[Mediation-information] Fwd: Question about results from R mediation package

[Kosuke Imai]

FYI.

Kosuke

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

Begin forwarded message:

> From: Teppei Yamamoto <teppei at mit.edu>
> Subject: Re: [Mediation-information] Question about results from R mediation package
> Date: April 12, 2012 8:50:58 AM EDT
> To: Alexander Weiss <alex.weiss at ed.ac.uk>
> Cc: Kosuke Imai <kimai at princeton.edu>
> 
> Hi Alex,
> 
> Yes. "Direct Effect_0" will correspond to the direct effect of the treatment when the mediator is fixed at the value it would take when treatment = 0, etc. Keep in mind that even if you set those to different values, the output will still be labeled "_0" and "_1".
> 
> Best,
> Teppei
> 
> Alexander Weiss wrote, (04/12/2012 08:43 AM):
>> Hi again,
>> 
>> One final (I hope) question. If my treatment variable in my accelerated failure
>> time analysis is continuous and I do not specify control.value and
>> treat.value, should I be getting separate direct effect estimates for 0 and 1?
>> 
>> Best,
>> 
>> Alex
>> 
>> On Thursday 12 Apr 2012 05:28:47 you wrote:
>>> Hi Alex,
>>> 
>>> Kosuke is right about the scale of the output. That is, the effect
>>> represents the average difference in the time until death between the
>>> treatment and control conditions, decomposed into indirect and direct
>>> components.
>>> 
>>> Looking at your code, I think you are not correctly using the
>>> "covaritates" argument in the mediate function --- you don't have to
>>> write out those variable names unless you want to make inferences
>>> conditional on specific values of those covariates. This is sometimes
>>> called the "moderated mediation" analysis by psychologists. See the help
>>> file, especially the example section.
>>> 
>>> In your message, you say at the beginning that your treatment is
>>> continuous but then you ask a question about a categorical treatment
>>> variable. Regardless, you can specify a specific contrast using the
>>> "treat.value" and "control.value" arguments. Again please refer to the
>>> help file.
>>> 
>>> Best,
>>> Teppei
>>> 
>>> Kosuke Imai wrote:
>>>> Hi Alex,
>>>> 
>>>>    The scale of the quantities of interest (the average causal mediation
>> effects and average direct effects) is the same as that of the outcome variable;
>> that is, the time to death in the survival analysis.  I'm ccing my
>> collaborators through our mailing list so that they can confirm: see https://r-
>> forge.r-project.org/mail/?group_id=1070
>>>> 
>>>> Best,
>>>> Kosuke
>>>> 
>>>> Department of Politics
>>>> Princeton University
>>>> http://imai.princeton.edu
>>>> 
>>>> On Apr 11, 2012, at 9:33 AM, Alexander Weiss wrote:
>>>> 
>>>>> Dear Prof. Imai:
>>>>> 
>>>>> I have recently been asked to determine whether the effect of cigarette
>> smoking
>>>>> (non = 0, former = 1, present = 2) mediates the effect of a personality
>> trait
>>>>> (Agreeableness, a continuous variable scaled as a z-score) on time to
>> death in
>>>>> an accelerated failure time model, i.e., a model of the type
>> survreg(Surv(y))
>>>>> with a Weibull distribution in R.
>>>>> 
>>>>> I used the package developed by yourself and your colleagues, though was
>>>>> a bit puzzled by the outcome as I do not understand the scale of the
>> parameter
>>>>> estimates. I specified the outcome model and mediator models as
>>>>> follows:
>>>>> 
>>>>> # AFT model for time to death
>>>>> model.y<- survreg(Surv(weekstodeath_ssn_dob,mortality_status_dob)
>>>>>              ~z_a + pat_sex + tribaseage + grade1 + diabetic +
>>>>>                cvd_broa + selfhealth_1 + adlbase + iadlbase + smoking2
>>>>>                + z_n + z_e + z_o + z_c,dist="weibull",data=ffi_mediator)
>>>>> 
>>>>> # Create polr model for predicting smoking from Agreeableness
>>>>> model.m<- (polr(data=ffi_mediator,Hess=T,method="probit",smoking2 ~ z_a
>> +
>>>>> pat_sex
>>>>>                  + tribaseage + grade1 + diabetic + cvd_broa +
>>>>>                  selfhealth_1 + adlbase + iadlbase + z_n + z_e + z_o +
>> z_c))
>>>>> 
>>>>> and used the following for the mediate() function:
>>>>> 
>>>>> smoking_agreeableness_covariates<-
>>>>>   mediate(model.m,model.y,sims=1000,treat="z_a",mediator="smoking2",
>>>>>   covariates=list("pat_sex","tribaseage","grade1","diabetic1","cvd_broa",
>>>>>   "selfhealth_1","adlbase","iadlbase","z_n","z_e","z_o","z_c"))
>>>>> 
>>>>> The results of the analyses are as attached as a text file.
>>>>> 
>>>>> My question concerns the parameter estimates obtained. First, I cannot
>> quite
>>>>> work out what metric they are on as they are not in the same as the AFT
>>>>> analysis. Moreover, I am not wholly clear on how to interpret the results
>>>>> given that they seem to be for cases in which the treatment variable is
>>>>> categorical.
>>>>> 
>>>>> I appreciate you are busy, but I would most appreciate your assistance in
>> this
>>>>> matter.
>>>>> 
>>>>> Sincerely,
>>>>> 
>>>>> Alex
>>>>> 
>>>>> 
>>>>> 
>>>>> --
>>>>> The University of Edinburgh is a charitable body, registered in
>>>>> Scotland, with registration number SC005336.
>>>>> 
>>>>> <mediation_results_11_april_2012.txt>
>>>> 
>>>> _______________________________________________
>>>> Mediation-information mailing list
>>>> Mediation-information at lists.r-forge.r-project.org
>>>> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/mediation-
>> information
>>> 
>>> --
>>> ====================================
>>> Teppei Yamamoto
>>> Assistant Professor
>>> Department of Political Science
>>> Massachusetts Institute of Technology
>>> http://web.mit.edu/teppei/www/
>>> ====================================
>>> 
>>> 
>> 
> 
> -- 
> ======================================
> Teppei Yamamoto
> Assistant Professor
> Department of Political Science
> Massachusetts Institute of Technology
> http://web.mit.edu/teppei/www/
> ======================================