Antarctic warming is robust

The difference between a single calculation and a solid paper in the technical literature is vast. A good paper examines a question from multiple angles and find ways to assess the robustness of its conclusions to all sorts of possible sources of error — in input data, in assumptions, and even occasionally in programming. If a conclusion is robust over as much of this as can be tested (and the good peer reviewers generally insist that this be shown), then the paper is likely to last the test of time. Although science proceeds by making use of the work that others have done before, it is not based on the assumption that everything that went before is correct. It is precisely because that there is always the possibility of errors that so much is based on ‘balance of evidence’ arguments’ that are mutually reinforcing.

So it is with the Steig et al paper published last week. Their conclusions that West Antarctica is warming quite strongly and that even Antarctica as a whole is warming since 1957 (the start of systematic measurements) were based on extending the long term manned weather station data (42 stations) using two different methodologies (RegEM and PCA) to interpolate to undersampled regions using correlations from two independent data sources (satellite AVHRR and the Automated Weather Stations (AWS) ), and validations based on subsets of the stations (15 vs 42 of them) etc. The answers in each of these cases are pretty much the same; thus the issues that undoubtedly exist (and that were raised in the paper) — with satellite data only being valid on clear days, with the spottiness of the AWS data, with the fundamental limits of the long term manned weather station data itself – aren’t that important to the basic conclusion.

One quick point about the reconstruction methodology. These methods are designed to fill in missing data points using as much information as possible concerning how the existing data at that point connects to the data that exists elsewhere. To give a simple example, if one station gave readings that were always the average of two other stations when it was working, then a good estimate of the value at that station when it wasn’t working, would simply be the average of the two other stations. Thus it is always the missing data points that are reconstructed; the process doesn’t affect the original input data.

This paper clearly increased the scrutiny of the various Antarctic data sources, and indeed the week, errors were found in the record from the AWS sites ‘Harry’ (West Antarctica) and ‘Racer Rock’ (Antarctic Peninsula) stored at the SCAR READER database. (There was a coincidental typo in the listing of Harry’s location in Table S2 in the supplemental information to the paper, but a trivial examination of the online resources — or the paper itself, in which Harry is shown in the correct location (Fig. S4b) — would have indicated that this was indeed only a typo). Those errors have now been fixed by the database managers at the British Antarctic Survey.

Naturally, people are interested on what affect these corrections will have on the analysis of the Steig et al paper. But before we get to that, we can think about some ‘Bayesian priors‘. Specifically, given that the results using the satellite data (the main reconstruction and source of the Nature cover image) were very similar to that using the AWS data, it is highly unlikely that a single station revision will have much of an effect on the conclusions (and clearly none at all on the main reconstruction which didn’t use AWS data). Additionally, the quality of the AWS data, particularly any trends, has been frequently questioned. The main issue is that since they are automatic and not manned, individual stations can be buried in snow, drift with the ice, fall over etc. and not be immediately fixed. Thus one of the tests Steig et al. did was a variation of the AWS reconstruction that detrended the AWS data before using them – any trend in the reconstruction would then come solely from the higher quality manned weather stations. The nature of the error in the Harry data record gave an erroneous positive trend, but this wouldn’t have affected the trend in the AWS-detrended based reconstruction.

Given all of the above, the Bayesian prior would therefore lean towards the expectation that the data corrections will not have much effect.

Page 1 of 2 | Next page