cc: "Malcolm Hughes" , "Keith Briffa" date: Mon, 31 Mar 2008 18:00:46 -0400 from: "Wahl, Eugene R" subject: RE: to: "Caspar Ammann" , Hello all: A clarification...by "truth" in the second paragraph below I don't mean to imply that critiques of reconstruction methods based simply on examinations of ensemble distributions should stand, per se. I only meant to say that recognition of the universe of possible reconstructions is a worthwhile addition of knowledge. Conceivably, doing this might possibly help us refine our validation schemes. Caspar and I have taken one look into this set of issues in the companion piece to the Wahl-Ammann paper in Climatic Change last fall (the Ammann-Wahl article there), to deal with critiques of validation methodology raised by MM. We revisited their ensemble approach (reconstructions driven only by the full-AR persistence structure in the proxies), but restricted its output with the kinds of calibration and verification criteria we use in actual practice (which MM did not do). The idea was to do exactly the kind of geophysical contextualization that Caspar mentions -- thereby incorporating the ensemble method, but also embedding the ensemble output into the real world decision-making structure we use with all our reconstructions. Interestingly, the results are quite similar to verification significance results based on small-lag AR structures in the target series itself, the general way this issue is approached in climatology. Peace, Gene From: Wahl, Eugene R Sent: Monday, March 31, 2008 11:02 AM To: 'Caspar Ammann'; mann@psu.edu Cc: Malcolm Hughes; Keith Briffa Subject: RE: http://www.cosis.net/abstracts/EGU2007/03128/EGU2007-J-03128.pdf?PHPSESSID=e Hi all: I think Caspar's on the money here. The statisticians have a point in that we are really sampling from noisy proxies that themselves are sampling from one of many possible realizations of climate for a given set of forcings (thinking of model ensembles, e.g., all with slightly perturbed initial conditions). However, we in the paleoclimate part of geophysics (and other disciplines that use similar or identical methods, such as econometrics) have clearly recognized the need to separate "wheat from chaff" in forecasting/hindcasting models, and thus the calibration and verification exercises we do. So, it seems to me (at least on a first pass) that there is "truth" in both perspectives on the problem. It would be interesting to explore what our validity screening procedures are in fact doing from a purely mathematical theoretical standpoint...what is the effect of the truncation of possibilities that our validation procedures entail in the underlying geometries we are examining? That could be one way to bridge the difference between the statistical and geophysical perspectives Caspar identifies. [Let me know if you think I've got something incorrect in this Q.] Happy spring to all. Even in Alfred winter is finally breaking! Peace, Gene From: Caspar Ammann [mailto:ammann@ucar.edu] Sent: Sunday, March 30, 2008 10:19 PM To: mann@psu.edu Cc: Malcolm Hughes; Keith Briffa; Wahl, Eugene R Subject: Re: http://www.cosis.net/abstracts/EGU2007/03128/EGU2007-J-03128.pdf?PHPSESSID=e Malcolm and Mike, I wouldn't read too much into this. I believe that all we are looking at is the difference between a statisticians approach and us in geophysics. The statisticians like to simulate many ensembles. I had the same discussions with our guys at NCAR. The tendency for them is to include all possible reconstructions and then describe the distributions. Our approach has been to throw away reconstructions that don't make sense or that don't pass verification. So its more philosophical than anything else. Though Mike might be right in the sense that the choices can lead some of these approaches astray. We had this with regard to the selections of uncertainty, what is actually independent uncertainty. There a good and strong check on reality is necessary. So we shall see in Vienna ... Caspar On Mar 30, 2008, at 8:32 PM, Michael Mann wrote: Malcolm, in short, this looks like nonsense. there is nothing magic about 'Bayesian' methods. Many of the methods we use can easily be recast as Bayesian approaches, the critical question comes down to what the "prior" is. For example, in RegEM, the prior is the first 'guess' in the iterative expectation-maximization algorithm. Of course, if the final result is sensitive to that choice, one becomes a bit worried, the pitfall indeed of many a Bayesian approach. mike Caspar Ammann wrote: Malcolm, are you referring to this? [1]http://www.cosis.net/abstracts/EGU2007/03128/EGU2007-J-03128.pdf?PHPSESSID=e Caspar Caspar M. Ammann National Center for Atmospheric Research Climate and Global Dynamics Division - Paleoclimatology 1850 Table Mesa Drive Boulder, CO 80307-3000 email: [2]ammann@ucar.edu tel: 303-497-1705 fax: 303-497-1348 -- Michael E. Mann Associate Professor Director, Earth System Science Center (ESSC) Department of Meteorology Phone: (814) 863-4075 503 Walker Building FAX: (814) 865-3663 The Pennsylvania State University email: [3]mann@psu.edu University Park, PA 16802-5013 [4]http://www.met.psu.edu/dept/faculty/mann.htm Caspar M. Ammann National Center for Atmospheric Research Climate and Global Dynamics Division - Paleoclimatology 1850 Table Mesa Drive Boulder, CO 80307-3000 email: [5]ammann@ucar.edu tel: 303-497-1705 fax: 303-497-1348