Latent variable analysis

[Chiara Tasselli]

Good morning,

I am trying to combine two variables ( measured ability & ability-self concept) to predict an unobserved variable (expectation of success: we have no data on that but, in our model, it originated from the two previous ones ( measured ability & ability-self concept)). I was thinking of working with latent variables but I am a little bit in trouble founding some reference guidelines. I was checking for some guide or literature online but the things that I found most are latent class analysis or working with SEM, but I don't think are what I need isn't it?

Can I have some suggestions about where I can look for a guideline or how to perform it??

Many thanks in advance for your time.

PS: an extra question: if measured ability & ability-self concept are indexes based on a set of sub-variable is better to introduce them in the model as a single indicator or as all their sub-component?

[Felix Bittmann]

You can either work in the SEM framework and run all your analyses there or you can generate a new latent variable for other analyses. For SEM, there are many examples to find. For generating a new var this would look like:

sem (new -> var1 var2 var3), latent(new)
predict newvar, latent

See also https://www.stata.com/manuals/sempredictaftersem.pdf

Potentially you could use all original variables and not the generated ones. But of course they should work together. You can use a correlation matrix, Cronbachs Alpha or a factor analysis to check this beforehand.