Hello, Working on a project using MIMIC to examine DIF across a set of dichotomous variables loading on a latent construct with a dichotomous covariate. Ran into a few questions I'm hoping someone can help with...
1) Can/should sampling weights be included? Our statistician says no because IRT usually ignores them. He gets very different/unrealistic numbers when running my input only when he includes sampling weights and we can't figure out why.
2) It seems that including an interaction terms between the covariate and the latent variable is essentially the same as testing non-uniform DIF despite the frequency with which I see publications assert MIMIC can only test uniform DIF. Am I off-base in this assumption or am I missing something?
3) Is there a way for me to generate the ICCs across levels of the covariate (differing in both difficulty and discrimination) broken down across levels of the covariate? I thought the Plot2 option did so, but appears not.
4) If no to #3, can I do it manually in excel? Its easy enough to plot uniform DIF using the provided thresholds as difficulty parameters for the reference group and adding the direct effect value to get it for the focal group. I think that should work, but I'm unclear how I'd handle the discrimination parameter (loading value + direct effect from interaction term?).
1) You can certainly include sampling weights. If the results are strange, perhaps the weights are not the right ones to use.
2) You are right. This possibility is less known and was opened up with the introduction of XITH.
3) Perhaps try Plot3 and the "Adjusted..." plot option.
4) You can do a plot of the probability conditional on both the covariate value and the factor value, varying them in their suitable ranges.
Jason Payne posted on Wednesday, August 22, 2018 - 5:10 pm
Hi Bengt/ Linda,
Is there anywhere I can find a worked example of a 2PL IRT with non-uniform DIF. I understand that with the introduction of XWITH, it is possible to specify the latent variable interaction. But I am just looking for a worked example to confirm that my own code is on the right track.
Do you mean non-invariant loading/discrimination? Yes, XWITH should work fine for that. Although I have tried it, I don't have an example handy. Send your code to Support along with your license number and we will check.