Why am I not surprised?

]]>I finally read the presentation. It is a very nice presentation in general. Some very pretty slides at the beggining that deliver a great story and allow you to talk without too much words on the ppt. Just using a visual aid.

Further on, there are some that are too technical (in keeping with audience and rest of talk).

Also, it seems it would take 3 hours to cover all this. In all seriousness, what was your time limit?

There is also one slide where you talk about PCA “mining for hockey sticks”. Since you have called the recon itself, “the hockey stick” and have been criticized for muddling pc1 versus recon to overemphasize the impact of mining, you really ought to stop such remarks, Steve. It is a slight dishonesty. I would say it to your face…

]]>The limiting case for the temp reconstruction is no proxies. So if one assumes the NH temperature series is ergodic gaussian wide sense stationary, the optimal estimate for the years 1500-1850AD is the sample mean (-.17C) of the 150 samples (1850-2000AD) of existing instrumental data with an estimation error variance equal to the sample variance for the 150 years of data (~.25C^2). Noting that this optimal estimate uses equations 1-3 from T. Schneider JC 2001 paper which I have seen called the LS method.

Now if one adds a standardized proxy with low r, than equation 1 justs adds some wiggles to the hockeystick and reduces the estimation error variance to (.25^2)*(1-r^2). Just a few thoughts.

]]>I have a hunch that proxy-based reconstruction is highly proxy-dependent, and that that is why the regional reconstructions tend to be fairly orthogonal and thus preserved through the PCA process. If the signal-carriers were true proxies, they should be highly non-orthogonal, i.e. correlated, under global cliamte change. That they’re not tells you something. Either the climate signals are truly regional (and there are millions of subtle teleconnections all around the globe), or the contamination is strong, such that all you can extract are regionally impure signals; the pure global (or hemispheric-scale) signal is non-extractable.

Remember that even the very best proxies have relatively low skill. I have a hunch that that’s the hole that the MWP has disappeared into.

]]>Then my #68 take is quite different, but still relevant. Each proxy type is going to be uniquely subject to its own peculiar brand of bias/error/contamination. If proxies vary from one data location to the next, then you have no way of separating those errors from the regional climatic signal. They become part of the regional signal. Whereas the intent of PCA is to factor out the common signal as large PCs, relegating the bias/error/contamination to small PCs. If there’s no spatial intermingling of proxy types (because the grid is sparse and proxy types are regionally unique), that isn’t going to happen. PCA becomes trivialized.

I’m sure the multiproxy people have figured this out – it’s kind of obvious. There’s not much you can do about it – except use caution in interpreting results. Which doesn’t necessarily happen when your modus operandi is alarmism.

]]>Now the assumption that there is a “signal” in the majority of MBH proxies is not a correct assumption. If you take the non-BCP proxies in the early network and do MBH regression (partial least squares), you don’t out-perform a similar operation using low-order red noise series. You get calibration r2 over 0.5; verification r2 of ~0.

If you add 1-2 HS shaped series (BCP plus Gaspe) into MBH regression, you get a tuned reconstruction with a HS shape. The bend in the HS is not as big as the bend in the PC1 but the PC1s tend to “overshoot” in RE terms. So by blending the HS with noise, you can tune the reconstruction to fit the NH temperature history and get a high RE result. So if you mix 1-2 synthetic HS series with networks of low-order red noise, you get the MBH verification statistic package – high calibration r2, verification r2 of ~0 and RE of about 0.4 or so.

MBH regression on large networks is an amusing little can of worms that I’ve alluded to here and really need to write up formally. IT comes nicely into focus in the Wahl and Ammann variation of MBH where (without knowing that this is what they’re doing) they do partial least squares regression of one temperature series of length 79 on 95 series with minimal common signal. Wahl and Ammann argue that their “fit” from this procedure entitles them to extend the proxies using the instrumental record. It’s pretty comic once you see what they’re doing. The Team are always full of such jolly pranks.

]]>If that’s what you’re saying, I agree. But I don’t think that’s what Steve is referring to. Steve doesn’t mention regression at all in the bit about PCA non-robustness. (Let’s let him answer. Maybe you’re right.)

(You know I was joking about the AUS bcps, right? ok. You never know.)

]]>I’d love to see some of those Australian bristlecone pines! There are some old trees in Tasmania, but not in the league of the BCPs. Perhaps TCO would like to come down and section them!

]]>