|
|
 |
| Multigroup model with TSCORES |
 
|
| Message/Author |
|
|
| Sanne posted on Thursday, July 02, 2020 - 7:41 am
|
|
|
Hi,
We are trying to run a multigroup multivariate growth curve model with the TSCORES option to account for biases in the growth factors related to the variation in age in our sample. However, we have some trouble finding the right way to modify our model to test for group differences appropriately. Modification indices do not seem to be available with the TSCORES option.
We have now tried to first run a fully unconstrained model. Then we tested a series of models by fixing parameters to be equal across groups, starting with parameters that showed the smallest differences between groups in the previous model. We continuously compared each model to the previous model using AIC and BIC. This way, we identified some parameters that differed between groups. However, we get the impression that the parameters that we can release depend on the order in which the parameters are constrained, so this does not seem like the most reliable method of constraining paths. Should we try a different approach? What would be a good approach to test for group differences between the different parameters in this case?
Thank you for your response! |
|
|
|
|
I assume you have looked at the approach of UG ex 6.18.
Yes, the order of restricting the models matter. There is no best way, except start from the end that is most likely closest to the best-fitting model. That is, either all unequal or all equal. Likelihood-ratio chi-square difference testing or Wald testing using Model Test can be used. |
|
|
| Sanne posted on Tuesday, July 14, 2020 - 4:19 am
|
|
|
Thank you for your response. We used the approach of ex 6.8, not 6.18, i.e., the TSCORES option.
When you start with the 'all equal' model, how do you decide which parameter to release? The reason why we started with the 'all unequal' model is because it allows you to see which parameters have the largest difference between groups. Would you take the parameters of the 'all unequal' model to decide which parameter you would release? |
|
|
|
|
| Looking at largest parameter estimate differences might be misleading. If you don't get modindices starting from all equal, you can still release one equality at a time and create likelihood-ratio chi-square differences (2 times the logL difference). And you can do this also when starting from all unequal. Where to start is always hard to say - start at the end which you think is closest to the best model. |
|
|
| Back to top |
|
|
