Mann 07 — Proxy Models Part 1
Posted by Jeff Id on July 15, 2010
I’ve been playing around with Mann07 which to my way of thinking is an important paper in Paleoclimatology. This is a very first step in the process as I only have a few hours into the analysis. The reason it’s important is that M07 is a demonstration of a lack of signal amplification I show in the hockey stick links above and many others have discussed. VonStorch and Zorita, who are treated like skeptics for their work on the topic.
I learned more from the SI and code than from the paper. Before we begin, those of us on the outside owe Mann a thanks for archiving everything for this work. Of all the climate work you can dig into, Mann is currently setting the standard and his colleagues should follow. There are people genuinely interested in the methods and data and in this case, he’s done another great job. Unfortunately for Mike Mann, Steve McIntyre gets points again because he’s the one who forced the situation, we must give credit where due.
All information for this paper, including pdf’s, data and code can be accessed without the paywall at this link:
Figure 5 of the supplimentary info is shown below. Mann is showing in this paper that RegEM, an obscure form of what is basically MV regression, does not create signal amplification of his proxy information. His contention is that this proves that his (and by proxy – some pun intended – paleoclimate) methods are extracting a consistent amplitude signal from the proxy data.
It’s an intriguing problem for me, in particular because the first time I realized what CPS is and later what these regressions are doing, I nearly blew a blood vessel. It cost me hundreds of hours of my life blogging and reading. Everything I found confirmed my initial belief that these methods de-amplify historic signal in relation to the modern one.
So, when I see Mann’s graph above, I ask – What the heck is going on? How come his results are so different from my own.
This will disappoint some here, but I’m not planning on running the full RegEM version any time soon. It’s a few hours of work with little payout. CPS correlation methods or simple regression are enough to figure out what is going on here. If there are enough of the original crowd still hanging around, you will get that both algorithms are methods of elimination and strengthening preferred information. Preferred being “the most temperature-like”.
My own results from proxies, came up with something like this:
From the link to my hockey stick work in this post above, you know that I used one signal plus noise. It’s sometimes a criticism of my post above that there isn’t any spatial difference in the signal, but I haven’t bothered doing a whole calculation on spatially noise filled signals because it makes no difference to the problem. Mann, however, managed to use model data to demonstrate spatial differences in signal and then applied noise before running his methods.
I’m certain his results should be the same as my own, but don’t have proof yet as to why they are not. Of course, I do have an idea though.
But for today, I’ll leave this post with a plot of what is for Mann07’s purposes the noiseless model data of temperature.
My first impression of this temperature network was that it’s awfully tame. Certainly not very proxy like but it shouldn’t be proxy-like if you believe tree rings are temperature. This plot is pure temperature so the ‘noise’ of the tree ring proxy needs to be added in, and that’s exactly what Mann07 correctly does. –The black line is the average.