Monday, December 21, 2009

Darwin and GHCN Adjustments

Willis Eschenbach created a stir a few days ago with his post at WUWT entitled The Smoking Gun At Darwin Zero. Darwin Zero is one of five duplicate Darwin records in the GHCN database. It is the longest, and the others seem to be fragments of it, although there are some variations.

The claim is that the homogeneity adjustments are so large, and upward, that they indicate manipulation. This stern statement was underlined:
Those, dear friends, are the clumsy fingerprints of someone messing with the data Egyptian style ... they are indisputable evidence that the "homogenized" data has been changed to fit someone’s preconceptions about whether the earth is warming.

The plot of adjustments to Darwin 0 is shown here:

There was a wide reaction, since these allegations came (by chance?) on the eve of the Copenhagen Climate Conference. Even political blogs like Kevin Drum, Volokh, and Megan McCardle chimed in. Drum kept his head - the others swallowed the fakery line.

Then the Economist weighed in with a forceful takedown. Willis responded; the Economist replied, and now Willis has put up something of a review post.

But the strongest riposte came in response to the first post. Giorgio Gilestro did the calculation that should have been done at the beginning. Instead of picking out one station, GG calculated for all the adjusted stations in the GHCN file the effect of the adjustment on the trend. For Darwin, for example, the trend rises by about 0.235 C/decade over the adjusted period, following adjustment.

As a digression, I should explain that not all stations are able to be adjusted (it needs a minimum record), and the length of record after adjustment is smaller. In what follows, the figures for a station relate to the time period available after adjustment. In my calculations, I look at duplicates, where they exist, independently, so in a few cases it may happen that more than one duplicate per actual station is counted. However, usually as with Darwin, the first one listed is the longest, and the rest are smaller fragments. Most stations do not have duplicates.

So GG showed the following histogram for the 6533 stations that survived a gentle quality screen:

The distribution is fairly symmetrical. Not exactly, though, and an important statistic is the mean - 0.0175 C/decade. Although there were almost as many adjustments adding a falling trend as rising, there was a slight bias towards a rising trend. It's noticeable relative to AGW, but not dominant. And of course, it's less than a tenth of the rise for Darwin.

So why is this such a strong riposte? Remember Willis' charge:

they are indisputable evidence that the "homogenized" data has been changed to fit someone’s preconceptions about whether the earth is warming.

But now there is almost no nett direction to the changes. The homogenization does not affirm anyone's preconceptions. You can't change a global plot just by fiddling with Darwin.

I wrote a R code which verified GG's result. You can find it on my repository thread here.

We can place Darwin in the distribution. For all 6736 stations (my screen let through more than GG's), Darwin's trend diff of 0.235 C/decade came in at number 576. A bit over 1 sd (0.189 C/dec) at the high end.

Darwin is really more unusual than this suggests. The extremes of the distribution are dominated by short fragments, where a modest adjustment gives a big trend change. If we restrict the stations to those with more than 80 years of adjusted record, Darwin's outlier status becomes clearer. Now there are 2074 stations and Darwin ranks #31. 30 stations have a more warming trend adjustment, and 17 stations are adjusted downward more than Darwin was adjusted up. The histogram, with Darwin marked, is here:

Now Willis suggests that Darwin's trend, post adjustment, is so high in absolute terms since 1940 that it must have been manipulated. Well, again, we can look at that. Here is the histogram of the unadjusted records (with at least 9 years in the record post 1940). Darwin is fairly central:

And here are the adjusted records. Darwin moves toward the tail, but not extremely so. It's #194 out of 6515.

GHCN homogenization adjustments are often misunderstood. They are trying to detect and adjust for discrete changes. A station gets moved - a screen replaced - a nearby tree removed. Plus the more widespread changes, such as the 1990's move toward thermistor MMTS. They don't correct for UHI.

They do have a reason. The reassuring consequence of GG's analysis is that, while they can be large (as with Darwin), they do not, as often alleged, dominate the warming signal in their nett effect. There is no reason to believe that their scatter is designed by someone to advance an AGW agenda.

Update (22 Dec)
At WUWT, Kevin M challenged me to prove it:
KevinM (08:11:16) :
Prove it! somebody show me one plot from one station with data that is downloadable and verifiable and shows the opposite pattern.

Well, I said above that of stations >80 years adjusted record, 17 were adjusted down by more than Darwin was adjusted up. Here they are:
211357000002 GUR'EV 
222234720001 TURUHANSK   
222246410000 VILJUJSK  
222255510002 MARKOVO    
414763930000  MONTERREY,N.L  
501947280000 COONABARABRAN NSW       
615076300000 TOULOUSE/BLAG
I didn't look up all the names, but you can find them, with details, on the v2.temperature,inv file.
I've plotted the Australian station, Coonabarabran. The net adjustment is shown in green (same scale, but zero at 14C)


  1. GG's analysis is flawed. Consider a simple example with only 2 stations. Both stations have an unadjusted trend of 0 degrees/decade. Station A has an adjusted trend of 1 degree/decade and Station B has a adjusted trend of -1 degree/decade. Both stations records begin at year 1. Station A's record is 100 years and Station B's is 50. The distribution of the trend difference is symmetrical. Now average the temperature over the 100 year period for both records. The effect is obvious.

  2. The analysis isn't overall flawed. You're objecting to the use of the simple mean as a summary statistic. That's true - a better statistic would be the mean weighted by number of years in the record. But with a large sample, and not much dependence on record length, it doesn't make much difference. I did the calc - the mean trend adjustment over all stations increases from 0.0175 to 0.0179 C/decade if you weight by record length.