trying to run a simple ESEM using two factors and the syntax below, and it is important that orthogonal rotation is used. However, I am getting the error "ROTATION=VARIMAX and ROTATION=PROMAX are only available for TYPE=EFA". Why is this? Is varimax rotation not available in ESEM?

MISSING = ALL (-999);
ANALYSIS:ROTATION = VARIMAX;
ESTIMATOR=MLR;
MODEL: f1 f2 BY y1-y100 (*1);
OUTPUT: STANDARDIZED RESIDUAL MODINDICES SVALUES TECH1;
 Bengt O. Muthen posted on Friday, May 08, 2020 - 9:54 am
Look at UG pages 678-681. Varimax has a more modern version which is obtained by saying:

Rotation = CF-Varimax(Orthogonal);
 Madison Smith posted on Friday, May 08, 2020 - 10:02 am
Thanks for this information! I did review pages 678-681, but was confused by the difference between Varimax and the Crawford-Ferguson Varimax rotations.

I ended up using GEOMIN (ORTHOGONAL) because I expect substantive loadings of my indicators on more than one factor. Do you think this is appropriate?

Also, my model is taking an exceptional amount of time to run. I added the line inegration=montecarlo (as I saw in the forums) to speed it up, but so far I haven't had any luck. Is this common for ESEM, or have I made a mistake somewhere?
 Bengt O. Muthen posted on Friday, May 08, 2020 - 10:17 am
Q1: It's also a good choice. You won't see much different I think.

Q2: How many dimensions of integration do you have (the screen printing and the output Summary shows it)? I assume you have categorical outcomes and many factors.
 Madison Smith posted on Friday, May 08, 2020 - 10:57 am
Thanks very much.

I'm not sure where to find dimensions of integration, I checked all the output and do not see it.

I am attempting to model either one or two factors and 101 indicators, 25 of which are continuous. The other indicators are a bit odd because they are aggregates of anywhere from two to seven 0/1 (binary) items, meaning that they can take on a range of values. But most often they will take one of 3, 4, or 5 discrete values based on their aggregation. I assume that ML or MLR can handle this because I also have continuous indicators, but I don't actually know that. And I'm not sure why it would affect the run time.
 Bengt O. Muthen posted on Friday, May 08, 2020 - 11:05 am
If you have only 1 or 2 factors, you have only 1 or 2 dimensions of integration. See the Summary of Analysis section of your output. With 1 or 2 factors, don't use integration = montecarlo; the default is probably faster. And use Proc=8.

Read about the different estimator choices and their pros and cons in the FAQ on our web site:

Estimator choices with categorical outcomes
 Madison Smith posted on Friday, May 08, 2020 - 11:28 am
Thank you. I anticipate fitting as many as 8 factors with the same indicators, so I was hoping that integration=montecarlo would make my task a bit less tedious.

I'm unfamiliar with proc=8; is there a resource on this I could consult?

From my reading of the article you suggested, it sounds like my use of ML with polytomous variables may be what is slowing me down, but that ML might still be more desirable because of some missingness I have in my data (that is probably not MCAR) and large number of observed variables.
 Bengt O. Muthen posted on Friday, May 08, 2020 - 2:07 pm
8 factors will be almost impossible with ML. Unless you have a very large sample, Bayesian estimation may be preferable. Bayes is as good as ML for missing data.

Regarding computations, see our Short Course video and handout for Topic 10, Part 1 on our website. With your heavy computations, you may want a computer with the fast i9 Intel CPU; see

http://www.statmodel.com/download/BayesProcBengt.pdf
 Madison Smith posted on Friday, May 08, 2020 - 2:22 pm
Thank you, it seems I might need to upgrade my computer to run the models that I want. I currently have an AMD Ryzen 3 3200U with Radeon Vega Mobile Gfx 2.60 GHz processor that I am working with.

I am using Goldberg's bass-ackwards approach to fit my models, so I'm not actually sure if I'll be fitting 8 factors, but that is my max. Do you think something like 5 would be feasible with ML (even if it does take a while to run)?
 Bengt O. Muthen posted on Saturday, May 09, 2020 - 7:55 am
That's very slow computer hardware.

You can try integration = montecarlo(5000). Both ML and Bayes speed will in your case depend on your sample size - the higher, the slower.
 Madison Smith posted on Monday, May 11, 2020 - 7:58 am
Thanks very much!
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: