[Mediation-information] Multilevel analysis for complex surveys

[Kentaro Hirose]

Hi Angel, 

> I don't understand why in med.fit the term related to the clustering units is (1|SCH_ID), while (1 + attachment|SCH_ID) is the term in the out.fit.

This is just fitting a varying-intercept model for the mediator and a varying-intercept-and-slope model for the outcome. You could also fit two varying-intercept models for both the mediator and outcome. And similarly, you could also fit two varying-intercept-and-slope models for both the mediator and outcome. It's all up to your theory. 


> My first question is if I should use mediation with glmer.

If your dependent variable is binary and you want to use a multilevel model, then you should use glmer. 


Best,
Kentaro Hirose

Postdoctoral Fellow
Department of Politics
Princeton University



________________________________________
From: mediation-information-bounces at r-forge.wu-wien.ac.at [mediation-information-bounces at r-forge.wu-wien.ac.at] on behalf of Angel Rodriguez [angel.rodriguez at matiainstituto.net]
Sent: Tuesday, August 26, 2014 12:49 PM
To: mediation-information at r-forge.wu-wien.ac.at
Subject: [Mediation-information] Multilevel analysis for complex surveys

Dear subscribers,

I want to carry out a mediation analysis with variables collected in a survey (for more information, see the thread "causally dependent mediators"). It was a complex survey with strata and individuals clustered in census sections. I do not have census section variables, but I want the models to be aware that individuals are not independent because they're grouped in census sections.

My first question is if I should use mediation with glmer.

In that case, my second question would be about the syntax to carry out this. In the example in Tingley's "mediation: R Package for Causal Mediation Analysis", the proposed models are:


R> library(lme4)
R> set.seed(2014)
R> med.fit <- glmer(attachment ~ catholic + gender + income + pared + (1|SCH_ID),
+ family = binomial(link = "logit"), data = student)
R> out.fit <- glmer(fight ~ catholic*attachment +
+ gender + income + pared + (1 + attachment|SCH_ID),
+ family = binomial(link = "logit"), data = student)

I don't understand why in med.fit the term related to the clustering units is (1|SCH_ID), while (1 + attachment|SCH_ID) is the term in the out.fit.

Thank you very much,

Angel Rodriguez-Laso
Research Project Manager
Matia Instituto Gerontologico