LGM in intervention research
Message/Author
 Arun Karpur posted on Tuesday, November 21, 2006 - 1:26 pm
I am researcher working at the Florida Mental Health Institute, and I used to work with Hendricks Brown's PSMG group a couple of year ago.

At this time I am facing a conceptual issue with LGM and its application in intervention studies. Firstly, I am new to LGM as a first line user. I went through Curran and Muthen (1999) article to understand the application of LGM in intervention studies and in this, the example is regarding one control and one treatment group.

At FMHI we are desiging a study to compare three treatment conditions (one control and two intervention)and we hypothesize that treatment two > one > control. We plan to LGM as an analytical method for the study and I wanted to know how could we apply LGM in this context? or at least how could we examine ordinal treatment effects?

I'd really appreciate your guidance with this issue.
 Bengt O. Muthen posted on Tuesday, November 21, 2006 - 4:56 pm
You can approach it as a 3-group analysis in line with Muthen-Curran, where you simply check if the treatment effects are ordered as you hypothesize. But the Muthen-Curran approach is really aimed at treatment-baseline interactions and is overly complex if such interactions are not hypothesized. A simpler approach is to have 2 dummy covariates for treatment1 against control and treatment2 against control and then let those dummy variables influence the slope after treatment starts.
 Arun Karpur posted on Wednesday, November 22, 2006 - 4:13 am
Thank you so much Dr. Muthen. Just to get this correctly, you recommend that in the LGM in addition to the intercept and slope, I can include two additional growth parameters - each representing the dummy variables. Or you are indicating that I can include these as covariates predicting the intercept and slope parameters for all the groups put together? Again thank you so much for educating me with this issue.
 Johnny Wu posted on Wednesday, November 22, 2006 - 8:14 am
here's an idea:

you could do a multiple group growth curve model with control = 0, tx1 = 1, and tx2 = 2.

hence, you could estimate a separate intercept and slope for each condition. further, you can regress the intercepts and slopes on intake variables and see if the effects are different across your treatment groups.
 Linda K. Muthen posted on Wednesday, November 22, 2006 - 10:01 am
Bengt suggested two options. The first option is multiple group analysis with three groups: control, treatment 1, and treatment 2. The second option is using two dummy variables to represent the three groups and regress the slope growth factor on those two dummy variables. There are no additional growth factors needed.
 Arun Karpur posted on Wednesday, November 22, 2006 - 11:36 am
Thank you so much for the recommendations.
 Raheem Paxton posted on Monday, December 08, 2008 - 4:35 pm
Hi folks,

I'm interested in determining if change in physical activity is associated with change in depression scores in a randomized control trial. I know that a stepwise multiple group approach (Muthen & Curran, 1997) is the way to go, but not quite what items need to contrained across groups. Please tell me what I am missing in my model. Oh by the way the distributions are non-normal and the analysis is taking a great deal of time to run. Thanks

grouping is rand (0=cont 1=trt);

model:
i s | cesd1@0 cesd2@1 cesd3@2;
i2 s2 | PA1@0 PA2@1 PA3@2;

[cesd1 - cesd3] (1);
[pa1 - pa3] (2);
[i@0];
i(3);
s(4);
i2(5);
s2(6);
i with s (7);
i2 with s2 (8);
[s] (9);
[s2] (10);

model control:
[s] (9);
[s2](10);
 Linda K. Muthen posted on Tuesday, December 09, 2008 - 8:47 am
I see one omission in your input. That is [i2@0]; See slides 69-70 of the Topic 4 course handout for further information.
 Raheem Paxton posted on Tuesday, December 09, 2008 - 11:23 am
Thanks for your response. Several questions: (1)Is it necessary to impose all constraint (2) If I'm exploring the relationship between growth of X on the growth of Y in the intervention and control group, how do you add the treatment effect. Slides 69-70 are only measuring the growth of 1 variable, but not the regression of one factor on another.
 Linda K. Muthen posted on Tuesday, December 09, 2008 - 11:28 am