There is indeed much literature of this kind. For
the last decade or so I have been following
non-technical rules of thumb without discernible
scientific loss as far as colleagues who want
to tell a scientific story or publications
are concerned. For "physical" naturally read
"economic", "medical" or whatever else is appropriate.
1. If it is too noisy to appeal as the physical story,
you should try smoothing more.
2. If it is very smooth, you might have smoothed away
interesting fine structure as well as noise, so try
smoothing less.
3. A smooth with a little visible roughness encourages
you to believe that you are not in situation 2.
4. I do not expect any formal criterion to match
my idea of physical sense, which will depend on
the physical situation, not just the characteristics
of the data.
5. Series to be compared should be smoothed in the
same way and to the same degree.
6. A trio of smooths, "about right", "rather more"
and "rather less", is often very helpful to indicate
the range and assure readers that your results
do not contain major artefacts.
7. Similar results should be expected from quite different
methods if the structure discerned is physically genuine.
Nick
[email protected]
Robert A Yaffee
> Hastie and Tibshirani in Generalized Additive Models have a nice
> treatment of
> this matter. They suggst a simulation and plot of the average mean
> square error or
> the average cross-validated residual against the bandwidth size. You
> can find
> the right tradeoff between bias and variance that way.
Joe Trubisz
> > Is there a 'best' way to determine what the optimal bandwidth
> > value
> > is for lowess?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
