Control Variables & moderated Mediation

#1
Hello,

I'm testing a moderated mediation with two dependent variables, so I'm running two tests with PROCESS SPSS.

I have included 5 control variables in my experiment and have some questions about control variables.

1.) Are there any assumptions for using control variables (i.e. they have not to correlate each other....?)
2.) Are there any psychological requirements for using control variables (i.e. they have to correlate only with DV, but not with the mediator?)
3.) Do you have some literatures and references about this issue?

:) thanks for answering!
 
#2
Testing for mediated moderation is quite complex, as it requires both strong theoretical grounds and robust empirical support. Make sure you build on the appropriate literature. Let me clarify -- two dependent variables -- means you have two separate (univariate) models or are you talking about a multivariate model? Now, answering your questions:

1. All assumptions of your estimator pertain to your controls in the model (e.g., there should be no multicollinearity between IVs, etc.)
2. Typically, controls come from literature (or sometimes from common sense :). If controls are not part of your mediation/moderation test, then you shouldn't worry much about them. However, if they are, for the "required" (or better say "desirable") relationships between IVs and mediator/moderator -- see the appropriate methodological literature.
3. Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of personality and social psychology, 51(6), 1173.
Hicks, R., & Tingley, D. (2011). Causal mediation analysis. Stata Journal, 11(4), 605.
Preacher, K. J. (2015). Advances in mediation analysis: A survey and synthesis of new developments. Annual Review of Psychology, 66, 825-852.
 
#3
Hello @kiton, thank you very much!

Correction: I have to test a moderated mediation :)

1.) Checking for multicollinearity: I've put my IV (not the moderator), the 5 control variables and both DV in a linear regression model (in SPSS) and checked the VIF - all values were about 1. So I think there is no multicollinearity (or did I something wrong?)

2.) My Master Thesis is a pilot study - so there are not many explanations for using control variables. But I've measured them just in case. I have to check helping behavior and attitudes against helpers so I've decided to check for empathy, social distrust, social desirability and so on. I thought that these control variables have an impact on my DV. But there are no many empirical support.

3.) If I put all my control variables in PROCESS my moderated mediation model is non significant (no direct and indirect effects). Did I something wrong?

Thanks :)
 
#4
Hello @kiton, thank you very much!

Correction: I have to test a moderated mediation :)
Well, no matter how you put it, the complexity of testing for it remains as high. Note, you will have to run multiple models to "distinguish" -- i.e., find evidence of -- the effect(s) you are looking for. On top of that, you will have to run several other tests and analyses specifically on mediator and moderator effects (e.g., at minimum -- a Sobel test; and a test if there is a sig difference between the slopes of the regression lines).

1.) Checking for multicollinearity: I've put my IV (not the moderator), the 5 control variables and both DV in a linear regression model (in SPSS) and checked the VIF - all values were about 1. So I think there is no multicollinearity (or did I something wrong?)
Approaches on this might vary. A plausible one involves: (A) analysis of correlation table that includes all IVs (including controls) -- any correlation higher than .5 should be a suspicious one; (B) checking the overall (total) VIF -- should revolve around 2 (conservatively); and also run a condition number test -- should be around 5-10. With (C) comes a limitation of SPSS -- consider R, Stata, or SAS if you really plan on doing data analytics in future.

2.) My Master Thesis is a pilot study - so there are not many explanations for using control variables. But I've measured them just in case. I have to check helping behavior and attitudes against helpers so I've decided to check for empathy, social distrust, social desirability and so on. I thought that these control variables have an impact on my DV. But there are no many empirical support.
No empirical support or not much? :) 1-2 citations may suffice. Regardless of that, there must be solid theoretical basis to assume an impact of your controls on the DV. Also, controls' effect on the DV should be significant (otherwise, what's the purpose of controlling for them?). You can use, for example, theories to explain why and how your suspected controls impact the DV.

3.) If I put all my control variables in PROCESS my moderated mediation model is non significant (no direct and indirect effects). Did I something wrong?
Please clarify what you mean under this. You can state your model as an equation (or series of equations, I should say). Note, assuming you want to test for the desired effects, you'd have to run at least 3 equations to test for mediation (i.e., paths "a,b,c"), and then additionally run the model with interactions to test for moderation. I may need to refresh myself on the simultaneous testing of those effects, but you won't get away with a single model anyway :)
 

hlsmith

Not a robit
#5
I would refer you to the work of Tyler VanderWeele (Epidemiology/Biostatistician) at Harvard. He has written extensively on this topic, including the 4-way decomposition.