Using MIRT results for SEM

[Martin Lacher]

Hi everybody,

I did some extended research about this question but didn't find a good answer: Is it ok to use MIRT results (factor scores) as input variable for (classical) SEM (lavaan) regression analysis?

Background:

In my educational (pre-)study I measured several psychological scales via IRT-methods (mostly GRM) and got valid models and factor scores. I also measured an learning achievement test which I evaluated with MIRT, too. Now I want to know which psych factor has the biggest impact on learning achievement - a classical SEM problem, isn't it? But can I use my IRT data directly for the lavaan-model or are there some caveats to consider?

I am also glad if you can refer me to some appropriate literature.

Thanks a lot in advance,

Martin

[Bo]

Why not just use SEM? It can deal with ordinal items. 

[Felix Fischer]

Hi,

a measurement model within a SEM and an IRT model is basically the same. if you have an established IRT measurement model, you can use it to estimate scores and then use these scores in a path model. the downside is that than your model assumes the latent scores to be true, ignoring their (known) error. there are ways to incorporate that as well, eg by imputing plausible values. another approach would be to run a latent regression in mirt and fixing the itemparameters to your values before. or you could reestimate the measurement parameters within a SEM. that might make the most sense if you're not entirely sure whether results of your pilot study are robust.

i think there are many roads to go here and each have pros and cons - depends a lot on your study.

best, felix