Item response theory with more than one latent variable

[Emma Zandi]

Hi!

I want to construct 5 latent variables from a set of 20 items, which are responded on a 7-score likert scale I was wondering if I can do that with -irt- or it should be done through gsem?

The items and their associated latent variables are pre-decided based on the theory, however, I do not know how can I specify the -irt- command to recognize that I want 5 variables each based on 5 determined items?

The reason that I am reluctant to use gsem is that when I specified the model in gsem, it took 2 days, and it was still on iteration 3.

Thank you
Emma

[Weiwen Ng]

No, IRT will only support one latent variable at a time (i.e. a unidimensional model). If you have 5 latent variables that you believe to be unidimensional, and each item loads on only one latent trait, then maybe you won't go too far wrong if you fit 5 separate IRT models.

[Marcos Almeida]

Taking two days for a gsem model to reach iteration 3 and... quite a long road to converge (or not)?

Maybe you should wish to read this thread.

Generally speaking, there might as well be a problem with the modeling itself, instead of the fact of having chosen the gsem approach.

[wbuchanan]

How large is the sample? There are some examples in the gsem manual of how to fit IRT models using the gsem commands. Under the hood, the irt commands are all using gsem anyway. I’d first suggest testing to see whether or not a single higher order factor fits the data better than a set of orthogonal latent variables. If it does, you can justify fitting a single unidimensional IRT model to all the items.