In the construction of his network, Loehle has done something very simple and very sensible that amazingly has never been done in a complete network in any previous temperature reconstruction. Something that neither bender nor JEG noticed; in fact even Loehle himself didn’t notice. Try to guess before you look at the answer.
To my knowledge, Loehle’s network is the first network to be constructed using series in which every proxy as input to the network has already been calibrated to temperature in a peer reviewed article. This is pretty amazing when you think about it. It’s actually breathtaking. Every prior network has included some, if not a majority, of uncalibrated proxies.
Here’s the distinction: proxy results from an original author sometimes come as temperature reconstructions or more often in native units, such as dO18, tree ring chronology units, pct G Bulloides etc. Ocean sediments are proportionally a greater part of the Loehle network than predecessor networks and here Stott and others calculate estimated SST in deg C often from Mg-Ca ratios, sometimes using other methods and it is these SST estimates that Loehle uses. By contrast, a Graybill tree ring chronology (for example) comes in dimensionless units and some ice core and coral studies are denominated by their original authors only in dO18 units. The distinction is not necessarily between ocean sediments and tree rings. Sometimes tree rings are used to make a temperature reconstruction (e.g. Tornetrask), but many important tree ring chronologies (and derivatives such as Mann’s PC1) have never been reduced to a temperature reconstruction. This can be important: there’s a reason why you’ve never seen a temperature reconstruction from Mann’s PC1 by itself – it would be too cold throughout its history, a point made long ago in MM2005 (EE), but insufficiently noted. In some networks, uncalibrated proxies were not necessarily even temperature proxies. MBH98 included instrumental precipitation series as “temperature” proxies, presuming that there would be some covariance somewhere with something.
I’ve done a mental survey of all the canonical predecessor networks (MBH, Jones et al 1998, Crowley and Lowery 2000, Briffa 2000, Briffa et al 2001, Esper et al 2002, Mann and Jones 2003, Moberg et al 2005, Rutherford et al 2005, Hegerl et al 2006, D’Arrigo et al 2006, Juckes et al 2007) and assure you that each one of them had a number of series that were in native currency, so to speak, i.e. in dO18 units, tree ring chronology units, etc. The only predecessor network that even had a high proportion of calibrated reconstructions was the low-freq portion of Moberg (which heavily overlaps the Loehle network.)
This selection criterion had an interesting knock-on effect for Loehle’s methodology, which accordingly seemed far too simplistic for JEG. In a CPS approach, popular with Team authors, the series will be standardized to a unit standard deviation, averaged and then re-scaled so that the standard deviation matches the standard deviation in the calibration period. (This latter can be expressed as a constrained regression.) Because the series are already in deg C, Loehle did not carry out either of the two re-scaling steps. Although JEG went ballistic about this, is there anything wrong with what Loehle did, given his network?
In practical terms, I doubt that the difference between the two methods amounts to a hill of beans. My guess is that the “topography” of Loehle’s reconstruction using CPS will be virtually identical to the series already calculated. Simply as a matter of prudence, I see no harm in doing a CPS calculation using Loehle’s network – the calculation is trivial – and that the calculation should be done to show the lack of difference simply because such CPS calculations are lingua franca in the trade. But I doubt that it will make a material difference to the result.
JEG vehemently criticized Loehle on this point and, in a way, it’s more interesting to try to understand exactly what underpins JEG’s vehemence. JEG:
This is a very different situation from usual multiproxy studies which use sophisticated methods to ensure that a proxy’s weight in the final result reflects its ability to record some variance in the temperature field (whether local or not). While there is merit in exploring a bare bones approach (arithmetic mean), it then becomes indispensable to demonstrate that each proxy is : a) a temperature proxy (not a salinity one…). b) a good one at that.
… Once again, such care would not be required when a climate-field-reconstruction or “composite-plus-scale” approach is employed, as the proxy’s ability to record temperature is implicit in the calibration therein. Since the author effectively treats the proxies are perfect thermometers (which is conceptually acceptable as long as it is explicitly justified), the lack of this discussion is unforgivable, and in my book, constitutes grounds for rejection any day of the week.
Or elsewhere JEG here :
either you care about LOCAL temperatures and you use something known as a Composite Plus Scale approach. Or you only care about the ability of a proxy series to record *some* climate information via teleconnections : that is the heart of climate-field reconstruction techniques, like MBH98 or the more sophisticated RegEM-based versions. … In either case you explicitly account for how proxies describe variance in the mean temperature ; either by regressing them against the mean, or against local temperature (and then average them), or against a subset of principal components that describe the large-scale features of the temperature field, then use that to reconstruct this linear subspace of the T field and then average it globally (the Mannian approach).
I think that JEG has got things backward here and that Loehle is actually on stronger ground than the Team on this particular issue.
The calibration of individual proxies to temperature and the demonstration of their validity is a critical issue in this area. Indeed, one of the persistent themes of this blog is precisely this: the failure by the Team to demonstrate the validity of key proxies like Graybill bristlecones or Arabian Sea G Bulloides. Because the Team has relied to a considerable extent on proxies that have not been individually calibrated in peer reviewed literature, the demonstration of the validity of any uncalibrated proxies should, in my opinion, be an item in the presentation of a multiproxy reconstruction; I think that it’s reasonable not to require this demonstration if the calibration has already been done in peer reviewed literature.
Emile-Geay argues that this demonstration of validity is unnecessary in Team articles because the “proxy’s ability to record temperature is implicit in the calibration”. This is simply untrue and completely neglects the pitfalls of spurious correlation and spurious covariance, pitfalls that can easily become worse in a data mining operation. I doubt that many MBH readers realized the degree to which they were relying on calibrations that had never been individually peer reviewed, and the degree to which the reconstruction relied on calibrations generated in a home made algorithm remote from scrutiny by peer reviewers. This problem extends to Mann et al 2007, where once again the same uncalibrated series are being calibrated in an algorithm whose workings are poorly understood, with the individual calibrations being remote from scrutiny by peer reviewers.
In contrast to the Team’s reliance on home-made calibrations remote from scrutiny by peer reviewers, Loehle relied exclusively on prior calibrations done in the sunlight of peer reviewed articles where the calibration could be individually scrutinized. JEG views this as a defect. I disagree – it’s an advantage.