Thursday, February 23, 2012

A(nother) climate sensitivity estimate using Bayesian fusion of instrumental observations and an Earth System model

The mode of the climate sensitivity estimate is 2.8C, with the corresponding 95% credible interval ranging from 1.8 to 4.9C.

Note that this is a 95% interval, meaning that the upper bound is at the 97.5% level rather than the commonly-quoted 95th percentile (for a 90% "very likely" range).

As I said previously, it will be interesting to see what approach the IPCC takes to this increasing number of "moderate" estimates appearing in the literature (bear in mind that a sensitivity of 3C still means a fair bit of climate change, but not as much as a sensitivity of 6C or 11C would...).

35 comments:

John Callender said...

Would you be willing to explain a little more (for the benefit of a non-scientist who prefers to get the implications described by someone who accepts reality, rather than by those who actively conceal it) what the implications of a lower sensitivity would be? In terms of warming scenarios, would lower sensitivity be expected to modify the projected future temperature rise, and potentially delay the arrival at tipping points resulting from positive feedbacks? If so, how much are we talking about? To what extent are the last batch of IPCC scenarios dependent on higher assumed sensitivity?

I'm not looking for you to produce a crystal ball. I'm just curious if you'd be willing to give more of your current thinking on this.

Thanks.

Carl C said...

I've been out of the loop for awhile but I thought a sensitivity of ~3C has been pretty standard stuff for quite awhile? hence all the mitigation efforts/limiting to 2C are "too little too late" etc?

reasonablemadness said...

"As I said previously, it will be interesting to see what approach the IPCC takes to this increasing number of "moderate" estimates appearing in the literature (bear in mind that a sensitivity of 3C still means a fair bit of climate change, but not as much as a sensitivity of 6C or 11C would...)."

James, have you read this paper at all?

I don't see why a estimate with a median value of 4.2 (uniform prior) or 3.0 (expert prior) from this paper would be any different than most of the estimates from AR4. The medians there where between 2.0 and 5.0, with a mean value of 3.3.

This is fully consistent with this study, as the authors say itself:

"Under the default assumptions of informative priors, the mode of climate sensitivity is 2.8°C, with the 95% credible interval from 1.8°C to 4.9°C. This mode is consistent with
many previous studies [..]."
(Quote from Olson 2012)

And even if you look at the upper tail, this study is not better, than most of the studies in AR4. The authors say:

"As in previous studies, the upper tail of the CS pdf is sensitive to priors."

And the AR4 said:

"The upper 95% limit for ECS ranges from 5°C to 10°C, or greater in different studies depending upon the approach taken, the number of uncertainties included and specific details of the prior distribution that was used."

So what has essentially happened since AR4? Nothing new.
If using uniform priors you still have an upper 95% limit of 6-8°C. This study is not better, it even has an upper 95% limit of 8.8°C (if cut of at 10K as in the AR4), higher than most studies from AR4 and most studies since then (e.g. 7.9°C from Royer 2007, 7.4°C from Forest 2008).

If using expert priors the upper tail is narrower and constrained around 5°C, but depends on the chosen priors. This was true in AR4, and is still the same. See Hegerl 2006 (1.5-6.1°C) or Forest 2006 (1.9-4.7°C).

So I don't understand what you expect to be different in AR5. I don't see that there has been substantial progress to better constrain the upper tail of climate sensitivity, despite all the effort that has been done to do so. The AR4 already said that CS is likely to be between 2°C and 4.5°C. CS values of 6°C and above were already unlikely with the knowledge from AR4.

Carrick said...

It seems to me an interesting related question is the variation in the warming across the world.

A higher number that is (hypothetically) less variable would be less dangerous than a smaller average value that was whoppingly large in some areas.

Do the models expect the geographical distribution of the relative rate of warming to remain the same from what we saw in the last century, or will it change (for example once the Arctic Ocean becomes ice free sometime during this century)?

David B. Benson said...

Carrick --- Regional change is certain; global averages aid only slightly in predicting important vaariables region by region.

James Annan said...

John (and Carl), the estimate here isn't actually low, compared to what the GCMs already simulate. However, some people had previously argued that the GCMs were (at least with some substantial probability) over-optimistic, and much worse outcomes were reasonably likely. So all this does is say that the models are likely to be broadly correct.

Carrick, it has been widely observed that even with a specified global mean change, there would be a lot of regional uncertainty. But there is a broad agreement on general matters such as land warming more than ocean.

reasonablemadness, the difference is that the AR4 clearly endorsed (as was widespread in the literature at that time) a uniform prior as the appropriate default basis for these estimates, and their "likely" sensitivity range of 2-4.5C (ie, 16% probability of a sensitivity greater than 4.5C) was supported by those analyses. Now as far as I can tell everyone agrees that a uniform prior is nonsense - I'm not aware of any counter-argument to our paper on the topic, and the default results presented by a range of researchers (including myself, Sokolov/Forest, and this group) no longer use it.

So it seems to me that if the IPCC authors wish to stick with their previous estimate in the face of this more recent and still growing literature it is going to have to come up with some pretty good excuses as to why this area of research (that it featured so highly in the past) is no longer considered reliable.

Carrick said...

Thanks for the response, James.

I assume this regional uncertainty is associated with a corresponding lack of regional skill of the models/

Is this something that can be addressed by improving the spatial (and maybe temporal) resolution of the models, or are there additional issues involved that would need addressing (e.g., improved description of the coupling of oceans to atmosphere, or biosphere involvement in land amplification of global warming)?

Thanks for any insight you can give, these questions are more along the lines of the sorts of things I'm pondering... what it would take to reduce the variability in the regional-scale predictions?

Carrick said...

I thought some people may not have seen these sorts of graphs before relating to latitudinal effect on warming, and might find them interesting. They relate the questions I was asking even if slightly OT.

Here's a comparison of land versus ocean trends derived from CRUTEMP/HADSST.

Notice that the oceans pretty much have a constant response to warming over multidecadal periods, at least between roughly the ±60° latitudes.

Here's GISTEMP split out by season.

(You can do this straight from their GUI interface.... pretty slick.)

One interesting implication of these results is that, if your geographic sampling changes enough, it can lead to a potential bias in your global temperature trend. (I did an analysis on the side on this at one point, what I found was that prior to 1950, the shift in the centroid of the geographical data was important enough to lead to a significant artifactual warming int the CRUTEMP algorithm. Here are my results, 7% is the amount of bias you are expected to see, it is just a statement that there is more land mass in the northern hemisphere than southern hemisphere.

Another is that if you have sites near marine-land boundaries, these will be heavily influenced by their proximity to the oceans. That makes it interesting to contemplate the effects of an inland shift over time of coastal stations.

reasonablemadness said...

James wrote:
"the difference is that the AR4 clearly endorsed (as was widespread in the literature at that time) a uniform prior as the appropriate default basis for these estimates"

Agreed, but as it was widespread in the literature at that time, I don't see why that would be a wrong choice back then. The IPCCs task is to summarize the literature, not to create a new one.


"and their "likely" sensitivity range of 2-4.5C (ie, 16% probability of a sensitivity greater than 4.5C) was supported by those analyses. "

But that's the point: I don't think, that if you look at all the estimates in AR4, that you could safely call 4.5°C the upper 66% limit. The average 66% limit of all the studies cited in AR4 is 5.4°C (mean) resp 5.2°C (median), nearly 1°C higher than the AR4 conclusion. So the IPCC was already a bit more confident, than the studies average suggested. It is not clear from the report, which methodology they used to come up with the 4.5°C limit. I could only think, that they lowered the value, e.g. *because* of the non-uniform prior estimates which did also exist back then and which were also cited in the report.

So, even if you look only at the studies which used non-uniform priors, I don't see how the conclusion of 2-4.5°C should change very much. If you compare the results from studies which are using uniform priors, you see..

http://img832.imageshack.us/img832/1736/climatesensitivityestim.png

... that the average likely-range with non-uniform priors is about 2-4°C, so the lower limit is exactly the same as in AR4 and even the upper limit is only 0.5°C lower.

And even if you look at pre/post 2007 estimates, to see how the estimates have changed, there is not very much progress at the 66% limit:

http://img32.imageshack.us/img32/9676/csestimatesnonuniform.png

The only thing that dropped clearly (by nearly 1°C) is the upper 95% limit. The 66% limit has dropped only minimally and is essentially the same. So I don't see that the AR5 has to deal with substantially more "moderate" estimates. The likely-range stayed nearly the same. It has to deal only with the change, that most studies don't use uniform priors anymore (as can be seen clearly in the right plot in the first link through the lack of those studies in the post-2008 era) and that most of these studies suggest, that we can rule out extremely high sensitivites above 5-6°C. So that is after all a good thing.

But as I said, the central estimate is still somewhere between 2-4°C. And this is still a huge range of uncertainty, when you look at how warm it will be getting, what changes we have to adapt to and how much time we have left to act.

James Annan said...

Carrick,

Thanks, that is interesting, though I'm not sure what calculation you have done. HadCRUT3 has large gaps at high latitude which certainly biases the trend downwards.

James Annan said...

reasonablemadness,

Certainly the IPCC summary was not based exclusively on this type of Bayesian estimate. Note however that even if only one accurate and trustworthy estimation is made, then the existence of poorer constraints would provide no justification for the IPCC to adopt a wider range - just as my inability to accurately determine the time based purely on the position of the sun in the sky does not mean that I should therefore conclude that my watch is probably more than 2 mins wrong.

I'm not trying to criticise the IPCC back then (at least, not here). However, the current literature seems quite clear both on agreeing that uniform priors are inappropriate and also on finding that when a non-uniform prior is used, a much tighter estimate for S results which does not have a long high tail.

I don't see how these results can be reconciled with their previous "likely" range. Even though they do not represent any great leap forward in terms of the underlying evidence (though there have of course been a few more years of data), they do represent a change in our approach to how these data should be interpreted.

Carrick said...

James, I used the CRUTEM3 algorithm to compute the temperature in each 5°x5° cell for each month starting from 1950 through 2009, computed the temperature in 10°-latitudinal bands by summing over all non-zero 5°x5° cell, then dividing by the number of non-zero cells.

I repeat this for each 10° band, in 5° steps (so -90...-80, -85..-75 etc). I then used an OLS algorithm to compute the trend for each latitude.

You can compare my estimate to GISTEMPs here. If there is interest, I can produce a plot of them on the same graph. I suspect within error bars, the difference isn't so great.

(Scroll down to see the plot of temperature trend versus latitude.

For people that aren't aware of it, There's a really cool widget here. You can choose land-only, sea-only or land+sea as well look at means versus trends.

With a bit of math and the assumption that the latitudinal amplification factor (I call it "Lambda") is constant over time, you can compute the effect of missing 5°x5° cells on the mean temperature trend.

This is the result I derived, which applies specifically to the CRUTEM3 algorithm (which averages only over non-zero cells).

Basically I think you need to do something akin to "pre-whitening" the data to remove the effect of a shift in mean latitude over time.

Hope that is somewhat illuminating, and sorry the discussion is a bit abridged.

Carrick said...

Also This is the formula for latitudinal bias I used (converted to a sum of course). Again N(theta) is the number of non-zero 5°x5° cells. It would have to be adjusted to compute the bias for GISTEMP, since they apportion the surface differently (and also for Nick Stoke's Voronoi tesselation method).

steven said...

James,

I recall you going to this conference. What did you think
of this presentation

http://www.newton.ac.uk/programmes/CLP/seminars/120812001.html

Carrick said...

Seems like there's an error in the GISTEMP link that I provided ... it appears to be interchanging zonal mean of temperature and trend (otherwise the trends are way too large... and would indicate a much larger warming than we've actually seen). If I use what I think is the correct figure, this is what I get

Curiously, GISTEMP is a bit lower over that period than CRUTEM3. (Note however this is my own reanalysis of the CRUTEM3 data, based on their 5°x5° gridded data and not an official product of theirs).

If you look at the temperature trends computed from the global means, it seems to bear this out, though (these are in units of degrees C per decade)


BEST 0.187
CRUTEM3GLl 0.152
CCC/GISTEMP 0.131
NCDC/land 0.187

It's my suspicion that the main effect of the difference in the sampling of high latitudes is on the "leakage" of short period climate fluctuations (which tend to have fairly short spatial wavelengths of course) into the global mean. In other word, the short-period noise of the global mean is more largely affected than the trend of the global mean.

So if you compare the last 10 years, where GISTEMP has the somewhat touted higher trend in temperature compared to the other series, that if you waited 30 years then did a 10-year trend, it might end up with the lowest temperature trend.

Of course that's speculation and would need a model to confirm. ;-)

Carrick said...

I should have mentioned I'm using land-only data here in this comparison.

James Annan said...

stephen, thanks for that - I didn't stay in Cambridge long enough to see it first-hand. Looks like a fairly typical approach, but they seem to get quite a tight (and low) result even with uniform prior on S. There doesn't seem to be a published paper though.

steven said...

aha, Carrick.

I think I get it now.

James: I liked your talk there. I think Iv'e sat through a good portion of the talks. I wish folks would post more of these types of things. For me watching a talk is way better than reading a paper.

James Annan said...

Carrick, I'm still not sure what case you are making here. If the gaps are predominantly at high latitudes (which I believe to be the case) then filling them in will lead to a higher trend but at no point will it lead to an overestimate of the true trend - it will just reduce the original low bias. Isn't this the case?

Carrick said...

James, not all of the current gaps are high attitudes (Africa contains some). I'm not really making a case about that, merely pointing out that over the period 1950-2009 GISTEMP in spite of its greater high latitude coverage paradoxically ended up with a lower trend. I also compared the CRUTEM3 to GISTEMP land only and showed they were pretty similar when overlaid against each other.

I then gave some speculation about why that might be and suggested that this might be worth exploring using models (semi-empirical variety because you need to be able to account for short-period fluctuations, which I think we've agreed the full AOGCMs have little skill with reproducing currently).

I also suggested that the recent departure from GISTEMP might be simply the difference in "leakage" of short-period noise into the global mean average (in the sense that James Hansen writes about). But that's speculation and an aside.

Previous to 1950, many of the missing land cells were actually at lower latitudes which can be seen by the plot of mean latitude of nonzero 5°x5° cells versus time. This is the part of the record I've been focusing on.

This movie is a bit sucky, sorry but it shows the coverage of the land-area by CRUTEM3 versus time (it would be much better if I put the years on there, but I did this for my own sake originally).

Prior to 1950, higher latitudes were oversampled and the Southern Hemisphere was almost not covered at all, meaning that a "naive average" would lead to an over-estimate of temperature trend for that period because higher latitudes have a larger latitudinal amplification
factor. [*] If that result held up it would probably be good news for the models, since as I understand it, they have a bit of trouble explaining all of the "observed" warming between 1905-1945 without "unrealistic" increases in solar intensity.

(The amount of bias depends intricately on the method used for the global temperature reocnstruction, but might be fixable by an algorithm similar to "prewhitening" that we use in digital signal processing).

I agree with you that an undersampling of high latitudes should have the opposite effect, the bias should be negative (if you squint you can see about a 2% effect in the CRUTEM3 algorithm due to this in recent years).

Hope this clarifies things a bit.

[*] For the activists who worry about public reaction to any potential findngs, this is nothing to sweat about---most of the warming prior to 1970 is thought to be of natural origin, due to competition between the cooling effects of anthropogenic aerosols and warming effects of GHGs. See for example GISS Model E assumed forcings....

Carrick said...

I focused on CRUTEM3 because algorithmically it is simpler to model than GISTEMP.

You could model the effects of land bias and the changing geographical distribution for GISTEMP of course, but it certainly would be more complex to deal with.

James Annan said...

Carrick,

Can you explain what the "bias" is in your graph - ie, bias of what relative to what? Is this really the difference between the mean lat of the nonempty land boxes in HadCRUT, and the mean lat of the actual land? The numbers seem implausibly high to me, given that the overall coverage is around the 80% level and has been for some time.


This pic is what I mean:
http://dl.dropbox.com/u/4520911/Climate/land_latitude_bias.jpg

Carrick said...

James, the represents the bias associated with the location of the geographical centroid, which is not the same as the bias error.

From my earlier comment

"7% is the amount of bias you are expected to see, it is just a statement that there is more land mass in the northern hemisphere than southern hemisphere."

So the bias error would be the difference

bias_error = theta_bias – 7%

Since 1950, the bias error in the data has been negligible, since the mean latitude of the data has not moved appreciably since then (note that the mean latitude is 15.5°N)

Since this is confusing, here is the bias error for the particular algorithm choice I made.

Figure.

Sorry for the confusion. Note that the "fuzz" in the curve represents the annual cycle (some stations don't report in winter.)

[Again, cautionary note, the value of theta_bias and the "expected value" of theta_bias depend on the details of the algorithm. This particular result is based on summing over all non-empty 5°x5° cells. But there are actually two CRU algorithms, in one version they sum the NH and SH separately then compute (T_NH+T_SH)/2. That doesn't mean the effect of missing cells can be neglected in other algorithms, just that it will be different than what I computed here.]

Carrick said...

Also, this figure

http://dl.dropbox.com/u/4520911/Climate/land_latitude_bias.jpg

just measures the mean latitude of the land mass. The "real" value is around 17°N if I recall (springs from the well known factor there is more land in the Northern hemisphere.)

Calling it a "bias" is a bit of a misnomer. It represents a physical bias in land mass, not a bias resulting from an algorithmic error.

Carrick said...

Actually the true geographical latitude is 15.5°N

Here's the raw data, first column is latitude, second is fraction of land area in that band.

This value is slightly different than what you would get if you computed it using 5°x5° cells and of course that in turn depends on how you weight partially empty cells (I think that does shift it a bit further north).

There's a bit of discussion of this in the comments here.

Carrick said...

Here's of a bit of an apples to oranges comparison.

CRUTEM3 non-empty land cells divided by total cells as function of latitude versus "true" fraction of land mass as function of latitude. An apples-to-apples comparison would require me to grid the latitude data in 5°x5°bins, using the same algorithm they use.

I think they take cells with any land data and use a 100% weighting factor (at least their published perl code does that). It's easy enough to do, so if somebody wants to see the comparison with available 5°x5° land cells, I can generate it.

Carrick said...

Missing figure.

reasonablemadness said...

James wrote:
"However, the current literature seems quite clear both on agreeing that uniform priors are inappropriate and also on finding that when a non-uniform prior is used, a much tighter estimate for S results which does not have a long high tail."

Well, I agree to that. But it's kinda obvious, that a non-uniform pior generates not such a long high tail, when you don't have as many model versions e.g. with extremly negative aerosol forcings that would allow a model with a very huge climate sensitivity to still simulate e.g. 20th century warming very well. A non-uniform prior still allows for such a extreme values, but makes them much less likely, which reflects our knowledge surely better. So I'm not saying you shouldn't use non-uniform priors. If chosen carefully, such estimates will certainly make a better guess on climate sensitivity. And I'm pretty sure, that the next AR will also come to a similar conclusion.

But I'm still not sure, how you came to this conclusion:

"I don't see how these results can be reconciled with their previous "likely" range."

Well, what would you say is a likely-range that would fit to the newer results? I just showed, that the non-uniform studies can't constrain climate sensitivity with a 66% change in a better range as 2-4°C. And that is pretty much the previous "likely" range of AR4.

Carrick said...

Offering a bit of a summary of what I think I've found:

1) Polar land amplification leads to artifactual warming or cooling, when the distribution of temperature stations shifts north or south. (Note I use the full formula based on number of non zero cells versus latitude to compute the bias, not just the variation in the mean geographical latitude of the stations over time. That's merely for illustrative purposes.)

2) The amount of (primarily natural) warming prior to 1950 is LIKELY overstated in CRUTEM3, and POSSIBLY the other series as well. This means that the amount of warming since 1970 looks even more extreme than it did, assuming the correction that I'm suggesting needs to be made, is made.

3) This only affects the assume natural portion of the radiative forcing, I think if I'm right, it makes it slightly easier to reconcile toe models and data prior to 1950, that is, it works in the same direction that the SST bucket problem does.

4) This analysis supports the conventional view that the reconstructed land and global temperatures are largely reliable since 1950.

5)I was thinking for a while it might suggest that the climate sensitivity would get inflated if natural variability were less than we thought, but I don't think it affects the estimates of climate sensitivity.

Hopefully this summary is somewhat helpful.

James Annan said...

rm,

well I have cited several papers that seem to point to a smaller range. It's certainly quite a change from 5 years ago when there were none. I'd not complain if they said 2-4C - there was plenty of noise last time when they shifted the lower end up by 0.5C. Also I suppose I didn't explicitly point to it, but the "substantially higher values cannot be excluded" also seems inappropriate IMO.

James Annan said...

Carrick, ok I think I see what you are saying - the older data is mostly high lats, so should (when converted to a global mean) have overestimated the trend (also, the more recent obs have a small southward bias). That's probably true in principle - but note that solar forcing (which was more important back then) does not generate as strong a polar amplification, and far enough back in time there wasn't much of a trend to overestimate anyway. Also, it's worth noting that any proper comparison with models - eg for the purposes of detection and attribution - is generally made on a gridpoint (or at least regional) basis, so the changing location of obs should not matter here. This is a point that many blogs posts etc do not make clearly enough IMO - I was already planning to post on this.

Carrick said...

James: "Carrick, ok I think I see what you are saying - the older data is mostly high lats, so should (when converted to a global mean) have overestimated the trend (also, the more recent obs have a small southward bias)."

Yep that's the basic idea.

" That's probably true in principle - but note that solar forcing (which was more important back then) does not generate as strong a polar amplification, and far enough back in time there wasn't much of a trend to overestimate anyway"

I agree the land-only trend for that period (1905-1940 or so) is only about 1/2 the trend of the 2nd half of the 20th century. I remember seeing criticism of the models ability to reproduce even that trend without fudging (as some people claim) the solar forcings.

I'm also not sure I understand why solar forcing doesn't generate as strong of a polar amplification as CO2 forcings. I'm wondering if you could explain? Actually I've never seen a really good explanation of the mechanism or mechanisms behind polar amplification (I thought one of them was surface ice loss, and my recollection is that was worse during the first half of the century).

[I also seem to recall a Hansen paper on temperature reconstruction using ice cores where he assumed the polar amplification was constant. Hm...]

"Also, it's worth noting that any proper comparison with models - eg for the purposes of detection and attribution - is generally made on a gridpoint (or at least regional) basis, so the changing location of obs should not matter here."

That sounds interesting... especially if the models can live up to the test on that scale.

Anyway thanks for the comments.

James Annan said...

As for solar vs GHG for polar amplification, note that any change due to solar forcing will be greatest at the equator (in terms of additional W/m^2) just due to geometry. CO2 forcing is roughly equal at all latitudes (though humidity and other details affect this a bit). So maybe it's not really the "amplification" that differs so much as the spatial distribution of the forcing. But I haven't looked into the mechanisms or details myself. I believe this is one of the basic limitations of solar shading as a means of geoengineering.

Most D&A stuff works in terms of a handful of EOFs really, so although these use the spatial pattern and should correctly account for the locations of missing data, it would be a stretch to claim that the models are really accurate at the grid box level.

James Annan said...

Looks like Aldrin has a new paper here (but I'm away from work for a few days and don't have access - come to think of it, we probably don't even get this journal anyway). The abstract doesn't say what the actual results are...

crandles said...

Have you looked at

Cohen et al
asymetric seasonal temperature trends
http://web.mit.edu/jlcohen/www/papers/Cohenetal_GRL12.pdf

Does this have implications in that you were unlucky not to know about this before deciding to make the record global temperature bet?