Once again, the Team has “moved on” so quickly that it takes some care keeping track of their movements. The criticisms in my most recent post apply to the still unpublished Tingley and Huybers 1200-year reconstruction at their website (that it uses Mann’s PC1, a second strip bark foxtail series, Yamal plus a van Engelen series that even the IPCC acknowledged could not be used as a “proxy”). This reconstruction (let’s call it TH2009) is a typical small subset reconstruction (14 series), in which the primary issue is data snooping – the re-use of data sets with known and even stereotyped properties – issues that were raised in my previous post
As I noted in the update to that post, it appears that there is another unpublished Tingley and Huybers submission covering only the past 600 years that isn’t posted at Tingley’s website and that it is this other unpublished unposted submission that is featured by David Appell in this month’s Scientific American. This other reconstruction appears to be the one presented by Tingley and Huybers at the PAGES 2009 conference here. Let’s call this reconstruction Tingley-Huybers 2010. Here is a plot of the 600-year TH2010 reconstruction from the PAGES PPT (reshaped here to facilitate comparison with other sticks.)
The TH2010 Network
The TH 2010 network falls into a different “family” of reconstructions, using an entirely different proxy network and methodology than TH2009.
Let’s start by trying to figure out the network from the sketchy information available in the PAGES PPT presentation, namely the following location map and legend which states that a total of 118 proxy series were used in the reconstruction (96 tree ring MXD series, 7 ice core O18 isotope series and 13 varve thickness series) with the locations shown below.
From the pattern and count, the MXD version used here appears likely to be the gridded version of the Briffa-Schweingruber MXD data derived in Rutherford, Mann et al 2005 (also used in Mann et al 2008). For a long time, Briffa refused to disclose which sites were used in his various articles, but, as a result of prolonged quasi-litigation, this information became available in late 2008 in the wake of Mann et al 2008 and we have some dividends from this for TH2010.
The various MXD networks are described at a CRU webpage here. The locations of the 105 gridded series discussed in Rutherford Mann et al 2005 are here; there are precisely 96 series north of 45N and their locations match closely to the locations in the PPT location map as shown below. So for now, it’s a reasonable guess that TH2010 used the gridded MXD series of Rutherford Mann et al 2005 located north of 45N.
Figure 3. Emulation of TH PPT Location Map – see text for explanation.
Before we try to decode exactly how the fancy “new” methodology works, it’s always a useful precaution to show a simple average of the data for each class. A simple average of the 96 MXD series is shown below, showing the familiar “divergence problem”.
As a crosscheck on the above figure, the information webpage also identified 340 different MXD sites from which the 98 gridded series were derived. 330 of the 340 MXD sites have versions at ITRDB – a few series, mostly south of 45N are missing from ITRDB despite the CRU statement that all the series are at ITRDB. I also calculated a simple average of these 330 MXD series as archived at ITRDB yielding a similar looking graphic.
The MXD “divergence problem” has always been a problem in Team reconstructions and is once again in the Tingley-Huybers version. 96 MXD series out of a total of 116 proxy series in the Tingley-Huybers network go down, but the overall reconstruction goes up. Hmmmm.
Ice Core Isotopes
Kaufman et al 2009 recently reported on a network which included 7 ice core isotope records, shown in the above location map. 5 of 7 series seem to match TH locations, with TH apparently using a Mount Logan series (probably the old Holdsworth version) and a Penny Ice Cap, Baffin Island version, while not using two Greenland series used in Kaufman.
Kaufman refused to provide the supposedly “publicly available” data that he used, including certain annual ice core data that is not “publicly available”; my request for this data is currently under quasi-litigation at Sciencemag. In the meantime, the figure below shows the average of the seven Kaufman ice core series (decadal averages), which also go down. As discussed previously at CA, Fisher’s relatively recent Mount Logan ice core series (not included in the average shown below) also goes down in the 20th century.
Varve thickness is something that we’ve discussed in the context of Kaufman et al 2009, which uses 9 series that one can count as varve thickness. TH2010 report the use of 13 varve thickness series (tho I can only locate 12 on their location map: perhaps a couple of sites overlay.) The Alaska sites seem to match Kaufman’s Iceberg Lake and Blue Lake; both studies have two sites in Baffin Island, with Donard in common, but TH perhaps having a different site in southern Baffin Island rather than Kaufman’s Big Round Lake; TH have a site in Svalbard, while Kaufman has a site in Iceland. Both have sites in Finland – I wonder whether TH use upside-down Tiljander where narrower varves are interpreted as evidence of warmth? TH have 6 or 7 sites in the Arctic Islands versus 2 in Kaufman. We’ve discussed problems with some of these studies already: e.g. inhomogeneity at Iceberg Lake and upside-down Tiljander.
Reviewing the Network
Tingley and Huybers develop a relatively complicated multivariate to extract a signal from proxy data – the classic Mannomatic situation.
The raw materials for the TH2010 reconstruction have the opposite problem from Yamal and Mann’s PC1 – they mostly suffer from the divergence problem. The 96 MXD series (out of 116) go down in the last half of the 20th century. The average of 7 ice core series also go down. The Tiljander series in its recommended orientation goes “down” (not due to climate). The Iceberg Lake series goes up but is plagued by inhomogeneity.
Something else must be going on in the algorithm and it will take a while to sort through this new algorithm to see what makes it tick. Tingley has provided code for it, but hasn’t provided data. But before doing that, there’s one other aspect of the Tingley code that we need to consider. Tingley-Huybers also use 249 instrumental series. Tingley-Huybers (in their second methodological article) compare their method to RegEM. Maybe their method effectively splices an instrumental data blade with a nondescript proxy handle.
Otherwise, it’s hard to see how their method – Bayesian or otherwise – can get from the nondescript proxy network to a HS. I’ll collate and post up a network that is close to the Tingley network and maybe readers can analyse it with Tingley’s Matlab code.