[Mediation-information] Sensitivity analysis for a GAM

[dustin tingley]

Hi Louise-

I see.

You mentioned rbounds several times, by which I presume you mean the Stata
module. Thats for sensitivity to random assignment of the treatment. Our
sensitivity analysis is to the sequential ignorability assumption, bc. we
are doing mediation analysis rather than just estimating treatment effects.
You probably know this, but I just want to be clear.

dt

Dustin Tingley
Government Department
Harvard University
http://scholar.harvard.edu/dtingley


On Mon, Dec 12, 2011 at 9:36 AM, Glew, Louise <Louise.Glew at wwfus.org> wrote:

>  Hi Dustin,****
>
> ** **
>
> Thanks for the suggestion. Unfortunately, I have a really nasty non-linear
> component which cycles on an irregular basis. Unless I ignore the
> underlying assumptions, I can’t fit a linear model to the data. The other
> alternative is to ignore the serial dependence, chop the data into a series
> of matched pairs (one at each time-step) and to run a sensitivity analysis
> on a Wilcoxon Matched Pairs test, as in rbounds. Either way, ignoring the
> underlying structure of the data seems to risk inflating Type I error, and
> finding the estimates of impact overly sensitive to a hidden bias.  In
> theory I could report this as the maximum sensitivity to hidden bias,
> acknowledging the (unknown) probability of a false positive but I suspect
> reviewers will foam at the mouth at this approach…. ****
>
> ** **
>
> Thanks so much for the response, very much appreciated! ****
>
> ** **
>
> Louise****
>
> ** **
>
> ** **
>
> ** **
>
> *From:* dustin tingley [mailto:dustin.tingley at gmail.com]
> *Sent:* Sunday, December 11, 2011 10:55 PM
> *To:* Kosuke Imai
> *Cc:* Glew, Louise; mediation-information at r-forge.wu-wien.ac.at
> *Subject:* Re: [Mediation-information] Sensitivity analysis for a GAM****
>
> ** **
>
> Hi Louise-****
>
> ** **
>
> Great question. ****
>
> ** **
>
> Yes, doing sensitivity analysis with GAM's as Kosuke mentioned is beyond
> what we can do right now. Its possible in theory, we think, but would be a
> major undertaking to figure out how to do. Our Stata package does not do
> this either (currently it supports a smaller set of models compared to our
> R package).****
>
> ** **
>
> Let me make another suggestion. When you estimate things with the GAM's,
> are you uncovering highly non-linear relationships (graph things out like
> what we do in the Psychological Methods paper). If you aren't, perhaps Mr.
> Linear is your friend, and you can report the sensitivity analysis based on
> the linear model as the robustness check?****
>
> ** **
>
> best,****
>
> Dustin****
>
> ** **
>
> PS: as an aside, we're so psyched someone from the WWF is using our
> software! Really cool!****
>
>
> Dustin Tingley
> Government Department
> Harvard University
> http://scholar.harvard.edu/dtingley
>
> ****
>
> On Sun, Dec 11, 2011 at 10:35 PM, Kosuke Imai <kimai at princeton.edu> wrote:
> ****
>
> Hi Louise,
>
>  Thanks for your email.  Unfortunately, our software, mediation, does not
> have sensitivity analysis for GAM.  We hope to expand the capability of the
> software in the future, though.  However, you should still be able to use
> GAM to calculate mediation effects.
>
> Best,
> Kosuke
>
> Department of Politics
> Princeton University
> http://imai.princeton.edu
>
>
> On Dec 10, 2011, at 11:54 AM, Glew, Louise wrote:
>
> > Dr. Imai,
> >
> > I’m working on a series of quasi-experimental assessments of the impact
> of biodiversity conservation interventions. I was wondering if I could ask
> your advice on how best to conduct a sensitivity analysis on a generalized
> additive mixed model.
> >
> > I’m analyzing a time-series of remotely-sensed imagery, to explore the
> impact of a series of conservation interventions against matched controls
> in northern Kenya. The properties of the data have precluded using a GLMM,
>  and are particularly challenging as they are both spatially
> auto-correlated and serially-dependent.
> >
> > When working with less complex data, I have used the ‘rbounds’ and the
> ‘mediation’ package to compute Rosenbaum’s sensitivity bounds. Reading the
> documentation for the mediation package, it became apparent that the
> current function can only handle linear models. The GAMM I’ve fit to the
> data is technically a penalized linear model, but the errors are not
> normally distributed (due to the spatial and temporal correlation). I was
> wondering if you could advise me of any alternative sensitivity analysis
> methods which may be able to cope with non-linear data please?
> >
> > Thank you,
> >
> > Best wishes,
> >
> > Louise
> >
> > Louise Glew (MEnvSci)
> > Monitoring and Evaluation Specialist
> > Conservation Science Program
> > World Wildlife Fund
> > 1250 24th Street, NW
> > Washington, DC 20037-1193
> > Office: 202-495-4184
> > Skype: louise.glew
> > Louise.Glew at wwfus.org
> > worldwildlife.org
> >
> > <image001.gif>
> >
>
>
> _______________________________________________
> 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
> ****
>
> ** **
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/mediation-information/attachments/20111212/de9500ff/attachment.htm>