Ammann, C. M. and Wahl, E. R. 2007. Importance of the geophysical context for statistical evaluation of climate reconstruction procedures. Clim. Change, doi: 10.1007/S10584-007-9276-X

Yes, i did note that, and i find it very entertaining that when Caspar Amman is struggling with publication deadlines, it deserves a “Where’s Caspar ?” entire thread, but that when Steve McIntyre is busy, his army stands still to hear what he has to say because they can’t form a judgment by themselves.
Julien

Julien entirely misses the point of this thread.

Ammann shouldn’t have been “struggling” with a publication deadline if his article had actually been accepted in February 2006 as he claimed to IPCC in order to make it eligible for citation in AR4. The issue raised in the Where’s Caspar post was how the published Wahl and Ammann would deal with citations to their rejected article – an issue raised a long time ago. The answer came soon after when we saw that the “accepted” article had been changed. The citations to the rejected article were changed to citations to an article accepted only in summer 2007 (at the same journal LOL).

The reason for the delay became clear: even a journal as compliant as Climatic Change probably drew a line in the sand at citing a rejected article (especially after we’d made it an issue here) and made Wahl and Ammann get the rejected results published somewhere. As it turned out, they had to publish the rejected results in Climatic Change itself. After they did this, the article originally “accepted” in February 2006 was actually published – after a publication delay almost unprecedented (or at least unusually long) in the recent history of the publishing journal.

If one examined the correspondence on this file, there’s little doubt in my mind that, after the rejection of the GRL article, CC’s acceptance of Wahl and Ammann became conditional on them getting the cited results into some journal or other and in the end, Stephen Schneider wasn’t going to embarrass the process and eventually had to hold his nose and publish the companion paper containing the rejected results himself.

Reply to Comments on Testing the Fidelity of Methods Used in Proxy-based Reconstructions of Past Climate

http://www.meteo.psu.edu/~mann/shared/articles/MRWAZoritaReplyJClimate07.pdf

Here’s one example (Fig 1) with SNR 0.4,

( http://signals.auditblogs.com/files/2007/09/pseudopr.png )

I wonder what verification R2 is in this case. ( I don’t know the answer to #3, just interested)

Take the SST raw data. Remove the Folland and Parker correction. Try again. That’s my advice.

JF

Are they really saying that it does not matter if the verification r2 is zero (series are orthogonal) if the mean is correct?!!?!

It is mostly panic-induced hand-waving, but let’s try to understand. First, if we assume that the reconstruction error is independent of the temperature, we’ll get a nice equation that relates (variance) Signal to Noise ratio and r2, r2=SNR/(SNR+1). Now, if signal variance goes down, but noise level doesn’t, r2 will be reduced. IOW:

very little change in energy balance actually happened

In addition, verification reference temperature is presumably a little bit noisier than calibration reference temperature. The problem is that signal variance doesn’t really go down that much, and r2 drops significantly ( 1820-1980 step from 0.76 to 0.06 *)

Another explanation would be that proxies first low-pass filter global temperature, then add independent noise to the result. See Wahl-Ammann Figure S1, middle. The problem here is twofold; the assumption is silly and high r2 during calibration remains unexplained..

* My emulation result, Wahl and Ammann preprint claims 0.76 to 0.2

I suppose they have a new reconstruction (shown on slides 31-32 here).

One symptom of overfitting, loss of r2 in verification is missing from the page 9

#29

They are similar but none of them match. It’s not just a linear ransformation as the correlations between versions are about 0.7 in nearly every case.

I just put that data through MBH machine, resulting 1000-1400 series has a correlation of 0.62 with the original (not exact correlation, but remarkable). Interestingly, in the calibration period, the correlation is 0.83.

Of course, the irony, as Jean S and UC realize, is that it doesn’t really matter what version you feed in to the Mannomatic.

That’s exactly what they mean by word robust.

It’s worse than ‘time varying’ though, isn’t it? I thought I saw a couple of references to Bristlecones being especially CO2 sensitive in the first place. Meaning they’re a truly daft choice for this particular problem.

Oh, I was only referring to that specific comment, i.e. “The Team” acknowledging the very thing that invalidates their method. By hypothesis, CO2 and temperature are correlated, but they are also inputs to tree growth, apparently moreso with BCPs. PCA is horribly unsuited to deal with correlated inputs. How on earth can you tell which of the two correlated inputs is generating the output? It’s absurd, absolutely absurd. Now we have Nychka actually publishing a paper in which he’s defending the very idea a panel he sat on refuted. What on earth was this guy thinking?

Mark