[Mediation-information] Inquiry about sensitivity analysis

[Kosuke Imai]

If you believe that you've controlled for many covariates, you might want to parameterize the sensitivity analysis by R^2 rather than correlation.  Have a look at this as all of our proposed methods are implemented via an R package: http://imai.princeton.edu/research/files/mediationR2.pdf

Best,
Kosuke

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

On Oct 24, 2012, at 11:33 AM, Baek-san Yu <postcentre at hanmail.net> wrote:

> 
> Dear, Dr. Imai
> I am a MA student in South Korea.
> I have read your "Identification, Inference and Sensitivity
> Analysis for Causal Mediation Effects", and I appreciate 
> your commitment. Since I am not much good at SEM, still
> I have few questions.I look forward to your advice. 
> 
> following your procedure, I could draw the following picture.
> 
> This picture shows that when rho overcomes .04  the significance of indirect effect
> will be lost and over .10, it would lead to opposite direction. In this sense, it appears
> that the indirect effects(.034, rho=0) are vulnerable to unobserved confounders. 
> 
> However, given that I controlled confounding variables which are well known, and 
> the estimated residual co-variance is about  .05, can I say the indirect effects still can be hold? 
> I am not sure that estimated residual co-variance between mediator and outcome variable in
> the model could be adequate reference. Or this result only can tell that the estimated indirect 
> effect is weak to hidden bias.
> 
> <output_1.jpeg>
> 
> 
> Thanks you for reading.
> I would appreciate your answer.
> 
> 
> 
> 
> 
> 
> Baek-San Yu. 
> The Sociology of Education. Master's Candidate 
> Department of Education, Korea University 
> Mobile: +82-10-2895-4280