Message/Author |
|
Olev Must posted on Wednesday, September 18, 2019 - 8:52 pm
|
|
|
I am a Flynn Effect (FE) researcher. This research means comparison of means of 2 cohorts. My current project involves an old study in years 1991 and 2001. FE comparison is possible across 9 countries. I have read several times the paper: Muthen and Asparouhov (2014) IRT studies of many groups: The alignment method. I tried to use analogy. I also defined 18 groups (9 x 2) and tried the alignment modeling. But unfortunately the share of non-invariant loadings and thresholds is high: more than 50%. This is much more than 25% as Muthen and Asparouhov proposed. I have here 2 sources of variance - across countries and across cohorts, therefore high non-invariance seems to be logical. I decided to follow the proposal for Monte Carlo simulation, but I met some problems. In the Supplementary material of the paper provide the Mplus input excerpts for Monte Carlo simulation. What is the source of used loadings and intercepts in this model? The corresponding values of loadings and thresholds from alignment model ouputs? Do I need repeat this model for each 18 groups ? What I should request for output to get the correlation between ordering of groups? Thank you very much for explanations! Olev |
|
|
You can find some more alignment montecarlo scripts here http://statmodel.com/download/Paper_Scripts.zip These are a part of this paper http://statmodel.com/download/RevisionSMnR.pdf For the montecarlo you will need to use the loading and threshold values from the original data run. You might find helpful the output:svalues option used in the original run so that when you setup the monetcarlo you can just copy and paste. Additional alignment output is available in output:alignment. If the sample size is large and you have plenty of indicators you might obtain good recovery in the montecarlo even with 50% non-invariance. Some things you might consider are - Is the factor model correct for each group, you can use estimato=wlsmv to check - is the alignment holding up better for 9 groups cohorts combined? If so maybe it is better to convert back to CFA following https://researchbank.acu.edu.au/cgi/viewcontent.cgi?article=10039&context=fhs_pub and using the cohort as a binary predictor - you can use the bayes estimator to estimate alignment if some residual correlations are needed - possibly explore BSEM as an alternative to alignment or in combination with alignment https://www.statmodel.com/examples/webnotes/webnote17.pdf |
|
Olev Must posted on Friday, September 20, 2019 - 1:18 am
|
|
|
Dear Tihomir, Thank you for advice! I tried to follow your suggestions and input files. The configural model across 18 groups was excellent (RMSEA = .043, CFI = .980, TLI =.972, SRMR =.038). But my model has some problems: *** ERROR One or more MODEL statements were ignored. These statements may be incorrect or are only supported by ALGORITHM=INTEGRATION. I use binary data. And the starting values are described by thresholds (for example [ t3$1*0.45809 ]) How to continue? Should I delete the reference on the thresholds? Greetings, Olev |
|
|
Send your output to Support along with your license number. |
|
Olev Must posted on Thursday, September 26, 2019 - 8:28 pm
|
|
|
Dear Mplus team, Is it justified to calculate the effect sizes (like Cohen d) between group means that are calculated in the alignment modelling? Olev |
|
|
Absolutely. |
|
Olev Must posted on Friday, September 27, 2019 - 9:07 pm
|
|
|
Thank you very much! Alignment in cohort and group comparisons is really excellent solution. Olev |
|
Back to top |