1
$\begingroup$

I am conducting a mediation analysis (model 4) in SPSS using the PROCESS macro by Hayes.

I have cross-sectional survey data and want to test the following relation:

  • Childhood adversity (X) -> Current loneliness (M) -> Current psychosis (Y)

I want to add three covariates to the analysis: age/sex/depressed feelings.

PROCESS gives me the option to add these covariates in models of both M and Y, and Y only (and M only, but not interested in that). I know for sure that I want to add these covariates to the model with Y, as I think that the covariates might influence the relation between X and Y, and between M and Y. However, I could also think of some reasons why they might influence the relation between X and M – I am not necessarily interested in this relation, but they could influence the estimate of the mediation effect.

My question is: should I only add my covariates to the model with Y as this was my a priori reason to include the covariates, or should I also add them to the model with M as there might be some relationship there which would be good to control for? (I.e., are there good statistical reasons to include covariates to both M and Y models, rather than Y only?)

$\endgroup$

1 Answer 1

1
$\begingroup$

You should definitely add (some of) those covariates to the X -> M relationship. You ideally want to control for all possible confounders of the X -> M relationship to estimate the X -> M path ina na unbiased way. You clearly have not collected enough variables to do that (e.g., SES, family structure, neighborhood characteristics, etc.), but you can get closer to a correct path by including the variables that you have that make sense. (Note: without controlling for these variables, you cannot estimate the effect of childhood adversity on loneliness; the association between the two could be due to any of the confounding factors I mentioned or others. I would be very hesitant about interpreting your findings as mediation rather than association).

Sex may cause both adversity and loneliness. Current age does not cause childhood adversity, but it acts as a proxy for year born, which might cause childhood adversity. Either way, both of these variables are fine to include in X -> M. Depressed feelings (if measured currently) cannot cause childhood adversity; rather, adversity causes depressed feelings. This would be inappropriate to include in the X -> M model. Rather, it's pretty clear that depressed feelings could be another potential mediator between childhood adversity and current psychosis or even a consequence of current psychosis. Where it belongs in the model is unclear, and how you include can wildly change your estimates and make them uninterpretable. See my answer here for why this is such a problem.

Basically, and I know this is a larger conclusion than what you asked for, you have not collected enough data or the right kind of data to perform the analysis you described and interpret any of the relationships as mediation. You have not collected enough variables to isolate the effects. You have not clearly delineated the temporal order of your variables (i.e., depressed feelings, loneliness, and psychosis are all concurrently measured so the causal directions are entirely unclear). Even the associations you find may not be validly interpretable as partial associations if you condition on outcomes.

$\endgroup$
4
  • $\begingroup$ (1/2) Thank you very much for your elaborate reply. I agree with your points concerning other potential covariates and temporal order of the relations, which are important to consider. Keeping these points in mind, but ignoring them for a minute so that I can ask my follow-up question, I’d like to come back specifically to the mediator depressed feelings which you mentioned, as that was the one I am mostly concerned about. $\endgroup$
    – Bru
    Commented Jun 18, 2018 at 9:05
  • $\begingroup$ (2/2) You are right that depressed feelings cannot cause childhood adversity. Yet, am I correct in thinking that in the X->M relation, adding a covariate controls for the shared variance between these variables and does not necessary “assume” a causal relation? I could think of a model where childhood adversity (X) increases depressed feelings, which may subsequently make feelings of loneliness (M) more likely to occur. In this case, would you still recommend that I add depression as a covariate to the X->M relation? Or does the covariate also have to influence X for this to be acceptable? $\endgroup$
    – Bru
    Commented Jun 18, 2018 at 9:05
  • $\begingroup$ It would not be appropriate to include depression in the X -> M model. The covariate must not be included by X if it is to be included in the X -> M model. Essentially the model you propose is a more complicated mediation model, where depression mediates the effect of X on M. Including depression in the X -> M model means your coefficient for X in that model corresponds only to the direct effect of X on M, but for your larger primary mediation model, you need the total effect of X on M. $\endgroup$
    – Noah
    Commented Jun 18, 2018 at 16:46
  • $\begingroup$ This problem has been called "overcontrol bias" in the literature. The primary psychology literature on mediation (Preacher, Hayes, Baron, Kenny) has not discussed these issues in enough detail and has been primarily focused on modeling and intuition rather than identification and causal assumptions. See Valeri & Vanderweele (2013) in Psych Methods for a good introduction to a more rigorous perspective on mediation. $\endgroup$
    – Noah
    Commented Jun 18, 2018 at 16:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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