Getting methodological information from Esper is a bit like dealing with Mann, a lot like dealing with Mann. It really makes me wonder whether there might be some clunker like Mann’s PC methodology lurking in Esper’s closet. Like Mann, instead of providing a comprehensive methodological desscription ideally with code as in econometrics journals, Esper would rather provide non-responsive answers. Right now, I have two outstanding methodological questions – one that I’ve been asking for a while: how he operationally allocates tree populations into “linear” and “nonlinear” trees; the other rose out of disclosure in February and March – not all trees were used from a site, so how did he decide which trees to use an which trees not to use.
Here were the questions that were put to Esper via Science:
a) In 4 cases (Athabaska, Jaemtland, Quebec, Zhaschiviersk), Esper’s site chronology says that not all of the data in the data set is used. This is not mentioned in the original article. What is the basis for de-selection of individual cores?
b) Esper et al.  do not provide a clear and operational definition distinguishing “linear” and “nonlinear” trees. As previously requested, could you please provide an operational definition of what they did, preferably with source code showing any differences in methodology.
Here is Esper’s non-responsive answer:
As described, in some of the sites we did not use all data. We did not remove single measurements, but clusters of series that had either significantly differing growth rates or differing age-related shapes, indicating that these trees represent a different population, and that combining these data in a single RCS run will result in a biased chronology. By the way, we excluded other sites because growth was too rapid, for example.
The split into linear and non-linear ring width series is shown in a supplementary figure accompanying the Science paper. The methods of this widely accepted, approach are described in the paper cited below and in the Science paper. It is possible to make this an operational approach, for example, by fitting growth curves to the single measurement series (e.g. straight line and negative exponential fits) and group the data accordingly. We didn’t do this in the Science paper, but rather investigated the data with respect to the meta information (i.e. for a particular site; data from living trees, and clusters of sub-fossil data), which I believe is a much stronger approach. This, however, requires experience with dendrochronological samplings and chronology development: Esper J, Cook ER, Krusic PJ, Peters K, Schweingruber FH (2003) Tests of the RCS method for preserving low-frequency variability in long tree-ring chronologies. Tree-Ring Research 59, 81-98.
First, consider Esper’s statement: “As described, in some of the sites we did not use all data.” I challenge anyone to locate any “description” or even hint in the four corners of Esper et al 2002 that they did not use all the data, let alone any reason for why they did not use all the data. There is no “description” or even hint in Esper et al 2002 that all the data was not used. The admission came only in response to my parsing through data that took nearly two years to get.
Esper now says that cores were de-selected to avoid a “biased chronology” and cited Esper et al 2003 as a suppposed authority for the procedure. However an examination of Esper et al 2003 provides no such authority. In fact, the closest thing in Esper et al 2003 to such a statement is the following, which I’ve quoted before:
Before venturing into the subject of sample depth and chronology quality, we state from the beginning, “more is always better”. However as we mentioned earlier on the subject of biological growth populations, this does not mean that one could not improve a chronology by reducing the number of series used if the purpose of removing samples is to enhance a desired signal. The ability to pick and choose which samples to use is an advantage unique to dendroclimatology.
Here Esper is talking about removing data to “enhance a desired signal”. Excuse me – that doesn’t sound like a way of avoiding a “biased chronology”; it sounds like a recipe for making biased chronologies – biased towards a “desired signal”. I think that readers are entitled to a better explanation of what Esper is doing.
As to the distinction between linear and nonlinear trees, it is simply not described in either publication. I challenge any of the people who usually disagree with me (Peter Hearnden, Steve Bloom, John Hunter) to read Esper et al 2003 and locate for me where this publication distinguishes between linear and nonlinear trees. The figure in the SI cited by Esper simply shows the number of “linear” and “nonlinear” trees. It does not explain how the distinction is made.
So how did Esper distinguish between linear and nonlinear trees? What effect does this classification have? An obvious question: if he didn’t make this distinction, does it affect relative MWP-modern levels? Did Esper exclude data to “enhance a desired signal”? If so, what was the “desired signal”? In fact, I’m not sure that these particular issues necessarily affect Esper’s results, but right now it’s impossible to replicate his calculations until one knows how these things were done. For all we know, maybe he used Mannian principal components.
There’s a little edge to Esper’s calculations because he has a high MWP as shown below.
The main reason why his MWP is high relative to usual Hockey Team fare is that he used the updated Polar Urals data (see my discussions of Briffa) which had elevated MWP values. This is the only Hockey Team study with updated Polar Urals data. Once Briffa realized the updated Polar Urals data set had a high MWP, he substituted the nearby Yamal data set which has a pronounced HS-shape and called it “Polar Urals”. The substituted data was quickly incorporated into subsequent HS studies – Mann and Jones 2003, Osborn and Brifa 2006, D’Arrigo et al 2006, etc.