Saturday, January 14, 2012

Not so fast!

It seems that reports of the death of global warming have been somewhat exaggerated. Thanks to commenter doskonaleszare on the previous post, a graph of the forthcoming updated HadCRUT4 data set has been located here, and here it is reproduced for your convenience:


Blowing up the right hand end, with a horizontal line to guide the eye:


Red is the new HadCRUT4 blue is HadCRUT3. It is quite clear that in the new data set, the 2010 value exceeds the 1998 value (as indeed does 2005), in line with the NCDC and GISS analyses. This seems to be attributed to the use of a larger set of high-latitude observations, which was long recognised to be a weakness with HadCRUT3.

42 comments:

GS said...

Good to see perfectgreybody on your blog ;) He's running on of the best scientific blogs in Poland, explaining climate science to laymen (and discussing other scientists) for three years now, here: http://doskonaleszare.blox.pl/html (BTW, "doskonaleszare" is Polish for "perfectgrey")...

He chooses to remain anonymous, but certainly is a professional, and was recently endorsed by one of the best known Polish climate scientists, prof. Krzysztof Haman, a specialist on cloud physics. Even though for most readers of his blog professor was just stating the obvious ;)

Carrick said...

Those years are the same within errors bars, aren't they (1998, 2005, 2010)?

James Annan said...

Those and many others besides, according to that graph. But the bet was never on whether there was a "significant" difference, despite Whitehouse's attempts at equivocation on that point - it was simply on whether the Hadley Centre analysis was hotter or colder.

Paul Matthews said...

That's "climate science". If the theory and data don't agree, go back and adjust the data.

James Annan said...

That's denialism - when the data are inconvenient, allege a conspiracy.

Paul S said...

Another successful prediction

Anonymous said...

Dr Annan
While I do appreciate your 'put your money where your mouth is' attitude this is a temperature series which is not entirely a random phenomenon (even given the short timescales involved in your bet). As you put it earlier, you had a >85% chance of winning the 2008 bet. So it is not just that you lost a bet the odds of winning were 17:3 in your favour. It is that, if you look at the same data in a different perspective one (say Whitehouse) is able to offer a different prediction.

In other words, you should not have lost the bet at all. You should have won by a wide margin instead of coming so close by a narrow margin and still losing the bet.

James Annan said...

It's not a matter of "margin". The question was whether there would be a higher temperature, or not. The new HadCRU analysis says there was.

Anonymous said...

You did not expect to win because of a revision of the temperature record bringing down past temperatures, did you?

:)

PLG said...

Laying the groundwork for the IPCC 5AR!

Paul S said...

Does James really win? Presumably the bet was made on the basis of HadCRUT3 results. The fact that HadCRUT4 shows 2010 as the warmest year leaves James with the moral victory but penniless, surely.

Otherwise how could you ever consider a bet closed?

Anonymous said...

Wow . . this is really important.

The Arctic is now almost as warm as it was in the 1930's & 40's.

Cool !

Paul Matthews said...

That's "climate science". When someone points out what's going on, resort to the 'd' word and the 'c' word, in the same sentence!

Electrical Engineer said...

Fred:

"The Arctic is now almost as warm as it was in the 1930's & 40's."

If you say so, oh omniscient one.

James Annan said...

nigguraths, you are right I did not expect to win in this manner. On the other hand, it has long been known that HadCRUT3 under-represents the Arctic warming, so it's hardly a surprise that (a) they should work to improve this, and (b) that this improvement increases the recent warming a little relative to earlier years.

Paul, there was no agreement to use HadCRUT3. We simply agreed to use the Hadley Centre analysis. It's not a very satisfactory situation. I did try back in 2008 to clarify the precise terms with David Whitehouse but he was unresponsive. Most established ad-hoc betting of this nature (eg the Foresight Exchange) would have a specified judgment date and (hopefully) a clear criterion but above all a named judge who has the final word. Of course, it is a rare case that is actually contentious.

Shibui said...

So, you mean Whitehouse is not going to pay up?

Carrick said...

James: it was simply on whether the Hadley Centre analysis was hotter or colder.

Then by my reading, absent a statistically significant difference, you don't know if it is hotter or colder, and the bet is nullified.

In any case, four years is a pretty short period to be placing any climate related bets (the variation over that short of a period is certainly dominated by natural variation, and you're rolling essentially unloaded dice). Even 10 years seems short to me, assuming that the NASA guys aren't right about the lack of pollution controls contributing to a long-term (circa 40 years) flattening of temperatures, as with the period 1945-1975.

James is certainly right that HadCrut3 under-represents the northern arctic, but it's not clear to me that adjustment for the sake of adjustment is good science either. (Making a "wrong" correction doesn't make your answer any more correct than making no correction. And if you're extrapolating into the unknown how do you know your correction is better than the uncorrected value.)

In any case, the effect of including that tiny additional area is really minute.

I realize this was for the sake of a bet, but only by ignoring the uncertainty in the reconstructed values is the recent differences between HadCrut3 and HadCrut4 even meaningful.

Or even HadCrut3 and GISTEMP for that matter.

Why is it that uncertainties seem to be so ubiquitously ignored in this field? Too many "citizen scientists" maybe?

James Annan said...

But the topic was never on the difference being "statistically significant" (which is not proof of anything anyway), just on whether the central estimate was higher or lower. I realise that of course this is of little scientific value, but it does appear to be of some public interest.

I don't think it is fair to say that they went looking for data in hot areas - there are additional grid cells filled in in colder areas too. It is not interpolation, but the use of more observations. However, given the obvious spatial coherence of anomalies, interpolation would be an obvious approach to take, and I'm rather surprised they haven't yet done this. They already average over 5x5 degree grid boxes, which relies on the same principle, only in a less precise and justifiable manner. There have been more accurate and sophisticated methods out there for decades. OTOH they may view bias correction and inhomogeneity as more of a priority than spatial averaging. The global mean is a bit of a byproduct that most scientists don't even use - direct comparison of the spatial pattern is more common and informative.

crandles said...

>"you are right I did not expect to win in this manner."

Does this mean you are claiming victory and refusing to pay up?


Meanwhile,
Arctic warming, increasing snow cover and widespread boreal winter cooling:

pdf

Should we expect further cooling of US & Eurasia over coming decade(s)? Is wondering if this could slightly reduce rate of global average temperate increase attaching too much importance to it? Or is there only so far the AO can be forced making future predictions about it rather speculative?

How important might unknowns like this have been in your win/loss?

Carrick said...

James: But the topic was never on the difference being "statistically significant" (which is not proof of anything anyway)
Well, agreed, it's not "proof" (otherwise there wouldn't be so many 3+ sigma results that turned out to be wrong). But "statistical significance" is our usual standard for making generalizations from data, e.g.. whether any warming or cooling is statistically significant.

I get the point of using the central value for the bet...but as you can tell I find that vexing. Mostly because it leads to an enormous amount of interest from "citizen scientists" in something that isn't at all interesting scientifically.

I don't think it is fair to say that they went looking for data in hot areas - there are additional grid cells filled in in colder areas too. It is not interpolation, but the use of more observations

That's a good way of thinking about this. Thanks.

James Annan said...

Chris, I'm not rushing to judgement. AIUI the HC hasn't actually released their analysis for 2011.

James Annan said...

Carrick, I don't believe it is in principle not interesting to scientists. There is lots of academic work on extremes and record-breaking, and by its nature this cannot generally consider significance in this way, or else there is never a new record since the current best is almost never broken by a "significant" margin.

Incidentally, it is not at all clear from the graph what the error on the difference between two years is, as many of the errors are temporally correlated.

Carrick said...

James: Incidentally, it is not at all clear from the graph what the error on the difference between two years is, as many of the errors are temporally correlated.


Well, of course that can be calculated (using reasonable assumptions about the underlying noise such as homoscedasticity), but this does expose a major weakness to using "records" (individual data points) to argue over trends (which I would argue is what eventually we really care about).

The uncertainty in the difference between the two data points in general doesn't improve by changes the separation between two points (there are exceptions, e.g., correlated noise that decorrelates as the interval between points increases).

On the other hand, if you compute the trend using OLS, for example, as the interval is increased, the rejection of short-period noise is improved. (This is the sort of curve I get using a Monte Carlo based approach... shown for comparison are results for four different methods generated by Lucia using e.g. ARIMA models of noise).

If I were to enter into what I considered a reasonable bet, it would involve (at the minimum) whether the sign of the temperature trend was positive or negative. In fact you could use this sort of model-based calculation of uncertainty to figure out your odds. If we assume 0.2°C/decade, even 20 years isn't a safe bet.

Carrick said...

Screwed up my link.

Heres' the Monte Carlo based error analysis I was referring to.

IcedVolvo said...

Gee lets think about whether we trust the Hadley CRU data; well lets see; they wont let anyone have access to the RAW data to verify the results, and I am pretty sure Jones reported that the original data was "lost". The "correction" algorithms are a state secret. The integrity of the dataset has been questioned over and over again with no detailed analysis as to what goes in and what gets left out. No surprises HADCRUT 4 supports the dogma!

When Hadley/Jones et al are completely transparent with the RAW data and algorithms then HADCRUT might be taken seriously by other than the IPCC converted.

John M said...

If I understand this correctly, despite a shortage of Artic temperature stations with at least 15 years of data between 1961-90 - that's all HadCRUT3 requires! - someone figures they can estimate what it would have have been for each month.

Mind you, all of the HadCRUT3 data is an estimate because it's not based on mean temperatures but just the mean of the mean maximum and mean minimum for the month - an averaging technique not used in any other branch of science.

Carrick said...

IcedVolvo: Gee lets think about whether we trust the Hadley CRU data; well lets see; they wont let anyone have access to the RAW data to verify the results, and I am pretty sure Jones reported that the original data was "lost

Lost data is not going to be a problem. From The Met Center:

The underlying data used to generate the new HadCRUT4 dataset will be free from any restrictions to public distribution when the dataset is released.

Maurizio Morabito said...

It's really really hard to avoid laughing when adjustments come out of thin air but anyway...let's consider HadCRUT4 less of a joke for a moment.

Since temperatures go up by including the upper Arctic, it is obvious that the rest of the world, and especially the inhabited regions, have not warmed as expected.

And a 5C increase in an area where the average is -20C is _not_ the same thing as a 5C increase where the average is +10C. The former is inconsequential, the latter a change in all seasons.

So HadCRUT4 simply confirms things are going as predicted only in faraway places where there are few measurements and nothing is really changing anyway.

Carrick said...

Maurizo: Since temperatures go up by including the upper Arctic, it is obvious that the rest of the world, and especially the inhabited regions, have not warmed as expected.

Em... there's a decent theoretical argument for why you'd expect more warming (or cooling) at the North Pole (arctic amplification). That's not to say the magnitude is fully understood, but the fact that it happens isn't unexpected.

Secondly, if you look at temperature trends, you find the dominant effect is from a) latitude and b) whether you are at a marine (or coastal) site versus and in-land site. Regarding latitudinal bias on land, what you find is below 30°N, there isn't much variation in trend.

Adding more sites at lower latitude may reduce your noise slightly, because it mashes down temperature variability associated with site selection, but it isn't going to change your trend estimate in any meaningful way. It is easy to see from these sorts of analysis, that the biggest uncertainty comes from undersampling latitudinal bands, especially those of the high arctic, so not under-sampling by adding in more stations does help improve the accuracy of the reconstruction slightly.

As I pointed out upthread, the effect of adding the additional sites on Hadcrut is actually quite minute. It may matter for bets made on central values (and parlor style arguments).

Here are the trends from the current "big four" reconstructions (°C/decade), 1980-2009 inclusive:

ECMWF 0.164
NCDC 0.167
HADcrut3gl 0.163
GISTEMP 0.165

The largest difference in trend amounts to only 0.004°C/decade or 0.04°C change in a century. This is likely inside the measurement uncertainty associated with the reconstruction algorithm. (As opposed to the larger uncertainty associated with the effect of short-period climate on your estimate of central trend, which is much larger, but doesn't affect the differences in reconstruction trends at first order.)

While it is interesting scientifically why HADcrut3 is lower than GISTEMP, I would argue there are no policy-related ramifications from the fact that they even differ.

Maurizio Morabito said...

Thank you Carrick...however my point was not about the effect of adding sites at lower latitudes, rather the effect of removing sites at higher ones...in fact if we take HadCRUT4 and remove the new Arctic and Russian sites, we get HadCRUT3 and a lower trend.

Carrick said...

Maurizo: in fact if we take HadCRUT4 and remove the new Arctic and Russian sites, we get HadCRUT3 and a lower trend.

I'd predict by maybe as much as 0.004°C/decade.

Interesting from a "seeing how many thermometers you can fit on the head of a pin" perspective, but policy wise, it's inconsequential...

which is really what I was suggesting.

It, is by the way, possible to compute the global temperature trend in a way that is robust against adding or dropping stations over time. (I'm thinking of something more than just 5°x5° gridded average.)

The trick is in deciding it is trend you are interested in measuring, and in recognizing and factoring in the latitudinal bias in temperature trend associated with arctic amplification. (Something similar to pre-whitening.)

Electrical Engineer said...

Since temperatures go up by including the upper Arctic, it is obvious that the rest of the world, and especially the inhabited regions, have not warmed as expected.

No, the rest of the world was expected to warm less rapidly than the upper Arctic.

IcedVolvo said...

Carrick: read the press release AGAIN! They are not releasing the RAW data; rather they are releasing the "massaged" data. Further no station ID or position data will be available!

Even disregarding the climate gate emails which clearly show collusion to "massage" data the lack of ability to verify results is a breach of all the basic principles of science!

So what I said stands: I DONT TRUST THE CRU!

Electrical Engineer said...

Iced Vovo:

When Hadley/Jones et al are completely transparent with the RAW data and algorithms then HADCRUT might be taken seriously by other than the IPCC converted.

Actually, HADCRUT is the dataset of choice by fake sceptics who like to say there has been no warming since 1998. They couldn't be choosing the dataset based on the result it gives, surely.

doskonaleszare said...

FIY, HadCRUT4 is PIP at JGR (acronym overload!)

http://www.agu.org/pubs/crossref/pip/2011JD017139.shtml

doskonaleszare said...

Ahh it's the land-only CRUTEM4... they haven't finished the merged product yet.

James Annan said...

Yes, that's not the Morice et al paper that we are all waiting for...HadCRUT3 is still the official record.

Hank Roberts said...

http://www.forbes.com/sites/petergleick/2012/02/05/global-warming-has-stopped-how-to-fool-people-using-cherry-picked-climate-data/

crandles said...

published in JGR?

http://www.bbc.co.uk/news/science-environment-17432194

Unknown said...

According to Hadcrut4 2010 has a temp anomaly of 0.53. So does 2005. Errors are 0.1.

Even if you retrospectively apply the climate bet to HadCrut4 (and who in their right mind retrospectively applies bets) I would have still won because no new record was achieved by 2011.

crandles said...

> "Even if you retrospectively apply"

Interesting phase. If you are suggesting that James has paid up but now want his money back twice over then you would have a strong point. If however the bet is still unsettled, then you either use the data specified or the latest data.

IMO David Whitehouse is in a much weaker position now. He should have pushed for settlement as soon as HadCRUT3 Dec 11 data was published.

Is it clear whether the wording was new annual temperature record or temperature higher than 1998 between 2008 and 2011 inclusive?

If it clearly was a new record then James is perhaps being naughty referring to it being all about since 1998. But if the wording wasn't sorted when James tried (assuming he did) then maybe DW should get what he deserves.

James Annan said...

I did actually pay up, because the HC did initially publish the HadCRUT3 analysis for 2011. Perhaps a more honourable opponent would consider the bet void, given the premise (of 1998 not having been beaten when we made it) is now clearly seen to have been invalid.

Of course, I'm not holding my breath.