Readers of this site are familiar with our concern over the use of bristlecones/foxtails in MBH98-99 and other multiproxy studies. The NAS Panel found in one place that "strip-bark samples" (which Graybill sought out in his bristlecone collections) should "not be used". They also reported that the MBH results were "strongly dependent" on "Great Basin data" – code for bristlecones and stated, somewhat inconsistently, that this should be reflected in confidence interval calculations. Because bristleconess/foxtails are used in many other studies (Crowley and Lowery 2000, Esper et al 2002, Cook et al 2004, Rutherford et al 2005, Moberg et al 2005, Osborn and Briffa 2006 and Hegerl et al 2006), this policy on bristlecones will have far-reaching consequences on the "other" studies, as I will demonstrate in due course (in a journal submission.)
For now, I’m going to review discusison of the first part of the issue – the validity of bristlecones as a temperature proxy – leading up to the NAS report. I’ll pick up the issue of the impact on MBH in another post.
We arrived at our understanding of the integral role of bristlecones in MBH98 rather indirectly. In Mann’s internet response to MM03, he identified different treatment of 3 series as leading to the different results: the North American PC1, the Stahle SWM PC1 and the Twisted Tree Heartrot Hill series. The latter two series didn’t matter, so the difference between results depended entirely on the North American PC1, which is why this has been such a battleground issue. For people who lack this context, re-read Mann’s 2003 response – see the links at the pdf category at right.
When the debate started, we knew only that we had been unable to replicate Mann’s North American PC1, but did not know about de-centering or bristlecones. In fact, there were a lot of inter-connected issues to disentangle with Mann’s PC series – the PC1 at the original URL was not only de-centered, it was spliced from 4 different steps. Not only that, they had incorrectly collated the data in the splice. After MM03, the MBH98 directory suddenly materialized in the public sector of Mann’s FTP site. We were able to quickly disentangle the splicing. The newly disclosed directory contained unspliced tree ring PCs (I note that no unspliced reconstructed PCs have been made available to this date). Mann condemned us for not replicating the previously undisclosed stepwise PC procedure, but refused to disclose the basis of determining the steps or retaining PCs.
Anyway when we replicated the stepwise procedure on the critical North American network, we still couldn’t replicate his results. It was really quite a puzzle. I started working through Mann’s iste, being thorough about these things, and noticed some Fortran code which proved to be his tree ring PC method. I transliterated the Fortran code into R, working through paragraph by paragraph, to try to reconcile results and, lo and behold, located a difference at the short-segment centering step. Thinking a little about the impact of short-segment centering, it was "obvious" that this would bias towards the weighting of HS-shaped series. We did some simulations (less sophisticated than the later simulations of our GRL article), wrote this up and sent an article to Nature, pointing out that one site (Sheep Mountain) had nearly 400 times the weight of Mayberry slough. At the time, we didn’t know anything very much about Sheep Mountain.
In his reply to our first Nature submission, Mann inadvertently opened up the bristlecone topic. In defence of his method, Mann said that, it wasn’t just Sheep Mountain that was highly weighted, there were 14 other sites that received significant weighting. So what were these sites? I look ed up the heavily weighted sites in Mann’s first eigenvector and then manually looked up the locations corresponding to the site identifications, ca534, ca530, ca529,… now all familiar id’s. The sites looked vaguely familiar – Sheep Mountain, Campito Mountain, Indian Garden, … I’d seen them somewhere before. Of course, it was in the Table from Graybill and Idso 1993 – Mann’s PC1 was, to all intents and purposes, Graybill and Idso’s strip-bark collection.
When the MBH98 directory was made public, there was a very puzzling directory entitled BACKTO_1400-CENSORED, you can imagine how that intrigued me. But there were no identifications to the sites or explanations of it. There were 20 fewer sites in the CENSORED directory (50 versus 70). Elsewhere there was a list of 212 sites and 192 "censored" sites. The 20 excluded sites were virtually all Graybill bristlecone sites. All reached back to 1400 and accounted for the difference between the CENSORED directory and the uncensored directory – Mann had done a sensitivity study on the effect of excluding bristlecones. As we’ve reproted, the PCs from this study did not have a HS-despite the data mining tendency of the MBH algorithm. (This is important to keep in mind – Mannian PCs do not ALWAYS yield a HS.)
While the Mannian PC method was biased, it was the interaction with the bristlecones that appeared to make it especially problematic. Also, even if the Mannian method was a lousy method for obtaining a climate signal, it was useful for identifying outliers. The other thing that seemed obvious: if the temperature history of the world supposedly turns on a few groves of bristlecones in the US Southwest, shouldn’t one have encyclopedic information ont hem? This is what engineering-level analysis would require.
Our GRL paper already covered other topics and was too space-constrained to provide a thorough discussion of bristlecones. There, we merely noted that the MBH98 PC1 was dominated by bristlecones and referred to our E&E article for a more extended discussion:
McIntyre and McKitrick [EE 2005] discuss, inter alia, problems relating to the interpretation of bristlecone/foxtail pine growth as a temperature proxy. (MM05a)
Our 2005 E&E article provided a fairly extended survey of different issues pertaining to bristlecone growth — and there were many interesting and intriguing aspects to bristlecones. Again, it’s worth re-reading the relevant section of our E&E article. We pointed out that Graybill had specifically said that he looked for strip-bark samples, which had a different chronology than full-bark samples:
Graybill specifically sought out strip bark samples and reported that strip bark forms had much stronger 20th century growth than entire bark forms at the same site [Graybill and Idso, 1993]. More recently, Bunn et al. [2003] confirmed higher growth in strip bark forms than entire bark forms at sites in Montana…. Brooks et al. [1996] also pointed out the impact of anthropogenic nitrogen on fertilization of high-altitude bristlecone pines, (EE, 81)
We also had a nice quotation from MBH coauthor Hughes, saying that 20th century bristlecone growth was a “mystery”. We missed finding a quote from NAS panelistBiondi in Biondi et al. [1999] (including MBH coauthor Hughes), but presented this in our NAS presentation:
The average of those sites [a network of high-elevation temperature-sensitive tree-ring sites in the Great Basin and Sierra Nevada of Hughes and Funkhouser, unpublished], plotted in Figure 5, is based on many ring-width series, each one being 500 years or longer, without individual growth surges or suppressions and from "strip-bark" five-needle upper forest border pines of great age. Such record is not a reliable temperature proxy for the last 150 years as it shows an increasing trend in about 1850 that has been attributed to atmospheric CO2 fertilization [Graybill and Idso, 1993]
While there has been much blog vituperation directed at us onother topics, there has been surprisingly little realclimate effort to show that bristlecones were a temperature proxy. The best attempt to justify bristlecones has been by Rob Wilson in a posting at climateaudit here . I engaged in a short dialogue with Rasmus, another realclimate coauthor — we’re dealing with an entire Hockey Team — at Roger Pielke’s blog, in which Rasmus said that citing Graybill and Idso was “no more reassuring than blaming the Flying Spaghetti Monster”
Rasmus: Two short comments: (i) The sentence ‘bristlecones, a series known to be potentially contaminated as a temperature proxy’ has been proposed by McKitric and MacIntyre (MM), but this claim has not, to my mind, been backed up by convincing evidence. Is it really known? How certain is this? …
Steve: 1) Rasmus, look at our paper in E&E for a discussion of bristlecones. Graybill and Idso, who published the chronologies which dominate the MBH98 PC1, stated that their 20th century growth pulse was not due to temperature. Even Hughes has said that the growth pulse is a "mystery" – a position hardly consistent with attributing them as a unique detector of world climate history….
Rasmus: Statement: ‘Graybill and Idso, who published the chronologies which dominate the MBH98 PC1, stated that their 20th century growth pulse was not due to temperature.’
Question: what is the evidence for this – in your own words if you please. Why? To just cite Graybill and Idso is not much more reassuring than blaming the Flying Spaghetti Monster…….
Rasmus: When I do a Google on ‘Bristlecone temperature’, I get 27500 hits, but when I include ‘flawed’ in the searxch string, I only get 297 (your name keeps cropping up…). Why such a difference if your claim about their quality were true?
I attempted to document my answer. Some of you may be interested in the dialogue, which went on for a while and gives a flavor of realclimate author debating techniques.
realclimate
realclimate has discussed the issue only slightly. One of their replies was that they had fully discussed the matter in MBH99, carrying out an adjustment which compensated for CO2 fertilization. We’ve pointed out elsewhere (including our NAS presentation) that this “adjustment” was not applied to MBH98 values; that the adjustment petered out at the end of the 19th century and did not affect the 20th century pulse that concerned Lamarche et al 1984 and Graybill and Idso 1993 , and that it remained very unsatisfactory that results should be critically dependent on flawed data about which you did not know whether the adjustment was valid or not.
The other realclimate reply was that the additional use of bristlecones improved the RE statistic (they used the phrase “verification statistics”, but it’s really just the RE statistic) in an MBH98 model. This is their primary position and one expounded in Wahl and Ammann:
Particular concerns with the "bristlecone pine" data were addressed in the followup paper MBH99 but the fact remains that including these data improves the statistical validation over the 19th Century period and they therefore should be included.
Note that the justification is not that the proxies have any physical validity – only that they improve the RE statistic. OK, but so do Dot.com Stock Prices, so this isn’t a good way of determining statistical validity.
Wahl and Ammann 2006
Wahl and Ammann 2006 is more or less an extended version of positions staked out at realclimate. The mention "bristlecones" 36 times. They acknowledge both that the North American PC1 is "a crucial source of information" (i.e. the reconstruction is not robust) and that PC1 is an alter ego for bristlecone/foxtails:
The temporal information captured by PC1 of the North American tree ring network is a crucial source of information in calibration for the 11th-14th century reconstruction (MBH99), and in verification for the early 15th century reconstruction in MBH98, as shown here (cf. section 3.2). A further aspect of this critique is that the single-bladed hockey stick shape in proxy PC summaries for North America is carried disproportionately by a relatively small subset (15) of proxy records derived from bristlecone/foxtail pines in the western United States, which the authors mention as being subject to question in the literature as local/regional temperature proxies after approximately 1850 (cf. MM05a/b; Hughes and Funkhauser, 2003; MBH99; Graybill and Idso, 1993). (p.9)
They point out that MBH methods do not require that bristlecones correlate with gridcell temperature, only that they correlate with a "climate field" somewhere in the world. They argue that because their information is "necessary" to increasing RE scores. See for example:
The failure to verify by scenario 5d, including only two PCs derived from unstandardized data in the MM centering convention, demonstrates that the bristlecone/foxtail pine records are important for the 1400-1449 network–the information they add to PC4 in this way of calculating the PCs is necessary for a meaningful climate reconstruction. Restricting the PCs in MM05a/b to only the first two (5d) indirectly omits the information carried by the bristlecone/foxtail pine records and thereby leads to a non-meaningful reconstruction. (p. 32)
Or again the following:
Given these observations, from a climate reconstruction point of view one can argue that, in general, the bristlecone/foxtail pine records do not introduce spurious noise and their inclusion is justifiable; or said more strongly, their elimination is not objectively justifiable. Their inclusion by standardization of the individual proxy records (independent of the centering convention) or, even if non-standardized series are applied, by using at least four PCs (until the resulting climate reconstructions converge), leads to reconstruction models that demonstrate skill in both calibration and independent verification. (p 34)
Or one more time:
The bristlecone/foxtail pine proxies from the Southwestern United States are shown to add necessary verification skill to the climate reconstructions for 1400-1449 when PC summaries are used and significantly greater verification skill to the reconstructions for 1400-1499 when no PC summaries are used–indicating that in these cases the records carry important climate information at the level of eigenvector patterns in global surface temperatures. These results are valid notwithstanding issues concerning these proxies’ empirical relationship to local/regional surface temperatures after 1850, noted by MBH in previous work (MBH99; cf. MM05a/b; Hughes and Funkhouser, 2003; Graybill and Idso, 1993). These results enhance the validity of the MBH assumption that proxies used in the reconstruction process do not necessarily need to be closely related to local/regional surface temperatures, as long as they register climatic variations that are linked to the empirical patterns of the global temperature field that the MBH method (and other climate field reconstructions) target. (p 36)
Note that none of these arguments rise above saying that they need the bristlecones to improve the RE score; they do not attempt to justify bristlecones as a temperature proxy through direct analysis. I’ll get to RE scores in another post – for now, please note that the RE benchmark that they use is 0.0 – a benchmark that we contest.
M&M NAS Pesentation
In Section 3.6 of our NAS Presentation, we discussed issues relating to bristlecones and it’s worth re-reading. It includes a reference to Biondi et al, 1999, not in our E&E discussion.
NAS Panel
NAS considered bristlecones in Chapter 4 — Tree Rings and Chapter 11 — Multiproxy Reconstructions and, while their recommendations were not entirely consistent, both have important implications for MBH and other reconstructions. In Chapter 4, they did not cite our E&E article (in fact, the article was strangely not cited at all by NAS), but their citations correspond to the citations in our E&E 2005 article and our NAS presentation. They stated
The possibility that increasing tree ring widths in modern times might be driven by increasing atmospheric carbon dioxide (CO2) concentrations, rather than increasing temperatures, was first proposed by LaMarche et al. (1984) for bristlecone pines (Pinus longaeva) in the White Mountains of California. In old age, these trees can assume a “stripbark” form, characterized by a band of trunk that remains alive and continues to grow after the rest of the stem has died. Such trees are sensitive to higher atmospheric CO2 concentrations (Graybill and Idso 1993), possibly because of greater water-use efficiency (Knapp et al. 2001, Bunn et al. 2003) or different carbon partitioning among tree parts (Tang et al. 1999)”Åstrip-bark’ samples should be avoided for temperature reconstructions, attention should also be paid to the confounding effects of anthropogenic nitrogen deposition (Vitousek et al. 1997), since the nutrient conditions of the soil determine wood growth response to increased atmospheric CO2 (Kostiainen et al. 2004).
While the publicity has attached to their use in MBH, they are also used in Crowley and Lowery 2000, Esper et al 2002, Cook et al 2004, Rutherford et al 2005, Moberg et al 2005, Osborn and Briffa 2006 and Hegerl et al 2006. Implementation of the above policy on bristlecones will have an impact past MBH.
In Chapter 11, the NAS Panel did refer to our mention of bristlecones as follows:
A second area of criticism focuses on statistical validation and robustness. McIntyre and McKitrick (2003, 2005a,b) question the choice and application of statistical methods, notably principal component analysis; the metric used in the validation step of the reconstruction exercise; and the selection of proxies, especially the bristlecone pine data used in some of the original temperature reconstruction studies. These and other criticisms, explored briefly in the remainder of this chapter, raised concerns that led to new research and ongoing efforts to improve how surface temperature reconstructions are performed. (106)
In chapter 11, the NAS Panel acknowledges the "strong dependence" of MBH on "data from the Great Basin region" and, somewhat inconsistently, merely says that this needs to be considered in the calculation of statistical uncertainties, although the Statistics chapter, again inconsistently, does not include a discussion of how these calculations should be implemented.
The more important aspect of this criticism [about principal components] is the issue of robustness with respect to the choice of proxies used in the reconstruction. For periods prior to the 16th century, the Mann et al. (1999) reconstruction that uses this particular principal component analysis technique is strongly dependent on data from the Great Basin region in the western United States. Such issues of robustness need to be taken into account in estimates of statistical uncertainties (106)
Ironically, they even cite Wahl and Ammann 2006 as authority for the fact that exclusion of subsets can have an impact on results. This is entirely in keeping with our interpretation of Wahl and Ammann under the rhetoric – that they confirm the non-robustness that we had previously reported (as they had confirmed the failed verification statistics that we previously reported). Are our feelings hurt that the NAS panel attributed these damaging claims about MBh to Wahl and Ammann as opposed to us? I guess we will manage to bear the pain.
While they accept Wahl and Ammann as authority for non-robustness, they squarely turn away from the argument of Wahl and Ammann 2006 – that the use of the proxies is justified by the increase of RE scores. In Chapter 4, they recommend not using the proxies altogether. In Chapter 11, they say that this would increase uncertainties. One of the impacts is that you get high 15th century results without bristlecones. In the discussion of this topic, the NAS Panel palced wieght on the fact that one could "get" a HS without using PCs, but, in my opinion, did not fully turn their minds to the role of bristlecones and a couple of other problematic series, to which I’ll return in another post.
I can see nothing about our position on bristlecones with which the NAS Panel disagreed and I see a repudiation of the Wahl and Ammann/realclimate justification for their inclusion as being the “correct” reconstruction on the one hand, and even for including them at all. Their findings here (as in the Statistics chapter which I’ll discuss on another occasion) are fundamental to considering the “other” studies.
Climate Audit 
53 Comments
Great, Steve. This reads like a mystery novel. Maybe you could re-read Raymond Chandler and have a tall blonde knock on your door one day to give you a tip to check out MBH’s Fortran code. But don’t have her in any way connected to Exxon.
I look forward to reading your journal submission. And by the way, what is this “Great Basin” stuff? Are they just avoiding the fact that it’s almost all California, i.e. in one state? Is Sheep Mountain part of the “Great Basin?” I’ll try to look it up, but if any geography experts want to comment, I would appreciate it.
Best of luck with this effort.
Wahl and Ammann 2006 is really funny. They have essentially said: “We gotta have these trees to make it work.” How could they have thought they would get away with this?
In reading through the quotes above, I was struck by the following:
I’ve always been bothered by this assumption in the absence of good evidence (or maybe even model evidence 🙂 ) for such a teleconnection. Reading it again here, it seems that such an assumption is really equivalent to performing multiple tests of correlation of a proxy vs. temperature at a location (e.g. testing bristlecones not just against local temp, but against Seattle temp, Denver temp, Okeefenokee temp, Sol Dav temp, etc). Because of this multiple testing, I would think that significance levels would have to be tightened up in proportion to the total number of temp locations in the analysis. Has this been discussed before on the site or by any of the researchers, or does the analysis already account for such an effect?
I think it has been discussed. At least, I recall Steve M. discussing the lack of gridcell correlation in a few threads. This, besides the lack of stationarity, is a key problem with correlating tree-rings against temperature.
Mark
RE: #1 – Technically, the following parts of California are part of the Great Basin – the thin strip north of Tahoe along the Nevada border, east of the Warner Mts / Sierra Nevada; the Tahoe Basin; the intermontaine valleys and basins south of Tahoe and East of the Sierra crest (e.g. Kirkwood, Priest Valley, etc); the Mono basin; the Owens Valley; the White and Panamint Mts; Death Valley; the Mojave Desert. All of these areas are characterized by extensional tectonics and high geological heat flow and fit in with the rest of the Great Basin. All of the Bristlecone sites are there, and all the Foxtail ones are either there or at the Sierra crest, along its western border.
Yes, the lack of gridcell correlation has been discussed, and MBH’s response (as in the quote above) has been that that isn’t necessary, since there may be teleconnections. My question is, assuming that MBH’s teleconnections are valid, don’t MBH then need to tighten up their significance levels to account for what, in effect, seems to be multiple tests of correlation?
Re#2, This is my favorite W&A joke, care of MM http://data.climateaudit.org/pdf/reply.ammann.pdf . It never gets old.
Re: #3
That’s because that statement of Mann is transparent bollocks. Yes, I’m using the climatological term for it. If we’ve ever hit a case of “faith-based” climate science then that is it.
Re:#8
Sorry, John, I’ll have to play TCO here and mark your response as “nonresponsive.” 🙂
LOL. Saying you can correlate the proxies with any gridcell is like saying you can use tree rings to predict stock prices, providing you select the appropriate trees for correlation over a certain period.
Teleconnections? Is this like Ouija or tree seance or Gaia? I suppose it is some technical term that I do not understand.
RE: #11 – Maybe it is related to remote viewing and lifeline analysis. Hey, who wrote “The Global Superstorm” after all? LOL!
Didn’t they say that reconstructions pre-1600 have unquantified levels of uncertainty ? Do you think that is one of the consequences of their conclusion on stip-bark pines ?
It is certainly a fairly harsh judgement on poor old MBH 🙂
cheers
per
Re:#10
No, the claim would be that, given the lack of a “true” reading for temps right at the tree site, perhaps there are other sites on earth that (because of the structure of the climate system) have temps that correlate with the tree site. As an extremely simplified example, insolation may be highly correlated between distant sites, and less so with a site in-between.
re: #11 and others,
It’s possible to imagine a real “teleconnection” of temperature as long as there’s another medium to supply the connection. The obvious candidate is rainfall. Thus if there’s a warming in an area and this leads to additional evaporation and then rainfall, this additional rainfall, whether local or regional or even global, might well result in larger rings in trees far from the warmed area. Of course it might also result in smaller rings if a larger snowpack and larger melt-time resulted in a shorter growing season.
But I think Armand is quite correct in any case that such a situation would result in an altered significance level. The linear-relationship sort of thing that’s assumed by Mann et. al. just won’t stand such a complicated analysis.
14. That’s one hell of a stretch, don’t ya think?
WARNING, EDITORIAL TO FOLLOW:
1. Steve, while the Team has many distressing flaws in their arguments, you are also not completely forthright. In many cases, you say things without actually taking a position. I’m pretty wise to this and see the subtlety…but many of your readers are not. I think to a certain extent the NAS even falls in here in how they “went a bridge further than you did”. Instead of cackling at the support, you might want to point out the difference!
That’s why I “smoked you out” on covariance and correlation matrices. You had made a bunch of pro-covariance and anti-correlation matrix statements. However, when pushed we see that your comments were selective–you were not standing behind one option as the better choice. Similarly on bristlecones, I think NAS has fallen into the trap that some of your readers do of “going a bridge further than you do”. I note that you’ve been very careful in discussion of CO2 fertilized bristlecones. Some of your readers think (and have told me so) that you are espousing that the ‘cones are contaminated. Actually you are not. It’s just a “concern to look at”. Instead of being so happy about the NAS as some sort of group that sanctions anti-bristlecone remarks, you ought to more thinkingly look at what they said and if you endorse it all and what basis they have for their judgements.
2. The methodology does not “interact” with the contaminated proxies. It (arguably) magnifies signals of a CERTAIN SHAPE. It’s irrelevant if that shape is caused by contamination or not. The issues of proxy relevance (as a valid thermometer) and methodology of regression ARE separable. I’m honestly not clear to what extent your issue here is sloppy thinking and to what extent it is a form of dishonesty (maybe a form that you self-justify…but still a form). On some days, I think heck…the guy is so damn smart–it must be dishonesty. On other days, I think, heck…that poster and powerpoint were very disorganized…perhaps its a logic gap. I wonder.
3. I actually understand and agree with the comments from Rasmus about that you should elucidate arguments yourself and not just appeal to authority. For instance your citation of people’s comments wrt “units” differences and other things in the covariance/correlation choice is tendentious. You know perfectly well that the issue is mathematical and that units MAY be a reason for the decision but that scale of variance versus the variable of interest is also relevant. But you didn’t say it. I think you know how to think about this issue from first principles and instead you picked the comments from authorities that helped your case (selective). That was not kosher. It was sophistry.
In addition, you have a pernicious habit of discussing PC1 versus the overall reconstruction. This is very misleading. Don’t blather to me about the EE paper. The EE paper was fine. Your other discussions are sophistry and overstating the case and misleading remarks made in a method so that you can still defend the error with the caveats. Just stop doing it! It’s bad science. Bad science is wrong. Read the Feynman article. Live those precepts. Don’t live the Mann precepts!
Steve:
Please stop responding to TCO.
It isn’t so much a matter of “don’t feed the trolls” as it is “don’t feed the troll wannabes.”
And PC1 is a selective amalgam of parts of the proxies. It’s not a brainwave that it is going to magnify some series contributions relevant to others. If it didn’t, it would be the AVERAGE. Sheesh!!!
TCO says
I do not friggin’ believe this guy
#11 Teleconnections was an approach that emerged in the 1980s and supposed you could anticpate events in one part of the world by measuring events in a distant location. Mann appears to be arguing that he can do this between adjoining gridcells. Maybe! The problem with teleconnections is the same as it is for comparison of any climate records. On a global scale and over a long time it might be true. For example, the entire world is now in an interglacial, but on a shorter time scale and in different regions the onset and termination vary considerably. You can see this problem with dates of onset and termination or even intensity of the Medieval Warm Period and Little Ice Age. I think there are problems with Mann’s assumption even for the length of record and the apparent proximity of the cells. Of course, there is little he can do about the lack of adequate data in space or time but I thnk there are better ways of accommodating the limitations. For example, the concept of relative homogeneity is a standard technique when comparing two station records, or in this case adjoining gridcells. This requires you identify and separate as much as possible what changes in each record are functions of local effects and which are those experinced by both sites due to larger scale changes. Consider the problem in comparing records from two sites when there is a climate change that causes one station to now receive more winds that now cross a body of water before arriving at the station. You record a temperature now modified by the water, but unless you also study the wind trajectory change and eliminate it as a cause of temperature change you are in danger of attributing the change to something else; such as a global warming.
Teleconnections and climate field response while interesting, seem very susceptible to data mining. The whole look at the one out of 20 models that has 95% significance. 🙂 I’m not a digital thinker, so I don’t want to shriek and say no one can ever do any work with these types of signals. Just that we should be very wary of this issue from these types of methods and have some systems/assurances that persuade us that the mining is not occuring as above.
Re:#23
Hence my original question in #3 about whether this had been dealt with or even discussed.
PS – what the heck is a “digital thinker”? If it means you’re an AI, that could explain a lot… 🙂
24: Agreed, these are points that I and we have made before. I was particularly distressed by the rather blithe off-hand remark by Bloomfield (name?) that one could deal with the Burger and cubasch issue of huge variations of the models (in methods) by picking one that worked the best by observation in the instrumental period. That sounds RIPE, RIPE, RIPE for the one out of 20 style error. I wonder if this guy is really an astute Tukey style thinker. Someone who you go to help you do a proper DOE in a tricky situation and who can watch out for logical flaws in assumptions. Or is he just a jock of a technician who knows one particular area and who tends to plug and chug without senior scientist physical insight.
To the folks commenting on the plausability of teleconnections: Recall the post (http://www.climateaudit.org/?p=697#more-697) that Steve put up just a few weeks ago on bristlecones. As I understand what he said there, when he regressed the foxtail series on the relevant grid cell temperature he found the “r2 is 0.002; the adjusted r2 is -0.006; the t-statistic for proxy on temperature is an insignificant 0.52 (without even worrying about autocorrelation) and the DW is 0.986.” How can the foxtails be a temperature proxy when they have this relationship to the local grid cell temperature?
The critical quote above on the mystical “global temperature field”, is:
“These results enhance the validity of the MBH assumption that proxies used in the reconstruction process do not necessarily need to be closely related to local/regional surface temperatures, as long as they register climatic variations that are linked to the empirical patterns of the global temperature field that the MBH method (and other climate field reconstructions) target.”
How can this be anything other than circular reasoning? The “global tempertaure field” that “the MBH method targets” is essentially the bristlecones. Of course the bristlecones have a high correlation with this target. How is that evidence that a mystical global temperature field” has been uncovered?
re: #23
TCO, I suggest that this post could have been limited to:
This was my thought on first reading about teleconnections here.
re: # 17
You should have concluded this post after:
We have heard the remainder of your message many times before and it becomes no more persuasive or clear on repetition.
Reminds me of the old joke about the prisoners who had told the same jokes so many times that they didn’t go to the trouble of actually repeating them each time. They just assigned each one a number and so someone would call out a number and everyone would laugh. One day a new prisoner came into the block and after being told how things worked he decided to try it out. “17,” he said, but nobody laughed. “What happened?” he asked one of the older prisioners. “Your timing was off,” he replied.
Perhaps we need to assign TCO some numbers for his usual laments.
Re:#26
Peter, their wording is rather dense. It’s not the bristlecones until *after* the reconstruction; the calibration is done on actual temperature records.
TCO, as I keep pointing out, you don’t sem to understand PCA at all, and in fact, you’re making a fool of yourself. I suggested a week or so ago that you actually familiarise yourself with the technique by taking some data and doing a few runs with different options (it’s easy), but you brushed me off. It’s very difficult to respond to some of you comments because they simply don’t make any sense. You seem to think that’s some sort of “gotcha”.
As for your covariance/correlation matrix issues, both Steve and Ross have adequately dealt with your comments. You seem to expect them to revisit statistical issues that were discussed and resolved decades ago. And at the same time you ignore the fact that results should be robust to either application if they’re any good.
And please stop calling on Steve M. for confounding PC1 with the reconstruction. The reconstruction is driven by the bristlecones whether they’re PC1 or PC4. The regression step doesn’t care. You seem to a smart guy, I don’t understand why you don’t see this?
Re #29 Armand
The actual temperature records are Phil Jones’ global composite, a secret blend of fudge and chicanery that no-one, outside of a select few in East Anglia, knows what’s in it.
Re #31, John A
John, is this secret blend also the source of the oft-quoted 0.6 C global temperature rise over the 20th century ?
Re #32
Yep. Nobody knows what Jones is doing to the data, as Warwick Hughes will testify when he tried to find out.
If there was another Steve McIntyre, I’d suggest the temperature record would be a good place to investigate.
Re: #29
OK — so the claim is that foxtails grow better in response to “urban heat island” temperatures and temperature changes in the irrigated central valley of California (see the Christy analysis in the post referred to above) and other such human influences on the Jones data (including data processing errors) than they do in response to temperatures in the grid cell in which they are actually found. How can anyone seriously propose that as a scientific hypothesis?
27-28: Good points.
30: My issue with the PC1 description is that it is an overstatement, by Steve, of impact of the off-centering in MBH, since it is an intermediate result, since other PCs (and non-PCs) are combined into the reconstruction, and since it is not just the bizarre (off-centering) that drive the mining, but standardization and PCA itself.
Man .. this is a great site! I am a researcher in Biology, but have gotten caught up in the global warming debate. I’m sure I’ve read on this site that the Bristlecone data do not correlate with the temperature gridcell that they are associated with.??, yet they are used to predict temperature in models. How is it that Climate Science can violate this very basic scientific princiipal regarding proxy data? It would seem that the entire scientific community would rise up and demand better practices, as it dilutes the validity of science as a whole. Where are all the authors of previous climate predictions and publications that were dismissed upon the publishing of Mann et al.?
I don’t think it’s a situation of competing statistical estimates. I think it’s a bit unfortunate (argumentative, advocate-style) when Ross or the like talk about the “previous consensus”. Yeah, Mann sucks as a study. But the previous work did too. I don’t see Ross or Steve lining up behind it as something that they think was correct. Their position is really more agnostic. Not that they believe in an MBH. Just that it is impossible to know one way or the other. Also, that one temperature picture in the 1995 IPCC is a joke. It’s a cartoon. And I don’t buy the “Mann had to be special to overturn the old consensus” argument. Mann sux. But so did the old stuff. So it could be overturned by demonstration of the flimsiness of statistical significance also.
36: welcome to the zoo. It is truly unbelievable, as you will find out if you read old posts or stick around.
TCO, you say:
“30: My issue with the PC1 description is that it is an overstatement, by Steve, of impact of the off-centering in MBH, since it is an intermediate result, since other PCs (and non-PCs) are combined into the reconstruction, and since it is not just the bizarre (off-centering) that drive the mining, but standardization and PCA itself.”
This is a perfect example of my earlier comment that your charges simply don’t make any sense (what are the “non-PCs”, BTW?) Further, you seem to think that there is no importance associated with the ordering of the PCs. The PCs are ordered according to amount of variance explained. Just because the regression step doesn’t care abut the ordering doesn’t mean there is no information there. Why not include PC9?
Re:#31
I know, John, I’m just playing Devil’s Advocate here. Perhaps you could add a little “devil” emoticon?
Re:#34
Peter, I’m not sure exactly what’s been proposed wrt teleconnections, but you *can* make plausible (that word again!) scientific arguments for them. Dave made one earlier on here, essentially that normal processes might convert a distant temperature signal into a local moisture signal. My thinking has been more along the lines that even within-gridcell temp records are possibly (likely) different from temps at a tree, given the large extent of a gridcell and varying topography. Thus, it may happen that another recording station is in a microenvironment more similar to that of the tree, and thus its temperature recordings correlate better with the tree’s rings.
The difficulty, of course, is if researchers just mention “teleconnections”, without investigating whether they are individually plausible and without statistically accounting for the multiple testing they imply.
To amplify what James Lane has said (#39), Ross has written a very clear explanation (sans linear algebra derivations) of PCA that is available on his site.
He shows that the off-centered NOAMER series did indeed have a very significant effect on MBH’s results, and that the hockey stick goes away without it.
As I understand the events, Mann then said that with proper centering you could get a hockey stick in the PC4.
This is nonsense. You run one regression and get a PC1. The variables are then run on the residuals to get a PC2. And on to PC3 and PC4. Which is by definition of fourth-order importance. And the PC4 may have some “significant” coefficients, but we know that based on how the steps were run, if you tested the PC4 against the original data instead of on the residuals of residuals of… there would be virtually nothing.
So Steve is entirely right to talk about the off-centering.
Re: #37, TCO says
wrong. The old “consensus” was not based on a cartoon. It was bassed on scores (today hundreds) of studies that showed that there was an event known as the MWP, and another event known as the LIA, and that these were, until proven otherwise, global events. Mann was attempting to overturn that consensus, and for a time he did. He should go down in history alongside the discoverer of Piltdown Man.
Dear readers,
everyone who is experienced with editing Wikipedia should try to fix the recent version of
http://en.wikipedia.org/wiki/Hockey_stick_graph
that claims that the NAS has just reinforced the insight that 1998 is the hottest in millenium etc. This just can’t be there.
If you are not sure how a more objective description works, look for the last version by Lumidek in history.
Best
Lubos
1. From a reconstruction standpoint in MBH, the different retained PCs have equal weighting. What this means is that if PC1 is egregious, but PC1 and PC2 are both retained, PC2 DILIUTES the impact of the egregious PC1. When Steve talks only about the mining in PC1 (or PC4 in the Colorodo post below), it is an overstatement of the problem. It is also rather imprecise. What if the PC2 had a very strong negative HS index? It might even be possible to have some example where PC1 was more then compensated by PC2. So that by looking at PC1 changes for a picture of how the method impacts the overall, you actuall get the opposite inference of what you should. Note: I’m not arguing that this DOES happen. I’m just pointing out the possible mishaps in looking at PC1 (intermediate, partial) vice reconstruction (final, what we care about).
I agree with your concern about the PC9. In the sense that PCA discards some of the PCs, it is weighting (like a 0 versus 1). I have not yet read a good validation of the rationale for PCA and throwing out some of the PCs in estimation problems. This is a general concern of mine. Of course, though, this is an aside and not related to the points about dilution.
In some cases, Steve (or was it Ross) have made the point that the PC1 got it’s own picture in the article and got some special discussion in the text. However, they’ve also pointed out that ALL RETAINED PCs are treated equally in the reconstruction. From a mathematical point, that’s what matters (retained or not retained). Sure, there is some special rhetorical attachment to the PC1. And if Steve and Ross want to criticize the PC1 in that aspect fine. What I have a problem with is misleading conflation of PC1 itself for the reconstruction. John Cross gigged Steve for this a while ago. So did Rasmus. And when I asked about it, the failure to answer (until I really pushed) seemed duplicious.
2. My understanding (from this blog, but if I got it wrong, please correct me, Steve) is that the reconstruction itself includes the PCs as major components of the reconstruction along with some regional averages or other very important individual proxies. So the impact of the PC1 on the reconstruction is diluted, not only by the other PCs, but by the non-PCs.
3. We’ve already beat the centering and the dividing by standard deviation thing to death. I don’t have a problem with examining all of this stuff. Complaining about it. Whatever. I do have a problem with ascribing the impact of standard deviation division to “off-centering”. I even have a problem with being non-precise and changing two variables at the same time (unless you show the full factorial) and then making comments about the impact of one of the variables or of both being seperate flaws in method. (If you only do the “change two variables at one time test”, then I guess you could make some point about the “two together” being a “single flaw” (not two separate flaws!)–but really why would you do something so ugly!?) You need to have enough equations for the variables in algebriac solution of simultaneious equations! Let’s be very clear and crisp.
4. A very simple numerical question is: given that I change the centering convention only*, how much does that change the hockey stick index of the final reconstruction? I think you will find the answer is SIGNIFICANTLY less then the change in the index of the PC1. Therefore, when Steve refers to PC1 instead of reconstruction, he is overstating the impact. I’m not the first to note this, btw. It’s a consistent criticism. I almost wonder if he has his own CENSORED directory where he did the exact calculation that I asked for, but failed to show it. For the sake of his ethics, I hope he doesn’t. For the sake of his brains, I hope he does!
5. I appreciate your belief that I could master the PCA methods and math. However, my arguments here are all matters of logic and of how the results of the PCA are used in the overall reconstruction. I may not have a law degree, but I can still weigh in and ask the relevant questions of the company lawyer. Same with the accountant. Same with the market researcher. Etc. Etc. That’s part of being a critical thinker. You can ask the hard questions and expose flawed thinking even in areas, that you don’t understand gnat’s ass. That’s why we went to college. That’s why we were junior officers.
If you ask the question and find out that you had a misunderstanding, fine. But you’d be amazed how many times, it’s not a misunderstanding and the expert (while knowing his topic better then you) has missed something. And the same thing of course applies in reverse. When I present a business plan and have it “murder boarded”, people point out flaws in my logic. Nothing wrong with that. That’s why Steve putting his thinking in writing and in real “play” would make both make it better and push others’ work to be better as well.
My old executive officer used to say that you should listen to an explanation and it may seem like a big ball of yarn, but with a little string hanging out the end. He said go ahead and be a pain and yank on that thread. You’ll probably get some further explanation (maybe even a dismissive one). But go ahead and give it a couple more yanks with followup. You’d be amazed how many times, the ball finally comes apart. And if it doesn’t…you learn something!
*By that I mean changing the mean used for normalization from the 20th century period to the entire period. If the standard deviation used was also over the wrong period, then we should change that at the same time. Basically we are fixing “wrong periodness” which was the thing that was distincitively Mannian and what has been thrown in his face. But we don’t change the aspect of dividing by a standard deviaton. Or if we want to, then we need to do a burger and cubash flavor style analysis or we need to make an argued position for why the covariance matrix is preferred over the correlation (and it’s still a seperate issue, so it’s one flaw at a time).
Re #41 Yes, but they are not talking about a better correlation to the temperature record of another particular grid cell. Neither do they give a physical explanation (of the sort you offer) for why that particular grid cell temperature might be better correlated with growth of these tree rings. They just say that the bristlecone tree rings reflect “temperatures in general” even though their unusual growth does not reflect temperatures in the neighborhood. Also, an argument like “normal processes might convert a distant temperature signal into a local moisture signal” would be testable by showing how the tree ring growth is highly correlated with local precipitation records, and then those records are in turn highly correlated with the global temperature signal. Your argument that “within-gridcell temp records are possibly (likely) different from temps at a tree, given the large extent of a gridcell and varying topography” might be tested by examining correlations with temperatures from different weather stations in the same grid cell. I seem to recall Steve had another post along those lines some time ago. I think he examined stations on different sides of a local ridge or something like that.
The whole “teleconnections” subterfuge is just stupid in this case. It’s merely another excuse for cherry-picking data.
I think some people over use teleconnections in trying to make sense of temperature variabilities in widely seperated regions. The most famous teleconnection, the El Nino Southern Oscillation covers such vast territory, and produces such extreme variabilites in temps, cloud cover, and precip, that it can cover almost any variability. The same holds true for the North Atlantic Oscillation. While smaller in size, it has also been blamed for everything from floods, blizzards, droughts, windstorms, to heatwaves. There are 3 other teleconnections. So if one teleconnection doesn’t fit the bill, there are 4 others that probably will.
Re: #46
Nice post, Peter! With regards to #41, I was just trying to point out that one *can* make a scientific argument for it, not that such an argument was correct in this case (or had even been made by Mann et al).
Re 44 (Lubos):
To those of us who unfortunately can’t access any URL that has “Hockey” (employer’s internet blocker thinks it’s a sports site), your link redirects to this article. The article currently includes this remarkable paragraph: “The report also confirmed the major points of the criticism by M&M: the statistical significance of the conclusions about the climate before 1600 is low; the bristlecone pines are not a good temperature proxy; the data and the software should have been made available; and the principal component analysis was not used properly.”
RE 46:
Peter H says:
Yeah, there’s a thread here from many months ago where a Bristlecone site had upper & lower populations sampled, and the ring-width trend was almost opposite between the two elevations. On the same mountain! ‘Course, the upturning-trend samples were chosen for the reconstructions, not the downturning ones.
I took a drive up to the Ancient Bristlecone Pine Forest in the White Mts yesterday to
get some idea of their environment. This group of BCPs is located east of the town of
Big Pine, CA. along hwy 395 about a gazillion miles north of anywhere. The White Mtns are
themselves ancient, an old sea bed that has been uplifted 9-10 kft, folded, and worn down
with time. The BCPs themselves have staked out a patch of the high mountains above 9,000 ft
elevation up to the treeline, which is about 11.5 kft. I noted the old 1973 report by Lamarche
that Steve referenced in another thread on tree lines. These old guys, I believe he was a
geologist, are very good at describing locations, environments. There is recent geologic
activity in the area, several volcanoes located south of Big Pine, but I didnt see much
volcanic activity up near the BCPs. The White Mts run North-South and run parallel to the
larger Sierra Nevadas to the west. The two are separated by the Owens Valley which is at about
4,000 ft ASL. Hwy 168 heads east out of Big Pine on a good, paved two-lane road out through
sagebrush as it climbs into the White Mtns. The road progresses for 13 miles up to about the
7,000 ft level before forking north on the way to Schulman Grove. At 6,500 ft level you run
into a nice mixed forest of Pinyon Pines and Junipers, with a few cedars. The road heads
north up along the spine of the White Mtn. Around the 9,000 ft level the vegetation starts to
thin out as the road starts to cling to the west side of the mountain. There is a great 270 deg
view at Crestview of the Sierras and Owens valley to the west, the desert lands to the east
and the spine of the Whites to the south. At this point there is a giant, steep canyon that
falls off to the west that terminates in the valley below about mid-way between Big Pine
and Bishop, just due east of the CAL TECH radio telescopes down in the valley. The mountain
ahead is really sparsly populated with vegetation, mostly the usual high desert scrub. Schulman
Grove of ancient BCPs is really a bench in the mountain. At the visitor center there are
two large hills north and south full of BCPs. The trees also grow out to the southeast around
the back of the mountain on the lee side of the hill, protected from the prevailing west winds.
The BCPs at Schulman grove include downed trees, some young growth, and trees with and without
strip-bark. The trees at the higher elevations in Sheep mtn and the older ones at Schulman
Grove along the Methusulah trail tend to be stripbark. I saw lots of two and three-trunk trees
some splitting right at ground level. Not sure how you would core those properly. I was surprised at the leaf structure of the trees since they looked more like monkey-puzzle trees to me. There is a noticable difference between them and the Pinyons down below, with which I am
more familiar. As I have said, there is lots of downed tress in this area, you would have material
to date further in time using the dead trees. That is something I noticed out in the National
Forest land below; you dont see much downed wood since it is permissable to collect fire wood.
You would have a hard time trying to extend Pinyon forest timelines back into the past, unless
you found buried logs. On the hill coming into to Schulman Grove a few hundred feet from the
road is an old mine adit and log structure. Sorry, the miners were here first! Didn’t see hide not hair of any living thing other than BCPs, low desert scruff and a few humans. Didn’t bring
Boxer the dog since I didn’t want him to unknowingly offend some 4,000 yr old tree. The view
from the top of the mountain is spectacular. You must be able to see several hundred miles
of the higher Sierra Nevada to the west including Mt. Williamson at 14,000 plus change. My
interest has been wetted to read up a bit more about these trees. One thing I did find out was that there are dead BCP logs up above the present tree line suggesting a much warmer condition
than today earlier in the Holocene.
Some additional comments on BCPs at Schulman Grove. The BCPs here appear to have plenty of
spacing between trees, don’t expect much competition for sunlight. The north-facing hill
is likely to be colder than h in winter, while the south-facing slope will get plenty of
sunshine, except at the base. The mountain pass at Schulman Grove means that there will be
some high winds in this area at times since this causes a natural venturi effect, much like
Tehachapi pass further south. I did not see the typical northwest lean that we see in trees in
the desert due to the prevailing southwest winds. Due to the shallow root system you might
find trees competing for water. I did not see any patches of snow or any streams in the area.
On the other side of the valley, the high Sierras continued to have high altitude glaciers
and runnoff at this late date. Sierra glaciers are north or northeast facing. Great potential
for rock slides in this area due to the broken rock (dolomite) and the steep hills. Expect
damage to the uphill side of trees. Right at Schulman Grove the BCPs are limited in how far
they can advance up the hills in warmer conditions since right now trees are maxed out at the
top of the hills. I believe at Sheep Mt and Campito trees have the opportunity to advance up the
mountains in warmer climates, and there is evidence of this due to dead wood 150 m above the
present treeline. The road to those locations is a good dirt road that heads out due north from
Schulman Grove. You would think that the oldest and most important trees in the world would have a corresponding weather station nearby to record the current wx conditions. I have found data for a wx station called White Mountain # 1 and # 2 but the records only go from
about 1958 to 1980. There may be more data available since the University of California
operates the Barcroft high altitude station up north of here, but I couldn’t find any data.
I next found data from Bishop Airport just down the hill from 1948 to present, and sure enough,
that data correlates more with the lower treeline ring widths here at Schulman Grove than at
the upper tree line BCPs that was found in Salzer et al Nov 2009 that is here on another
thread. The later shows a steep uptick in growth lately. In that report I notice they dont
specify which Wx station is used as their reference and they construct a model to simulate
the temperature at the Sheep Mt altitude.