cc: Tim Osborn <t.osborn@uea.ac.uk>
date: Fri, 23 Mar 2001 16:38:58 +0000
from: "Tim Osborn" <t.osborn@uea.ac.uk>
subject: Re: quick question
to: srutherford@virginia.edu

>>Sorry to bother you again, I know you're swamped at the moment.  I'm 
>>still having trouble reproducing your comparisons; even with the 
>>corrected reference period.  Maybe the difference is the time 
>>dependant mask.  In any case, I don't think we need anything else for 
>>the EGS talk, I just want to be clear on a couple of things.  In your 
>>results when you compare your reconstruction, say 0-90 and our 
>>0-70masked.  Can you tell me exactly what goes into both?  What's the 
>>difference between 0-70masked, treeboxes, and treeboxes masked?

Scott,

I'm away from the office today (meeting in London) so I still can't get the
time dependent mask to you.  But I can explain the differences:

We have two distinct reconstructions.

First is our regional and quasi-hemispheric reconstructions, where we
average (with some appropriate weighting) our tree-ring density series into
regional averages (in the case of the quasi-hemispheric series, all density
series are averaged), and *then* we calibrated against regional-mean
warm-season temperatures.  We prefer to do this because the temperature
signal is stronger at the regional scale than at the grid box scale. 
However, we have only done this for certain pre-defined regions and thus it
is not useful if you want to know the past variation in temperature averaged
over some region that we've never used!  For the quasi-hemispheric series,
where we average all the trees together, we have then calibrated this
against two different observed temperature series.  The first is the
temperature obtained by averaging all instrumental temperatures from grid
boxes that contain tree-ring sites.  In our papers we called this
reconstruction "ALL".  The second is the temperature obtained by averaging
warm-season temperatures from all land grid boxes north of 20N.  In our
papers we called this reconstruction "NH" and this is the time series that I
sent you I think.

The second reconstructions we have are the gridded fields of warm-season
temperatures that we've reconstructed by a combination of local linear
regression and principal components regression.  But the coverage is
incomplete because we only reconstruct those boxes that had a reasonable
independent verification correlation.  And this coverage varies through time
because the number of predictors (trees) decreases, and with less predictors
we can reconstruct less boxes with skillful verification.

Using the gridded fields, we can of course reconstruct many different
regions etc.  When we reconstruct the exact same regions for which we
already have regional averages, then we reproduce them reasonably well from
the gridded data.

So, for *our* reconstruction, the black lines give the average of our
gridded fields over various regions.  0-90 is an average over all our grid
boxes (since we don't try to reconstruct any southern hemisphere boxes), but
it isn't complete coverage since we only have some boxes to average
together, and as mentioned above the coverage also varies with time.  For
the REG-EM case, 0-70 is an average of all boxes (which has virtually
complete coverage - all but 6 boxes).  So this is interesting, but isn't a
like-with-like comparison.  0-70masked is an average of the REG-EM
reconstructions over all boxes, but after the same time-dependent mask is
applied to the REG-EM data as we have for our gridded fields.  So this is a
like-with-like comparison (because we have very few boxes north of 70N; if
we had more we'd have to truncate our fields at 70N, to avoid more bias).

land20-90 is then an average of all land boxes north of 20 in our fields,
with time-varying incomplete coverage.  land20-70 is average of all land
>20N from REG-EM, with complete and time-invariant coverage - not a
like-with-like comparison.  On this I think I overlaid the blue curve which
is our regional "NH" series.  A like-with-like comparison is possible by
applying our implicit time-varying mask to the REG-EM data and then sampling
land boxes north of 20N (i.e. land20-70masked cf. our land20-90).

Finally, treeboxes is when we average only those boxes that contain trees. 
But in our reconstruction this coverage still varies in time, whereas the
REG-EM treeboxes uses the time-invariant mask I sent to you - again, not a
fair comparison.  I overlaid in blue our "ALL" regional reconstruction here.
 treeboxesmasked is when I sample REG-EM at only the boxes with trees *and*
make the mask time-varying, so we can compare with our treeboxes series
(which already has the time-varying mask in it implicitly).

I think maybe the latter point is the confusing one - because our fields
have time-varying coverage anyway, there's no need to apply the mask to
them, so I didn't add the word "masked" to those series.  Perhaps I should
have called these series "masked", because the time-varying mask is there
implicitly.

Hope this long explanation is now completely clear!

Tim


