Willis writes in:
While researching ocean drill cores at the WCDC, I stumbled across Mann’s borehole data. One of the proxies used for historical temperature reconstruction is "borehole temperature", the temperature down in the ground. In 2002, Michael Mann et al. published a study called Optimal surface temperature reconstructions using terrestrial borehole data. It is available here, for $9.00. In it they use all of the available weapons to construct the temperature proxy “¢’¬? EOFs, PCA, and of course that perennial favorite, "optimal fingerprinting", viz:
We employ a spatial signal detection approach that bears a loose relationship with “Å”Åoptimal detection” approaches used in anthropogenic climate signal fingerprinting [Mitchell et al., 2001]. In such “Å”Åoptimal detection” approaches, one seeks to identify, through generalized linear regression, the estimate of a target signal (as predicted by a model) in empirical data. Detection is accomplished through rotation of the empirical data, in EOF state-space, away from the direction of maximal noise (as estimated from, e.g., a control model simulation).
In our approach, an independent estimate of noise is not available. Rather, we employ an EOF rotation of the information in the borehole dataset toward an independent estimate of the target spatial SAT signal from the instrumental record, based on ordinary (potentially weighted) least squares spatial regression. Once an optimal rotation is found that provides maximal (and statistically significant) agreement between the spatial information in the borehole and instrumental record during the 20th century, the associated eigenvector rotation is used to project the estimated borehole SAT signal back in time.
So I decided to see how well this "optimal detection" works compared to plain old linear regression. I regressed several of their results against the HadCRUT3 temperature dataset. I compared my own area-weighted () average of their raw gridded data, their calculated area-weighted average of their "optimal" gridded data, and their final "optimal reconstruction". Here are the results:
A couple things of note. First, the difference between just regressing the plain old raw average and their "optimal" result amounts to 0.06 degrees five hundred years ago … seems like they wasted a lot of time checking for fingerprints.
Second, this is as boring a paleo record as I could imagine, very little detail.
Finally, the lack of detail pertains to a very interesting sleight-of-hand manoeuvre … notice the kinks in the borehole reconstructions every hundred years in the graph above? Those kinks are present in Mann’s paper as well, only he’s hidden them by squishing the graphs down flat. But once you know where to look, you can still see them. Take a look … the black arrows show the big ones. The 1900 kink is the most prominent one, but they’ve made it so flat that even it’s hard to see.
Why are there kinks in the reconstruction? Because they’ve infilled the data. In each grid cell, there are actually only six data points, one each for the years 1500, 1600, 1700, 1800, 1900, and 1980. The other 474 data points are just linearly filled in between those six points.
Now, consider the true statistical significance of their data. Any average temperature will only have six data points … do they have a significant trend? No way, including autocorrelation you’re down to only about three data points, that’s one degree of freedom …
None of their results have any statistical significance whatsoever. For the period 1900-1980, for example, they are regressing two data points (or EOFing and PCAing two data points) in each grid against the actual temperature.
Grrrr … the Mannomatic truly can chop, slice, and dice anything. Most likely, they’re now using the data from this "reconstruction" in their latest hockeystick emulation …
Questions? The data is here, read’em and weep …
Since there’s only six data points per gridcell, we really don’t have enough data to say if there is a temperature signal or not. Not one of them has a statistically significant trend, either up or down, there’s simply not enough data.
Here’s a sample of a few of the boreholes …
These just happen to be the first ten gridcells in the dataset. Notice that each one only has six data points. Now, do you see a temperature signal in there?
I don’t think there’s much of one, in part because they are all so radically different. Look at gridcell 3, it has cooled 2 degrees over the period of the record … is that a temperature signal?
I also don’t think there’s a temperature signal there because of the monotonic nature of the signal, increasing every century. No other proxy dataset (including Mann’s ill-fated "hockeystick") shows a rise every century. All of them are coldest about 1700. So no, I don’t think there’s a discernable temperature signal there.
My graphs are just graphs, not sleight of hand. Mann never mentions in his paper that there are only six temperatures per gridcell, one per century, and he squeezed down his graphs until nobody notices. That’s sleight of hand …