Tuesday, June 09, 2015

That hiatus thing

The hiatus is no more, apparently. Or rather, it never was. “Nature Hiatus” might have to change its name, or at least its main focus of publication.

I never really understood why the “hiatus” was such a thing. Whether or not the warming trend since some carefully chosen date is positive, negative, and “significantly” so or not, is mostly an exercise in cherry-picking and the abuse of significance testing (The Difference Between “Significant” and “Not Significant” is not Itself Statistically Significant), not to mention the sort of “gotcha” that belongs in the political domain if anywhere. What matters is how well the obs agree with model projections, and there is no particular threshold of zero trend that has any special importance in that respect. Furthermore, whether or not there is an interval with zero or negative trend, no-one with any clue would dispute that we will continue to see warming in the long term, with some natural variability overlaid on top of that.

Lots of people have blogged this in some detail. RC, Doug MacNeall (Doug can you please change your name to a sensible spelling?) and Stoat, to name but three. I am still somewhat unconvinced by some of the model-data comparisons which smack rather too much of move-along-nothing-to-see-ism for my taste. The models do generally overestimate the trend over quite a long period, it is pretty marginal to claim that they agree with the data, and we've been waiting for the long-promised acceleration in warming for some time. As recently as 2006 or so, many prominent scientists were proclaiming an expected warming of 0.72C over the inerval 2000-2030. That's not looking too likely right now. Even with the multiply-promised El Ninos (and 2015 looks set for being another record year), the recent warming is no more than steady.

20 comments:

Tom C said...

Thanks for your honesty. Seems to me that "pretty marginal" is a euphemism.

Carrick said...

James: What matters is how well the obs agree with model projections, and there is no particular threshold of zero trend that has any special importance in that respect.

In practice, I think most climate scientists were really discussing “ a period of a slower rate of increase of the global mean surface temperature (GMST), the globally average land and sea temperature at the bottom of the troposphere" (wiki) and not a true "pause" (zero or negative trend).

Part of the trouble has been the language hasn't been very precise. I would suggest here that "global warming slowdown" is a much better description of the apparent feature in the data than "pause".

Whether or not the warming trend since some carefully chosen date is positive, negative, and “significantly” so or not, is mostly an exercise in cherry-picking and the abuse of significance testing

Following my previous comment, what I think you'd really want to look at, though, is whether the trend from say 2000-2014 is statistically distinct to that from say 1970-2000. Most of us would, I think, agree there was a visual change in the pattern of warming circa 2000 AD. We can get tricked though, because our eyes are designed to pick out patterns, and will do so gleefully even when it's just noise that we're staring at.

I made a few attempts at a better definition than what I thought were very poor efforts in the IPCC: Depending on the observational data set, the GMST trend over 1998–2012 is estimated to be around one-third to one-half of the trend over 1951–2012 (Section 2.4.3, Table 2.7; Box 9.2 Figure 1a, c). For example, in HadCRUT4 the trend is 0.04°C per decade over 1998–2012, compared to 0.11°C per decade over 1951–2012.

This just seems to be a very crappy definition to me, at best. The period 1951-1970 is noted for a marked lack of warming. It seems like some effort is being made to dilute the difference in trends between the period 1970-2000, which saw a marked increase in rate of warming, and the period 2000-now, which has seen a somewhat slower rate of warming, albeit possibly not significantly so.

James Annan said...

Tom, no I really do mean pretty marginal. It depends how you do the comparison, and of course choosing the parameters after seeing the data (as everyone has done on all sides of the debate) completely invalidates the whole statistical theory anyway. A couple of warming years (2015 looks like a hottie so far) and the agreement will look a bit better.

James Annan said...

Carrick, as above, you have to be very careful with the testing when you choose the analysis after seeing the data. Plus, there is no reason to expect the forced component of the two trends to be precisely the same (though to be fair most would expect the second to be slightly higher). Plus, even if you determine there is a "significant" difference, so what? There is no plausible theory for a reduction in the forced response, you are left with either being unlucky with natural variability (possible even with a robust 99.5% significance failure) or some other external forcing that we have no plausible hypothesis for.

That said, I do agree with much of what you say about the IPCC trying to brush the problem under the carpet.

Victor Venema said...

What matters is how well the obs agree with model projections

Where one should look at the model uncertainty and not at the multi-model ensemble spread. They are not the same, I do not think that anyone would make the claim that climate models model natural variability realistically and many components of the models are the same or similar.

Where one should look at the observational uncertainty and not only at the sampling uncertainty. This determines the error bars for most global temperature dataset. HadCRUT takes some aspects of biases due to inhomogeneities into account, but far from all. Especially the uncertainty due to the transition to Automatic Weather Stations is missing in all datasets.

Carrick said...

James: Carrick, as above, you have to be very careful with the testing when you choose the analysis after seeing the data.

Absolutely. To study this properly, IMO, you'd need a good metric for features of the sort we are discussing, and then you need a model for the natural fluctuations over that same time period. My guess (which is why I won't spend any time on this) is when you get done, you'll still find a result of marginal significance that, as you suggest, doesn't still teach you very much about how climate works.

There is no plausible theory for a reduction in the forced response, you are left with either being unlucky with natural variability (possible even with a robust 99.5% significance failure) or some other external forcing that we have no plausible hypothesis for.


I agree of course that the anthropogenic forcing is increasing (though there are people who are suggesting that recent volcanic activity may be playing a role in some of the moderation in the rate of warming).

One candidate for natural variability that would explain the recent apparent slowdown in warming would be a ~60 year AMO cycle with a physically significant amplitude in global mean temperature. I don't think it's implausible that this is at work here, though I obviously lack the expertise to credibly defend the presence of such a 60 year oscillation with a significant amplitude either.

Were it to be present and important, there are obvious implications of this for e.g. the period of rapid warming from circa 1970-2000. See e.g. this "toy" model result:

figure

Shown is the OLS of a polynomial (secular trend) plus sinusoid. The amplitude, frequency and phase of the sinusoid were all fitted parameters of the model.

Given the limited number of oscillations, this is nothing more than a curve fitting exercise. But I think the figure is still useful as an indicator of the significance of a "real global warming reduced warming period" associated with a 60-year AMO global mean temperature cycle.

Carrick said...

Victor: Where one should look at the model uncertainty and not at the multi-model ensemble spread.

Yes, but that's rather hard to do in practice.

Among the issues you'd face is the likelihood that the modelers did not explore the full parameter space of their models when generating an ensemble of model runs.

Getting a reliable measure of the real model uncertainty is I think really tough, at least until somebody builds a computer where you can do hundreds of simulations per year at a credible resolution where you aren't replacing real physics with toy models.

Anonymous said...

James,
You're probably right that something like 0.72C over the period 2000 - 2030 is unlikely. However, given that we could increase anthropogenic forcings by 1W/m^2 (or maybe a little more) by 2030, and given that we currently appear to have a planetary energy imbalance of something a bit bigger than 0.5W/m^2, something close-ish isn't physically implausible. If I look at the AR5 radiative forcing diagram, it appears that we increased anthropogenic forcings by 1W/m^2 between 1980 and 2011, and the temperature datasets suggests we warmed (globally averaged) by just over 0.5C over that time period.

Of course, it will depend on our actual emissions, but if we do continue to follow a high emission pathway, it certainly wouldn't surprise me if we did warm much more rapidly in the coming couple of decades than we have in the last 15 years or so.

James Annan said...

Carrick,

Well irrespective of the nature of the internal variability it is still the case that a period of negative contribution is probably preceded by a period of positive contribution. But overall the best estimate of the trend is still given by the full time series.

ATTP, well people have been saying things like this for a long time now, dating back at least to Hansen 1984...we are still waiting...

Anonymous said...

James, you say: "I am still somewhat unconvinced by some of the model-data comparisons which smack rather too much of move-along-nothing-to-see-ism for my taste."

Could you elaborate on this? Schmidt et al (2014) and Huber and Knutti (2014) are (IMHO) credible reconciliations of models and observations. Those analyses take into account both coverage bias in observations and updated forcings. Thus, they represent a much-needed corrective for AR5 Chapter 9, which unfortunately greatly exaggerated model-obs discrepancies.

HR said...

I guess first off if people are talking about the long promised acceleration in warming in the future then really what they are not doing is talking about the slowdown today.

Secondly ...... I've forgotten 2ndly! Ah yes if you think about the obs comparison with each model individually rather than just the ensemble then for a good number of those models ( specifically the hottest ones) I would suggest "pretty marginal" starts to look like 'somewhat wonky'.

James Annan said...

DC, it wasn't so long ago that the consensus was clear that the decadal-scale forced trend was insensitive to scenario. Also (and this is also related to HR's comment) the more sensitive models are too sensitive specifically to GHG forcing according to D&A analyses that look at the fingerprint (stott et al), so this takes other forcings out of the equation. OTOH it is always risky to disagree with Gavin :-)

Anonymous said...

" DC, it wasn't so long ago that the consensus was clear that the decadal-scale forced trend was insensitive to scenario."

Sure, but changes to the (anthropogenic) scenarios are a smaller part of the Schmidt et al analysis. And most of the update to natural volcanic forcings occurs pre-2005.

Correcting for coverage bias implies a short-term 15-year linear trend of ~0.1C/decade, instead of ~0.05C reported by IPCC. The full range of 5 updated series linear trends for 1998-2012 is 0.085C/decade [-0.06 to 0.23], using ARMA(1, 1) noise model to account for serial autocorrelation. For 2000-2014 that rises slightly to 0.094C/decade [-0.04 to 0.22].

That's still a slowdown, but not a "hiatus". (I am most certainly *not* in the "nothing-to-see-here" or "no slowdown" camp). It does falsify the IPCC claim that more than 95% of model run trends were above the surface temperature trend for 1998-2012.

On the model side, sure, there are definitely some models that are too warm. IIRC, attribution studies need to downscale the hindcast model ensemble by 10%.

I'd have to go back and check the details, but apparently Gavin Schmidt did mention at Spring AGU that he was involved in efforts to assess CMIP6 model reliability, and move beyond "one-model, one-vote" analyses.

Anonymous said...

James,
ATTP, well people have been saying things like this for a long time now, dating back at least to Hansen 1984...we are still waiting...
Well, yes, but that's why I was suggesting that it was physically plausible, rather than trying to make an actual prediction/projection :-)

To follow up on that point, though, are you aware of any estimate (or way of estimating) the planetary energy imbalance we could sustain (for decades say) without the surface warming exceeding something like 0.1C/decade? If the variability in how fast we can accrue energy in the deep ocean is large, than maybe we can build a large planetary energy imbalance without an acceleration in surface warming (which would have to happen at some stage, but could be delayed). If the variability is not large, then presumably there is a limit to how large the planetary energy imbalance can become without accelerated surface warming. Hence, one could estimate an emission pathway along which accelerated warming becomes quite likely.

Well, that was my thinking at least, but maybe there's some reason why this doesn't make sense, or the system is too complex to think of in this way.

Carrick said...

DC: That's still a slowdown, but not a "hiatus".

Actually the IPCC uses "hiatus" to mean a slowdown, which is very clear from their definition "the reduction in [global mean surface temperature] trend during 1998–2012 as compared to the trend during 1951–2012." I've seen argument about whether this is appropriate word choice on their part—and it's why I prefer using "slowdown"—but truthfully I don't see that question very relevant (they have clearly labeled what they meant by the word, we can agree there are better choices and choose to use those instead).

The trouble I have with Gavin and others work is that it smacks of confirmation bias. I don't doubt there are enough uncertainties that, if all of the tweaks you make in your analysis work to reduce the discrepancy between the series, that you'll eventually succeed in doing so.

This reminds me of a experiment an acquaintance did in college measuring Newton's constant of gravity "G". They were measuring the force between a cylinder mounted on a torsion bar compared to a larger cylinder on a table.

The class average for big "G" was really close to the then current published value. The problem is that the attraction of cylinders gravitationally isn't just G* m1 * m2/r^2 (which is what they all assumed). It's got a more complex form.

Nobody should have measured the published value, except through measurement uncertainty. I'm sure most people were able to "honestly" arrive at near the "right" (but actually wrong) answer innocently. It's just a statement of confirmation bias: You know the answer you should have measured, and you work to correct errors that tend to push you away form the right one.

The fact there was a discrepancy was a sign there was real physics that weren't being accounted for, and had you simply accepted this as a fact, perhaps at least one of the students would have had a moment of insight and realized you needed to use the formula for two attracting cylinders (which is a published result, though it's one people with a decent mathematical background can derive themselves).

I think this may be the same issue here: My reading of the apparent discrepancy between the models and data is that the models underestimate the influence longer-period natural fluctuations. If I'm right, then in less than 15 years at at the latest, we be back in track with an period of increased warming. After all, the trend over 15 years is interesting to us as humans, but it isn't very interesting for climate change.

When I first got into this "game", people were saying you needed 30 years of data to test models against data. I don't see any reason we need to change that definition now.

Anyway, there are likely economically interesting outcomes associated with an increase in natural variability over what the models predict (especially if the natural variability is associated with pseudo-cycles, and is then somewhat predictable). So it's probably worth improving the models to more accurately predict these shorter period fluctuations, though I suspect it's probably a challenging thing it do given the current state of computational resources.

Steve Bloom said...

So in the '90s we had an acceleration in the trend of confidence in decadal predictions, followed by...

Anonymous said...

From IPCC AR5 Chapter 9:


Box 9.2 | Climate Models and the Hiatus in Global Mean Surface Warming of the Past 15 Years

The observed global mean surface temperature (GMST) has shown a much smaller increasing linear trend over the past 15 years than over the past 30 to 60 years. Depending on the observational data set, the GMST trend over 1998–2012 is estimated to be around one-third to one-half of the trend over 1951–2012. For example, in HadCRUT4 the trend is 0.04ºC per decade over 1998–2012, compared to 0.11ºC per decade over 1951–2012. [Emphasis added; internal citations omitted]


So the "hiatus" was explicitly defined by the IPCC as a linear trend closer to 0 than to the trend since 1951. That simply no longer holds up even for 1998-2012.

Carrick said...

DC, I don't follow what you're trying to say here. You quoted the IPCC as saying "the GMST trend over 1998–2012 is estimated to be around one-third to one-half of the trend over 1951–2012". That's explicitly the same thing as "slowdown".

You say So the "hiatus" was explicitly defined by the IPCC as a linear trend closer to 0 than to the trend since 1951. .

Well yes. That's the condition for a slowdown in warming. But not a stoppage.

You then say [t]hat simply no longer holds up even for 1998-2012.

Of course it holds true. It explicitly is the case that the trend for 1998-2012 is less than that of 1951-2012.

So I'm not sure what you mean here. You seem to be contradicting yourself.

MikeR said...

Could others help me understand how important this result is, in context of the other sets of temperature measurements? Is it still true that (for instance) satellite measurements, land-based measurements, and ARGO measurements are all showing trends significantly less than the model ensemble?
If that is true, why would the correct response to this result be, "Great - no more 'hiatus'!", and not, "This is really confusing and interesting; why are sea surface temperatures different from the rest?"

James Annan said...

Mike,

It's not always that I agree 100% with RC, but Stefan's latest post is one such case. It is such a small change that it doesn't really affect anything at all.