[R] Population abundance, change point
marchywka at hotmail.com
Wed Nov 24 17:34:36 CET 2010
> To: gavin.simpson at ucl.ac.uk
> CC: carusonm at gmail.com; marchywka at hotmail.com; r-help at r-project.org; r-help-bounces at r-project.org
> Subject: Re: [R] Population abundance, change point
> From: jdaily at usgs.gov
> Date: Wed, 24 Nov 2010 10:59:17 -0500
> I agree that SiZer is not the ultimate answer to all changepoint analysis,
> but that is why there are so many changepoint detection methods used. I
> will clarify, though, that my understanding of SiZer (which may be wrong)
> was that the smoothing splines are just a vessel for finding the
> changepoints, and made no assumptions about the continuity of the
> changepoint itself.
> One thing that would certainly help, especially with the confidence
> intervals about 0, is some bandwidth selection standard, though choosing
> that standard would be a difficult process to say the least.
I had a rather long draft response to earlier post on this thread,
quoting Einstein and Homer Simpson, but it seemed to long on philosophy
and was becoming a rant. But in response to this, Standards are good,
so good in fact everyone should have their own, LOL.
If you find yourself agonizing over the selection of arbitary parameters,
like cutoffs of various types say p<.05, it may simply be that
there is no sacred answer and the only answer is "get more data"
or at least discuss the implications of various choices.
> Jonathan P. Daily
> Gavin Simpson wrote on 11/24/2010 09:15:55 AM:
> > [image removed]
> > Re: [R] Population abundance, change point
> > Gavin Simpson
> > to:
> > Jonathan P Daily
> > 11/24/2010 09:16 AM
> > Cc:
> > Mike Marchywka, r-help, r-help-bounces, carusonm
> > Please respond to gavin.simpson
> > On Wed, 2010-11-17 at 09:17 -0500, Jonathan P Daily wrote:
> > > Indeed I have looked into various non-standard changepoint analysis
> > > methods. I figured the OP was more interested in traditional methods
> > > you have to spend less time justifying your methodology. Wavelets are
> > > potential nontraditional method, as is Significant Zero Crossings (R
> > > package SiZer), which fits arbitrary-degree smoothing splines over a
> > > of bandwidth parameters and looks for changes.
> > ...By looking to see if the derivative of the fitted curve is different
> > from 0 (given a suitable confidence interval on the derivative. My
> > problem with all of this is that these data are time series and SiZeR
> > doesn't take this into account (AFAICS) when computing the confidence
> > intervals - they are certainly too narrow for examples I have run.
> > Also, if these things are using splines, aren't we already assuming that
> > the underlying function is smooth and not a discontinuity? So which
> > technique the OP chooses will depend on how they think about the type of
> > change taking place at the "changepoint" - a point I think you made
> > earlier Jonathan.
> > Don't mean to be too negative, this has been a very useful discussion
> > that I am coming to late after a spot of time in the field.
> > All the best,
More information about the R-help