|
|
 |
| LCGA for randomized control trials |
 
|
| Message/Author |
|
|
| Sisi Ma posted on Thursday, January 18, 2018 - 1:05 pm
|
|
|
Hi,
I am trying to identify latent trajectories from RCT data to uncover potential differential response to treatment. I would like to identify the latent class from all treatment arms combined(similar to the Uher 2010 study that you are a co-author on). So I can later identify the variables that influence trajectory membership with a separate model. One reason I want to do this in a two-step fashion is that I have a large number of candidate variables I want to examine, so I will probably use some sort of features selection+supervised learning.
I fitted a linear growth trajectory (I only have three time points). According to BIC and LO–MENDELL–RUBIN, my model resulted in a two class solution. One class have high intercept and small slope and the other have low intercept and large slope. Different treatment arms have different proportion of class 1 and class 2 individuals (I conducted chi^2 test between class membership and treatment arm, it is significant).
The difference in intercept among the latent class plus different proportion of class1 and class2 made me think that different arm have different average intercept, which should not happen since it is a RCT and the variable I am using is randomized.
Am I looking at this correctly? Thanks for your advise! |
|
|
|
|
| Perhaps you have 1 pre-treatment occasion and 2 post-treatment occasions. In which case you can define the growth intercept at time 1 and test if the intercept growth factor means are equal across treatment arms (as they should be in that case). |
|
|
| Back to top |
|
|
