I’m a bit late to the party, but I’ve now had an opportunity to look over the Spencer and Braswell 2011 paper, along with the criticisms from Real Climate. I’d like to keep this technical amid all the controversy, and clearly I have some catching up to do with the Dessler 2011 paper (along with Spencer’s responses on his blog), which I’d like to look at in a different post. The script I’ve used to reproduce the model in the paper is available here.
From my reading of SB11, there seem to be three main points:
1) Using a simple model, it can be shown that unknown radiative forcings lead to underestimates of the climate sensitivity.
2) The lagged signatures of the observations don’t match up well with the lagged signatures of the global climate models.
3) The lagged signature of the simple model with radiative forcings matches up well with the lagged signature of the observations, suggesting that there are significant radiative forcings over the period.
The main technical criticisms laid out at RealClimate seem to be:
1) The differences between the models and observations in SB11 result from a combination of dataset choices (CERES SSF and hadCRUT), the choice of models to compare against, and noise.
2) The SB11 model is too simple, excluding ENSO, which may in itself be the reason for the lagged signature without needing to invoke radiative forcings, and the SB11 match with observations may be the result of tuning.
3) The ENSO variations are not the result of cloud forcings, and so the SB2010 argument #1 is moot.
To me, Spencer’s model makes sense for illustrating his simple first point (that unknown cloud forcings could cause a misdiagnosis of the climate sensitivity). My reproduction of his first figure using monthly steps is available here:
I don’t believe this point is in dispute — the only dispute is whether it is actually relevant here (that is, whether there exist unknown radiative forcings over the period).
For SB’s point number two, things are not nearly as clear-cut. I find the differences interesting, but as we’ll see later, I think figure 3 (shown below) actually undermines SB’s later point. And as the Trenberth and Fusullo post at Real Climate points out, it is important to see to what degree the differences are the result of dataset and GCM choices. On the other hand, adding the error bounds the way that TF does could also be misleading…what we essentially care about here is the amplitude of the variations in the lag regression plots, and so if all of the model runs have essentially the same shape but are merely shifted up or down the y-axis, it gives the impression of more uncertainty regarding the "shape" than actually exists. The bottom plot in the TF post, which uses all models, shows so much uncertainty that it is hard to find observations that wouldn’t fit inside it.
The weakest part of the SB10 paper to me is point number three. Yes, the simple model matches up well with observations, as they highlight in figure 4, and which I have attempted to reproduce below:
However, note the shape of the non-radiative forcings line, and then compare it to those lines of the climate models from SB figure 3 above.
To make the case that the climate models are off because they assume variations are not radiatively forced (in the 21st century), you would expect the climate model lines to look like the non radiatively forced line in figure 3. Of course, the plot also uses runs from the 20th century, which includes periods where the GCM variations were largely radiatively forced, so the result is somewhat muddled. Nonetheless, because the GCMs don’t match the non-radiative forcing line, and because they have a shape that is at least somewhat more similar to the observations than that line, it suggests that the simple model might be missing something in this lead-lag relationship that IS present in the GCMs. That brings it to the main issue that others have raised — are there any other physically reasonable models (e.g. those that include ENSO) that can reproduce the observations? My understanding is that Dessler 2011 focuses on whether GCMs are up to the task.
Now, with respect to the SB model and tuning, the equation seems pretty straight-forward, but with 3 major choices: 1) the depth of the ocean mixed layer, which determines heat capacity, 2) the choice of lambda (the feedback response), and 3) the choice of noise models for both the radiatively and non-radiative forcings. The figure below shows the 70% radiatively forcing, but with lambda and the ocean layer depth changed:
At first glance the tuning doesn’t seem to be much of an issue. The choice of ocean layer depth does not strongly affect the amplitude, whereas the sensitivity does…which is the point of the exercise. SB are trying to show that those models with high sensitivity (lambda=1) yield the flatter lines in the regression, whereas those with low sensitivity (lambda=3) get more amplitude, and more in synch with the observations. This perhaps explains why SB stratified GCMs in terms of climate sensitivity in their lag regression chart. However, as shown in the TF post, the GCMs don’t necessarily break down that way.
If the "simple model" with no radiative forcings could simulate the lagged signature of GCMs in the early 21st century, I think SB11 would have a much stronger case. As it is, the evidence presented in the paper leaves one wondering if there are other models with non-radiatively forced ENSO variations that can also explain the lagged signature.
However, just because I don’t find SB point #3 conclusive does not mean I necessarily agree with TF’s point #3. That clouds respond to ENSO does not mean they can’t result in the type of misdiagnosis that SB refers to, as I explained in a comment at the Air Vent:
Consider the hypothetical scenario where winds associated with El Nino blow clouds in a specific region from over an area of low surface albedo to one of high surface albedo, thus creating a downward positive flux anomaly of X over what we would expect otherwise. Now, at the same time a global temperature increase, dT, has occured due to El Nino, causing a radiative feedback response of Y.
What we are trying to determine is the radiative response to the temperature increase, or Y/dT. But we only have the TOA flux measurement, which, since X and Y are in opposite directions, will be equal to TOA = Y – X. So obviously the (Y-X)/dT will be an underestimate, depending on the size of X.
Now, you can label X whatever you want, as it may be driven by ENSO. But X is not a response to the temperature increase of ENSO (which is what we care about WRT CO2 sensitivity), and causes an underestimate of the actual response to temperature if it is unknown or ignored.
This point seems to be missed amid all the arguments over the definitions of "forcing" or "feedback". And yet, if this “raditively forced” portion from clouds is indeed tiny, it won’t make much of a difference. From what I gather, Dessler 2011 dives deeper into the particular analysis of what percentage of the variation is radiatively forced. I’m looking forward to seeing what comes out there, hopefully in another post.
As always, any comments explaining what I have wrong here are appreciated.