Boris Asks About Gavin’s Spinning.
In comments, Boris asked for clarification of yesterday’s Figure 4. I’ll provide that today. To support the discussion, I have extended time period to show the level of agreement both before and after 1999. (Yesterday’s graphs imitated Gavin’s choice of illustrating the comparison only after 1999.)
Here is a graph showing the IPCC model runs compared to the observations since 1980. All anomalies are zeroed relative to the IPCC AR4 reference period from Jan 1980-Dec 1999:
To create the traces I
- Downloaded monthly averages of GMST for individual realizations. (For models, a realizations corresponds to an individual run.)
- For model data, I computed the average temperature for January by computing the average over the 20 Januaries from 1980 to 1999. I subtracted this from all January average temperatures. I repeated for Feb, March etc. This turned the data into anomalies and with both monthly and annual averages matching zero for the period from Jan 1980-Dec 1999. (This is the IPCC AR4 reference period.)
- The GISS and HadCrut observations are already in the form of anomalies. But, they don’t average to zero during the AR4 baseline period. So, to set to the corresponding baseline, all 240 months from Jan 1980-Dec 1999 and subtracted that.
I plotted those data.
As you can see:
- From 1980-2000, the observations oscillate above and below the model average. Superficially the trends appear to match. (In fact, the mostly do match.)
- After 2000, the observations flatten, while the average of projections rises. The observations currently fall below 1 sd of the average over rund. (Bear in mind two things:1) The AR4 uncertainty intervals corresponded to ±1 sd, but the wording says of model. It’s not clear to me whether they mean the average over all runs or the average over the individual models. These are different because some models have 1 run, some have 7. 2) I have used the data available at The Climate Explorer. A few runs were missing; there were also a few “extras”!)
- Generally, speaking, someone looking at this graph would say the models are now lie above the observations. This occurs despite the fact that the everything was baselined to enforce agreement on average during 1980-2000.
How does the data look with Gavin’s choice
Gavin chose to zero the anomalies such that every model run and the observations all show a zero anomaly in 1990. He didn’t provide details, but basedon his graph, I assume he used some reasonable period to define anomalies, then found the 1990 temperature for each realizations and subtracted that from each realization.
This is kinda-sorta his graph, but I use 12 month lagging averages, and I also show time back to 1980:
As you can see, Gavin’s subjective decision to zero each individual realization to zero forces everything through zero in 1999. It also shifts the ‘observed’ temperature up relative to the models.
Note that if someone only shows the data after 1999, as Gavin elected to do, the observations would seem to oscillate around models. This is because the GISS and HadCrut observations were magically “heated up” by rebaselining using a single year when the earths temperature fell below the model average.
Note that if we reveal the data prior to 1999, we realize the apparent good agreement after 1999 occurs at the expense of having poor agreement before 1999.
Of course, as Gavin pointed out, the choice of start year is somewhat subjective. Also, the decision to show or not show data prior to 1999 is also subjective. Still, it is rather well accepted that some subjective choices are fairer and more evenhanded than others.
For example, as I pointed out yesterday, we could could achieve the opposite of the “Gavin the warmer effect” by selecting 1998 to zero data. There is no criterion by which one can claim that choice is worse than 1999, and one might expect the more tendentious of climate bloggers to suggest it. Yet, wo far, not even the frostiest of stone cold cooler has resorted to this sort of sordid cherry-picked re-baselining to support their case.
So, just as I would criticize the choice of 1998, I would suggest that selecting 1999 to zero the data, crowing about good agreement after 1999 and claiming one gets “similar” graphs with other choices of zero years is…. well… cough, cough….
Zeroing the IPPC AR4 models using the 20 year reference period specified in the IPCC AR4 may be subjective relative to some other possible choices. But it is a far better and…. dare I say it… more “robust” choice than picking the single year that creates a graph Gavin would prefer the world to see.
Written by lucia.Comments Closed: If you would like them re-opened, Contact Lucia
Comments
Boris (Comment#7321) December 18th, 2008 at 11:41 am
Okay. Thanks for explaining.
jeff Id (Comment#7335) December 18th, 2008 at 6:09 pm
I actually can’t believe he tried this. It’s nothing but propaganda. Wow.
paminator (Comment#7336) December 18th, 2008 at 7:51 pm
I seem to remember that Hansen’s forecast from 1988 does much worse when compared with measured data since 1988. In that case, there is no opportunity to fudge the model results. These current model results have been continuously tweaked using historical data, and really are not much more than expensive curve-fits.
Marcus (Comment#7338) December 18th, 2008 at 8:36 pm
I do think showing a full 28 year timeline is better than a 9 year chunk. I think it fairly likely that Gavin would agree that the longer timeline was an improvement (and possibly also that zeroing to a longer term baseline is better than zeroing to a single year).
However, I would also point out: 1998 exceeded 1 std. dev warmer than the model mean. So having 2008 exceed 1 std. dev below the model mean is not proof that the models are not getting the long term climate trends right.
One could argue that the 1998 deviation from mean was due to a super-el Nino year, and what is the equivalent explanation for the 2008 deviation from mean? Well, we did have a decent La Nina cooling the first half of the year. We’ve also had that absence of sunspots that has led much of the skeptic blogosphere to predict dangers of a new ice age. The next time we have sunspots and an el Nino together, I would expect that we’ll go above the model mean again.
I would also note that the last couple months, absent any el Nino or la Nina but still including a sunspot effect, lie fairly near the model mean again. Not that two months is a good judge of anything, but… *shrug*
A proper Bayesian analysis would, of course, lead to a lower prediction of climate sensitivity after every cold year (and a warmer prediction after every warm year) if the only data you had was global mean temperature from GISS and HadCrut and the models. Of course, we have more data than that, such as information about La Ninas, sunspots, stratospheric temperature change, sea level changes, Greenland mass balance, ocean heat content, etc. etc.
In any case, the next two to three years will provide more data. If (absent major volcanoes or 3 consecutive La Ninas) temperatures remain around the 2008 average… well, that would certainly lead me to think that future warming was likely overpredicted. But in my opinion, 2008 by itself is well within expectations given the circumstances. For comparison, the Arctic ice retreat in 2007 was more outside the bounds of expectations, but I was willing to ascribe it to “a temporary change in winds” and thought it was fairly likely that summer peak ice loss would return to the longer term trend (which was still negative). However, Arctic 2008 ice did not return all the way… again, if in the next couple years Arctic ice still hasn’t returned to the trend, I’ll have to assume that Arctic ice models _underpredicted_ ice loss.
(and interesting exercise might be to look at all deviations from the model mean and try and come up with explanations: is there a sunspot cycle in the deviations – though I presume some of the models include a guesstimate to the cycle, so I would presume this signal would be small? Is/will there be a PDO-shift signal in the deviations?)
lucia (Comment#7343) December 19th, 2008 at 6:25 am
Marcus-
Correct. This visual method has very little statistical power. You need to use more powerful statistical tools (like t-tests) to detect differences (or show they don’t exist.)
But it’s stil important to show graphs that don’t skew the results in favor of a claim you are promulgating.The graphs with more data and/or the graphs with appropriately chosen baselines are better choices.
I’m pretty sure they are still low. Early this year, monthly means were out 2 standard deviations. But we expect that to happen 5% of the time. So, one of two pop-outs like that also doesn’t prove anything. That’s why trend analysis, which uses multiple samples, and has many degrees of freedom is better than just looking at end points.
Yes. The next two years will give us more data.
MartinGAtkins (Comment#7347) December 19th, 2008 at 8:53 am
Lucia
IMHO You should put a place holder on this post. It shows nicely how to dig through the data and the methodology used to approach it.
Nice one.
The Big Lie Of Anthropogenic Global Warming (And Why It Is So Worrying…) « The Unbearable Nakedness of CLIMATE CHANGE (Pingback#8252) January 7th, 2009 at 6:25 pm
[...] The fabled IPCC reports thread tentatively on the matter of present evidence of global warming, with the AR4 dedicating to it just a single chapter, mostly focused on listing changes that are “compatible” with global warming. The temperature readings are still in ranges that can be easily reverted by relatively modest volcanic eruptions, and everybody admits that even decadal trends do depend on what reference values are used. [...]