Adding a control increases the impact of focal predictor -- sign of what?

kiton

New Member
#1
Hello dear forum members!

I have recently been working on a mediation analysis, in which inclusion of an additional variable (say, C1) implies reduction of the effect of a focal predictor (X).

Now, moving further, I add a new variable into the model (say, C2). It's impact on Y is negative; however the impact of the focal regressor (X) on Y is increasing -- what is this indicative of?

Assuming inclusion of C2 decreased the impact of X1, I would suspect it acts as a suppressor (considering C2 has negative impact on Y). But I am a bit confused by C2 boosting X's impact. Note, there are no signs of multicollinearity. Predictors are correlated at .3 max (both, VIF and condition number tests are at minimums -- 1.5 and 3, respectively).

Your comments are greatly appreciated :)
 

Karabiner

TS Contributor
#2
A suppressor suppresses the impact of a predictor on another variable.
If you incorporate the suppressor in your predictive model, you control
for that effect, and the impact of your predictor on the DV is no longer
suppressed by the suppressor.

With kind regards

Karabiner
 

hlsmith

Omega Contributor
#3
Kiton,


It is usually best to also try and draw out a causal graph to visualize the relationships and temporal ordering of variables. This helps to process relationships.
 

kiton

New Member
#4
Kiton,
It is usually best to also try and draw out a causal graph to visualize the relationships and temporal ordering of variables. This helps to process relationships.
Thank you for the comment, hlsmith. Let me elaborate here a bit more. Initially, following Selig and Preacher (2009), I constructed and tested the attached cross-lagged panel mediation (CLPM) model. Path a is positive; Path b is negative (direct [total] effect of R -> Y decreased, therefore, I concluded "suppression"); a>b (i.e., proximal mediation); c'(indirect effect) was found significant for some years, but very small in effect size (I used kappa squared introduced by Preacher and Kelley(2011)).

Now, a colleague of mine introduced another variable -- a substitute for my current M variable (mediator) in the attached model. It behaves similarly, but with a major difference that the impact of R in a controlled model increases almost two times. And I am trying to understand the nature of this situation.