I think you're approaching matters from quite the wrong angle.
You seem to want a firm, fixed technical answer where one is hardly
possible.
What we're looking at here is, despite its complexities, just another
exploratory method. What is best will be what works for your data given
your theory- and experience-based expectations about what kind of
relationship there should be, or could be, within your data, given also
various ideas about the limitations of the data. It will yield a graph
that you'll want to think about, to show others, or to put in a report.
A rule-of-thumb method is always so called because, in the author's
experience, it works quite well given the datasets and simulations
tried.
Often there's an attempt to balance the variability of the data against
the number of sample points. But usually the global variability of the
data is only indirectly relevant in any case.
Also, within the smoothing industry, somebody may work on a method
intensively for weeks or months and then publish with their
"recommendations", which may get enshrined in some other literature. If
you look carefully, you often find the same experts are now often
working on something quite different....
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Ana R. Rios
Sent: 23 January 2008 20:57
To: [email protected]
Subject: Re: st: RE: lpoly
Austin and Nick,
Thank you for your advise.
When choosing bandwidth, is the rule-of-thumb
estimator the best option? How is the optimal
bandwidth selected (i.e. what criteria is followed)?
Regards,
Ana Rios
--- Austin Nichols <[email protected]> wrote:
> Ana--
> I would advise degree 1, for local linear
> regression, and note that
> the kernel makes little difference, whereas the
> bandwidth is crucially
> important. With a large enough bandwidth, you will
> get -tw lfit- and
> with a small enough bandwidth, you will get -tw
> line-.
>
> On Jan 23, 2008 2:52 PM, Nick Cox
> <[email protected]> wrote:
> > I don't know what best to advise, but more
> importantly there's a very
> > large associated literature that gives lots of
> guidance. What I do
> > notice is that there are three main choices, the
> kernel, the bandwidth
> > and the degree.
> >
> > I've found restricted cubic splines, as
> implemented, in Stata 10, within
> >
> > -mkspline-, much easier to handle. It's true that
> under that the knot
> > positions (and so the number of knots) need to be
> specified, but the
> > default positions for a given number of knots in
> my experience alweays
> > work well.
> >
> > There is a wrapper program to make it easier on
> SSC as -rcspline-.
> >
> > Nick
> > [email protected]
> >
> > Ana R. Rios
> >
> > I was wondering if there is any guideline for
> choosing
> > the degree of the polynomial to be used in the
> > smoothing (degree(#) option in lpoly).
*
* 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/
