2
$\begingroup$

I did a survey, using many intervall-scaled items representing several constructs. I want to use these constructs to predict an outcome (decision: yes/no).

Thus, I got (latent) predictors to explain a manifest, categorial outcome.

Can you suggest an adequate SEM-framework? I considered using a latent class analysis. Is this the right way to go?

$\endgroup$
2
  • 3
    $\begingroup$ Couldn't you just take the estimated values for the latent predictors and use these as predictors with a logistic regression approach? $\endgroup$
    – Stijn
    Jan 18, 2016 at 12:55
  • $\begingroup$ Thanks for this advice, this sounds like something you can rely on! $\endgroup$
    – Jens
    Jan 19, 2016 at 11:03

1 Answer 1

2
$\begingroup$

Latent class analysis assumes that the manifest variables are categorical, but the latent is continuous. In your case, it's the predictor that is categorical. So latent class analysis does not apply to the SEM portion of your model.

You could use a latent class analysis on the measurement piece, by modeling the ordinal manifests as such. People often pretend that the likert type items (or whatever you have) are "normal" and proceed in the usual way. But you could use latent class analysis here, especially if the number of options for each manifest is small (say, 3).

Latent class analysis requires a larger sample size to work than the usual SEM model, because it has more work to do. Basically, you assume that each ordinal manifest represents a cut-point from an underlying normal distribution, and you need to estimate those cut-points. I sometimes find that more difficulties emerge from estimating the additional parameters of the latent class model, than by applying normal theory assumptions to ordinal data.

But back to your situation. I agree with @Stijn. I would estimate the factor scores from the predictor variables and throw them into a logistic regression model.

$\endgroup$
1
  • $\begingroup$ Sounds plausible! $\endgroup$
    – Jens
    Jan 19, 2016 at 11:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.