I recently re-visited an article in Science (Briffa and Osborn 1999), that, together with Jones et al 1999 (Rev Geophys), were the first bites of the poison apple of hide-the-decline. I observed that key conclusions in Briffa and Osborn 1999 depended on the rhetorical effect of deleting the decline from their spaghetti graph.
I’ve been looking at the subsequent development of this graphic and, in the process, noticed another curious feature of the figure in Briffa and Osborn 1999 – shown below.
As noted previously, Briffa data was deleted after 1960. (Smoothing was done after the deletion further accentuating the impact of the deletion of post-1960 data.)
In addition, and this point has not been previously discussed, Briffa and Osborn did not show data prior to 1550 for the Briffa MXD reconstruction. I’d previously noticed that an archive for Jones et al 1998 (surprisingly) contained a Briffa version that is linearly related to the Science graphic – the match is shown as a dotted line. (The basis of the linear relationship is not reported and not known to me at present, but can nonetheless be used empirically to show the extensions.)
In the graphic below, I’ve shown (in magenta) not just the hide-the-decline extension, but the deleted data prior to 1550. Take a look. (Update note- see yesterday’s post for provenance.)
Obviously, the deletion of the pre-1550 portion of the Briffa reconstruction version also makes a large contribution to the rhetorical impression of coherence between reconstructions. Ironically, Briffa and Osborn observe in their running text:
An uninformed reader would be forgiven for interpreting the similarity between the 1000-year temperature curve of Mann et al. and a variety of others also representing either temperature change over the NH as a whole or a large part of it (see the figure) as strong corroboration of their general validity, and, to some extent, this may well be so.
Needless to say, one of the reasons for the reader being “uninformed” is the deletion of adverse data (both before 1550 and after 1960) to give the impression of “corroboration” of the “general validity” of the reconstructions. Note that smoothing after deletion enhances the impact of the deletion – look at the strong divergence pre-1550 as well as post 1960.
The provenance of the Briffa reconstruction was tersely described only as follows ( in the caption): “NH tree-ring densities [1550-1960, from (3 – Briffa et al 1998(Nature); Briffa et al 1998 (Proc Roy Soc London)), processed to retain low-frequency signals]”. Appendix A of Briffa et al 2001 contains some useful further information, showing that this version arose from a composite obtained from averaging MXD chronologies in the large Schweingruber network (scaled over 1700-1994, then opportunistically calibrated over 1880-1960):
In Figure 4, we show different temperature reconstructions of the NH (all land area temperatures north of 20N), each based on a different way of using MXD predictor data. One curve was produced by performing the age-banding procedure on all chronologies in the data set and by using an unweighted mean of all banded series from all locations. This is similar to the curve from 1650-1960 [sic – actually 1550-1960] presented by Briffa and Osborn  (although we have since made very minor modifications to the age-banding procedure and the input data set.) All other curves in Figure 4 were obtained by prior averaging of the age-banded density series into the nine subregions (as defined by Figure 1.)
The next figure shows the Briffa et al 2001 (JGR) spaghetti graph – showing the Briffa 2001 version in limegreen (together with the decline, expurgated in the original figure) and the Briffa-Osborn 1999 version in magenta (including both pre-1550 and post-1960 data):
Figure 2. From Briffa et al 2001 (JGR) Plate 3, also showing the pre-deletion data used in Briffa-Osborn 1999 (magenta). The Briffa 2001 version ends in 1960 and has been extended using data from Climategate emails (showing the decline).
The changes from the 1999 Briffa version to the 2001 Briffa version are instructive for several reasons.
Briffa et al 2001 uses virtually the same population of sites as Briffa and Osborn 1999. The B2001 population was 387 sites, while the Briffa et al 1998 (Nature 393) population (cited in BO99) was 383 sites – immaterially different. The Briffa et al 2001 site count was 19 sites in 1550, 8 in 1500 and only 2 in 1402, but there were enough for Briffa to report a reconstruction. (Readers should bear in mind that the Jones reconstruction, for example, was based on only 3 proxies in the 11th century, one of which was a Briffa tree ring site with only 3-4 cores, well under standard requirements.)
So why were pre-1550 values shown in Briffa et al 2001 and not in Briffa and Osborn 1999? The only reason that I can deduce is that the Briffa 2001 reconstruction had a rhetorical similarity to the Mann and Jones reconstructions in the 1400-1550 period – and therefore was shown, while the Briffa and Osborn 1999 version showed a major discrepancy – and was therefore not shown.
How did Briffa get from the 1999 reconstruction version (magenta) to the 2001 version (limegreen)?
This is an interesting exercise that I’ll describe in more detail on another occasion. But I can’t resist a quick preview. The Briffa-Osborn 1999 version was based on averages of all available sites – a sensible enough procedure. Changes in the Briffa 2001 methodology include the calculation of regional averages followed by stepwise principal components. (The methodological description in Briffa et al 2001 is very sketchy and, unfortunately, the Climategate computer dossier didn’t include Briffa’s programs.)
The effect of using principal components on regional averages is to change the weights for individual sites, including the possibility of negative weights i.e. flipping the regional MXD series. In particular, the closing uptick in the Briffa 2001 reconstruction may well depend on the flipping of data – a point that I’ll try to examine in the future.