| Message/Author |
|
|
|
|
Dear forum,
I am checking if an ECEC instrument can be validated across countries. For this I want to use 2 analysis steps:
1. An invariance check if the same dimensions load on the same factors for both countries (no other constraints)
2. An invariance check if the magnitude of the relationships between dimensions and factors is equivalent across counties (factor loadings are constrained)
Model fit seems okay for step one, yet there is a Heywood case (negative variance) for one of the items - for only one country. Thus when I want to continue to step 2, it seems I need to constrain the Heywood case manually (instead of just using MODEL = configural metric scalar) for both countries.
Questions:
1. Do I have to constrain the Heywood case for both countries?
2. If so, is this the right way?:
TEACHER BY Zposeng Zteacom;
PEER BY Zpeersoc Zpeercom Zpeerass;
TASK BY Ztaskeng zselfrel;
CONTROL BY Zteacon Zpeercon Zbehcon;
Zteacom@0; !heywood case
MODEL NL:
[TEACHER-CONTROL@0];
[Zposeng-Zbehcon];
Zteacom@0; !heywood case
Thank you, |
|
|
|
|
1. No
2. Yes |
|
|
|
|
Dear Bengt Muthen, thank you for the fast reply.
Why are the models allowed to differ? I assumed that with any invariance testing one should keep the models the same as much as possible in terms of other constraints and parameters (apart from the specific factor loadings constraints for example). |
|
|
|
|
Dear Bengt Muthen,
In the meanwhile I ran into some other questions:
1. See previous post
2. Somehow the factor loadings are not the same for the groups if you see the standardized model results, that does make sense right? The factor loadings of the model results are the same though.
3. The model does not fit too well, see below. Say I want to define partial metric invariance and let the factor loadings of the CONTROL latent factor be freely estimated for both countries, how should that be defined? (thus the part of CONTROL BY Zteacon Zpeercon Zbehcon;).
4. Would you know an example of a results section on how to report on (partial) configural/metric invariance?
5. I experience some difficulties with calculating the Satorra-Bentler Chi square. Is there an example output of which it is calculated?
RMSEA: 0.139
CFI: 0.879
SRMR: 0.084
Satorra Chi Square/TRd: still to be calculated |
|
|
|
|
1. Restrictions to residual variances don't affect the invariance testing of loadings and intercepts very much if at all.
2. Standardized loadings will be different because factor variances differ across groups.
3. Just say e.g.
MODEL NL:
CONTROL BY Zpeercon Zbehcon;
You need to leave out the first loading. We explain all this in our Short Course video and handout on our website, Topic 1.
4. There is the early paper by Byrne et al. For more sources, ask on SEMNET.
5. See the left margin of our home page regarding Chi-square diff:
http://www.statmodel.com/chidiff.shtml |
|
|
|
|
Thank you.
1. Good to know.
2. Makes sense, of course.
3. This didn't work for me, but then I also forgot to follow the modification indices. That's what I did now, and starting from weak factorial invariance (the metric model) I added two WITH statements based on the modification indices. Unfortunately, I got another heywood case (n.s), for one variable of a added WITH statement. I kept the WITH statement though; does that make sense? And what does two heywood cased I have to account for about my model?
4. Thanks a lot. If necessary I will use SEMNET, thanks for the reference.
5.To what chi square do I compare the current (final) partial metric invariance model with? The configural one?
Thank you, |
|
|
|
|
3. Q1: Yes. Q2: Question suitable for SEMNET.
5. Both the full metric and the configural would be of interest. |
|
|
|
|
| Clear. I posted the questions in SEMNET. Thank you greatly for the support. |
|
|
|
|
Dear Bengt,
Another question: when using modification indices for a configural model, I end up with 3 modification suggestions. One for Country 1, two for Country 2. I cannot find how I undo the added WITH statement for Country 2 if I add a covariance for Country 1 (ZBEHCON WITH ZSELFREL), see the arrow in the syntax below.
MODEL: !Model for Country 1
TEACHER BY Zposeng Zteacom;
PEER BY Zpeersoc Zpeercom Zpeerass;
TASK BY Ztaskeng zselfrel;
CONTROL BY Zteacon Zpeercon Zbehcon;
[TEACHER-CONTROL@0];
Zteacom@0; !heywood case
ZBEHCON WITH ZSELFREL; ---> this one automatically is also set for Country 2, how do I make the exception for Country 2?
MODEL NL: !Model for Country 2
TEACHER BY Zteacom;
PEER BY Zpeercom Zpeerass;
TASK BY zselfrel;
CONTROL BY Zpeercon Zbehcon;
![TEACHER-CONTROL@0];
[Zposeng-Zbehcon];
Zteacom; !to undo heywood case
ZBEHCON WITH ZPEERASS;
ZPEERASS WITH ZPEERSOC;
Thank you. |
|
|
|
|
| What you call "! Model for country 1" is actually the Overall model for all countries which you can then modify in country-specific model parts. So if you don't want a WITH statement in a country, just fix the WITH at zero. |
|
|
|
|
Thank you Bengt, that worked.
Kind regards |
|
|
| Back to top |
