A few questions about DIF, negative LRT statistics

113 views
Skip to first unread message

Erin717

unread,
Oct 11, 2020, 9:25:37 AM10/11/20
to mirt-package
Dear Phil, 

Hope everything is going well!
I have a few questions about multiple group DIF analysis. I got some odd results that I could not figure out why.

Question 1: the model has mixed item types. Is it still okay to use the DIF() function? 

Here is what I did: First I estimated a fully constrained model,
spec.invar.full <- c(names(tam.dat[,-c(1:3,24)]), "free_var", "free_means")
mg.full <- multipleGroup(dat[,-c(1:3,24)], model=1, itemtype = c(rep("2PL",14), rep("gpcm",6)),  SE=TRUE,group=group, invariance=spec.invar.full,survey.weights=dat$wt,
                         technical=list(NCYCLES=2000))

Then I used the DIF() function 'drop' scheme to identify potential anchor items
dif.full <- DIF(mg.full, which.par = c("a1","d"), 
                scheme="drop",p.adjust="BH",technical=list(NCYCLES=2000))
However, because the item type is mixed, I tried the following too, 
dif.full.dich <- DIF(mg.full, which.par = c("a1","d"), items2test = 1:14,
                scheme="drop",p.adjust="BH",technical=list(NCYCLES=2000))
dif.full.poly <- DIF(mg.full, which.par = c("a1","d1","d2","d3"), items2test = 15:20,
                     scheme="drop",p.adjust="BH")
Do they look sensible?

Question 2: the LRT has negative statistics for all the dif models. 
For example: the result from dif.full.dich resulted large X2 statistics. Did I do somethign wrong here or does it mean the chi square test is wrong? It seems really odd to me. 


Question 3: Does it have anything to do with convergence issues? Because I got warnings that goes like "EM cycles terminated after 2000 iterations" and "Log-likelihood was decreasing near the ML solution. EM method may be unstable".

Question 4: I also tried to specify a configural model in which everything was freely estimated. Then I ran an anova() between the fully constrained model and configural model, it was statistically not significant. Does it mean there is no non-invariance? However, it is counter-intuitive because according to RMSD_DIF I did get quite some high RMSD values (e.g., > .30). Is it because of high df?
mg.config <- multipleGroup(dat[,-c(1:3,24)], model=1, itemtype = c(rep("2PL",14), rep("gpcm",6)), SE=TRUE,group=group, survey.weights=dat$wt, invariance = c("free_var","free_means")) ###may not be identified, just to have a look
anova(mg.config, mg.full)
      AIC   AICc    SABIC      HQ     BIC  logLik    X2  df   p
1 160.461 44.389  -51.887   7.360  87.259 -24.230   NaN NaN NaN
2 265.161 41.126 -151.949 -35.572 121.373 -22.581 3.299  54   1
The change of AIC looks substaintial, but small X2. Why would this happen?

Applogise in advance if these questions are naive. Thank you for all of your help!

Best,
Erin

Erin717

unread,
Oct 11, 2020, 9:28:40 AM10/11/20
to mirt-package
Sorry, looks like the screen shot to Question 2 was not attached successfully. I have attached it here.
DIF results.JPG

Thank you!

Phil Chalmers

unread,
Oct 26, 2020, 12:37:01 PM10/26/20
to Erin717, mirt-package
Hi Erin,

I can't quite see what's going on in the code without some data, so unfortunately I can only guess based on your responses (feel free to provide a reproducible instance of your problem). 

1. Input looks fine to me in terms of setting up an anchor-constrained model with user-defined weights. But I don't think the weights are being used in DIF(), which could explain the weird results (I'm skeptical that likelihood ratio tests should be used alongside survey weights anyway). 

2-3. I'd need to reproduce this first hand to give a more definitive answer, but it sounds as though it could be an estimation problem where the models do not converge to a reasonable MLE location. That can cause all kinds of weird problems.

4. That's correct, your configural model is not identified when you free the latent mean/variance. So comparing to the constrained identified model doesn't make much sense.

HTH, and sorry for the late reply.

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/750089e4-5f53-4e70-8489-71a0914357d0n%40googlegroups.com.
Reply all
Reply to author
Forward
0 new messages