0
$\begingroup$

I have a question concerning the inclusion of control variables into my research. I have data from an experimental, randomly distributed social protection program. I want to test whether a moderating variable influences the relationship between treatment and outcome. Also, I want to test for other confounds. However, the experimental design should eliminate the effect of all third variables, right? But in this case, I want to focus on the moderating variable, not the treatment itself. Is it ok to include control variables or is it rather a mediated moderation?

Thanks, m

$\endgroup$
  • $\begingroup$ Can you please write down the models that you are considering explicitly? Also please provide more details regarding what kinds of variables you want to include in your model. These details would make it much easier to give some helpful advice. :) $\endgroup$ – StoryTeller0815 Jul 30 '19 at 13:14
  • $\begingroup$ The treatment variable is the reception of a social protection program (cash transfer etc.), the outcome variable is 'political participation', and the moderating variable is 'traumatic war experiences' I was planning to test for other third variables, for example 'risk preferences' Thanks for your quick answer $\endgroup$ – Malte Hiekel Jul 30 '19 at 13:18
  • $\begingroup$ Yes, but how many categories does your treatment have? Do you assume that the other variables are continuous? Do you assume normality? (It is extremely important to have such information to judge the complexity of your situation.) And the moderator is observed not manipulated, right? $\endgroup$ – StoryTeller0815 Jul 30 '19 at 13:20
  • $\begingroup$ The treatment variable is a dummy variable. The outcome variable of 'political participation' is standardized (z-score) and the moderating variable is continuous. And yes, the moderator is observed. Does that help? $\endgroup$ – Malte Hiekel Jul 30 '19 at 13:25
1
$\begingroup$

It is not "forbidden" to enter further control variables to a model with interaction terms. It just makes the model larger (more complex). What is referred to as mediated moderation or moderated mediation is just a certain linear model (see e.g. here). Whether this model represents your theoretic beliefs can only be judged by you but I strongly recommend the books by Andrew Hayes on that.

Independently from the exact model you are aiming for, you can enter further variables. (If you see it from a path model perspective, you could even decide which of the variables you want to control for your control variables.) You only slightly change the interpretation of your model in that all coefficients of the moderated describe the imaginary case that the other variables are held constant.

You mainly need to consider this when you illustrate your interaction effect. That is, when you plot your conditional regression coefficients you should also state which level of the control variables you conditioned on (typically "mean in all control variables").

Hope this helps - feel free to ask further clarification questions so I can optimize my answer.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you so much for your detailed answer! Also for the literature suggestions. Mainly, my supervisor for the master thesis was concerned that I cannot simply include control variables into my model, because of the experimental design. She said that the experimental design possibly eliminates all effects of control variables. However, some baseline characteristics could influence the outcome and the moderator. Therefore, she suggested it might be a form of mediated moderation.. $\endgroup$ – Malte Hiekel Jul 30 '19 at 14:13
  • $\begingroup$ If your randomization works properly, the effects of the control variables on your treatment variable are controlled for - that is, your treatment variable does not correlate with anything but the outcome. However, it may still have power advantages to include the control variables because you control for some unlucky randomization. There are some simulation studies on that in the context of ANCOVA. $\endgroup$ – StoryTeller0815 Jul 30 '19 at 14:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.