There are some important findings at Climate Audit today. Once again I have tried to set out a layman's version of the discussion there.
One of the perennial problems with temperature reconstructions has been a lack of data covering the southern hemisphere - the Hockey Stick itself was a northern hemisphere reconstruction, although the IPCC billed it as global in extent. However, a recent paper by Gergis et al sought to partially remedy this by presenting an Australasian temperature reconstruction for the last millennium, based on 27 proxy records, primarily from tree rings and corals. The headline was, perhaps not unexpectedly, that late twentieth century warming was unprecedented:
The average reconstructed temperature anomaly in Australasia during A.D. 1238–1267, the warmest 30-year pre-instrumental period, is 0.09°C (±0.19°C) below 1961–1990 levels.
Gergis et al reported that they had used 27 proxy series, but that they had selected these from a larger body of data. This was a critical step in the process and one that could well have led to a bias in the results. Many of the studies in the field of temperature reconstructions rely on "sorting" or filtering the data in some way, either choosing only proxy series that correlate well with their local temperature or alternatively weighting them according to how well they correlate. On the face of it, this is a reasonable approach, as the argument might well be made that if there is no correlation then the series is clearly not a proxy for temperature. However, the problem with this approach is that it amounts to a circular argument; it could be that the correlation between temperature records and proxy data is coincidental. This flaw has been demonstrated by studies in which the proxy series are replaced with dummy data series that wiggle up and down at random. In these studies, it has been shown that most of the time the resultant "reconstruction" is a hockey stick.
However, the good news was that Gergis et al were apparently aware of these issues and they declared that they had found a way around them. According to the paper:
For predictor selection, both proxy climate and instrumental data were linearly detrended over the 1921–1990 period to avoid inflating the correlation coefficient due to the presence of the global warming signal present in the observed temperature record.
In other words, when doing the correlation calculations to determine which proxies they would use, they first removed the twentieth century trend (actually, the 1921-1990 trend) from both the proxies and the temperature records, so that instead of filtering for series with a twentieth century uptick, they were picking records that appeared to match in their year-to-year wiggles.
So far so good, but this would have been a calculation that had to be done with immense care and a replication of it could therefore have been illuminating. However, in order to do this, it would be necessary to obtain all the Gergis data including the series that had been filtered out and not used, and unfortunately Gergis was not playing ball:
The compilation of this database represents years of our research effort based on the development of our professional networks. We risk damaging our work relationships by releasing other people’s records against their wishes. Clearly this is something that we are not prepared to do.
We have, however, provided an extensive contact list of all data contributors in the supplementary section of our recent study ‘Southern Hemisphere high-resolution palaeoclimate records of the last 2000 years’ published in The Holocene (Table S3):
This list allows any researcher who wants to access non publically available records to follow the appropriate protocol of contacting the original authors to obtain the necessary permission to use the record, take the time needed to process the data into a format suitable for data analysis etc, just as we have done. This is commonly referred to as ‘research’.
The refusal to release data was troubling, as it prevented anyone ever being able to replicate Gergis's work. And when Gergis's blog was unearthed and it was revealed that she was a committed environmentalist, the alarm bells became louder still.
Despite the flat refusal to allow replication of the paper, it was still possible to verify certain aspects of the filtering process. In particular, the data for the 27 proxies that had been used was available and so it would be possible at least to replicate the calculation that showed that these had significant correlations to their local temperature once the 1921-1990 trend had been removed. This task was taken up by statistician "Jean Sibelius" but rather remarkably he found himself unequal to the task:
Steve, Roman, or somebody, what am I doing wrong here? I tried to check the screening correlations of Gergis et al, and I’m getting such low values for a few proxies that there is no way that those can pass any test. I understood from the text that they used correlation on period 1921-1990 after detrending (both the instrumental and proxies), and that the instrumental was the actual target series (and not the against individual grid series).
Sibelius's difficulties were confirmed by others, including Steve McIntyre, but perhaps most significantly, by CSIRO's Nick Stokes, who is no sort of a sceptic. Stokes agreed with Sibelius that, when detrended, the correlations for the 27 proxies used in the Gergis reconstruction were insignificant, completely contradicting Gergis's paper. However, extraordinarily, Stokes also ran the calculations without detrending and found correlations that were significant.
I’ve run Steve’s code with and without detrending, and with and without the Quenouille correction. Without detrending (but with zero mean) or AR1 correction all (exc maybe Madang) proxies do seem significant
This seems to suggest that Gergis's declaration that the correlations were based on detrended data was false and that she and her co-authors had indeed fallen foul of the circular argument noted above. The finding of unprecedented warmth reported in the Gergis paper appears as though it is a function of the methodology used rather than of the underlying data.
Paul Matthews points out another serious problem with the Gergis et al paper:
Worse still, proxies were selected with positive or negative correlation. In other words, some were used 'upside-down'.