Sunday, March 25, 2012

2829.txt

date: Fri Jul 21 19:00:20 2006
from: Keith Briffa <k.briffaatXYZxyz.ac.uk>
subject: RE: confidential
to: "Wahl, Eugene R" <wahleatXYZxyzred.edu>

Gene
your comments have been really useful and reassuring that I am not doing MM a disservice.
I will use some sections of your text in my comments that will be eventually archived so
hope this is ok with you. I will keep the section in the chapter very brief - but will cite
all the papers to avoid claims of bias. I really would like to discuss the whole issue of
the reconstruction differences at a later , less stressful time. I completely accept the
arguments about the limitation in the r2 and the value of capturing longer-term variance .
I think I will have to stop now as the temp and humidity are killing here.
Thanks a lot again
Keith
At 18:39 21/07/2006, you wrote:

Hi Keith:

I'm sorry that there is a bit to digest...although I know it is just a result of the
nature of things.

By the way, copied below is a synopsis that I sent this morning to a person in DC who is
working on all this with regard to the House of Representative hearings. Evidently,
there is to be at least one more hearing next week, and Mike Mann will go. The person I
sent this to is trying to understand the importance of the proxy PC issues --especially
how, no matter what way the PC extraction is done, the reconstructions converge if the
structures actually present in the data are not tossed out by truncating the number
retained PCs at a too low level. What I've copied is this synopsis. I think it is
straightforward -- maybe a bit dense, but at least brief.

Also, let me know if I can help on the issue of RE vs r^2. I could write a few brief
sentences as something for you to look at if you would like. Wahl-Ammann show very
clearly that there is objectively demonstrated skill at the low-frequency level of the
verification period mean for all the MBH segments, although the earlier MBH segments do
have really low r^2 values (indicating very little skill at the interannual level). Our
argument that to throw out the reconstruction completely based on the fastest varying
frequency, when it has objectively demonstrable meaning at lower frequencies, is to me
quite reasonable. That it is some how entirely ad hoc, as McIntyre claims in one
(more?) of his comments, is neither logical nor factual in my perspective. The idea of
frequency dependent skill/non-skill is not new to the literature, and the independent
re-reviewer that Steve Schneider had look over Wahl-Ammann said s/he had experienced
this issue in his/her work. G.


****************************** COPIED TEXT ******************************

What it boils down to in the end is as follows:
1) The different reference periods used to calculate proxy PCs from N. America
(calibration only for MBH, full period for MM) only have the effect of re-arranging how
the hockey stick shape appears across the rank ordering of PCs. In MBH it is
concentrated in PC1. In the full-period method, it is spread over PCs 1 and 2. If one
adds PCs 1 and 2 (either arithmetically or as vectors) from either convention, you get
an essentially IDENTICAL time series, only the amplitudes are a bit different. [Note
that the input data were centered AND standardized before being put into the PC
calculation algorithm. This is important, as shown below.]
WHEN ACTUALLY USED IN THE RECONSTRUCTION, THE DIFFERENCE IS MINISCULE -- MBH is
colder over 1400-1449 by 0.05 degrees!
2) IF the data are centered but NOT standardized and are input into in a PCA algorithm
using the variance-covariance matrix and not the correlation matrix (the way MM did it),
then the hockey stick shape shows up in PC4. MM in fact reported this first in their
2005 Energy and Environment article. In effect, the first two PCs are ARE ACTING TO DO
THE STANDARDIZING OF THE DATA not done as a pre-processing step. [When the correlation
matrix is used instead in the PCA algorithm, then the standardization is in effect done
by the algorithm, because all the correlations are "standardized" by construction--they
all range between 0 and 1.]
When 4 PCs from this calculation method are used rather than 2 PCs calculated
as above, then the RECONSTRUCTION CONVERGES TO THE SAME AS ABOVE.
3) Thus, all the different "flavors" for PC extraction have essentially no effect on
reconstruction when one does the exercise of adding PCs sequentially from 2 to 5 for any
flavor. In the case of (1), the reconstructions converge by the second PC. In the case
of (2), they converge by PC4. They don't change with higher order PCs added.
THIS SHOULD BE EXPECTED FROM FIRST PRINCIPLES. That is, the same underlying
information is there in all cases, it is only how the structures present in these data
are spread across the rank order of PCs, as explained. The simple exercise of taking
the reconstructions to convergence across the number of PCs used shows this clearly.
4) In fact, MM essentially say all this in the 2005 EE article--INCLUDING ABOUT THE
RECONSTRUCTION RESULTS -- but they strongly claim that the movement of the hockey stick
shape to the 4th PC shows it is not a leading pattern of variance as MBH claim, and thus
should not be used. This might be logical if their analysis was an apples-apples
comparison, but it is not, due to the PCA method they use and applying it on
NON-standardized data.
THESE TWO DIFFERENCES (which one can only fully get from their actual code,
not in the articles published) DRIVE THEIR ENTIRE ARGUMENT ON THIS PARTICULAR ISSUE.
What they do not say is that convergence to something like the MBH result is expectable,
and indeed MUST happen given the data used, because the hockey stick shape is actually
IN the data, it is NOT an artifact of PC calculation procedure.
5) FINALLY, note that all of this rests on the foundation that keeping the bristlecone
pine records in the data is appropriate, which Caspar and I find can be reasonable
presumption. If one believes that the bristlecone data should be removed, then the
1400-1449 reconstruction does not pass verification testing with the RE statistic, and
the MBH reconstruction should commence from 1450 on out.
Although there are a number of reasons to keep the bristlecone data in, maybe the most
compelling reason they are a NON-ISSUE is that, over the common period of overlap
(1450-1980), the reconstruction based on using them from 1400-1980 is very close to the
reconstruction based on omitting them from 1450-1980. Since the issues about the
bristlecone response to climate are primarily about 1850 onwards, especially 1900
onwards [KEITH -- PLEASE LET ME KNOW IF I AM NOT ACCURATE IN THIS], there is no reason
to expect that their behavior during 1400-1449 is in any way anomalous to their behavior
from 1450-1850. Thus, THERE IS NO REASON TO THINK THAT THE BRISTLECONES ARE SOMEHOW
MAKING THE 1400-1449 SEGMENT OF THE MBH RECONSTRUCTION BE INAPPROPRIATELY SKEWED.
****************************** END OF COPIED TEXT *******************

Peace, Gene
Dr. Eugene R. Wahl
Asst. Professor of Environmental Studies
Alfred University

607-871-2604
1 Saxon Drive
Alfred, NY 14802
________________________________
From: Keith Briffa [[1]mailto:k.briffa@uea.ac.uk]
Sent: Fri 7/21/2006 4:51 AM
To: Wahl, Eugene R
Subject: RE: confidential
Gene
thanks a lot for this - I need to digest and I will come back to you.
thanks again
Keith

--
Professor Keith Briffa,
Climatic Research Unit
University of East Anglia
Norwich, NR4 7TJ, U.K.

Phone: +44-1603-593909
Fax: +44-1603-507784
[2]http://www.cru.uea.ac.uk/cru/people/briffa/

No comments:

Post a Comment