cc: "Keith Briffa" <k.briffa@uea.ac.uk>, <sandy.tudhope@ed.ac.uk>
date: Tue, 30 Nov 2004 11:24:41 -0000
from: "Rob Wilson" <rjwilson_dendro@blueyonder.co.uk>
subject: Re: e-mail problem?
to: "Tim Osborn" <t.osborn@uea.ac.uk>

   Hi Tim,

   thanks for the data access.

   I will play around with all the data over the next few weeks and try and cobble together a
   terse report/explanation of what I have been doing. I think I may have confused people by
   just sending figures without explanation. Sorry! I just wanted to get some preliminary
   figures out showing that I had finally started.



   Some brief answers to your comments:

   Now to some brief comments:
   (1) The calibration and verification statistics appear to improve for
   1635-1702 for the simple means method.  I first thought that this was
   because of the Palmyra data, but then I couldn't think how it could improve
   the stats when it doesn't overlap with the instrumental data against which
   the skill is measured.  Also, in your simple means method, do you average
   the O18 data themselves, or do you normalise/anomalise each series first
   and then average them.  If the latter, what did you use for a reference
   period, and how did you cope with the difficulty of different record
   lengths - especially the Palmyra data?

   I did not average the raw O18 values. For each nest, I normalised the data to the common
   period of the records in that nest.



   For the 1635-1702 part of the reconstruction (actually two nested steps (1635-1643 and
   1644-1702 )), the living Palmyra data were included and the data were normalised to the
   common period (e.g.1644-1700/1886-1981) period and then the calibration/verification stats
   made.

   For splicing all the sections together to make the final reconstruction, the reconstructed
   data were all re-scaled to the mean/variance of the most replicated nest over the period
   1897-1981.

   Therefore, as the living portion of the Palmyra data were included for the 1635-1643 and
   1644-1702 nested models, the calibration statistics were improved for that portion of the
   final record. This is what made the averaging approach more appealing.

   Of course, a similar approach was not possible for the PC regression approach.
   (2) I don't have the corrections yet for removing the artificial step
   between the HadCM3 "all" and "nat" runs.  I hope to have it soon from
   Simon.  But I guess it won't be large in the tropics, since it arose from
   an incorrect surface roughness in regions where there had been an
   anthropogenic change in land use (cropland, deforesting, pasture etc.)
   between 1750 and 1990 - hence mainly the mid latitude land areas.  Still, I
   guess there could be a response of the oceans.

   Fair enough.
   (3) The comparison with HadCM3 seems excellent - r(ALL,RECON)=0.54
   !!!  This seems a very publishable result.  I guess much of it is related
   to the warming trend, though?  Both the model and recon will contain forced
   variability and internally-generated variability (plus model will contain
   errors in forcing, and recon will contain errors), but ONLY the forced
   variability should be in common, r=0.54 is very high - this supports
   Simon's analysis that the relative contribution of forcing is greatest in
   the tropical oceans!  It would be interesting to correlate the
   reconstructions with a combination of NAT for pre-1750 with ALL for
   post-1750, because this is (supposedly) the best representation of reality.

   will make a full period comparison, but the quality of the coral recon does decrease
   markedly prior to 1850.
   (4) As well as the correlation, a comparison of the magnitudes of
   variability and trends is important.  Allowing for the different vertical
   scales, these seem qualitatively similar, with perhaps stronger variations
   in the model - but wouldn't you expect that, given the recons don't capture
   all the variability (or have you subsequently scaled them to have the same
   variance as the actual data?).

   For the simple comparison, I did not scale the records to a common data-set/period. The
   problem is, if I understand this correctly, we would have to scale the records to the
   period prior to anthropogenic impacts (i.e. 1750-1850), but this is precisely the period
   where we should start being cautious about our interpretation of the reconstruction. I need
   to think on this more.
   (5) It would be nice to see the comparison and correlations with the ECHO-G
   runs.  Also, Sarah Raper and I are making runs with the simple MAGICC model
   driven by the same forcings - we could provide those time series when
   they're ready and compare with those too.

   The more data the merrier.
   (6) Keith and I will talk a bit more, but Keith's already made some good
   comments about whether some sub-regions of the SST might be more skilfully
   reconstructed, especially with fewer corals early on.

   Keith and I talked about this briefly yesterday.

   My main concern with this approach is that some records are much stronger than others and
   if we go to smaller regions, we may not be able to develop reconstructions that would
   verify very well. They may also not be very long. The reality is that those records which
   are best (i.e. correlation strongly with local or large scale temperatures) are also the
   shortest.

   I think we should very much take the line that this study is a first effort and the results
   appear promising, but the reality is, the current coral data-base is simply too short, and,
   if possible, future sampling needs to focus on finding longer coral records and/or
   extending the living records with fossil material (as Cobb). Although it is tempting to
   over interpret these reconstructions, I think we need to be very cautious with the pre 1850
   period. I don't see any way around this.



   regards

   Rob
