From	Michael E. Mann	Date	Fri, 16 Mar 2001 14:25:57 -0500 (EST)
To	Tim Osborn
CC	Scott Rutherford, Michael E. Mann
Subject	Re: comparison with our existing reconstructions
Dear Tim, Scott On the road w/ tenous email connection so have to be brief. This sounds good. Hoping we can have age-banded connections by the end of tnext week so I can show in Nice! Scott: can you rectify the comparisons that Scott is producing w/ your own comparisons that show more of a discpreapancy ? Thanks, mike Fri, 16 Mar 2001, Tim Osborn wrote: > Dear Mike & Scott, > > Attached is "traditional.ps", comparing the 1404-1855 (i.e. > pre-instrumental) REG-EM reconstruction with our existing Osborn et al. > maps and Briffa et al. quasi-hemispheric series (see refs below). Neither > the REG-EM nor the existing reconstructions use the age-banded trees, so > low frequencies are suppressed. [Scott - thanks for the new age-banded > results, but I probably won't get to them till next week due to other > commitments.] > > The time series comparisons are, as you see, quite good - thought you'd > expect this as we're comparing two methods but identical data! Red is > REG-EM, black is from the Osborn et al. existing reconstructions (then > averaged into quasi-hemispheric means), while blue is from Briffa et al. > (where we average the tree-density into regions/hemisphere *before* > calibrating against regional/hemispheric temperature). Blue & black agree > quite closely, so all correlations and % var explained are between red and > black. > > Timeseries are: > > '0-90' = full spatial average over each of our existing maps. > '0-70' = full spatial average over each of the REG-EM maps. > 'masked' indicates REG-EM maps are masked by the time-dependent coverage of > our existing maps. > 'land20-90' or 'land20-70' indicates only land grid boxes north of 20N are > averaged. > 'treeboxes' indicates only those grid boxes that contain tree-ring sites > are averaged together. > > The pattern correlations range from 0.2 to 0.8, with a mean of 0.6 > (approx). Fairly consistent then. The pattern of temporal correlations is > reasonable, ranging from 0.0 to 0.9, with a mean of 0.6 (approx). > > Comments: > (1) Time series generally have less variance in REG-EM, especially early > on, though masking of data brings them closer to our time series. > (2) Getting the mean level correct (I've converted REG-EM to behave like > anomalies from 1961-90 mean) helps with the %variance explained considerably. > (3) The temporal correlations are poorer for boxes containing trees than > those that do not! > > The decreased variance early in the REG-EM [comment (1)] is, I guess, > because the fewer the records with data, the earlier the > truncation/weighting function kicks in etc. and therefore the less the > variance that is reconstructed. As the 'skill' of REG-EM decreases, the > more the values are filled in with something near to their mean, I seem to > recall. This raises the question that the early values might be biased > towards the observational mean? If so, it might be better to replace box > values by missing values when their expected 'skill' becomes fairly low. > > Comment (3) can be explained two ways. In the non-tree boxes our two > methods (REG-EM and principal component regression) have similarities, and > given the common input data, one would expect similar reconstructions - > which the high correlations indicate. In the tree boxes, however, the > difference is our approach uses only local information, while REG-EM still > uses non-local information too. So, either (i) our reconstructions are > poorer *because* we're ignoring non-local information, or (ii) REG-EM > reconstructions are poorer *because* real local variations are partly > masked by regional-scale variations. It might be possible to choose either > (i) or (ii) as a preferred explanation, using verification or other > consideration, but I'd prefer to stick with (i) and (ii) as being equally > possible and therefore justifying both approaches. This is politically > better too! What I get out of the comparison is that the REG-EM is > producing variability that is highly correlated with our method, given the > same input data. The main concern is the difference in variance and hence > absolute anomalies. We should look at this again when I've compared the > age-banded stuff too. > > Another long e-mail, but I hope that this is useful (especially for EGS) > and will form the basis of a comparion of methods paper. > > Have a good weekend! > > Tim > > > > > > > > Briffa KR, Osborn TJ, Schweingruber FH, Harris IC, Jones PD, Shiyatov SG > and Vaganov EA (2001) Low-frequency temperature variations from a northern > tree-ring-density network. Journal of Geophysical Research 106, 2929-2941. > > Osborn TJ, Briffa KR, Schweingruber FH and Jones PD (2001) > Annually-resolved patterns of summer temperatures over the Northern > Hemisphere since AD 1400 from a tree-ring-density network. In preparation. > _______________________________________________________________________ Professor Michael E. Mann Department of Environmental Sciences, Clark Hall University of Virginia Charlottesville, VA 22903 _______________________________________________________________________ e-mail: mann@virginia.edu Phone: (804) 924-7770 FAX: (804) 982-2137 http://www.evsc.virginia.edu/faculty/people/mann.html