Thursday, April 14, 2011

Quiet time

McIntyre, Mann, and the Gaspe cedars



Sixteen years later, people are still arguing about Mann's 1998 Nature paper (MBH98) on multiproxy (112) reconstructions. McIntyre and McKitrick (M&M) wrote a paper in Energy and Environment in 2005, and I find myself still arguing about that.



The 2005 paper was a major resource for Congressman Barton's 2006 inquisition of Mann. One of its featured criticisms was the Gaspe dataset, number 53, and in particular the fact that Mann had padded some missing data from 1400 to 1403 with the 1404 value. And Barton directed them to explain.



One thing not often now mentioned is that in that paper, M&M actually did what many other critics should have done. They repeated the calculation with the criticism made good, to see what effect it had. This was in their Figure 1. They showed the effect of marking those four Gaspe years as missing, and then the effect of using a centered mean rather than Mann's famous calibration mean. They got a surprisingly large difference, which has been much cited in recent days. This post reports on my investigation of that surprise.



Barton got statistician Wegman to report, and his report was heavily based on the M&M 2005 papers. He said of this

"In MBH98, the first four years of both of these series were extrapolated. The extrapolation has the effect of depressing early 15th century results, and was not disclosed by Mann et al. until a later paper, "

This depressing effect was based on the M&M calculation reported in their Fig 1.



I'll show that the effect of padding those four years was limited to just those four years, and is minor. The effect Wegman describes was due to a coding rule in which M&M in effect removed the effect of Gaspe not just the four years in question, but all years from 1400-1450.



I have previously criticized the use Wegman made of the M&M 2005 GRL code for showing the effect of noise. There there was a clearly deliberate selection of the top 100, ordered by "hockey stick index", from which the displayed samples were chosen. In the case described here, there is no indication that the amplification of the effect was deliberate. In fact, M&M say they are following a policy of Mann's. However, the effect as presented is misleading, in that the change shown is due to the application of a rule rather than the actual effect of the padding.





M&M Fig 1, seen in perspective



Here is the original M&M05 Fig 1



The top panel shows their MBH emulation, which they say appears to be completely accurate. The second shows supposed effect of replacing the 1400-1403 years of proxy 53 (Gaspe) by NA, the R marker for missing. There is a moderate deviation in the years 1400-1450.



The bottom panel shows the effect of changing the offest mean to centered. I'll say more about that in a future post.



I'm rerunning the M&M code, so I'll show that my version does indeed reproduce that plot. I've used a bit of color:





Now I'll plot the same data but superimposed. This emphasises that the Gaspe deviation affects the years 1400-1450.



I've made the panel 3 plot faint, since it isn't the current subject.



For the remaining plots, I'll focus on the years 1400-1450. Here is again the data from panels 1 and 2



Padding with numbers



Mann used the 1404 value (0.723) to pad, and as Wegman suggests, this was a low point. The range from 1404-1450 was from 0.618 to 1.351. So I tried using the maximum value, and got this:



This compares Mann's padding with the effect of padding with that max. It isn't very different at all, and if you focus on the annual non-smoothed values, only the 1400-1403 values are changed. Much less effect than with NA.



What the M&M code does.





The recon work is done in a function NHbeta() in the file ee2005functions.r. They loop over the 11 periods, of which the last is 1400-1450. For each period they set up a "roster" of proxies that will be used. The statement defining it is:

roster<-!is.na(proxy[period[k+1],])



Here period[k+1] points to the first data point in the period, ie year 1400. So by setting the 1400 value to NA, the program removes Gaspe entirely for that period. They might as well have set the values from 1400-1499 to NA. In fact, I checked that doing that has the same effect.



Conclusion



Padding as Mann did had a very small, local effect. The M&M sffort to replace with missing values triggered a large response. But it was an artefact.










0 comments:

Post a Comment