QUESTION Re Mediation Analysis (part conceptual/part STATA related)

[Tara Jones]

I previously posted about this but no one answered my question. I am reposting as I am hoping someone out here will be able to provide an answer or at least an opinion. Please note, I will also cross post at the other stata forum, CV.


I have 3 variables, T (independent variable), M (mediator), Y (outcome), all of them binary.

I fitted the following:

Y <- T + covariates (coeff for T is positive and significant)

M<- T+ covariates (coef for T is negative and significant)

Y <- T + M + covariates (coeff for T is positive and significant; coefficient for M is negative BUT not significant)

Now, following Baron and Kenny, I would conclude M is not a mediator of the Y and T association, since M is not "significant".

Here are my questions:

1) Does the rule by Baron and Kenny (that both T --> M and M-->Y are significant) apply in the case when you have a binary outcome and mediator?
2) Do the associations have to be significant at the 5% level or can I use a different cut-off? (the p-value for the M--> Y association = .13)
3) I ran a mediation analysis, using Stata's medeff, and it said that there was 12% mediation by M. Is this just a nonsense value that I should ignore? Or is there actually mediation, even though M is not signifcantly associated with Y?

I hope someone can provide some insight.

Thanks

TJ

[Clyde Schechter]

I think you may have gotten no response because mediation is a thorny and controversial concept, and people may have been reluctant to wade into the swamp. Many statisticians reject the Baron and Kenny framework of mediation entirely. Some would argue that your model needs to be estimated as a (generalized) structural equations model, with estimation of the direct and indirect paths. You don't say anything about the design of the study that led to your data, but some would further argue that absent appropriate randomization of both T and M nothing at all can be said about mediation.

Different approaches to mediation reign in different disciplines (and in some disciplines no one approach has general acceptance and controversy reigns.) I think you would be best advised to consult with experts in your discipline about what is likely to be regarded as credible by your intended audience.

[Tara Jones]

Thank you so much for your response. I found it very helpful. As regards the design of my study, it is a cross-sectional study. I realize that that is a problem, but I don't want to go into this here. Further, my treatment isnt a "treatment", it is gender, so not sure how that impacts on what you said about not being able to say anything about mediation (since there is no randomization).

My study is an epidemiological study, (does M mediate the association between gender and Y), and I used causal mediation analysis (Imai et al) using Stata's medeff to do the analysis. As I understand it, causal mediation analysis is based on generalized SEM's and doesnt rely on the Baron and Kenny framework.

From what you wrote, I now wonder whether I am mistaken in combining these two different frameworks (B & K on the one hand, and Causal Mediation analysis on the other) in trying to make sense of my results? Should I just rely on the results of medeff? In "causal mediation analysis," does the mediator not have to be "significantly" associated with the outcome for there to be mediation?

Well, if I ignore the B& K framework, and just use what medeff tells me, then my results are as follows (see bolded lines).

Effect Mean [95% Conf. Interval] ACME1 .0189176 -.005018 .0438004 ACME0 .0160403 -.0041128 .0379756 Direct Effect 1 .1274808 .0903369 .1646811 Direct Effect 0 .1246035 .0863965 .1633032 Total Effect .1435211 .1134255 .1722113 % of Total via ACME1 .1313853 .1098513 .1667854 % of Total via ACME0 .1114022 .0931433 .141418 Average Mediation .017479 -.0045654 .0409513 Average Direct Effect .1260421 .088528 .1640466 % of Tot Eff mediated .1213937 .1014973 .1541017 So, 12 % mediaton. Except, "Average mediation" is not significant. How would one interpret this result?

As regards asking an expert in my field, unfortunately I have no access to any experts in this area. Hence my post here.


Thanks for your help!!

TJ

[Tara Jones]

Sorry about the formatting, here are relevant numbers:

Effect Mean [95% Conf. Interval]
----------------------------------------------

Total Effect .1435211 .1134255 .1722113

Average Mediation .017479 -.0045654 .0409513

Average Direct Effect .1260421 .088528 .1640466

% of Tot Eff mediated .1213937 .1014973

[Tara Jones]

% of Tot Eff mediated .1213937 .1014973 .1541017

[Nick Cox]

Please post the exact URL of any cross-posting. No reason to make any interested person Google for answers elsewhere.

[Tara Jones]

https://stats.stackexchange.com/ques...ation-analysis

[Tara Jones]

I just noticed my question was burried in that mess of numbers:


Effect Mean [95% Conf. Interval]
----------------------------------------------

Total Effect .1435211 .1134255 .1722113

Average Mediation .017479 -.0045654 .0409513

Average Direct Effect .1260421 .088528 .1640466

% of Tot Eff mediated .1213937 .1014973 .1541017




So, 12 % mediaton. Except, "Average mediation" is not significant. How would one interpret this result?


thanks again!