[Mediation-information] Appropriate use of package "mediation"

[Tormod Bøe]

Dear Kosuke and Dustin,

Thanks for your prompt and thoughtful replies. In my case there is 
evidence of a linear "social gradient" in that symptom scores of outcome 
increases with decreasing levels of SES. However, the association with 
outcome is stronger in the most extreme categories of "treatment" (i.e. 
for "very poor" economy and for "basic" parental education levels).

If I am understanding your responses correctly, this suggest that the 
results of the mediation analysis would be different for different 
levels of treatment (e.g. for comparing "poor"[=0] with "average"[=1] 
vs. "poor"[=0] with "very good"[=4] economy).

Yours sincerely,
Tormod

On 15.09.2011 04:27, dustin tingley wrote:
> Hi Tormod-
>
> To follow up, we do allow you to set two values of the treatment other
> than 0 or 1. Make sure you're using the latest version of the package
> (3.1), in R 2.13. As Kosuke mentions, it won't matter when everything is
> linear, but when things are non-linear, then the ACME from 0 to 1 might
> different from a 3 to 4 change. This is typical of any time you use
> probit/logit etc. due to the nature of the response function in those
> models.  Also, the sensitivity analysis, is only worked out for the t=0
> and t=1 case. Note that things like MPlus won't calculate sensitivity
> analyses...Keep in mind the strong assumptions you are making when
> conducting mediation analysis. For us, that means running a sensitivity
> analysis if possible, or thinking about alternative research designs.
>
> best,
> dustin
>
>
>
> Dustin Tingley
> Government Department
> Harvard University
> http://scholar.harvard.edu/dtingley
>
>
>
> On Wed, Sep 14, 2011 at 10:19 PM, Kosuke Imai <kimai at princeton.edu
> <mailto:kimai at princeton.edu>> wrote:
>
>     Hi Tormod,
>
>       I'm ccing the listserv we have so that my collaborators and others
>     can give additional insights.  The mediation software doesn't have
>     the functionality to fully accomodate the continuous treatment.
>       However, when the software will calculate the average causal
>     mediation effect of changing the treatment from 0 to 1.  If you use
>     a linear model for both mediator and outcome, then this should be
>     sufficient because the ACME would not change regardless of the base
>     value you choose for the treatment.  If you have a nonlinear model,
>     however, the ACME you get is still valid but only for the scenario
>     where you change the treatment value from 0 to 1.  If you are
>     thinking about different scenario (e.g., changing the treatment from
>     3 to 4), then you will get a different answer.
>
>     Best,
>     Kosuke
>
>     Department of Politics
>     Princeton University
>     http://imai.princeton.edu
>
>     On Sep 14, 2011, at 8:24 AM, Tormod Bøe wrote:
>
>      > Dear Associate Professor Imai,
>      >
>      > I am currently doing a PhD at the University of Bergen. We have
>     conducted a large-scale population based study of childhood mental
>     health.
>      >
>      > I am in the process of writing at paper where we investigate to
>     which extent sleep problems mediate the association between parental
>     socioeconomic status and children's mental health problems.
>      >
>      > We have two sleep measures (one rated on a three point scale, and
>     two that are binary[0=adequate sleep time, 1=long/short sleep
>     time]), parental SES is continuous (years of education and family
>     income rated on five point scales) and outcome is continuous (range
>     0-30).
>      >
>      > I have made some attempts at running mediation analyses using
>     your R package mediation. I have made one model m where I regress
>     family SES on sleep problems, and model y where I regress family SES
>     and sleep problems on mental health problems outcome. I use sleep
>     problems as "mediator" and have used SES variables (e.g. family
>     economy) as treatment variables. However, is this an appropriate use
>     of your package?
>      >
>      > I am especially concerned with regards to the use of a continuous
>     "treatment" variable. The models run without any errors, and the
>     pattern of results (with regards to significant interactions)
>     resemble those I obtain when running path analyses in Mplus.
>      >
>      > If you could please provide some advice in this manner it would
>     be greatly appreciated.
>      >
>      > Yours sincerely,
>      > Tormod Bøe
>
>     _______________________________________________
>     Mediation-information mailing list
>     Mediation-information at lists.r-forge.r-project.org
>     <mailto:Mediation-information at lists.r-forge.r-project.org>
>     https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/mediation-information
>
>