Control variables in longitudinal design
Message/Author
 J_F posted on Sunday, January 26, 2014 - 2:34 am
Dear Professor,
I’m running a longitudinal design with two latent constructs: “IG” and “IS”. I have two measurement points T1 and T2 (IG_T1,IG_T2 and IS_T1, IS_T2) and the hypothesis is that IG_T1 predicts IS_T2 (even when controlled for IS_T1) and that IS_T2 itself is predictive of IG_T2.
The syntax is as follows:

IG_T2 on IS_T1 IS_T2 IG_T1;
IS_T2 on IG_T1 IS_T1;

The problem is that I get a very unexpected negative path between IG_T2 and IS_T1, even though the path between IS_T2 and IG_T2, the path between IS_T1 and IS_T2 and the path between IG_T1 and IS_T2 are positive. There’s no logical reason why IS_T1 should have a different effect than IS_T2 and I wonder whether it is correct to account for the effect of IG_T2 on IS_T1 when the same construct at the second measurement point (IS_T2) is part of the equation.
I would appreciate any help. Thank you very much.
 Linda K. Muthen posted on Sunday, January 26, 2014 - 10:38 am
This could be caused by multicollinearity. Are ig and sg at time one highly correlated. Are the two slope growth factors highly correlated.
 J_F posted on Monday, January 27, 2014 - 12:43 am
The correlation between IG and IS at T1 is about .65 and the correlation between the two slope factors is almost the same. If this is a problem of multicollinearity, would it be rational to exclude the paths between IG_T2 and IS_T1 when the equation accounts for IS_T2?
Thank you very much for your help.
 Linda K. Muthen posted on Monday, January 27, 2014 - 12:27 pm
It's hard to say what to do. The coefficients are partial coefficients so the sign may not be the same as the sign of the correlation.