CFA not converging (correlated factors, overlapping items)

General help and assistance with jamovi. Bug reports can be made at our issues page: https://github.com/jamovi/jamovi/issues . (If you're unsure feel free to discuss it here)

by Marco Ventura » Sat Nov 21, 2020 4:42 am

Hi

I am working on a research paper using a questionnaire with 237 items, grouped in 15 factors (developed using an EFA), and now we want to do a CFA, for publishing the paper, but Jamovi got stuck (2,266 people in the dataset). All 15 factors have overlapping items (items that belong to more than one factor). My computer has 8 GB RAM (Win 10) and I am using Jamovi stable version 1.2.27.

Could anyone please help me with some directions on how to proceed? Maybe a larger dataset is needed to be able to perform a relevant CFA, or there is a technical issue I am missing?

Thank you very much
Last edited by Marco Ventura on Fri Nov 27, 2020 9:50 am, edited 2 times in total.
Marco Ventura
 
Posts: 4
Joined: Sat Nov 21, 2020 4:22 am

by jonathon » Mon Nov 23, 2020 10:44 am

hi,

could you just clarify what you mean by "stuck"?

i don't know much about factor analysis, but hopefully someone else will be able to chime in.

cheers

jonathon
User avatar
jonathon
 
Posts: 1483
Joined: Fri Jan 27, 2017 10:04 am

by Marco Ventura » Mon Nov 23, 2020 11:01 am

Hello Jonathon

Well, after I introduced the factors (15), with 50-100 overlapping items each, the report on the right side either displays as working and working without any results being displayed (I left it for a few hours one day), or it displays the message "Iavaan ERROR: fit measures not available if model did not converge". I even tried to allocate all the items in one factor, to see how Jamovi behaves on the same sample of 2266 people, and it displayed an error related to something like the vector has 24 GB or so.

So, I am wondering maybe the RAM of the computer I am using is not enough, or maybe I did something wrong in the analysis configuration. Or, maybe the sample is too low, although it should be enough according to the various websites I read, but even if the sample it too low, then I suppose JAMOVI would have to calculate at least other statistics, beside fit measures? The file has retained the factor allocation, so I thought maybe someone could open the file and see if Jamovi delivers to report, or what happens.

Thank you
Marco Ventura
 
Posts: 4
Joined: Sat Nov 21, 2020 4:22 am

by mcfanda@gmail.com » Wed Nov 25, 2020 2:17 pm

Hi
I think it is a problem of convergence, not a hardware issue. You should check if the model is identified. Try a model with uncorrelated factors, one item per factor, no correlated errors. If it works, your hardware is fine. Free a few secondary loading at the time, and see how far you can go.
User avatar
mcfanda@gmail.com
 
Posts: 217
Joined: Thu Mar 23, 2017 9:24 pm

by Marco Ventura » Wed Nov 25, 2020 7:51 pm

Hello

Yes, it's convergence it seems. The software works well. Thank you.

Now, about convergence, to test if the sample size is not enough, I duplicated the sample just to see if a larger sample would solve the convergence issue, but even using a dataset of 8,400 entries the model did not converge.

Hmm. Any ideas on how to proceed? Looks like these overlapping items and highly-correlated factors are hard to analyze using standard settings of CFA in Jamovi.

Perhaps I should talk to a professional in statistics to play around with the data, or there is some setting in Jamovi that could facilitate the converge in this situation?

Thanks
Marco Ventura
 
Posts: 4
Joined: Sat Nov 21, 2020 4:22 am

by jonathon » Wed Nov 25, 2020 9:48 pm

i don't really know much about convergence (hopefully marcello weighs in again), but we do use the lavaan package to perform our CFA. you may find it helpful to reproduce this analysis in R with lavaan, and ask the lavaan community about it. these sorts of analyses are their bread and butter.

cheers

jonathon
User avatar
jonathon
 
Posts: 1483
Joined: Fri Jan 27, 2017 10:04 am

by Marco Ventura » Thu Nov 26, 2020 11:57 am

Hello

Well, I discussed my dilemma on the Lavann group, and it seems that we need a dataset of 28,000+ to get a really relevant CFA analysis using standard procedures, for this type of complex analysis we are looking for.

Thank you very much
Last edited by Marco Ventura on Sat Nov 28, 2020 10:23 am, edited 1 time in total.
Marco Ventura
 
Posts: 4
Joined: Sat Nov 21, 2020 4:22 am

by jonathon » Thu Nov 26, 2020 11:44 pm

ah cool. glad you could figure it out.

cheers

jonathon
User avatar
jonathon
 
Posts: 1483
Joined: Fri Jan 27, 2017 10:04 am


Return to Help