Message/Author 


I wanted to see if either of you thought the following was a defensible alternative for modeling baselinetreatment interaction effects: I am working with a dataset where there are three intervention conditions that are effectcoded into two variables. But I did not want to recreate group indicators externally (out of laziness) in order to set up a multiplegroup, additive treatment model in order to model baseline tx interaction effects (as in Muthen/Curran 97 or Khoo 01). With the availability of type=random for interaction effects between latent variables, I was thinking that the following code would give you equivalent results to the triedandtrue BTI approaches: cva cvb on int2 int3; cva with cvb; s  cvb on int2; s on cva; s@0;[s@0]; r  cvb on int3; r on cva; r@0;[r@0]; where cva and cvb are growth parameters, int2 and int3 are the effectcoded variables for the 3 groups, and s & r are the random effects capturing the extent to which the intervention effect(s) vary across cva (estimated initial status on the outcome). Does this look reasonable? 


I'm afraid I don't understand this approach. First you define a random slope s for cvb on int2, and then you regress s on cva  what does that mean? I would take the more transparent approach. 


I was thinking that with this setup, the main tx effect (cvb on int2) could vary as a function of initial status (cva)  analogous to the regression of Bt on Ac you illustrate in Figure 5 of MC 97 under the additional growth factor framework. But I'll stick with the more transparent approach as you suggest...... 


I see what you mean now. I am not sure how well that would work. 


I compared it to MC97 and the estimates are not close enough for me to be confident that it works (though I would have made the same inferences on all parameters across either framework). BTI using type=random might be intuitive  but probably not mathematically equivalent to MC97.... 

Back to top 