Guest Post by Willis Eschenbach
Bob Tisdale has a detailed post on the new 2014 paper entitled “The Atlantic Multidecadal Oscillation as a dominant factor of oceanic influence on climate” by Chylek et al. Nic Lewis also did a good analysis of the paper, see the Notes below for the links. I have a different take on it than theirs, one which centers on the opening statement from their abstract:
ABSTRACT: A multiple linear regression analysis of global annual mean near-surface air temperature (1900–2012) using the known radiative forcing and the El Niño–Southern Oscillation index as explanatory variables account for 89% of the observed temperature variance. When the Atlantic Multidecadal Oscillation (AMO) index is added to the set of explanatory variables, the fraction of accounted for temperature variance increases to 94%. …
They seem impressed with a couple of things. The first is that their four aggregated forcings of greenhouse gases (GHGs), aerosols, volcanic forcings, and solar variations, plus an ENSO dataset, can emulate the global average temperatures with an adjusted R^2 of 0.89 or so. The second thing that impresses them is that when you add in the AMO as an explanatory variable, the R^2 jumps up to 0.94 or so … I’m not impressed by either one, for reasons which will become clear.
There are several problems with the analysis done in Chylek 2014. Let me take the issues in no particular order.
PROBLEM THE FIRST
Does anyone but me see the huge issue inherent in including the Atlantic Multidecadal Oscillation (AMO) Index among the explanatory variables when trying to emulate the global surface temperature?
Perhaps it will help if I post up the explanation of just how the AMO Index is calculated …
From their link to the AMO dataset (see below)…
The [AMO] timeseries are calculated from the Kaplan SST dataset which is updated monthly. It is basically an index of the N Atlantic temperatures. …
Use the Kaplan SST dataset (5×5).
Compute the area weighted average over the N Atlantic, basically 0 to 70N.
Detrend that time series
Optionally smooth it with a 121 month smoother.
In other words … the AMO is just the temperature of the North Atlantic with the trend removed.
So let me ask again … if we’re trying to emulate the “global annual mean near-surface air temperature for the period 1900-2011″, will it help us if we know the detrended North Atlantic temperature for the period 1900-2011 … or is that just cheating?
Me, I say it’s cheating. The dependent variable that we are trying to emulate is the global surface temperature. But they have included the North Atlantic temperature, which is a large part of the very thing that they are trying to explain, as an explanatory variable.
But wait, it gets worse. The El Nino index that they use is a fairly obscure one, the “Cold Tongue Index”. It is described as follows (emphasis mine):
The cold tongue index (CTI) is the average SST anomaly over 6N-6S, 180-90W (the dotted region in the map) minus the global mean SST.
There are a number of El Nino indices. One group of them are the detrended average of the sea surface temperatures in various areas—El Nino 1 through El Nino 4, El Nino 3.4, and the like. There is also the MEI, the Multivariate ENSO Index. Then there are pressure-based indices like ENSO, based on the difference in pressure between Tahiti and Darwin, Australia.
There’s an odd wrinkle in the cold tongue index (CTI), however. This is that the CTI is not detrended. Instead, they subtract the global average sea surface temperature (SST) from the average temperature in the CTI area of 6°N/S, 180° to 90° W.
But this means that they’ve included, not just the average temperature of the CTI area, but also the entire global SST as a part of their explanatory variable, because:
CTI Index = CTI Sea Surface Temperature – Global Mean Sea Surface Temperature
I ask again … if you are trying to emulate the “global annual mean near-surface air temperature for the period 1900-2011″, will it help if an explanatory variable contains the global mean sea surface temperature for the period 1900-2011 … or again, is that just cheating?
I have to say the same as I said before … cheating. Using some portion of this year’s global temperature data (e.g. North Atlantic SSTs or CTI SSTs or global SSTs) to predict this year’s global temperature data is not a valid procedure. I’m sure my beloved and most erudite friend Lord Monckton could tell us the Latin name of this particular logical error, but Latin or not … you can’t do ‘dat …
Which is why, although the authors seem to be impressed that including the AMO increased the adjusted R^2 up to 0.94, I’m not impressed in the slightest. You can’t use any part of what you are trying to predict as a predictor. See how the AMO index (bottom right, Fig. 1) goes down until 1910, then up until 1940, down until 1970, and then up again? Those are the North Atlantic version of the very swings in temperature that we are trying to explain, so you absolutely can’t use them as an explanatory variable.
PROBLEM THE SECOND
Let’s look at just the forcings used in the climate models, setting aside the ENSO and AMO variables. Chylek 2014 uses the GISS forcings, which are composed of the following separate datasets:
Now, for anybody that thinks that e.g. ozone levels in the atmosphere actually look like that … well, seems highly doubtful. But while that is a problem in and of itself, that’s not the problem in this context. The problem here is that all of these are measured in watts per square metre (W/m2). As a result they should all have the same effect … but Chylek et al. do a strange thing. They add together the well-mixed ghgs plus ozone plus stratospheric H2O into one group they call “GHGs”. Then they put reflective aerosols, aerosol indirect, black carbon, and snow albedo into a second group they call “Aerosols”. Volcanic forcing are treated as a third separate group, solar is the fourth, and land use is ignored entirely. This grouping is shown in Figure 1 above.
Then each of these four groups (GHSs, Aerosols, Volcanoes, and Solar) gets its own individual parameter in their equation … but this means that a watt per square metre (W/m2) from aerosols and a W/m2 from solar and a W/m2 from GHGs all have a very, very different effect … they make no effort to explain or justify this curious procedure.
PROBLEM THE THIRD
Here’s an odd fact for you. They are impressed that they can get an R^2 of 0.88 or something like that (if they cheat and include the entire global SST within the “explanatory” variables of their model). I can get close to that, 0.87. However, let’s start by calculating the R^2 of a much simpler model … the linear model. Figure 2 shows the GISS Land-Ocean Temperature Index (LOTI), and a straight-line emulation. The odd fact is the size of the R^2 of such a simplistic model …
Note that the R^2 of a straight line is quite high, 0.81. So their correlation of 0.88 … well, not all that impressive.
In any case, here are a few more emulations, with their corresponding adjusted R^2. First, Figure 3 shows their group called “aerosols” (AER) along with the volcanic forcing (VOL):
Now, even this bozo-simple (and assuredly incorrect) emulation has an adjusted R^2 of 0.854 … or, if you don’t like the use of aerosols, Figure 4 shows the same thing as Figure 3, but with GHGs in place of aerosols:
There are a couple of issues revealed by this pair of analyses, using either GHGs or aerosols. One is that you can hardly see the difference between the two red lines in Figures 3 and 4. Obviously, this means that getting a good-looking match and a fairly impressive-sounding adjusted R^2 means absolutely nothing about the underlying reality.
Another issue is the difference between the strengths of the supposedly equivalent W/m2 values from GHGs, aerosols, and volcanoes.
Having seen that, let’s see what happens when we use all of the Chylek forcings except the cheating forcings (ENSO and AMO). Figure 5 shows the emulation using the sun, the aerosols, the volcanoes, and the greenhouse gases:
Note that again, watt for watt the volcanoes are only about a third of the strength of the GHGs. The solar forcings are quite strong, presumably because the solar variations are quite small … which highlights another problem with this type of analysis.
So that’s the third problem. They are giving different strengths to different types of forcings, without any justification for the procedure. Not only that, but the variation in the strengths is three to one or more … I see no physical reason for their whole method.
PROBLEM THE FOURTH
Now we’ve seen what happens when we’re not cheating by using a portion of the dependent variable as an explanatory variable. So let’s start cheating and add in the ENSO data.
As I said, I couldn’t quite replicate their 0.88 value, but that comes close.
Now, before I go any further, let me point out a shortcoming of all of these emulations in Figs 2 to 6. They do not catch the drop in temperatures around 1910, or the high point around 1940, or the drop from around 1940 to 1970. Even including all of the forcings, and (improperly) giving them different weights, Figure 6 above still shows these problems.
However, all of these global average temperature changes are clearly reflected in the corresponding temperature changes in the North Atlantic ocean … take another look at the bottom right panel of Figure 1. And so of course when they (improperly) include the AMO as an explanatory variable, you get a much better adjusted R^2 … duh. But it means nothing.
PROBLEM THE FIFTH
All of the above is made somewhat moot by a deeper flaw in their analysis. This is the lack of any lagging of the applied forcings. IF you believe in the forcing fairy, then you have to believe in lags. Me, I don’t think that the changes in global average temperature are a linear function of the changes in global average forcing. Instead, I think that there are strong emergent temperature regulating mechanisms acting at time scales of minutes to hours, largely negating both the changes from the forcing and any associated lags. So I’m not much bothered by lags.
But if you think that global average temperature follows forcing, then you need to do a more sophisticated lagged analysis involving at least one time constant.
• I find the analysis in Chylek 2014 to be totally invalid because they are including parts of the dependent variable (ENSO and AMO) as explanatory variables. Bad scientists, no cookies.
• As is shown by the examples using either GHGs or aerosols plus volcanoes (Figs. 3 & 4), a good fit and an impressive adjusted R^2 mean nothing. We get equally strong and nearly indistinguishable results using either GHGs or aerosols. This is an indication that this is the wrong tool for the job. Heck, even a straight line does a reasonable job, R^2 = 0.81 …
• Giving different weights to different kinds of forcing (e.g. volcanic, solar) is a novel procedure that requires strong physical justification. To the contrary, they have not provided any justification for the procedure.
• As you add or vary the explanatory variables, their parameters change. Again, this is another indication that they are not using the right tool for the job.
• The lack of any consideration of lag in the analysis is in contradiction to their assumption that changes in the global surface temperature are a linear function of changes in global average forcing.
Best to everyone,
De Rigeur: If you disagree with something I or anyone else says, please quote their exact words. That way, we can all be clear on exactly what you are objecting to.