modification indices

[Makrem Ben Youssef]

Hi phil

sorry for my stupid question, but is there any way to get modification indices in mirt package as it is the case with Mplus or Amos

tanhks a lot for your help!

Makrem

[Phil Chalmers]

In theory yes, but no mirt does not have this feature (and to my knowledge, no IRT software has this feature because MIRT models are not that cheap to estimate over and over again). I'll think about adding this feature in at some point, but for now it doesn't exist. Cheers.

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.
For more options, visit https://groups.google.com/d/optout.

[Phil Chalmers]

Thanks for the tip, Ryne. I was under the impression that one had to re-estimate the model in a similar way to profiled-likelihood confidence intervals. I'll try to learn a bit more about how SEM handles modification indices and try generalizing that method to IRT. 

I think modification indices would still have their use as MIRT models are often analogous to confirmatory factor analysis models (when the rotational indeterminacy is controlled) but for categorical indicators. So if a model is formed to have a relatively simple structure, in the Thurstonian sense, then cross-loadings could be freed one at a time to test their contribution. I suppose bifactor models can have some use here too, though perhaps to a lesser extent. Cheers.

Phil

On Wed, Apr 29, 2015 at 12:00 PM, Christopher Ryne Estabrook <resta...@northwestern.edu> wrote:

Models being too expensive to iteratively re-estimate is actually where modification indices come from. MIs theoretically allow you to avoid unnecessary estimations. Whether they do so in practice is a question for the literature.

A barrier to implementing modification indices within an IRT framework is selecting the additional parameters to test. My understanding of the relationship between MIRT models is that there are rarely comparisons between models that differ by only one parameter across all items, instead adding additional traits (and thus additional discrimination parameters) or additional item parameters, each of which would add multiple parameters at once. What single parameters would you consider adding to an existing MIRT model?

ryne

----

Ryne Estabrook, Ph.D.
Assistant Professor
Department of Medical Social Sciences
Northwestern University
633 N. St. Clair St., 19th Floor
Chicago, IL, 60611

Phone: 312-503-1459
E-mail: resta...@northwestern.edu

[Makrem Ben Youssef]

Thank you Phil for your quick answer, i'm using mirt for confirmatory analysis purpose and my model dosen't fit well (RMSEA= 0.097, CFI =0.75 and TLI= 0.75) , i tried all the possible alternatives for the specification of my model but it was the best that i can get. I know that for binary non-normal data the cutoff for adequate fit  is about 0.089 (Olivares and Joe,2014), that's why I'm really confused , could I report these values and accept the model or is it just unacceptable!!!?

[Phil Chalmers]

That's a question that has to be answered outside of the statistics. Your model fit isn't terrible by any means, but of course it could be better. If this is the first study using your test, or the first time MIRT has been fit to the data, then these results are probably just fine. Worse case, they provide a benchmark for improvement for others to reference and improve upon. 

As an aside, you could also check out where the residual variances are the largest with the M2(mod, residmat=TRUE) argument. These are standardized residuals used to compute the SRMR statistic, and so large values indicate a lot of covariation remaining between the items. Might help to illuminate why the fit isn't closer perfect in the second-order tables. Cheers.

Phil

--

[Makrem Ben Youssef]

Hi Phil

Thank you for the argument and the possible interpretations of the results, that's really helpful !!! even if it's a newly developed instrument , i'll try to find high covariation and try to enhance the value of SRMSR.

Thanks again for your support, I'm really grateful

Have a nice day

Makrem

Le mercredi 29 avril 2015 15:09:06 UTC+2, Makrem Ben Youssef a écrit :