1
$\begingroup$

I‘m doing mediational analysis on clustered (AKA nested/multilevel) data. The question I have is not specific to mediational analyses but applies to any kind of multilevel analysis.

I‘m assessing whether the effect of SES (socioeconomic status) on an outcome is mediated by another variable. SES is a continuous individual predictor and the other two variables are continuous (Likert scores transformed to percentages). This data comes from schools so the nested structure is children within classes within schools. Meaning that the analysis needs to account for the possible similarities in responses of children who are in the same school and in the same class.

The problem is that since the schools were selected on the basis of the SES status of the neighbourhood, the school and the individual SES predictor are correlated. This means that modeling the effect of school does not seem to make sense in this case (as it would presumably dramatically reduce the effect of the predictor). Would it be safe here to simply ignore the school level of the hierarchy and model the effect of class membership? Any advice would be greatly appreciated!

$\endgroup$
1
$\begingroup$

I would say that this sounds like an additional variable on the causal path.

NSES -> SchoolSES -> M -> Y

(Where NSES is neighorhood SES, M is mediator, Y is outcome). So I agree with you that you should not add NSES as a predictor to the model, but you might think about analyzing SchoolSES as a mediator of the effect of NSES (although that adds a lot of complexity to the model, and I probably wouldn't.)

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Good idea. Done. $\endgroup$ – Jeremy Miles Jan 30 '17 at 17:23
  • $\begingroup$ Maybe I can rephrase and explain my problem more clearly. As I understand it, hierarchy potentially introduces dependency in the data but the levels are not typically directly relevant to the predictor(s) or the hypothesis. But in this case I would expect the order of one of the levels (school) to correlate strongly with the predictor (the individual SES) and thus both of these to correlate in the same way with the outcome. So it seems that modeling the effect of school by allowing the intercept to vary would incorrectly weaken an expected correlation between the predictor and the outcome. $\endgroup$ – Eldur Jan 30 '17 at 21:06
  • $\begingroup$ Yes, that's what I was trying to say. $\endgroup$ – Jeremy Miles Jan 30 '17 at 21:26

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.