gibbon lab posted on Thursday, May 12, 2011 - 12:48 pm
Hi Dr. Muthen,
I was running an SEM with some control variables. There are two structual equations in this model:
y on x1 x2 z1-z8; U on x1 x2 z1-z6;
where y is the manifest dependent variable, U is the latent dependent variable, x1 and x2 are the two main manifest predictors of interest, z's are the manifest control variables. The results of the above model look fine. But when I add two more manifest control variables (say w7 and w8) in the second equation (with U being the dependent variable), it completely changes the path from x1 to y (from near positive significant to a negative path). How can adding control variables in one equation changes the other equation so much? This puzzles me because in a regular regression, the coefficients in an equation will not change if the control variables for that equation do not change. It is unrelated to other equations. Thanks a lot.
When you add predictors to only one equation your model says that those added predictors do not influence the DV in the other equation at all. And if they actually do, your model is misspecified and the estimates not dependable. Also make sure that the residual covariance is free between y and u.