As I recall Steig admits that he gets the Peninsula trend wrong, but waves that off by stating it is small price to pay for getting the rest of Antarctica correct and besides we know that the Peninsula is heating up due to all the measurements we have of that area.

The main thesis of the Steig paper, in my judgment, was that counter to treating the Peninsula of Antarctica as a special and isolated climate area, one could show that the warming of the Peninsula had spread into West (especially) and East Antarctica.

Showing that the Peninsula is warming at a much faster rate than the remaining 95% of the Antarctica area and to even the exclusion of much warming in that larger area would not have made the cover of Nature. If the authors artifically spread some of that Peninsula warming to West Antarctica their conclusion as captured by Nature is misleading.

In any event, Steig’s results are obviously wrong — given this data, there has to be more trend in the peninsula than in W. Ant., but he has it the other way around.

I ended up modifying Steig’s literal regional definitions a little to make them more reasonable — I included the tip of the peninsula and its entire west coast in “Peninsula” rather than cutting it off at 60 dW and above some latitude as Steig said he did. I also included a few degrees in all directions from the S pole in “East Ant”, rather than allowing “W Ant” to come right up to the pole as Steig said he did. I think I used 3 degrees, but maybe 5.

See the BAS map of “Antarctic Peninsula” at http://wattsupwiththat.com/2009/11/14/every-cloud-has-a-silver-lining-antarctica-glacier-retreat-creates-new-carbon-dioxide-store/ for a conventional definition. They go all the way down to 75S (not 72S like Steig), and then impose an 80W boundary. I just used Steig’s 72S, plus an eastern boundary somewhere to exclude the tip of Thurston Island.

If Steig really did impose the 65dS north boundary that he describes. that would have eliminated the strongly trending King George Island stations, and might account somehow for his weak Peninsula trend. (Just a guess.)

As it happens, the key W. Ant. trend for their full period of study was still significant (though less strongly) after this was corrected (as replicated in the subsequent Steig Corrigendum), so I can’t be said to have reversed their conclusions.

Of course, you’re welcome to cite discussion here by Kenneth Fritsch of the robustness of Steig’s results to shorter time periods, and by Ryan O, the two Jeffs and RomanM on Steig’s use of RegEM, but please attribute those findings to their respective authors.

I was able to get some interesting results using Steig’s AVHRR data plus the manned stations, but without using RegEM or PCA, simply by computing a covariance matrix from the avhrr file, associating each of the 42 (or fewer — see below) manned stations with its avhrr cell, and then projecting the value of each avhrr cell for each month from that month’s available stations.

Unfortunately, my scripts went down in a disk failure (backup! backup!), and I’m too busy with something else to reconstruct them now, but here are some of my observations:

1. Several of the stations are not within 50 km of any of the 50-km AVHRR gridsquares representing the ice cap. Several (6 or so) are on K George Is in the S. Shetland Is, about 100 km away from the coast, so I just assigned them to the nearest gridsquare, which happened to also contain O’Higgins on the Peninsula. Others (S. Orkneys etc) were 700-2000 km from the nearest avhrr grid square, so I declared these unusable.

2. Stations sharing gridsquares effectively are multiple observations on the same temperature by my model, and so may as well just be averaged together to a single record for that gridsquare (as available).

3. The only covariances that are required are those between the (somewhat fewer than) 42 gridsquares with station data and the 5509 gridsquares. Therefore at most only a 42×5509 covariance matrix needs to be in memory. This is fortunate, since 5509 x 5509 is getting pretty unwieldy.

4. Plotting the covariances of several key stations — Byrd, Scott-Amundsen, McMurdo, O’Higgins, Russkaya, Adelaide, etc with the entire continent shows plausible and useful patterns — there is generally close correlation up close, with decaying correlation as distance increases, but the AVHRR correlations clearly show that the Transantarctic Range is a big barrier to weather — Byrd correlates with W. Ant, but not S. Pole, while S. Pole correlates with E. Ant but not W. Ant. O’Higgins (plus S. Shetlands) doesn’t correlate much with W. Ant, but Adelaide (which shares its cell with another station whose name I don’t recall) does correlate at least with the adjacent part of W. Ant., and has a long record. So Steig’s AVHRR strategy does have a lot to be said for it over simple Kriging. Lubos did find that distance has a lot of explanatory power for correlation, but the AVHRR correlations should be even better.

5. Ordinary demeaning of the station data will virtually eliminate any continental time trend, since there are very few stations that cover the entire period (Scott-Amundsen at the S pole being a notable exception). If there is a continental trend (which we must be open to if we wish to determine empirically whether or not there is one!), demeaning each station relative to its own observation period will bias the results towards no trend (either way). In order to preserve any trend signal that is present (while allowing the trend to vary with location and therefore region), it is necessary to compute what I call trending anomalies, ie. fit each station’s data to a time trend passing through 0 at the middle of the period (month 300.5), with seasonal dummies adding to zero for each month of the year, and add the residuals to the trendline. (sentence corrected 10:57 AM EST)

A station like Byrd, which has a distinct uptrend during its 15 years or so of reporting early in the 50-year study period, will therefore end up having distinctly negative trending anomalies, but would have anomalies averaging to zero using simple demeaning. Since it dominates W. Ant while it existed, this will make a big difference for the regional trend.

The resulting extrapolated trending anomalies will contain a lot of spurious noise, but that will just be part of the noise in the final reconstruction, and should just average out or cause big CI’s if it is not for real.

6. The peninsula trend ends up much stronger than Steig found, since his 3 PC RegEM approach blurs the detailed covariance information obtainable from the AVHRR file. W. Ant ends up with a bigger point estimate than Steig found, but weaker than the peninsula, and much more localized toward the peninsula and away from McMurdo than Steig finds. I didn’t get around to scrutinizing its significance, but it probably was significant. E. Ant wasn’t doing much.

But of course replicating and scrutinizing Steig’s findings with his own RegEM methodology, which is your goal, is also worth doing and publishing.