Latent Change Score BY GROUP

46 views
Skip to first unread message

Iris Menu

unread,
Feb 19, 2019, 4:27:17 AM2/19/19
to lavaan
Hi everyone!

I'm trying to fit my data on R with lavaan according to Roger Kievit's codes (https://osf.io/ch3rf/)

I have one latent variable, COG, combination of to cognitive scores X1 and X2, and another variable NEU.

I know that all the measures are impacted by gender AND by the type of training (control vs. test).
I created a variable 'Groupe' with 4 modalities : Control-Male, Control-Female, Test-Male and Test-Female.

Is it enough to make the analyses by group (i.e. adding the argument group in the fit command, see below)?
Or should I do a model by gender and another one by type of training (i.e. am I allowed to combine these two variables?)?

Then, should I compare the results to another model to make my conclusions?

Thanks in advance,

Iris Menu.

fitBYGGterULCS <- lavaan(terULCS, data=modelData, estimator='mlr',fixed.x=FALSE,missing='fiml', group='Groupe')
summary(fitBYGGterULCS, fit.measures=TRUE, standardized=TRUE, rsquare=TRUE)

Terrence Jorgensen

unread,
Feb 19, 2019, 4:58:47 PM2/19/19
to lavaan

I have one latent variable


Then make sure the indicators have invariant measurement parameters across groups.

Is it enough to make the analyses by group (i.e. adding the argument group in the fit command, see below)? 

Or should I do a model by gender and another one by type of training (i.e. am I allowed to combine these two variables?)?

Then, should I compare the results to another model to make my conclusions?


Multigroup models can only take one grouping factor, so yes, this is the way to estimate models independently across groups of different factors (sex and condition).  Assuming you want to compare the mean of the latent change score across groups, the model that places no equivalence constraints on the means across 4 groups is equivalent to a regression/ANOVA model that estimates both main effects and their interaction (i.e., each cell in the design has a uniquely estimated mean).  You can compare this model to a model in which the means are constrained to equality across all 4 groups.  If that model comparison is significant, you could then label the means in the unrestricted syntax, and test any particular pairwise comparison you like by defining differences as parameters in your syntax.  

Terrence D. Jorgensen
Assistant Professor, Methods and Statistics
Research Institute for Child Development and Education, the University of Amsterdam

Iris Menu

unread,
Feb 20, 2019, 8:18:02 AM2/20/19
to lavaan
Thank you a lot for the answer, it's really helping!

When you say that I should compare my model to another model in which the means are constrained, can I do this by adding 'group.equal="means"' in argument? And then testing an anova(model1,model2) to conclude?

Best regards,

Iris Menu, second year Master's degree student.

Terrence Jorgensen

unread,
Feb 21, 2019, 3:33:17 PM2/21/19
to lavaan

can I do this by adding 'group.equal="means"' in argument?


That constraints all latent intercepts.  I would recommend using labels to constrain exactly the parameter(s) you want to constrain in the syntax.

Reply all
Reply to author
Forward
0 new messages