cc: Edward Cook <drdendro@ldeo.columbia.edu>
date: Sat, 20 Oct 2007 10:01:31 -0400
from: Edward Cook <drdendro@ldeo.columbia.edu>
subject: Re: GILGIT and ASTORE
to: Tom Melvin <t.m.melvin@uea.ac.uk>, Keith Briffa <K.Briffa@uea.ac.uk>

   Hi Tom and Keith,

   Here is my review of your "signal-free" paper just sent in to the journal and Connie
   Woodhouse who handled your paper. Please accept my sincerest apology for being so slow in
   doing this. Between two international trips, two proposals, and a congressional report on
   drought to do under tight time constraints with Richard Seager and others, I simply let the
   review slide. No excuse, just a bit of mea culpa explanation.

   The paper is very good and clearly needs to be published. As indicated below, the biggest
   issue I had related to convergence, how to determine it, and its apparent failure using a
   tree-ring data set from Gaspe. I have attached the ring-width data file for you to play
   with if you wish in order to see if you get the same unstable results that I got. Maybe I
   am doing something wrong, but other test data sets seemed to work fine in a signal-free
   sense.

   Cheers,

   Ed

   My review:

   This paper by Melvin and others is a significant new contribution to the basic development
   of tree-ring chronologies for studies of climatic and environmental change. It introduces
   and demonstrates clearly the potential problem of trend distortion caused by the fitting of
   growth curves to ring-width series, which can be influenced by the common signal of
   interest itself. Since this problem primarily manifests itself at the ends of the
   chronologies where interest lies concerning changes in the growth environment, trend
   distortion is a critically important matter to investigate and correct for. Melvins paper
   provides one approach to the correction of trend distortion through signal-free
   standardization.  It is based on iteratively adjusting the individual detrended series to
   remove the effects of the common signal on the curve fitting procedure. That the
   signal-free approach can work, as expected, is indicated by a number of synthetic examples
   and examples using real tree-ring data at the end. While the results are in many ways
   compelling, experiments done by me on different tree-ring data sets using the signal-free
   algorithm provided in the Appendix (using a simple length-N spline for detrending in each
   case) produced results suggesting that convergence is might occur very slowly, if at all,
   in some cases. Depending on the data set, sometimes convergence occurred after 4-6
   iterations or maybe only after 20+ iterations. The latter pathological case related to very
   slow growing trees in the 19th century (mean ring width ~0.2 mm around 1850) followed by a
   near-linear increase in growth up to the last year of growth in 1982 (mean ring width ~0.8
   mm by that time). Each iteration increased the steepness of the growth increase from 1850
   to 1982 with no apparent end in site after 20+ iterations and this was accompanied by the
   addition of a long-term negative trend being added up to 1850, which was suspiciously like
   the common long-term negative age trends in the raw data up to that time. Assuming that I
   have applied the signal-free method correctly (e.g., does use of the length-N spline
   matter?), this may be a problem. I suspect that this problem may relate to the growth
   curves getting too close to zero in the iterative procedure, a circumstance that Melvin
   rightly warns against in the Discussion. It would be useful for Melvin to discuss the issue
   of convergence, possible criteria for determining it, and under what conditions it might
   not occur very quickly, if at all. I am also a bit puzzled by Melvins statement in the
   Discussion concerning the growth curves falling below zero. Why is 3% indicated as a
   threshold? I would have thought that ANY growth curve that falls below zero anywhere in the
   sequence would be invalid to use. A bit more explanation would help here. One last thing.
   The wording in the bottom paragraph on pg. 19 is a bit ambiguous to me: the combination of
   the segment length curse and trend distortion severely restricts the value of chronologies
   produced using curve-fitting methods, for comparing the magnitudes of past and current
   warming. Does this refer to classical detrending methods before signal-free improvement or
   also the signal-free method as well? I suspect both. I would also prefer that the term
   severely either be removed or moderated because I am not sure such a sweeping stern
   generalization is completely justified in every case. Isnt restricts enough?

   Attachment Converted: "c:\eudora\attach\St_Anne_River_THOC.rwl.zip"


   ==================================
   Dr. Edward R. Cook
   Doherty Senior Scholar and
   Director, Tree-Ring Laboratory
   Lamont-Doherty Earth Observatory
   Palisades, New York 10964  USA
   Email: [1]drdendro@ldeo.columbia.edu
   Phone: 845-365-8618
   Fax: 845-365-8152
   ==================================

