# [Mediation-information] Dear Dr. Imai:

Kosuke Imai kimai at Princeton.EDU
Wed Sep 21 03:45:58 CEST 2011

```It's the latter, and in the scenario below, A is more sensitive than B.  Be sure to look at this docs for examples and interpretations: http://cran.r-project.org/web/packages/mediation/vignettes/mediation.pdf

Kosuke

Department of Politics
Princeton University
http://imai.princeton.edu

On Sep 20, 2011, at 12:19 PM, Glenn Walters wrote:

> Dear Dr. Imai:
>
> I've read through all of the papers you recommended and just when I start thinking I understand sensitivity analysis I come across something that makes me question whether or not I really do understand it.  Do the R^2 values represent the proportions of original variances explained by the unobserved confounder OR do they represent the magnitude at which there is no longer a mediation effect (i.e., mediation = 0).  Let me put it this way, from the sensitivity results listed below for A and B, which mediational effect is more robust?
>
> A
> rho at which AMCE = 0 of .06
> R^2_M*R^2_Y* at which ACME = 0 of .003
> R^2_M~R^2_Y~ at which ACME = 0 of .003
>
> B
> rho at which AMCE = 0 of .12
> R^2_M*R^2_Y* at which ACME = 0 of .02
> R^2_M~R^2_Y~ at which ACME = 0 of .02
>
> Thanks.
>
> Glenn
>
>
> >>> Kosuke Imai <kimai at Princeton.Edu> 5/2/2011 9:55 AM >>>
> Dear Glenn,
>
>    First, let me point you towards a few papers we've written on this
> topic.   The software is designed to implement the methods proposed in our
> paper: http://imai.princeton.edu/projects/mechanisms.html
>
>    The easiest paper to read is the working paper titled "Unpacking..."
> though this one is targeted to political scientists.  Given your
> background, the Psychological Methods paper may be the one you might want
>
>    Next, we have just released a new version 3.0 to CRAN with lots of
> additional functionalities.  We also have a mailing list where you can
>
> http://r-forge.r-project.org/projects/mediation/
>
>
> On Mon, 2 May 2011, Glenn Walters wrote:
>
> > Dear Dr. Imai:
> >
> > I just started using your causal mediation program for R and had a few
> > questions.
>
>   Have a look at this vignette, which is an updated version of our Springer
> chapter:
> http://cran.r-project.org/web/packages/mediation/vignettes/mediation.pdf
> This details the software and its use with examples.
>
> > 1. I noticed that the independent variable in all of your examples was
> > some form of treatment in which participants were randomly assigned to
> > conditions.  Is it legitimate to use your program with a non-manipulated
> > independent variable like history of mental health problems?
>
>   Yes in principle but of course in observational studies it's important to
> adjust for pre-treatment variables so that the treatment and control
> groups are comparable.
>
> > 2. What do the output values for the mediation effect, direct effect,
> > and total effect represent?  I am thinking that this represents amount
> > of variance explained ( R-square) but I wanted to make sure.
>
> Please have a look at our Unpacking or Psychological Methods papers.  It's
>
> > 3. How do I interpret the sensitivity results?  The two R-square values
> > at the bottom of the output window seem to make the most sense, but does
> > a low value (e.g., .0033) indicate good or poor robustness to violations
> > of key identifying assumptions, or does this depend on the absolute
> > level of the mediation effect?
>
> Again, please have a look at the Psych Methods paper and the software
> vignette.
>
> > Thanks for any light you might be able to shed on these issues.
> >
> > Glenn D. Walters, Ph.D.
> > Psychology Services
> > FCI-Schuylkill
> > Minersville, PA  17954
> >
> > This message is intended for official use and may
> > contain SENSITIVE information. If this message
> > contains SENSITIVE information, it should be
> > properly delivered, labeled, stored, and disposed
> > of according to policy.”
> >
> >
>

```