Tue, 03/08/2011 - 12:38

Hi,

I ran a piecewise latent growth model with 9 timepoints and the knot at the 6th timepoint. The model converged normally but the chi-square, p and rmsea were all NA. Is there any reason this should be the case or am i missing something?

Regards,

Yonghao

This is intentional, assuming you're using raw data. Those statistics depend on comparison to a saturated model, which in SEM is most commonly a model with totally free means and covariance matrices. When you input a covariance matrix or covariance matrix and means as data, then that saturated model is easy (read: so quick you don't notice that we do it) to do automatically, so we do it for you. When you have raw data with missingness, then it can take some time to estimate the saturated model under FIML. We don't do it for you so your model won't take longer. This also means that you can run the saturated model once, then run all of the comparison models you want without redoing any work.

You can specify the saturated model yourself and add it to the summary of your fitted model. The summary function has an argument for SaturatedLikelihood; just supply the fitted saturated model to that argument and fit stats like RMSEA and the chi-square will be populated.

Hi, Your post is from March 2011. Is this still the way to proceed? Or do more recent versions of OpenMx compute these fit indices for me? -- M

This is still the case, and is an intentional feature of OpenMx. If you want fit stats for raw data models, you'll have to provide the saturated model you want your model compared to. We intentionally don't make assumptions about your models, leaving you to specify exactly what models you want to compare.

If you have complete data (i.e., no missingness), you'll get fit stats automatically when you analyze a covariance matrix (and means). We've also added support for users to add their own specified independence models for stats like CFI (independenceLikelihood and independenceDoF arguments in the summary function).

Dear Ryne,

Thank you for your reply. I have also just gotten the answer by searching through the forum and reading earlier posts. Should have done my homework before i post my question. I do not have any missing data so i guessed inputting the covariance matrix and mean vector would be much more convenient.

Once again, thanks a lot.

Regards