[R] Sensibility analysis

[Peter Rabinovitch]

Before you embark on a "vary-one-parameter-hold-others-constant", have a
look at "One-Factor-at-a-Time Versus Designed Experiments" from the
American Statistican, available at
http://www.amstat.org/publications/tas/czitrom.pdf

Micheall Taylor wrote:
> 
> I believe the english term you are looking for is sensitivity analysis -
> sometimes also called a stress analysis.
> 
> I'll have to leave it to others to discuss the R portion as I am pretty new
> to that, but may help some in the other areas.
> 
> There are no rules concerning the parameters and and step variances that
> should be used in this form of analysis.
> 
> The parameters that you chose to vary are those that you are concerned may
> 1) vary in the "real-world" 2) affect outcomes.  For example, modeling
> elections in the U.S. one may "stress" the model with a variances in
> weather as bad weather will cause Americans to stay home and vote in fewer
> numbers.  One would be less tempted to do so in France as the french voter
> tends to vote in elections (compared to Americans) and weather isn't quite
> as varied as in the U.S. (I would imagine - milder climate - smaller
> geography, etc)
> 
> What step to use?  What level of precision to you want?  There is no right
> and wrong. Being sort of a data guy, I tend to you smaller steps to create
> more granular precision - the resulting graphs look better :)
> 
> Your vary-one-parameter-hold-others-constant is a fine strategy most of the
> time. One word of caution.  If you are doing a nonlinear mixed-effects
> model, following such a technique may not show you the real sensitivity of
> the model.  You might want to consider a model in which all parameters are
> allowed to vary randomly within a known and fixed distribution. This sort
> of model is much harder to intrepret in the end, but can produce facinating
> spikes in model outcomes that can be very informative.
> 
> I've built such a model for "stressing" investment models. The non-linear
> and joint "stresses" produce radically different outcome contours than
> single variable variances.  These outcome contours better match reality.
> 
> ==================================
> Michaell Taylor, PhD
> Chief Economist, Reis, Inc., New York
> Professor of Political Science, NTNU, Trondheim, Norway
> Adjunct Professor, UofD, South Africa
> 
> On Wed, 18 Jul 2001 11:50:14 +0200 (MEST) brunels at free.fr wrote:
> 
> > Hi everyone,
> >
> > i'm actually working on a nonlinear mixed-effects model, and beginning the
> > study of its sensibility.
> > It takes 35 input and gives a n*35 matrix of output as it's a growth model (n
> > is the number of days of the growth period).
> > Well, I have to analyse the variation of the output in relation to the
> > variation of the parameters of this models (first univariate then multivariate
> > variation)  :  it's a sensibility analysis ( a french word-by-word
> > translation ).
> > So, I choose one parameter, increase its value, do it one more time and again,
> > and for each of the 10 responses I have a curve of sensibility which i have to
> > calculate numerically the derivative.
> >  I have read some books but none of them treated this scope in detail, so i
> > wonder :
> >     - what are the numeric differentation methods available in this case. Are
> > they available on R?
> >     - which criteria to choose for the variation step of the parameters ?
> >     - do you have any reference of books i may consult ?
> >
> > I'd be glad to receive any help from you all. Thanks.
> >
> > the student i am bless your mailing-list
> >
> >
> >
> >
> >
> > -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
> > r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
> > Send "info", "help", or "[un]subscribe"
> > (in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
> > _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
> >
> 
> -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
> r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
> Send "info", "help", or "[un]subscribe"
> (in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
> _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._