The current status of the falsification of the IPCC AR4 projection of 2 C/century is: Falsified. I first discussed this falsification in IPCC Projections Overpredict Recent Warming. That discussion included some caveats. I have been addressing criticisms as they arise. Today, I am addressing a discussion by Tamino, who decided to reanalyze the existing data using an inappropriate analytical technique that is less precise than the method I used when applied to messy data.
The main result is: Tamino’s entire discussion of calculating uncertainty intervals applies to the imprecise method Tamino used to estimate the temperature trend based on available data. His discussing has nothing to do with the based method I use, which is more precise. Tamino’s method fails to falsify the IPCC results because the method is crude and approximate
It is simply a fact that, in statistical analysis, one of the reasons one can fail to falsify a claim that is actually false and which can be shown to be false given the best available data and methods, is to use unnecessarily imprecise data or use an unnecessarily imprecise analytical technique. Doing one or the other or both results in failure to falsify because one’s uncertainty intervals are too large. This is justifiable when it happens due to limitations in existing data. It is not justifiable when one stubbornly refuses to use more precise analytical techniques. Tamino may be accustomed to using OLS with pumped up error bars, but it is know to be imprecise and result in larger than necessary uncertainty intervals. Tamino may be accustomed to not averaging over data sets, but that also results in larger than necessary uncertainty intervals (though that effect is smaller).
So, to save the IPCC projections, Tamino is not only hugging the error bars, his method has really big ones.
Using the more precise Cochrane-Orcutt, data through the end of February, we still get this:
The central tendency for the short term warming rate predicted by the AR4 was 2C/century and is shown in brown. It falls well outside the uncertainty intervals calculated using Cochrane Orcutt, using temperature data averaged over five data centers. This is discussed more here. The caveats are discussed in here.
Background
Suppose we expect the global average temperature of the earth’s surface (GMST) during a particular time period can be described by a mean trend, which is a linear function of time, and some random element, due to weather, and that the measurement of this temperature also includes some random errors. That is, the monthly average temperature measurements can be described using this function
where Ti and ti are the temperature and time of the ‘ith measurement, and ei is some sort of “noise” which could be due to weather or measurement uncertainty.
- determining the magnitude of trend, m, from a series of temperature measurements,
- making sure that our estimate of the trend, m, is the best estimate possible
- knowing the uncertainty in our estimate and
- doing the best possible job testing a hypothesis about the trend magnitude of the trend. In particular, I have been examining the IPCC AR4’s hypothesis about the trend; they predict a trend of 2.0 C/century.
How might we go about doing all four? Is there more than one way? Are some ways better and worse when judged by our goals of hypothesis testing?
It turns out that, yes, there are better and worse ways to accomplish our goals.
Today, we will discuss the method described by Tamino (OLS) and explain why, when applied to the monthly temperature series Tamino and I are both analyzing, it is less precise than the method I have been using. The method I use is called Cochrane-Orcutt. (I should also note that the analysis I have done so far indicates that ARMA(2) would be even more precise, than Cochrane-Orcutt. )
The method Tamino uses might be called “OLS with plumped up error bars”.
The remaining portion of this post discusses uncertainties that result when using OLS gives more incorrect “fails to falsify” when we set the incorrect level of incorrect “falsifications” to 5%. This will involve cartoons and text. ( In this article, to avoid over cluttering the cartoons, I will discuss the IPCC’s AR4’s prediction for the central tendency, but I have discussed their stated uncertainty intervals elsewhere.)
Uncertainty Intervals and the Dartboard Analogy.
When we try to determine trends from time series, we generally can chose among a variety of methods. We can look at the data, slide a ruler through the points, say “looks good to me!” and say we’ve found a trend. We can use Ordinary Least Squares. We can use Cochrane-Orcutt, ARMA(2), or any number of methods that fall inside the class called “Generalized Least Squares”.
Now, let us think of each of the possible methods used to calculate the trend, ‘m’ from a set of data as a being a player in a game of darts. Think of getting the exacctly correct value of “m” in an experiment as hitting the bulleyes in the image to the left.
In that image, I have drawn some circles around the bullseye. For any given reasonably decent dart player, we should be able to draw a circle that encloses the region in which roughly 67% of the darts hits. We can draw a larger circle that encloses 95% of the dart hits. More skillful players will have tighter circles for their “dart hits”. In our analogy, players are “statistical methods” and more skillful methods will have tighter circles for their “uncertainty intervals” than less skillful methods.
Naturally, with regard to testing the IPCC projections for temperature, most people would think it best to use the most skillful method available given the data we have. That is, if we want to hit close to the bullseye, we’d like to pick a method that gives a tighter spread of hits.
Cochrane-Orcutt closer to bullseye than Ordinary Least Squares
Oddly enough, Tamino picked Ordinary Least Squares to estimate the magnitude of the trend. This is odd because we know that when when OLS is applied to fit a trend to the monthly data both Tamino and I are analyzing, the residuals in the fit exhibits something called “serial auto-correlation”.
Because this serial auto-correlation exists, using OLS to estimate the magnitude of the trend does not give the tightest spray pattern when estimating the trend.
How do we know this? It’s widely know, but it also happens to be mentioned on page 241 of Lee & Lund 2004 cited by Tamino. Just after equation 1.4, the text says this:
Translated from journal-ese, this means: If you use a generalized linear regression with an appropriate design matrix, your estimate of the trend will have a tighter spray pattern around the true mean value.
So, we know, based on analysis of the monthly mean GMST data that Cochrane-Orcutt gives a tighter spray pattern around the true mean than does OLS.
So, since in statistics, we don’t actually know where the “bullseye” is, it’s generally wiser to use the method that gives the tighter bullseye: that answer is more likely to be close to the mean. This means, unless we are lazy, or find doing additional computations over-taxing, we should use Cochrane-Orcutt, not OLS to estimate the mean trend.
That’s why I picked it.
What happens if we use OLS with pumped up uncertainty intervals?
In contrast, for some reason, Tamino uses OLS. In some sense, this is less time consuming than using Cochrane-Orcutt. However, after using OLS one must do some arithmetic to adjust the uncertainty intervals spit out by standard application of OLS.
Having decided to use a poor method that generates larger than necessary uncertainty intervals as his “base method”, Tamino, thens explain how to estimate the humongo-nourmous uncertainty intervals for the method he uses.
Mind you, his long discussion of how to fix up his own error bars has absolutely nothing to do with the uncertainty intervals in the method I use. Tamino needs big uncertainty intervals using the method he chose because he picked a method with gives larger uncertainty intervals that more appropriate methods.
To the left, I have placed a grey dot to illustrate a hypothetical “dart hit” (or estimate of the trend.) This would be the “best estimate of the mean” an analyst might calculate based on the data he has. I picked a spot just inside the 95% uncertainty intervals for OLS.
It turns out that if you analyze a time series in EXCEL, and use “LINEST”, it will provide an estimate of the trend “m” using OLS. It will also provide an estimate of the uncertainty. If there is no serial autocorrelation in the residuals, this estimate of the uncertainty is fine. But, if there is serial autocorrelation in the data, EXCEL will underestimate the scatter in the data.
The result is an analyst using, OLS may get an ‘best estimate’ where I show the grey dot, but he will then tell you 95% probability the mean falls inside the tiny little grey circle.
Those error bars are too small. Notice the mean doesn’t fall inside the grey circle? This means that, when OLS is used to estimate the mean trend , ‘m’, and one then uses the uncertainty intervals provided by EXCEL, the anaylist will often claim to falsify a null hypothesis when it is true. This is called “alpha” (α), or ‘false positive’ error. In the “blog-climate-war’ context, it is the chance that we would say the IPCC is wrong when they are right.
A mistake of this sort or any sort, are considered a bad thing. But the fact that OLS used inapppropriately can result in this error has nothing to do my analysis, because my base method is Cochrane-Orcutt, not OLS. This potential mistake affects Tamino’s analysis because he didn’t switch using a more appropriate method of estimating the trend.
But Tamino thinks it’s somehow ok to stick with this poor method because he thinks he can estimate correct size of the uncertainty intervals for his method, exactly. He says:
I’ll also use the exact formula for the impact of autocorrelation on the probable error in an estimated trend rate from OLS (see Lee & Lund 2004, Biometrika, 91, 240).
Tamino doesn’t state the equation number, but it appears likely Tamino used equation 3.6 on page 242;
Interestingly enough, the formula in Lee & Lund is not exact. The approximation they describe was suggested in an unpublished NCAR technical report by Nychka et al. According to L&L, the form of the equation is justified as asymptotically correct: asymptotically correct means the equation is approximate, but approaches the correct value as the number of samples “n” goes to infinity. L&L also tell us the 0.68 n1/2 terms are justified based on simulations: it’s not exact. This means equation is 3.6 is useful for redusing α errors when applying OLS to estimate trends the specifics cases that Nychka et al. considered. The method may, or may not be result in correct uncertainty intervals in slightly different circumstances.
So, I wanted to do a quick and dirty computation, and didn’t care about false negative (β error) and only cared about false positives (α error), and generally didn’t care about precision, I might use the method Tamino is using. This happens in industry, and even science, from time to time.
Nevertheless, the truth is: there is no “exact” method to fix up the uncertainty intervals for OLS. Even if there were, when precision is required, you would not wish to use it when analyzing data similar to the monthly temperature data both Tamino and I are using. Using the currently available data, Cochrane-Orcutt gives more reliable answer tha nOLS because a) it accounts for the autocorrelation and gets a better estimate of the trend in the first place and b) if there is no AR(2) noise, you don’t need to fix up the uncertainty intervals further. (Note: For our case C-O is also an approximation, becaue there is AR(2) noise. This means we need to use ARMA; I plan to do so. However, I haven’t learned it yet! I mentioned this issue of needing fancier analytical methods in the caveats of a previous post.)
How might ‘estimates’ of truth using two methods compare?
To the left, I have illustrated how the estimates using the two trends might compare, using the dart board analogy. In both cases, I have located a “bullseye”, which is the analogy for the location of the “true trend” and surrounded this with the uncertainty intervals for the trends we might estimate using either OLS or CO.
I placed a hypothetical green data point for the C-O result: that’s the method I use. Notice I put this on a boundary near the 95% confidence for C-O. Then I placed a hypothetical data point for the OLS result. I also placed this near the 95% confidence interval for OLS. Then I drew uncertainty intervals for each case. Notice, that in this hypothetical, the “true” mean ends up inside the 95% confidence intervals for the experiment. That’s pretty much what we expect to happen 19 times out of twenty when we use 95% confidence intervals. (That’s what 95% confidence intervals mean.)
What if we test the IPCC estimates?
Now, let’s consider what could happen during a falsification test. To do a falsification, we examine someone’s hypothesis of where the “true mean” lies. Since it’s a hypothesis, we don’t expect it to be precisely right, but we hope it lies near the true mean. For the purpose of this discussion, I’ve shown that as yellow dot. I’ve placed that dot in a hypothetical location that is consistent with the results obtained using my method (Cochrane-Orcutt) and the results Tamino gets with his cruder “OLS with pumped up uncertainty intervals”.
Where is the yellow dot?
As indicated by my previous analyses, It is outside the range of possible results that are consistent with the data. To indicate this, I placed the yellow dot outside green circle described using Cochrane-Orcutt. This means that according to Cochrane-Orcutt, the more precise method of the two, the IPCC AR4 projections are inconsistent with the data, to a confidence of 95%. We reject the IPCC AR4 projections: they are falsified. This is true no matter what we find with OLS because OLS is a cruder method. So, Cochrane-Orcutt trumps OLS even if we estimate the uncertainty intervals for OLS correctly!
What do we find using the cruder method? To illustrate this, I also placed the yellow dot inside the blue outline for the uncertainty intervals calculated using OLS where Tamino claims they go. (I’m not entirely sure he is correct, as the equation in Lee &Lund is approximate. So, Tamino’s uncertainty intervals may be too large; they may be too small. Who knows?)
So, what does Tamino’s failure to falsify man? It means that because OLS become imprecise when there is serial autocorrelation, the “shots” go wild. Because we know the shots go wild, we slap large uncertainty intervals around the large shots. So, we have now done a crude hypothesis test that fails to falsify claims that are wrong more often than necessary based on the available data.
It’s the statistical equivalent of insisting a right handed dart player shoot darts with his left hand. He gets bad results, not because he’s a bad dart player, but you forced him to use a bad technique.
Summary of Current Results: IPCC Falsified.
If we examine the results obtained using Cochrane-Orcutt, we find the IPCC projection of 2C/century is rejected; when measurement from 5 agencies are pooled the margin is substantial. This isn’t a nitpick where 2.0000 C/century is falsified but 1.99999 C/century is not: trends above 1.1 C/century are falsified. Moreover, the 2.0 C/century is falsified with data from 4 out of 5 of the respected agencies, using the data on their site on March 22, 2008. (The monthly data are updated monthly, though the web pages don’t always indicate the date of the most recent update.)
This is a summary of my results, for a a previous post. I have removed the OLS estimates, which I had previously included for comparison purposes. Cochrane-Orcutt is the more appropraite method: When I learn how to do ARMA, I will include that.
| Best Fit Trend <m> with 95% confidence uncertainty. Method: Cochrane-Orcutt. | Reject 2.0 C/century to confidence of 95%? (α=5%) | |||
| Method | C/century | |||
| Average all, then fit trend. | -1.1 ±2.2 | IPCC Projection Rejected | ||
| Fit trend to each, then average. | -0.9 ± 1.6 | See note. | ||
| Individual Instruments | ||||
| GISS | -0.4± 2.2 | IPCC Projection Rejected | ||
| HadCrut | -1.6 ± 1.8 | IPCC Projection Rejected | ||
| NOOA | -0.3 ± 1.7 | IPCC Projection Rejected | ||
| RSS | -1.4 ± 2.1 | IPCC Projection Rejected | ||
| UHA | -0.8 ± 2.9 | Fail to reject | ||
| Note: 1 ‘Method 3’, that is taking the average of the 5 individual trends results in ‘reject/reject’ for the IPCC 2C/century trend. However, as I noted, that is meaningless, as the uncertainty intervals only include the variation due to measurement uncertainty and fail to properly include weather. | ||||
Caveats
Once again: Could the falsification of the IPCC AR4 projections be overturned? Sure.
As I have often said, things that happen 5% of the time, happen. It could turn out this weather is an outlier, the climate will warm as the IPCC AR4 projected, and those who not only believe in AGW, but who are rooting for the infallibility of IPCC estimates will have their faith rewarded.
In fact, due to the low power of these test, it is likely that even if the IPCC AR4 is too high we will see the falsification reverse at least once before turning around and confirming again. That’s actually the way these falsifications generally work when data trickle in.
More importantly, it may be that if someone applies a more sophisticated method, these results would be over turned right now, today. It seems to me, and my readers agree, that one should apply ARMA(2). I do not yet know how to do ARMA(2), have not applied it to the data. I haven’t included the effect of uncertainty in the autocorrelation coefficients– either has Tamino. (However, I did a little exploratory statistics early on, trying higher and lower values, and generally, this method is robust with regard to falsifying.) I am planning to increase the number of statistical methods in my toolbox, and will be reporting on this.
But currently: Cochrane-Orcutt shows falsification. Tamino cannot be overturned this falsification by stubbornly clinging to a cruder, more imprecise method.
Anyone can fail to falsify by using unnecessarily imprecise methods. Lots of imprecise methods exist: if an analyst picks a sufficiently imprecise methods, he can pump up the uncertainty intervals to include every conceivable weather outcome that could occur, not only on earth, but Venus and Mars.
But using unnecessarily imprecise methods not the a good way to test hypothesis. That’s the way to trick yourself and use lots of math to formalize your confirmation bias!
lucÃa:
sorry, but in the paragraph: “Today, we will discuss the method described by Tamino (OLS)…)after “even more precise” tere is a “1”… have you forgot something?? i know is a very long post…
lamento ser “el maestro ciruela”!!
great, great blog!!!! and thank you in advance
Thanks Jorge–
I put the information in the footnote in the text. I wanted to capture the fact that one could, hypothetically, get a different outcome on the hypothesis test using a better more precise method. Possibly one that accounts for AR(2) noise properly. But one can’t do it by using the cruder more approximate method Tamino uses. He gets a different outcome by picking a less precise method that is subject to more β error.
Also… one of these days, I’m going to have to remember my Spanish. Mom insists I just pretend I don’t speak it, but I did have to look that up! 😉
lucÃa:
in my country, uruguay, “maestro ciruela” is someone who makes pedantics remarks, about a very complex, very well done and intelligent post and blog…
thank you!!
Cochrane-Orcutt estimation is a fine method, which gives good error estimates, IF AND ONLY IF the random part of the data is an AR(1) process. That’s the assumption behind the C-O method. This does NOT include all autoregressive processes.
If the noise is AR(1), then you can’t do much better than C-O. The error range estimated by OLS with AR(1)-compensation will be comparable (which it is), and the range estimated by generalized least squares (GLS) will be even tighter. But if the noise is NOT AR(1), then the error range estimated by C-O will be JUST PLAIN WRONG.
And indeed the noise is not AR(1). This was pointed out by a reader commenting on my blog, before I even noted it myself.
So the real reason my error range is so much larger than yours, is that yours is just plain wrong. The noise is *not* AR(1).
Tamino:
You are selecting a method with lots of uncertainty when you know that is unnecessary. If you want to show these ranges are wrong, do the AR(2) to deal with the relatively small remaining serial autocorrelation. Everyone reading my blog knows I always said I haven’t done that, and the results could change when I do. But you resorting to such a sloppy method, with unnecessarily large uncertainty intervals is silly.
Oh–
For those wondering, after adding the data for February, and throwing in NOAA data, the serials autocorrelation for the OLS was 0.59. (These can only fall between -1 and 1. So, ρ= 0.6 is a large residual that cannot be neglected. That’s why I deal with the AR(1) noise using CO.
After for performing the CO fit, I examined the serial correlation of the residuals of the CO transformed fit ρCO =-0.077; this is rather small. As Tamino notes: these must be small for CO uncertainty intervals to be correct. (A number of papers on errors with CO suggest that a rule of thumb to limit it’s use to cases where |ρCO| < 0.2, but the correct approach is to check whether the remaining correlation is statistically significant. In any case while ρCO is small, it isn’t zero. Moreover, ρCO was larger in magnitude before the February data came in and before I added the NOOA data. So, I will be learning to deal with the AR(2) noise going forward in time.
Why do these discussions with Tamino always degenerate to ‘my statistical technique is better than yours’?
Frankly, the choice of 95% as a falsification measure is arbitrary and fiddling with the statistical technique does not change the fact that the actual weather has diverged significantly from the IPCC projections. We also know that warmers would be screeching about how the end of the world was coming if the ‘weather noise’ was in the opposition direction (Rahmstorf said as much when they concluded that the actual temps were above the TAR projections).
More importantly, there are other pieces of data that support the Lucia’s statistical analysis (no warming in the oceans according to the argo floats, evidence of strong negative cloud feedback from the aqua satellite and a strong cooling trend that corresponds to a supposedly irrelevant solar cycle).
I think we need to wait and see. The next 5-7 years will tell us a lot.
I think you may be even more shocked when the you see this
http://discover.itsc.uah.edu/amsutemps/
go to channel 05 4.4 km and 06 UAH 7.5km data for March 1 to 31, 2008 . This is were it is supposed to be “warming”. Poor ol Tamino will have to start praying that the dam temps start going back up LOL
@Raven–
On Rahmstorf:
Oddly enough, Rahmstorf’s analysis requires a statistical method, but it’s glossed over. One must decide where the “true” temperature for 1990 was. To rebaseline, all measured temperature anomalies are adjusted to make that “true” temperature zero.
If we don’t account for serial autocorrelation in data, and use the simple 11 year average temperature centered on 1990, the 95% confidence intervals are about 0.07 C. Rebaselining up or down ±0.07C makes a huge difference in whether the subsequent temperatures look high or low– depending on what you select.
If we used “Tamino” error bars, the 95% confidence intervals would be at least twice as large as I said before. (I’m not going to calcualte the correlation for a comment– but say at least 0.13C, but possibly 0.15C.)
The Rahmstorf paper doesn’t even discuss the uncertainty in the shift. Their method does not make it magically go away.
On the “my statistics are better than yours”:
All statistical methods involve making assumptions. All. Tamino is choosing to use a method that results in a) the largest scatter possible and b) the largest error bars he can justify– and which is clearly wrong in at least one important limit.
For example: If we examine the formula he calls “exact”, it’s wrong in the limit that ρ=0. That is, when OLS actually works, the supposedly “exact” formula replaces “n”, the number of samples, with neff=n(1- 0.68 n-1/2)/(1+ 0.68 n-1/2). (That’s equation 3.6 with ρ=0)
Using this would dramatically widen the uncertainty intervals on OLS.
But OLS with ρ=0, the real value of “n” is the exactly correct number to use. For samples the size I am using, this would widen the error bars by 10%. That doesn’t seem like much, but when we are dealing with hypothesis tests, very small values start to make the difference between 95%, 99%, 99.5% confidence etc.
But what’s even odder about Tamino’s use of that formula is this:
Tamino complains my error bars only apply exactly when the noise is AR(1), exactly. Yes. I admit it, and have done so in the text of my blog post. There is no “discovery” by his readers– I said it. But all statistical treatments involve assumptions.
But what of Tamino’s formula and justifications? The entire justification for the form of equation 3.6 he uses is that as n->infinitiy, it takes the OLS error bars, and multiplies them with a factor that results in the error bars for a process that is AR(1). This is discussed in the text between equation 3.6 and 3.7.
So, based on the text of the paper he cites, the basic formulation for his method of inflating the error bars for OLS is based on the exact same assumption I make. It assumes the ‘noise’ is AR(1). The difference between his method and mine is that while “accepts” this assumptoin as getting correct error bars for him, he then does his analysis to find the trends and neglects the AR(1) noise entirely when getting the trends!
If statisticians really believed the formula in Lee and Lund was exact for all cases, everyone would use all the time. No one does. Why not? It’s an approximation.
Lee and Lund are discussing what you can do to fix up your error bars if you want to use the crudest method possible. It’s ok if you don’t care about making β error, and only worry about α error.
(Those in datamining often don’t care about β error. The philosophy is: if an effect isn’t strong enough to pick up, it’s probably not strong enough to invest time for further investigation. So, they only worry about α error, and set it to a low value like 0.1%)
Lucia,
Tammy has a guest on RC about AR(1).
Referenced by St.Mac here.
http://www.climateaudit.org/?p=2086
Of note, the IPCC might appear to disagree with Tammy. I say that with reservation. It deserves some inquiry
here is the Link to Tammys RC guest post.
http://www.realclimate.org/index.php/archives/2007/09/climate-insensitivity/
Steve,
Oddly enough, I too commented on Tamino’s “proof” at RC. In my case, I was examining Schwartz model. But, the fact it Tamino made a hash of that. How? He mistook Schwarz’s claim that the global means surface temperature (a climate variable) can be modeled using the approximate method discussed by, ahem, the IPCC for a claim the temperature measurements will be modeled that process.
This may seem like an odd statement, as the measurements of temperature are intended to measure temperature. The difficulty is there is always some instrument uncertainty. In the case of measuring global mean surface temperature, all measuring agencies (GISS, Hadley, NOAA, etc.) admit there is uncertainty for a variety of reasons.
It turns out that if one assumes there is some unknown amount of noise in the data, one gets results presicely like the one Schwartz got.
The only quibble I have with Schwartz is he did a “visual” on the data, and plotted it in a way that makes it difficult to pick out the time constant, as you have to try to pick an assymptotic value.
My post is here:
http://rankexploits.com/musings/2007/time-constant-for-climate-greater-than-schwartz-suggests/
And the post explains how if climate were an AR(1) process, and forcing looks like white noise, and we plotted the correlation in residuals as a function of lag on a log-log graph we get a straight line with a non-zero intercept. The non-zero intercept is due to the measurement noise.
This is how the graph looks if we plot that.
The scatter about the straight line would be expected due to the statistical uncertainty our measurements of the autocorrelation. (It doesn’t converge quickly.)
Oh, and guess what? When I estimate the measurement uncertainty from the intercept, it’s in the same ballpark the measurement groups claim for their measurement uncertianty. Go figure?
So basically: Tamino’s “analysis” assumes there is no measurement uncertainty. Mine assumes there is an unknown amount, and provides an estimate based on measurements. It returns the amount we’d expect based on how much Hadley, NOOA, GISS’s own estimates of the noise.
So, Tamino may think he proved something about climate. I think he proved something else.
mismatched subscript tags in comment 1468?
erik– I don’t see mismatched tags. But I know there were some when I first entered. Maybe you saw a cached version. If you still see them, suggest where you think it starts– maybe I’m blind.
It’s not a big deal. It’s just hard for me to read the small font.
It’s in comment 1468, in the last paragraph. “In any case while Ï^CO is small, it isn’t.” Subscript starts for the CO, but never ends in all subsequent comments.
I see in under linux:firefox and under windowsXP:explorer and windowsXP:firefox.
Although, like I said. It’s not a big deal. I just keep coming back to check comments because I find your work the last few days so interesting.
-erik
Well put Lucia.
I recall the first time I did analysis for a co worker working on their PHD,
Funded by the company. I had just come up with a nice approach that reduced CI dramatically.
It was really simple and clear.
They frowned. They liked the old method. Moshpit: ” but the old method has HUGE CIs? ” I walked through
the math with them. And then they suggested Non parametric methods. I could not figure out
why they wanted less precison in the tests. WTF?
Then I got the picture. “not rejected, meant TRUE” to them.
steve moscher:
It is not uncommon for researchers to prefer to fail to falsify their own hypothesis. People always recognize this tendency in others but often fail to notice it in themselves. This behavior isn’t peculiar to climate science: it’s everywhere!
In science, over the long haul, this all works out because people who didn’t create a hypothesis inevitably test it, and third parties inevitably notice when a hypothesis didn’t hold up. (Sometimes this can take a long time though.)
What’s puzzling me is this:
“Everyone” supposedly agrees that magnitude of warming that will occur is one of the key uncertainties. It’s also clear the IPCC doesn’t have a 100% successful track-record of precicting the magnitude of the warming. For example, the AR4 said the IPCC FAR projected 3 C/century, and it’s quite obvious that didn’t occur. It’s not even controversial to say that prediction was falsified: later panels recognized that was well off, and revised their predictions.
Also, some are willing to believe the IPCC TAR was off, but on the low side based on an analysis that was nothing more than “slide and eyeball the data”.
Both the IPCC third report ( TAR) and the IPCC second report (SAR) were for lower warming than the fourth report (AR4) predictions.
So, why is the idea the new, higher, IPCC predictions might be inconsistent with the data and on the high side at all controversial? This doesn’t overturn the basic theory of AGW. Heck, it agrees with the general observation that it’s difficult to predict the magnitude of warming using current models!
Why is it so controversial?
Its a social phenomenon. Both sides, in many cases because the individuals doing the debating are not capable of grasping the scientific issues in any specificity, have resorted to attaching their views on AGW to political and social attitudes to which they are irrelevant. It is much easier to talk about denialists, or green communists, than it is to address the detailed issue of whether MBH98 was right or wrong. Or what exactly the recent ocean probe data shows. Or whether feedbacks are on balance positive or negative, why, and what the evidence is, and whether Spencer is right about sampling intervals. Or whether Schwartz is right. Its much easier to attack Spencer for his religious beliefs than to address the sampling interval issue.
The Gore movie made this a lot worse, because it used all the techniques of propaganda, and so could only infuriate skeptics and inflame believers. The Swindle move did not help. Oreskes has not helped at all. What happens then is that a party line emerges. This defines orthodoxy and moral righteousness. We then discover the ‘doctrine of insufficient praise’. It becomes threatening to accept only part of the platform, or to accept it less than enthusiastically. Over on Tamino one finds people defending MBH to the bitter end, and simultaneously proclaiming its irrelevance to AGW. The rhetoric of ‘denialism’ is part of this, the claim that the ‘science is settled’ – without specifying exactly what propositions are settled, the tone of many of the editorial contributions on Real Climate, with their mixture of contempt for dissenters and condescension. The apparent desire of many of the lay contributors to blog comments to identify righteousness, the Democratic Party in the US, care for the environment, and belief in everything any of the movement fathers have ever written, is the result. We see the same thing on the right, whose rhetoric rather than being authoritarian left, is more of the McCarthy sort, but equally unattractive.
You are falling foul of a hardening of attitudes and a drawing of the lines around the party line. You are doubting what has become, in the last few years, part of Scripture. The Internet accelerates the process. The coalescence of doctrine into dogma in the early Church took centuries, and now it takes less than a decade. You’ve only attracted rage from the AGW wing so far, but don’t worry, the skeptics will turn on you sooner or later as long as you think you can pick and choose on doctrine.
A man is stopped in an alley in Northern Ireland during the troubles. The group asks him menacingly if he is Catholic or Protestant. He replies he is Jewish. “Aye, but are you a Protestant Jew or a Catholic Jew?”
The result is that lay people start forming their views on AGW from the people and parties whose beliefs they do understand. They will then refuse to adopt AGW, because they do not trust the Democrats and Environmentalists. Or, they will start to think that all skeptics are in the pay of big business and Republicans. They will then flail around trying to make up their own crackpot theories and do terrible analysis. This in turn will accentuate the bitterness of their opponents, and the hostility and ridicule directed at them will convince them they are being harrassed by enemies….
What should you do? Focus on the science. Do not be led into the hysterical blather about evolution, creation, tobacco, big oil…all the rest of the crazed nonsense of the blogosphere. Maintain an even tone, though it will not be reciprocated. In fact, pretty much carry on what you’re doing. You are making a real contribution to understanding. In the end, we will get to the bottom of this, and whether right or wrong, AGW the hypothesis will have contributed greatly to the development of our understanding of climate. Unfortunately on the way it may also contribute to our understanding of some of the darker aspects of human nature. But that’s life!
Lucia,
Yes none of this stuff helps. The rhetorical devices you see employed today are the devices that
divide lines and create tribes. That’s why no one, except lukewarmers, can admit an error.
I won’t pick out examples on either side, but the reluctance to admit a simple error, and the penalty
you pay when you do is rather stunning.
Now, oddly, I’ve found, lukewarmers can admit mistakes. why is that.?
Because we have large stones.
TCO nailed that one.