Science has published the “Correction & Clarification” to Kaufman et al 2009, but curiously the reconstructed temperature proxy has changed yet again between the draft clarification and the published clarification, to “hide the decline”? from the temperature anomaly 2000 years ago that in the draft was shown to be about the same as present.

http://hockeyschtick.blogspot.com/2010/02/kaufman-correction-before-after.html

]]>ftp://ftp.ncdc.noaa.gov/pub/data/paleo/paleolimnology/northamerica/canada/baffin/donard_2001.txt

it seems that the sign is correct in Kaufman. It just happens to be ‘significantly’ correlated with the inverted Arctic-wide JJA temperature. Interestingly, the local calibration done in Moore01 is ICE ( regression of temperature on proxy); this biased estimator easily gives you false impression on accuracy. Past temperatures should be well within the range of calibration temperatures for this one to work. Where does this prior information come from??

re Jean S

http://www.climateaudit.org/?p=7360#comment-363136

has anyone tried to do Brown&Sundberg type multivariate calibration for local (“grid cell” wide) temperatures?

Probably there’s been efforts, but 10 degrees C wide CIs turned them away 😉

]]>like so:

]]>They find in Figure S4 of the new SOM that these changes have only a small effect on their temperature reconstruction. This is fine, but if there is a right and a wrong way to do a calculation, you ought to do it right, even if you get a similar answer doing it wrong.

They defend their original approach on the grounds that

When applying a single calibration equation to the average of the records available for each 10-year interval, we assume that the average

summer temperature of the remaining sites reflects the temperature of the Arctic as a whole.

While this introduces some uncertainty, it also takes advantage of the most robust calibration,

which is based on all available sites rather than a subset of sites.

It’s true that if you have say several ice cores from the same ice cap, or several tree cores from the same stand of trees, then the several series may as well be averaged together (in natural units, not z-scores) without much concern about which are present or not, since they are all measuring the same thing, whatever that is, with idiosyncratic noise that cancels out across the series.

However, in the case of Kaufman (2009), they have several very different proxies — treering widths, d18O ratios from ice cores, varve thickness, varve density, biogenic silica content, etc — that don’t even have the same units, let alone the same response (if any) to temperature. There is no reason to expect a composite index made from these to have the same relation to temperature when the makeup of the index changes.

While I’m pleased that Kaufman and coauthors have taken the time to respond to my comment, it would be customary in such a circumstance to acknowledge the person making the suggestion, if only in the revised SOM where the issue is addressed. Likewise, the published Corrigendum definitely should identify Steve McIntyre, who originally pointed out that Kaufman was using the Tiljander series upside down, and who (with Ross McKitrick) actually published a comment in PNAS on an earlier paper that had made the same mistake.

On another issue, which I had made only briefly in my headpost, Kaufman’s website linked above states,

This synthesis includes all available proxy records (as of 3/09) that are (1) located north of 60°N latitude, (2) extended back to at least 1000 AD, (3) resolved at annual to decadal level, (4) published with data available publicly, and (5) shown to be sensitive to temperature.

Item (5) is disturbing, since it admits that they have cherry-picked proxy records that correlate with temperature (even if they have done a poor job of it per UC #28 above). As David Stockman showed back in 2006 (Aust. Inst. of Geophysics News, http://landshape.org/enm/wp-content/uploads/2006/06/AIGNews_Mar06%2014.pdf), this can lead to almost exactly the same type of picture Kaufman comes up, just using serially correlated randomly generated series.

Brown’s approach to multiproxy calibration, mentioned by UC above, does require all the proxies to have a significant correlation to the target state variable. However, if some are considered and set aside by this criterion, the confidence levels should be adjusted appropriately — either assuming independence or using Bonferroni’s less restrictive bound perhaps. The DOF adjustments in the final model should also be adjusted appropriately for the omitted series.

(Note to S. Mosher — a Bonferroni is sort of like a Zamboni, except that it tears the ice up instead of glossing it over, and occasionally resets the scoreboard to 0.) 😉

]]>Re: UC (#28),

[Use Nelson’s voice from Simpsons]

Ha ha!

Re: UC (#17),

According to Brown, any self-respecting calibrator will use only statistically significant calibration slopes. With this rule (sl 0.05), the proxies that can be used are

2

5

6

9

10

12

14

15

18

22

23

If we take serial correlation into account, with Ebisuzaki (sl 0.05) the remaining proxies will be

2

10

12

14

22

The interesting one is proxy # 10 (Donard Lake) that survives both tests, but with different sign that is used in the Kaufman composite.

]]>In the original SI, they gave the wrong formula for the autocorrelation adjustment, with r1^2 where they should have had r1, but Roman has confirmed that they must have used the right formula for their actual calculation, so this was just a slip in the SI writeup.

In the new SI, they have corrected the formula and used it to compute a CI for the slope coefficient, but then add a second formula, that in principle is a little more precise, and use it to test for significance of the slope coefficient. This is a little odd, since the slope is significant at the 5% level if and only if 0 lies outside the 95% CI. They should just use one or the other, and leave it at that.

The newly added second formula actually goes back to Bartlett (1935), who took into account the serial correlation of x in addition to the serial correlation of the errors (the latter being the same thing as the serial correlation of y under the null of no correlation). In principle these interact, but if the regressor is a just time trend, its first order serial correlation is 1 in large samples, so it drops out, as in the simpified Santer-Nychka formula. The same would be approximately true for a variable with a strong trend or drift during the regression period, like temperature, so again the Santer-Nychka simplification should give very similar results.

]]>