mediation analysis

[Sandra Virgo]

Hello there

My team are interested in doing a causal mediation analysis using data from the Demographic and Health Surveys.


Specifically we are wondering if we can find out how much of the effect of Region on whether or not a woman delivers her baby in a facility (or not) is actually mediated by:

• education
• wealth
• urban/rural residence
• highest level of facility in the cluster
• religion
• ethnicity
• level of obstetric risk
• other....

Surveying the different mediation packages on Stata, this is where we're currently at:

medeff: this won't work as it can't handle multiple mediators
khb: this can handle multiple mediators but as it doesn't do a sensitivity analysis this could be a problem
ldecomp: this looks appealing, but again, no sensitivity analysis
paramed: it needs a continuous or binary mediator, and things like religion or ethnicity are not that kind of variable
gformula: have been reading about this but seems very complex, though possibly right for us

Conceptually we are also thinking of the fact that many of our mediators above could actually affect the region someone lives in (i.e. the mediator and the exposure - Region - could affect each other bidirectionally). Therefore is mediation analysis even possible with what we have? We did a Directed Acyclic Graph to clarify our thinking, but it got very complex.

Do you have any advice?

Thanks very much!
Sandra

[Doug Hemken]

You might consider using sem/gsem.

[Sandra Virgo]

Hello Doug
Yes, the thought of doing that had crossed my mind as well. I am trying to remember whether it can cope with bidirectionality.

Thanks for the thought.

Sandra

[wbuchanan]

SEM can handle non-recursive relationships, but to identify the causal effect you would still need to have an instrument for that equation. Based on the limited information available it isn't clear what specifically your goal is and/or what the model would look like. Additionally, if you are trying to model a combination of mediators measured on numerous measurement scales you may want to reconsider a more parsimonious model for the sake of interpretation of the decomposed effects. The model is also made more complex given the levels at which the mediators do/do not vary. Based on the goal that is listed in the post above it seems like you are more interested in whether or not the effect of region varies across the different values of the "mediators" which would be a "moderated" relationship.

[Dick Campbell]

I'm not sure that mediation is how you want to think about this. Disregarding bidirectionality, a mediation analysis assumes that the exogenous variable(s) cause the mediators. In your case, it's hard to see how, for example, region can cause religion. In fact, I would argue that the regional differences you see reflect compositional effects of one kind or another, e.g. regional differences in the distribution of ethnicity or whatever. Perhaps you simply want to look at the behavior of coefficients reflecting regional differences as you introduces on or more of the variables you have listed as mediators. In any case, you want to think about whats really exogenous here.

[Sandra Virgo]

Hello all - thanks for all your very helpful comments and apologies for my delay in responding.

We have now decided to scrap the mediation analysis and just do this comparing a model with just region and the response variable to a model including region and everything else. So keeping it simple.

In one of our datasets we do have evidence that the woman has lived in that region all her life, so there is temporal precedence (notwithstanding influences before birth). But I take your point that a mediation analysis probably isn't appropriate.

Thanks very much.

[Dick Campbell]

You may want to think about decomposing between region differences. Ben Jann's oaxaca program might be quite useful for this. You can get Jann's Stata Journal article on the method free at http://www.stata-journal.com/article...article=st0151.