Starting values for mixture IRT models

[Shelley L.]

I have a couple questions about the mirt package please:

1)      I’m fitting a mixture graded response model. If I use GenRandomPars=TRUE, then the group membership changes for different seeds, and the findings don’t appear that stable. In Mplus one can allow for many random starting values to avoid being stuck at local maxima. Is there something equivalent in mirt?

2)      How are starting values determined if GenRandomPars=FALSE? Is it recommended to use GenRandomPars=FALSE?

Thanks for your help,

Shelley 

[Phil Chalmers]

Hi Shelley,

Mixture models can be tricky, and are often prone to local minimum issues. When you fit these models do the final log-likelihood values differ? If so, you likely have local min problems, and should consider what you are seeing (even when picking the 'most likely' model) with some skepticism as there are multiple good ways to fit the same data under the same model. From my experience whenever the mixture model has difficulty converging it is because the mixtures are not sufficiently identifiable, and though they may exist the data does not allow for their structure to be stably detected. And yes, the GenRandomPars=TRUE argument is effectively the same as what MPlus does in the same model.

For your question 2), you can pass pars = 'values' to see how the starting values look, and if you like feel free to modify this data.frame of values and re-supply it to the pars input to use something you believe is better. HTH.

Phil

--
You received this message because you are subscribed to the Google Groups "mirt-package" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mirt-package...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mirt-package/3f771374-2775-4ecd-8b3d-85a05cd38e1en%40googlegroups.com.

[Shelley L.]

Great, thank you Phil!