Guest Post by Willis Eschenbach
Over at Climate Audit, Steve McIntyre is engaged in the slow public defenestration of the latest multi-proxy extravapalooza, a gem of a paper yclept “Inter-hemispheric temperature variability over the past millennium” by Neukom et al, hereinafter Neukom2014. Not content with simply creating a thousand-year temperature reconstruction for the Southern Hemisphere, they top it off by comparing that reconstruction to a thousand-year temperature reconstruction for the Northern Hemisphere … the paper is paywalled but the supplemental Information is available, and covers many of the topics (16 mb PDF).
On that Climate Audit thread, someone asked Steve McIntyre why, since he is arguably the world’s foremost expert on the minutia of the various proxies, he doesn’t make his own reconstruction. He replied:
I don’t see any point in dumping a lot of stuff into a hopper and making a weighted average. What specialists need to demonstrate first is that there is consistency among proxies. … I have long urged specialists to examine proxies for consistency as a precondition to presenting a reconstruction. If one has consistent proxies, then you get similar reconstructions regardless of the methodology. Specialists have either ignored the idea or sneered at it. Instead, they prefer to throw increasingly complicated and poorly understood multivariate methods at the problem, yielding poorly interpreted squiggles.
Like Steve, I am a huge advocate of using the tried and true mark one eyeball to examine the proxies for consistency. It’s detail work, but it is of extreme importance. It’s also often very interesting. So with that in mind, let’s look at what Neukom2014 calls “an unprecedented network of terrestrial and oceanic palaeoclimate proxy records”. It is indeed unprecedented, although perhaps not for the reasons he believes … here are the first 24 proxies.
Figure 1. Twenty-four of the 111 proxies used in Neukom2014. Colors indicate the type of proxy. Datasets standardized by Neukom authors, units are standard deviations. Dashed gray lines show ± 3 standard deviations. Note that the proxies cover different time spans. Thick transparent gray lines are loess smooths of the data.The scarlet letter (H) indicates that the data is heteroskedastic (Goldfield-Quant test) at p less than 0.05. (“Heteroskedastic” is a lovely word, almost a skat rhythm, that simply means that the variation changes over time.) Click for larger version.
Dang … already you can see that McIntyre’s “consistency” is not exactly a prominent feature of this dataset. So, only way to do it, jump in, start from the top … onwards.
The first four proxies (top row) are lake sediment proxies (marked “l sed”), which I’ve colored brown like the sediment. The first is a record from Lake Challa in Kenya, where obviously the sedimentation is very low and quite variable. This is evidenced by the flat bottom of dataset. Remember that annual sedimentation rates can vary but they can’t go below zero. Obviously, when the data was standardized, the dataset was moved downwards, so the flat bottom which represents almost no sedimentation is now below zero.
The second proxy, rather than sedimentation rates, is measuring d18O, the change in an oxygen isotope linked to temperature. Also heteroskedastic, but lacking the flat bottom seen in Lake Challa.
The next two lake sediment proxies obviously are not teleconnected. One says no change, dead flat … and heteroskedastic. The other is not heteroskedastic, but drops seriously 1300 -1400, then rises again by 1800 … which one to believe? What are they measuring? In Laguna Aculeo, it is “pigment reflection”, and in Lago Puyeheye, it is thickness of the sediment layer.
The first problem with these lake sedimentation proxies is indicated by the scarlet letter. You can see that the last half of the data has a decidedly greater variance than the first half.
I have to confess, I’m not a big fan of lake sediment as a temperature proxy, for several reasons. The first is that it is what I call a “double proxy”. By that I mean that it is NOT a proxy for temperature. Instead, it is a proxy for some other variable, which is in turn connected to the temperature. In other words, it is a proxy of a proxy. In this case, lake sediment is often a direct proxy for rainfall, and thus only indirectly connected to temperature.
But not always, of course … that would be too simple, and climate is rarely that. For example, in Eagle Lake, Alaska, the “varve thickness” (annual layer thickness) is inversely related to both rainfall and temperature … go figure. It seems to be related to the fact that the lake is glacier-fed. The tongue of a glacier is always moving, one direction or the other. When it is retreating you get less sediment. But when it is advancing (cooling) it acts like a bulldozer, pushing up the soil and creating sediment. So rather than get more sediment from rain, this lake gets more sediment from falling temperatures.
We also came across another large problem with the Eagle Lake varve thickness. This is that the rate of sedimentation, as you might imagine, is greatest near where the creek or river empties into the lake, and decreases outwards from there. The problem is, often a delta builds up where the river enters the lake, and then after a flood or just after some time, the river changes its course through the delta (as they all do eventually). This can lead to huge variations in the varve thickness for a given sampling site. I corresponded with Mike Loso, the author of the study, regarding this question, and he was later generous enough to mention my contribution in his acknowledgements in a paper. In any case, the wandering of the location of the incoming water is a big problem for lake sediment proxies.
Next difficulty with lake sediments is that they are often exquisitely dependent on the local vegetation. If the vegetation dies from any cause, whether fire, humans, bark beetles, or something else, erosion is guaranteed to increase. And the varve thickness will increase with the erosion … so lake sediments are not unlikely to be proxies for things that have only a vague connection to either temperature or rainfall.
In the case of the first lake sediment record, one or the other of the many possible confounding variables has so contaminated the record so much that if we take a running average (yellow line), we end up with a gradually raising average over 2000 years. But that’s not real. You can see the change in the character of the record in about the year 1250. Don’t know what happened then, perhaps the river mouth shifted, but since then the record has a distinctly different appearance than the record before that. And it is exactly that spurious shift that is picked up by the heteroskedasticity test and results in the red “H” …
The second lake sediment proxy is also heteroskedastic. For a thousand years, almost no change. Then for the next thousand years, down and then up. I don’t know what the lake is measuring … seems doubtful that it’s temperature, though. Anyhow, that’s why I’m not a big fan of using lakes as thermometers.
Next, we have a whole bunch of corals. Figure 2 shows the rest of them.
Most of them are on the order of only three or four hundred years in length. Most but not all of them are gradually descending over the time of the record. They are measuring different things in the coral—strontium/calcium ratios (Sr/Ca), change in oxygen-18 isotope (d18O), thickness of annual layers. What are they actually proxies for? I don’t know. If it’s temperature, they are almost all going down over the period. Are they negatively correlated with temperature? Did the authors use them right-side-up? Is “right-side-up” for d18O the same direction as for Sr/Ca? I do not know if the authors have answered all of these questions … but given the number of proxies used upside-down in past reconstructions, I’m not sanguine about the answer.
Next, we have a number of “documentary” records, shown in gray. These are estimates of a climate variable from written records of the past. Of course, these are almost all since the Spanish conquest. A number of them appear to be discrete variables that take one of only a few possible values (e.g. “very high”,”high”, “no mention”, “low”, “very low”).
The bizarre thing about the documentary records is that they document the following subjects in order: precipitation, ENSO, precipitation, precipitation, snow depth, precipitation, snow occurrence … but not a single one of them is a documentary record of temperature! Not one? How crazy is that?
From there we go to the ice cores, in Figure 3 below.
Like the coral proxies, various things are measured in the ice. Mostly it is d18O, but some are measuring annual accumulation rates, sea salt levels, and change in deuterium (dD). The only one of these ice core proxies that is not in Antarctica is Quelccaya, which is in Peru. Now … if all of those records are showing the temperature in Antarctica, why are they so different? Are we to believe that some places in Antarctica there has been no change in temperature in the last 400 years, in other places it has warmed steadily for the last 250 years, and in other places it has been cooling over the last 250 years? Because I don’t believe it. In any case, seven of the seventeen ice cores are heteroskedastic.
Next there is one lonely ocean sediment proxy, measuring the magnesium/calcium ratios. Not sure why there is only one, as there are certainly other ocean cores … but what do I know, I was born yesterday. Did the ocean temperature actually take a jog up in 1500 and then drop to 1700? Looks possible, I suppose. Does this mean the medieval warm period (MWP) was warm in sunny Cariaco? Quien sabe, señor …
Then we have a reconstruction, shown in gold. A reconstruction of what? … well, of rainfall on the great barrier reef. Since 1890. Why? who knows.
Then a speleothem, which is generally either a stalagmite or a stalactite in some cave. This is measuring annual lamina (layer) thickness. It’s another double proxy, because lamina thickness is related to rainfall rather than temperature. In addition, lamina thickness is a function of how much water is coming into the cave, how much of that water makes it to the specific spot, the ever-changing underground geometry as the limestone collects and changes and wears away … not a simple story.
And after that? Trees, trees, and more trees, almost 40% of all of the proxies are of tree ring width or density.
To start with, about 40% of all the tree proxies are heteroskedastic. Next, according to the trees, the southern hemisphere is warming, cooling and staying exactly the same over the last 250 years or so … go figure.
I’m sorry, but the tree ring datasets don’t impress. The problems with them have been covered extensively in the past. The basic problem is that trees simply don’t make very good thermometers.
So there you have it, all 111 of the “unprecedented” proxies used in Neukom 2014. Steve McIntyre advised looking for proxy consistency. The problem with the Neukom proxies is that there is little consistency to be found anywhere, either between the groups of proxies or within each group of proxies.
Now, if you average all of those proxies together, you get … well, no hockeystick, as shown in gray below. But of course, the authors have their secret sauce which is guaranteed to prepare a hockeystick … Figure 6 shows the average of the proxies, along with their final results.
How dey do dat? Well, that’s a fascinating topic in itself, but for another day …
Conclusions? I can only echo the quote I put up earlier:
I don’t see any point in dumping a lot of stuff into a hopper and making a weighted average. What specialists need to demonstrate first is that there is consistency among proxies.
I can’t say it any better than that …
My best wishes,
AS ALWAYS: If you disagree with me (line forms to the left) please quote the exact words you disagree with … that way we can all be clear regarding just what you think is wrong.
DATA: All of the Neukom14 data is online as .csv files here.