Comments on: Gajewski, Mann and Warm Ice Age Arizona

By: richardT

richardT — Sat, 26 May 2007 18:30:23 +0000

#54
There are several implementation of MAT available, if the same options are chosen they should all give the same result. The main differences are in the variety of distance metrics available, and how the number of best analogues is selected, and in the amount and type of diagnostics available. MAT is easy to implement, the only problem comes with dealing with large training sets on computers with a finite amount of memory (the distance matrix created is huge). I can send you my code if you want it.
The modern analogue technique is based on the premise that the modern sample with the most similar biota to a fossil sample has the most similar environment. There are many distance metrics that could be used to measure this similarity (the squared chord distance is the most popular). Gavin Simpson’s analogue package on CRAN offers the widest choice of metrics (these are all implemented in R, rather than in C, so it runs slowly, and I doubt it would process the ~5000 sample North American pollen training set).
While there are other methods, normally the same number of analogues are used for all fossil samples, and this number, typically 5-10, is selected to minimise the training set’s cross-validation error (a procedure that can seriously underestimate both the error and the optimal number of analogues if the training set environment is spatially structured). Viau et al is unusual in that it uses only a single analogue, and interpolates the biota between the pollen samples, rather than making reconstructions directly from the raw data, and interpolating these temperatures.

There are some ‘extensions’ of MAT, called RAM and SIMMAX, devised and used mainly by the foram community that I would not recommend (see Paleoceanography vol 19, PA4014).

I have not worked with the N. American Pollen Database – I intend to, but am still testing the performance of the training set. The data collated before 2000 are available in a Paradox database, so can, in principle, be read into R with RODBC.

Sawada’s code is at http://www.geomatics.uottawa.ca/mattools/ but the package appears to have been compiled for an old version of R.

By: Steve McIntyre

Steve McIntyre — Sat, 26 May 2007 11:03:58 +0000

#53. Richard, from my own experience with other proxies, I agree 100% with your choice if it’s a beauty contest between the two alternatives. There’s an R package for analogue techniques by another person here
http://cran.r-project.org/doc/packages/analogue.pdf , but I didn’t notice Sawada’s code being online anywhere in a first google. Would you be interested in writing a post on what the differences are between the various implementations? Also do you have a digital listing of site identifications so that one could emulate the reconstructions in an automated way (i.e. an R script reading data from WDCP – a procedure that does not cause consternation there)?

By: richardT

richardT — Sat, 26 May 2007 08:39:26 +0000

#52
Most of the reconstructions are probably not archived. The fossil pollen and modern training set are the important things to archive – in this case they all are. The reconstructions can easily calculated from these using the Modern Analogue Technique.
I am far happier with this situation than the contary, where only the reconstruction is archived, but not the raw data. I would not want to use someone else’s reconstruction – I always want to check the reconstruction diagnostics.
Sawada, one of the authors on this paper, has published R code for calculating MAT (I’ve not used his code as I have my own implementation)
Sawada M. 2006, An open source implementation of the Modern Analog Technique (MAT) within the R computing environment, Computers & Geosciences, 32: 818-833.

By: Earle Williams

Earle Williams — Fri, 25 May 2007 21:36:02 +0000

Re #50

RichardT,

The NAPD provides pollen counts and chronologies (where they exist) for an impressive number of locations. Are you aware of any archive (NAPD or otherwise) that shows paleotemperature reconstructions for the individual sites used in Viau et al, 2006?

http://www.ncdc.noaa.gov/paleo/napd.html

P.S. Which are the sites used in the compilation?

By: Steve McIntyre

Steve McIntyre — Fri, 25 May 2007 20:32:24 +0000

#51. What I meant by this is that that the individual site time series constructed by Viau Gajewski are not archived as a cross-check against this. You’re right about the pollen data (or at least most of it). I’ll amend this.

By: richardT

richardT — Fri, 25 May 2007 19:00:26 +0000

Steve,

You write that

The underlying data sets are not archived.

This is not correct. The caption to figure 1 in the paper states that the fossil data all come from the North American Pollen Database. This dataset is freely available to download. Pollen analysts have worked hard to archive their data, and deserve credit for doing so.

By: Armand MacMurray

Armand MacMurray — Thu, 24 May 2007 10:44:03 +0000

Re #44

I notice that 1000 years during the MWP that spruce percentages in the Arctic islands were 5-20% and were up to 5% in west Greenland – in locations that are modern 0%.

According to the legend, white areas represent “no data”, not 0%. Based on that, I don’t see any significant parts of the areas you mention that are green (>5%) at 1000bp, but gray (0-5%) on the modern map. You would think that the sites should have modern counts reported — perhaps it’s a case of “bring the pollen up to date!” (I found the “display sites” and “compare images” settings very useful on this)

By: richardT

richardT — Thu, 24 May 2007 05:51:23 +0000

#44
Long distance transport of wind blown pollen into areas of low pollen productivity (such as arctic tundra) is a well known problem in palynology. The modern analogue technique uses the taxonomic distance between the fossil sample and the samples in the modern training set to find the best match. Because this uses the entire assemblage, rather than taxon-wise, it is not usually select inappropriate analogues on the basis of a single taxon, unless it dominates the assemblage. The only way to check if inappropriate analogues have been selected is to inspect which analogues have been chosen for each level.

By: bender

bender — Wed, 23 May 2007 23:13:29 +0000

Re #41

Of course, it is possible that some part of Quebec was not glaciated at all.

This is, apparently, the case for Anticosti Island.

By: bender

bender — Wed, 23 May 2007 23:02:27 +0000

For the record, thanks go to Phil Leduc.