date: Tue, 15 Aug 2006 20:21:50 -0400
from: Gabi Hegerl <hegerlatXYZxyze.edu>
subject: Re: Figure 5, plus "executive summary"
to: Martin Juckes <m.n.juckesatXYZxyzac.uk>
First, many apologies for being so late!! (There is this four letter
word thats taking lots more of my
time than I anticipated... sorry...)
Martin, I just crosschecked your email and realized that you wanted this
this morning... sorry....
This is very very nice and useful paper, and I really enjoyed reading
the MM vs MBH discussion
and really liked it.
There are some things though that I am worried about that refer more to
the other techniques than MBH,
the comparison figure, and IPCC. A lot of it is quite self serving,
sorry for that, becuase of time constraints,
I plowed particularly into segments referring to IPCC or my stuff...
sorry for that!
The one aspect I worry about is that in figure 1, the CH-blend (cited as
hegerl et al., Tom would
prefer calling it CH-blend because he is the main recon guy, I am just
calibrating and detecting - but ok either
does not look the same as in our figure in the paper (figure 2). Also,
there is several versions, 3 segments
of different length, all using same amount of data, for land only,
30-90N, and 2 for land and ocean.
I am not sure which one you use, the one you show may be the land and
ocean one, while the one our figure 2
shows is the land only. I am happy showing either one,
it would be nice to use the full length, you can just use the "long"
version before the regular version starts.
However, the comparison shown in figure 1 is a bit misleading then,
since there are reconstructions for
land only extratropics (eg the boreholes) plotted into reconstrucitons
of land and ocean all NH (eg Moberg),
and all kinds of things in between (Briffa 2001 I think is growing
season land only, for example).
I think this is ok if its explained, but without explanation gives a
idea of the level of disagreement.
The boreholes for example are if compared over the same time and space
domains consistent with CH-blend
(and probably other high-variance recons too) but this is not clear from
the comparison in the figure.
This could be clarified by labelling what the different reconstructions
represent, and adding a sentence
to both caption and text discussing it that says
that part of the difference in amplitude is due to the difference in the
physical domain reconstructed.
I am also not quite sure what the reconstruction you show further down
(Union etc) does represent,
is it 0-90N land and ocean?
Another minor worry is that the discussion of reconstruction methods
other than MBH gives different results
from other work I am aware of, and I think it may be partly due to what
exactly is done. For example, the
total least squares approach is very similar to the inverse regression
approach, and can work pretty nice, at least does
in my paper and another work I heard about. I tried inverse regression,
too, and its very similar to total least
square, and worked very nice for me unless I tried calibrating the
records to local temps using it.
I think the one used here (please bear with me if I didnt read this
careful enough!! sorry in taht
case, I was trying to sneak this in between IPCC stuff, final draft
deadline approaching phew...) may not be
exactly the same as one that only calibrates a (weighted or not) average
of the paleo records "paleo" to the hemispheric
mean using inverse ols by saying paleo = beta * instrumental + noise,
and then using 1/beta.
If I am right here, it would be good to say so, if not, then maybe it
could be made a bit clearer. Tls might also
be mentioned as the case where noise in both is equally considered, but
the prize is more estimates / assumptions.
If you feel like trying out something, I would find it really
interesting to try using inverse regression and
variance matching in comparison for going from the composite to the
final reconstruction scaled to temperature.
This is where in my view tls or inverse regression is nice, since it
does not assume that the proxy based composite
has the same amount of noise as instrumental data, although the latter
are much more tightly sampled and
have much less non-temperature variability etc superimposed, all reasons
to think that they would have lots less
error. On the other hand, it could be that the errors are small relative
to the variability in temperature, in which case
both would perform similarly. The assumption that the proxy timeseries
should jiggle a little bit about the
(less noisy) instrumental timeseries due to its extra noise variance
seems like a good one to me...
But doing or not doing this is totally up to you!
So specifically to executive summaryES text sent out about a week ago:
1. The first page after "perturb our climate": It may be good to add
there something like:
Also, conclusions of the IPCC report (IPCC, 2001) that "most of the
observed warming over the last 50 years is likely to have been due to
the increase in greenhouse gas concentrations� were based nearly
entirely on studies analyzing the instrumental record and distinguishing
and estimating the signature of temperature response to greenhouse gas
Therefore, this conclusion is not affected by uncertainties in
reconstruction techniques and data (Mitchell et al., 2001).
Mitchell, J.F.B., D.J. Karoly, G.C. Hegerl, F.W. Zwiers, and J. Marengo,
2001: Detection of climate change and attribution of causes. In: Climate
Change 2001. The Scientific Basis. The Contribution of Working Group I
to the Third Assessment Report of the Intergovernmental Panel on Climate
Change. [J.T. Houghton, et al. (eds.)]. Cambridge University Press, New
York, NY, USA, pp. 695-738
The reason to add this would be that some part of this debate is trying
to sink all conclusions about greenhouse warming
with the hockeystick, which is a total stretch since the hockeystick
doesnt contribute that much to that conclusion
at all! Actually, nearly nothing!
2. Section 5 of ES: Are these techniques known in exactly the form they
are used here? It looks like the technique
known as inverse ols uses a scaling factor that varies from proxy series
to proxy series, and I suspect that the
usefulness of that technique depends very strongly on how its done. It
will do poorly if there is lots of noise (like
i tried calibration of local records to temperature) and will be better
if there is high correlation (eg if there is
already some proxy reconstruction that just needs to be matched to the
instrumental in amplitude, as in my paper,
in otehr words, if instead of variance matching, tls or inverse ols
would be done).
So can we qualify the end of the first para of section 5 ES by saying
using two techniques of different complexity,
The simpler method is known as "Composite... matching", is widely used
and gives robust results. The second
method is based on inverse regression. Here, the reconstruction and its
amplitude is determined using this method, and
this variant of inverse regression is found to be sensitive to give less
robust results than C...M.
This sensitivity is less important if there is relatively good
correlation between the target and the reconstruction timeseries, and in
those cases it has been shown to work well. In cases where the amount of
sampling and random noise on proxy data is substantially larger than on
instrumental data, inverse regression produces more reliable amplitude
estimates for the reconstruction, which can be important if the
amplitude of externally forced signals in proxy
reconstructions are used, for example, to estimate climate sensitivity.
(Last sentence is just a suggestion!)
3. Last section of last page in ES: I am not quite sure what the little
hitch about climate sensitivity is saying, but
a certain nature paepr that came out in April (sorry self serving)
showed that indeed the constraints from
data for the last few hundred years alone provide quite wide pdfs of
climate sensitivity. (Hegerl et al, nature, see
J Climate paper for reference).
If this information is combined with (also wide) pdfs for instrumental
data, where the problen is the poorly known
ocean heat uptake etc, our understanding of climate sensitivity can
narrow a bit. But thats not necessary to say here
unless you want, it may be a nice pitch for paleo to say taht if
instrumental estimates are combined with those from
proxy data, climate sensitivities outside the IPCC range of 1.5 to 4.5
become substantially less likely.
One could also add that a further interest for the proxy reconstructions
arises because the detection of greenhouse
warming depends on model estimates of climate natural variability.
Recent simulations of climate variability in
the last few millennia are generally in agreement with reconstructions
(e.g., Tett et al., others - can find) and
suggest that the level of natural variability produced by climate models
is reasonable (see also Zorita et al., Hegerl
et al., 2006b JCLimate paper).
MAIN PAPER (long version, very early August):
Abstract: see point 2 above, can this also qualify the inverse
regression a bit referring to one particular variant tested
p.3, see above suggestion 1, would be a good place to add this after the
first paragraph (...beyond dispute.)
p. 3, beginning of 3rd paragraph: Not sure I understand this!
p.4, beginning of last paragraph: HCA is not 0-90N, there is one version
tahts 30-90N, and one thats 30-90N land.
I am not srue if ECS is 0-90, I was under the impression its also
extratropical - do you use the cook-scaled version?
p. 6, footnote: I don't agree with the pitch for confidence level vs
likelyhood. The overall, expert assessment of
the likelyhood that something is true like the 90ies are the warmest
period in the millennium, includes both an
assessment of the robustness of the method, and of remaining
uncertainties. Confidence stuff is still not much used
in 4AR drafts at least of my chapter, we tried to use in one case and
reviewers didnt like it much.
Also, there could be the very confusing result of something being "very
likely" true with low confidence (which
may well be what you'd conclude from some paleo stuff), which is just
confusing. So teh expert assessment is
supposed to try to account for all remaining uncertainties.
p. 12: Wasn't there a detrending issue with Zorita et al, vs MBH?
p. 13, end of first para: If you want, you could add after
"under the assumption that the instrumental noise is known" the method
then accounts for the uncetainty due to unknown noise in proxy data.
(this is something taht not many seem to catch, so I try pitching in for
p. 14, detection stuff: We do multiregression, not correlations alone,
and we use response to forcings as
determined by an Energy Balance Model. (this is important to me,
volcanic forcing wouldnt correlate very
well with response). So say Hegerl et al. (2003, 2006b)
use a multiregression detection and attribution method to determine
the fingerprints of temperature response to solar, volcanic and
greenhouse gas forcing in a variety of reconstrucitons.
They find... variance, namely more than 50% of the decadal variance in
all records explored.
I need to check what Nanne did, but I think it was a bit different, I
remember it was a nice paper.
P15, end of first paragraph: This might be a nice place to refer to
Osborne et al., nature, for the unusualness of
the overall pattern of warming in the proxy records.
p. 16, top: Is the IPCC stuff that is referred to the finding that the
nineties are unusual, or that most of the observed
warming... greenhouse gases? I think the sceptics are (maybe on purpose)
murky about this, so we should be
very clear and separate between which finding is dicussed. So maybe
p. 18, 2nd last before section 4: typo, remove "and".
p. 21, top: Moberg et al., I THINK also find that the time around 16 was
colder than the early 19th, and
CH-blend suggests that also, although less clearly
p. 22, 2nd paragraph from bottom: Its fascinating that union has more
variance than the others - is there any explanation
available? Its really interesting!
This would also be a good place to crossreference that the INVR approach
used here is one realization of many possible
ways of doing inverse regression.
table 3 should get a qualifyer that CH-blend is an outlier because of
its lower (decadal) resolution.
If possble, it would be nice to explicitly mention it as a
reconstruction that uses the same number of records throughout
(at least the individual segments from it are).
p. 25, 2nd line: "this" reconstruciton is Union, right?
Greetings, I'd be happy to more thoroughly read this again, if you want
to - if already submitted, maybe
we can fix thse things later..
Martin Juckes wrote:
>As Anders pointed out, one curve is not visible in figure 5 -- this is because
>it is virtually indistinguishable from others: a better version of the figure
>is attached. I'll also modify the caption to give a more complete description
>of how each curve is generated.
>Also attached is an "Executive Summary" for the Netherlands Environment
>Assessment Agency. This is still a little rough in places, but any views on
>the opinions expressed and general layout would be welcome.
Division of Earth and Ocean Sciences,
Nicholas School for the Environment and Earth Sciences,
Duke University, Durham NC 27708
Ph: 919 684 6167, fax 684 5833
email: hegerlatXYZxyze.edu, http://www.env.duke.edu/faculty/bios/hegerl.html