Multigroup invariance testing with mi... PreviousNext
Mplus Discussion > Confirmatory Factor Analysis >
 Jen posted on Thursday, April 10, 2014 - 10:36 am

Using a multigroup model, I am testing invariance across two groups where

3 items are shared
2 items are assessed only for Group 1
3 items are assessed only for Group 2

I obviously am only interested in the invariance for the shared items, but want the latent construct to be informed by the additional info available in both groups.

Is there a workaround for this? Mplus does not like items to be missing for everyone in a group.

My current workaround is to generate random values from a N(0,1) for the missing items and fix the loading to 0, intercept to 0, residual variance to 1 for those items. Not sure if this makes sense or not.

I am interested in loadings and the factor variance in addition to means and intercepts, so the MIMIC model is not a good option.

Other ideas?

Thank you!
 Bengt O. Muthen posted on Friday, April 11, 2014 - 6:07 am
There is a FAQ called Different number of variables in different groups" that suggests what to do. Another approach is to use Type=Mixture, letting the groups be represented by Knownclass.
 Jen posted on Wednesday, April 16, 2014 - 5:41 am
Hi Bengt,

Thank you for referring me to that note. TYPE=MIXTURE seems problematic because of the lack of fit indices, but with continuous outcomes, setting the residual variances to .0001 (along with supplying some starting values) seemed to work well.

I wondered if you had any suggestion about what to do when the variables missing in one group are ordinal (resulting in a "less than two categories" warning). Do I do something with the thresholds?

Thanks again!
 Bengt O. Muthen posted on Wednesday, April 16, 2014 - 12:11 pm
There is a * option shown in the UG on pages 543-544.
 Justin posted on Tuesday, November 22, 2016 - 2:35 am

I am running a Monte Carlo simulation fitting a factor model to two groups with sparse item overlap, similar to that described above. The two groups have mostly unique sets of (categorical) items generated from a single factor with a couple of overlapping items, so some items are observed in all individuals but other items have blocks of missing data flags.

I used the KNOWNCLASS option with TYPE = MIXTURE. When I run the external Monte Carlo script on 1000 generated datasets, I get the message "FATAL ERROR: Class 1 has 0 observations." However, this shouldn't be the case, and the error message appears for different repetitions of the simulation, e.g., the first run I received the error for the 10th dataset; without changing anything, I ran the simulation again and it completed 150 repetitions before giving me the error.

If you have any suggestions for getting this model to run, it would be helpful. Thanks in advance.
 Bengt O. Muthen posted on Tuesday, November 22, 2016 - 5:32 pm
Please send the 2 outputs with differences for your 2 runs to Support along with your license number.
 Bengt O. Muthen posted on Wednesday, November 23, 2016 - 11:06 am
You can generate the data in a first step using the multiple group approach which holds the numbers in each group fixed. In a second step you do an "external, step 2" Monte Carlo run (see e.g. UG ex12.6 step 1 and 2) where you use Mixture Knownclass.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message