Rob Wilson has referred us to Wilson et al 2007. In addition to being an example of site selection, Wilson et al 2007 uses a type of principal components on a tree ring network – something that should be of interest to many CA readers – and an interesting illustration of non-Mannian statistical methods within the tree ring community.
Update (2016): In comments to this post, there was considerable criticism of the failure of the authors to archive data at the time of publication, with one of the authors taking offence. Unfortunately, as critics had feared, the missing measurement data was not archived until five years later (2012), with one site-species dataset still not archived.
The sites in Wilson et al 2007 are located along the Gulf of Alaska, well to the south of the Brooks Range at about 68N where Wilmking collected samples. On the left is a location map from the article; on the right is a location map within Alaska.
Right panel – red shows sites with archived data.
The selection protocol is described as follows:
For this study, we selected from the region 31 ring-width data-sets, that came up to at least 1986, from the international tree-ring data-bank (http://www.ncdc.noaa.gov/paleo/treering.html) as well as newly sampled data-sets for further investigation (Fig. 1; Table 1).
Wilson et al do not provide ITRDB identification numbers – something that should be part of any SI – but do provide names and coordinates for all 31 sites considered. From this information, I was able to locate 16 series within the ITRDB, none more recent than 1995, and was unable to locate 15 other series, which are presumably unarchived. In the map at right above, the located sites are marked in red. While the statement of selection criterion suggests that the data was selected from ITRDB data, my inspection of the data provenance indicates that unless co-author Wiles was involved in the collection, the site was not included even if it was in ITRDB. So this is really a publication of Wiles’ collection rather than a statistical sampling from ITRDB – which is fair enough, Wiles did a lot of work collecting the data. (One of the problems for data collectors in archiving – perhaps the major problem – is getting scooped by multiproxy scavengers – that Mann or Crowley or that type would steal the limelight.)
Here’s something a little ironic in the selection of sites, that I can’t help but tease Rob about. Rob objected elsewhere at CA to using data collected by Wilmking, who he described as “dendroecologist” who was interested in positive and negative responders.
This whole blog [thread] started with me criticising Steve M about averaging 14 sites without expressing where the sites came from and why they were originally sampled. Your 30 chronology mean is no different I am afraid. Not all the data are from the Jacoby group. Many of these series are from Wilmking. Again, I do not know the details of the individual sites, but Wilmking is a dendroecologist and has been study the divergence’ issue in recent years with respect to positive and negative responder trees.
I didn’t think that his criticism was well-founded, but here’s the irony: four of the sites in Rob’s list – Middle Telaquama, Fish Trap, Lower Twin, Portage Lake – are reported in Driscoll et al 2005,, also see ppt here, coauthored by Wilmking, which was previously discussed in CA in the post Positive and Negative Responders and are type examples for Wilmking’s positive and negative responders. (Also see previous discussion of Wilmking et al 2004 here .) Since Rob did not cite Driscoll et al 2005, he was presumably unaware of this connection.
First, Rob reduced the population from 31 sites to 22 sites, eliminating sites with a low correlation to temperature. The population was originally 31 sites, consisting of 18 TSME (mountain hemlock), 8 PCSI (Sitka spruce), 4 PCGL (white spruce) and 1 (CHNO) yellow cedar. 3 of the 4 white spruce sites were eliminated in this stage. The only remaining white spruce site, Middle Telaquama, one of the Driscoll sites, has a reported correlation of -0.01 with the other series. I consider this elimination step as part of the multivariate methodology since it gives a coefficient of 0 to these 9 sites in the reconstruction.
After reducing the population to 22 sites, Rob carried out a type of principal components analysis on the remaining 22 site chronologies described as follows:
A rotated (varimax) PC analysis (Richman 1986) using the remaining 22 chronologies identified five principal components (PCs) with an eigenvalue greater than unity.
I’ve collated the network of 16 sites for which data is available here and have experimented a little with it. Obviously without all 31 sites being archived, it is impossible to replicate any of the actual reported results.
Given all the publicity about correlation versus covariance PCs – an issue which I’ve always regarded as a distraction as the two methods merely provide alternative weighting of site chronologies, neither of which is “right” in any meaningful way as far as I can tell – climateaudit readers will probably wonder which Rob chose. It’s hard to say from the description. The “rotated (varimax) PC analysis” appears to be the same procedure as factor analysis . My guess is that he used something like the factanal function in R with the default option of rotation=”varimax” , which also has covariance matrix default but a correlation matrix option. It’s hard to say and, in the absence of a complete data archive, it is impossible to replicate calculations to find out. Yes, it’s possible to email Rob and ask, but a good SI would have all this information and ideally the source code for the calculations.
Some readers will probably wonder what the difference is between the “rotated (varimax) PCs” and correlation/covariance PCs. The difference, as I understand it presently, is as follows: if you do a PC analysis, you often end up with weightings that are hard to interpret (for tree ring networks, dare I say, this is often because the PCs probably don’t mean anything very much at all.) If (say) 5 PCs are selected (truncation) – 5 is the number chosen in Wilson et al without saying why, factor analysis (varimax rotation) attempts to find an orthogonal rotation within the subspace that loads the weights as much as possible on groups. In the Wilson et al case, if you re-inspect the left panel in the first figure above, you’ll see 5 different colors – these represent the 5 groupings from the varimax rotation. In this case, “rotated (varimax) PCs” boil down to being a high-falutin way of simply making 5 “regional” averages. In the case of the Mannian network, while I haven’t tried a varimax rotation, it seems highly likely to me that it would find the bristlecones as a distinct pattern and load them as a distinct factor.
Rob then regressed temperature against the first 5 rotated PC series. (He considers both current and lag-one temperatures as regressands which I won’t discuss in this quick review).. January-September average temperature was selected as a target (presumably others were examined.) This resulted in the following formula:
Let’s leave the lag-one issue to the side for a moment and reflect on what’s being done here. In effect, Rob is doing a multiple inverse regression on 5 weighted averages of the proxies (with the weightings chosen to be orthogonal.) If you do the linear algebra in a different order (and the operations are obviously associative) and can be done in a different order, you could reduce Rob’s expression shown above to”
where is not the OLS coefficients, but a different set of coefficients from this procedure and is the data matrix.
The obvious question in any of these sorts of analyses is whether there is any justification for not just using an average with at most some regional weighting. This issue was raised by the NAS Panel (and discussed by both Huybers and ourselves.) However, Wilson et al don’t reflect on the matter. The trouble, of course, is that, in the absence of any theoretical reasons why one site chronology within the group should be better than another site chronology, the fitting exercises carried out here seem to me to be simply exercises in overfitting and, if an r2 or other statistic arises from this sort of method, I’m inclined to be mistrustful of it if it can’t be substantiated in a simple average.
In passing, the underlying statistical literature on factor analysis (Ripley has his usual excellent discussion in his text), the inspiration comes from psychological testing, where testers try to associate skills into patterns “linguistic”, “mathematical”. As practiced by tree ringers, it is a type of inverse factor analysis. I’m having trouble thinking that there’s any particular theoretical justification for the procedure or any reason to adopt it, but, at the end of the day, it appears to simply generate regional averages and may not do any immediate harm. The regression step seems more problematic to me – I really have trouble with the idea of multiple inverse regression of temperature against 5 tree ring indexes. I realize that this just assigns weights, but I find it hard to think of any reason not to prefer averaging. Weighting by calculating regression coefficients seems ripe for overfitting.
In any event, here is the Wilson et al temperature reconstruction. Once again, I have trouble discerning a HS shape in this reconstruction. I’m also a little puzzled as to why Rob and Mike Pisaric were so quick to conclude that an average of white spruce sites was “flawed!” given what appear to be some commonality to this plot.
As an exercise, I calculated a simple average of the 16 chronologies where data is available. Yeah, yeah, I’m sure that it is somehow “wrong” to simply use data that is available and I’m sure that this analysis is “flawed!” in Pisaric’s phrase, maybe even “flawed!!!!”, but here it is anyway. I’ve only shown values since 1500 matching what Rob showed, but the values as plotted in the early portion are very elevated. The latest value in this average is 1995. The Wilson results shown above are not archived but, guessing, I’d say that the 1995 uptick in the average corresponds to the last uptick in the series (near where the solid smooth ends) and that values have declined since 1995. In both cases, one has rather elevated values in the 1920s, 1930s and early 1940s. The last strong peak (value-1469) is in 1944. There is a marked low in 1974. The simple average of these chronologies ends on a somewhat lower noted than the [arguably overfitted] version in Wilson et al 2007, but the impression of the two versions is obviously similar.
What does it all mean? Well, there are some interesting points of commonality between the multivariate methodology of Wilson et al 2007 and MBH98. Both carry out principal components analyses of tree ring networks – MBH using Mannian principal components; Wilson et al 2007 using rotated varimax PCs. In each case, the PC analysis is followed by a regression analysis against temperature (or temperature-equivalent PCs), in Rob’s case with an OLS regression; in Mann’s case with a PLS regression (though obviously not described as such in MBH or anywhere other than CA). Mann also throws in a pile of other “proxies” in addition to the tree ring PCs. Rob also uses a stepwise procedure, as does MBH. It’s interesting to make these comparisons to see exactly what lies within tree ring traditions and what were Mannian innovations praised by Bradley for enhancing the detection of a “faint signal”.
But always, always – keep firmly in mind that these operations are all linear and at the end of the day, weights are assigned to the individual proxies and ask yourself whether there is any valid reason to weight one proxy more than another.