Above: map of mean temperature and departure by state for February 1936 in the USA, a 5 sigma event. Source: NCDC’s map generator at http://www.ncdc.noaa.gov/oa/climate/research/cag3/cag3.html
Steve Mosher writes in to tell me that he’s discovered an odd and interesting discrepancy in CRU’s global land temperature series. It seems that they are tossing out valid data that is 5 sigma (5σ) or greater. In this case, an anomalously cold February 1936 in the USA. As a result, CRU data was much warmer than his analysis was, almost 2C. This month being an extreme event is backed up by historical accounts and US surface data. Wikipedia says about it:
The 1936 North American cold wave ranks among the most intense cold waves of the 1930s. The states of the Midwest United States were hit the hardest. February 1936 was one of the coldest months recorded in the Midwest. The states of North Dakota, South Dakota, and Minnesota saw the their coldest month on record. What was so significant about this cold wave was that the 1930s had some of the mildest winters in the US history. 1936 was also one of the coldest years in the 1930s. And the winter was followed one of the warmest summers on record which brought on the 1936 North American heat wave.
This finding of tossing out 5 sigma data is all part of an independent global temperature program he’s designed called “MOSHTEMP” which you can read about here. He’s also found that it appears to be seasonal. The difference between CRU and Moshtemp is a seasonal matter. When they toss 5 sigma events it appears that the tossing happens November through February.
His summary and graphs follow: Steve Mosher writes:
A short update. I’m in the process of integration the Land Analysis and the SST analysis into one application. The principle task in front of me is integrating some new capability in the ‘raster’ package. As that effort proceeds I continue to check against prior work and against the accepted ‘standards’. So, I reran the Land analysis and benchmarked against CRU. Using the same database, the same anomaly period, and the same CAM criteria. That produced the following:
My approach shows a lot more noise. Something not seen in the SST analysis which matched nicely. Wondering if CRU had done anything else I reread the paper.
” Each grid-box value is the mean of all available station anomaly values, except that station outliers in excess of five standard deviations are omitted.”
I don’t do that! Curious, I looked at the monthly data:
The month where CRU and I differ THE MOST is Feb, 1936.
Let’s look at the whole year of 1936.
First CRU:
had1936
[1] -0.708 -0.303 -0.330 -0.168 -0.082 0.292 0.068 -0.095 0.009 0.032 0.128 -0.296
> anom1936
[1] “-0.328″ “-2.575″ “0.136″ ”-0.55″ ”0.612″ ”0.306″ ”1.088″ ”0.74″ “0.291″ ”-0.252″ “0.091″ ”0.667″
So Feb 1936 sticks out as a big issue.
Turning to the anomaly data for 1936, here is what we see in UNWEIGHTED Anomalies for the entire year:
summary(lg)
Min. 1st Qu. Median Mean 3rd Qu. Max. NA’s
-21.04000 -1.04100 0.22900 0.07023 1.57200 13.75000 31386.00000
The issue when you look at the detailed data is for example some record cold in the US. 5 sigma type weather.
Looking through the data you will find that in the US you have Feb anomalies beyond the 5 sigma mark with some regularity. And if you check Google, of course it was a bitter winter. Just an example below. Much more digging is required here and other places where the method of tossing out 5 sigma events appears to cause differences(in apparently both directions). So, no conclusions yet, just a curious place to look. More later as time permits. If you’re interested double check these results.
had1936
[1] -0.708 -0.303 -0.330 -0.168 -0.082 0.292 0.068 -0.095 0.009 0.032 0.128 -0.296
> anom1936
[1] “-0.328″ “-2.575″ “0.136″ ”-0.55″ ”0.612″ ”0.306″ ”1.088″ ”0.74″ “0.291″ ”-0.252″ “0.091″ ”0.667″
had1936[1] -0.708 -0.303 -0.330 -0.168 -0.082 0.292 0.068 -0.095 0.009 0.032 0.128 -0.296> anom1936[1] “-0.328″ “-2.575″ “0.136″ ”-0.55″ ”0.612″ ”0.306″ ”1.088″ ”0.74″ “0.291″ ”-0.252″ “0.091″ ”0.667″
Previous post on the issue:
CRU, it appears, trims out station data when it lies outside 5 sigma. Well, for certain years where there was actually record cold weather that leads to discrepancies between CRU and me. probably happens in warm years as well. Overall this trimming of data amounts to around .1C. ( mean of all differences)
Below, see what 1936 looked like. Average for every month, max anomaly, min anomaly, and 95% CI (orange) And note these are actual anomalies from 1961-90 baseline. So that’s a -21C departure from the average. With a standard deviation around 2.5 that means CRU is trimming departures greater than 13C or so. A simple look at the data showed bitterly cold weather in the US. Weather that gets snipped by a 5 sigma trim.
And more interesting facts: If one throws out data because of outlier status one can expect outliers to be uniformly distributed over the months. In other words bad data has no season. So, I sorted the ‘error’ between CRU and Moshtemp. Where do we differ. Uniformly over the months? Or, does the dropping of 5sigma events happen in certain seasons? First lets look at when CRU is warmer than Moshtemp. I take the top 100 months in terms of positive error. Months here are expressed as fractions 0= jan
Next, we take the top 100 months in terms of negative error. Is that uniformly distributed?
If this data holds up upon further examination it would appear that CRU processing has a seasonal bias, really cold winters and really warm winters ( 5 sigma events) get tossed. Hmm.
The “delta” between Moshtemp and CRU varies with the season. The worst months on average are Dec/Jan. The standard deviation for the winter month delta is twice that of other months. Again, if these 5 sigma events were just bad data we would not expect this. Over all Moshtemp is warmer that CRU, but when we look at TRENDS it matters where these events happen
There’s a very good reason to toss 5-sigma data.
The result is data that ‘looks’ less noisy, yielding graphs which are more convincing in appearance.
You have got to be kidding me. Outliers are indicators of either some important behavior or bad data. Either way you have to investigate. Ad hoc application of a statistical cut-off is not good data analysis policy. You are just an apologist.
Wonder of NE Oregon is one of those areas where the extremes are dropped from the record. Meacham is notorious for extremes, and August saw two inches of snow fall on summer tourists up the top of the tram at Wallowa Lake. That is nearly two months ahead of the official start of Autumn! Let alone Winter!
Geoff Sharp says:
The 1930′s in the USA has the lowest lows and the highest highs. I think there is good reason for this, in particular 1936. The PDO was about to flip into negative, but more important is perhaps the solar position. SC16 which peaked around 1928 was weak with a today count of around 75 SSN. Take off the Waldmeier/Wolfer inflation factor and this cycle would be close to a Dalton Minimum cycle. 1936 is near cycle minimum, so with the already weak preceding solar max the EUV values would be similar or less than today.
Missed this earlier, my money would be on local events for the cold winter weather. I’d look at Black Sunday 1935 and it’s like for a cause.
tty says:
September 5, 2010 at 12:29 pm
This seems to be a very goodc example of a classic type of statistical mistake. The mechanics of statistic analysis are deceptively simple, particularly today when we have tools like Excel, SPSS, Matlab etc.
However to actually use these tools correctly you need to have a profound knowledge of the mathematical basis of statistics, of the characteristics of the data you are analyzing…..
_________________________________________________
One of my major pet peeves with the readily accessible statistical programs and the “instant PhD in Statistics” seminars pushed by the “flavor of the month” salesmen/consultants.
I would not be surprised if this was done so the error bars of the data was narrowed significantly. If the error is too large the small warming signal is lost in the noise. Since we are talking a signal of less than one degree per century they had to get the error lower somehow.
AJ Strata shows a 1969 graph: the CRU computed sampling (measurement) error in C for 1969. Graph at http://strata-sphere.com/blog/index.php/archives/11420
AJ Says
“They start at 0.5°C, which is the mark where any indication of global warming is just statistical noise and not reality. Most of the data is in the +/- 1°C range, which means any attempt to claim a global increase below this threshold is mathematically false. Imagine the noise in the 1880 data! You cannot create detail (resolution) below what your sensor system can measure. CRU has proven my point already – they do not have the temperature data to detect a 0.8°C global warming trend since 1960, let alone 1880.”
So when might we see the actual data upon which the speculation wrt global warming is based? No data= no science. All claims are opinion without the missing? data. What say you Phil Jones?
Jaye says:
September 6, 2010 at 8:37 am
Phil. says:
September 5, 2010 at 8:57 am
Seems prudent
You have got to be kidding me. Outliers are indicators of either some important behavior or bad data. Either way you have to investigate. Ad hoc application of a statistical cut-off is not good data analysis policy.
I think this is at least the fifth time that I’ve pointed out that this wasn’t what was done, not to mention Rod and Mosh in his most recent post!
You are just an apologist.
No I just read the material, perhaps you should do likewise?
Nice. Good catch. It would be interesting to take your top 100 months in terms of positive/negative error and plot them by year. Is there any change in the occurrence of outliers? More extremes as the climate warms? Oh wait it would be simpler to look at the variance by year. Never mind.
Phil.
I think a lot of people should read more carefully. Hmm. I thought I was pretty clear at a couple points that the trimming happened in both warm and cold, that IN THE END the difference was minimal and that somebody should double check the stuff. Ron pretty much explained that the CRU 5sig screen does little and that the key lies in the processing post GHCN. since I’m trying to match CRU that was huge relief for me. Anyways hopefully when Im done there will be a “standards” based way for people to quickly run a CRU type analysis. And then people can fiddle about with the trimming of outliers and see that the difference really makes no difference.. execpt at the margins where angels dance on the heads of pins.
I’m wondering if people will clue in.
1. heres a small difference, oh makes no substantial difference
2. heres another small difference, oh makes no difference
etc.
My experience in finding these is this. The more I find, the more confidence I have that the peculiarities of men and methods dont amount to a hill of beans in this matter. the warming is there. anybody who wants to question it should look at data integrity. thats some serious grunt work.
Steven Mosher says:
September 6, 2010 at 11:16 pm
Phil.
I think a lot of people should read more carefully. Hmm. I thought I was pretty clear at a couple points that the trimming happened in both warm and cold, that IN THE END the difference was minimal and that somebody should double check the stuff.
I thought you were pretty clear too.
Ron pretty much explained that the CRU 5sig screen does little and that the key lies in the processing post GHCN. since I’m trying to match CRU that was huge relief for me.
Right, as I understand Ron the difference in the datasets appears to be in the homogenization step (at GHCN?)
The trouble is that if the model was soundly based there should not be many outliers of 5 sigma+. It makes no diffference that the outliers are equally spread between warm and cold. The fat tails indicate that the smoothed model is not a good fit for the data. Unless all or nearly all the outliers can be put down to errors in data gathering, the Gaussian distribution is not an appropriate one on which to base the model and another more appropriate one should be sought.
Solomon Green says:
September 7, 2010 at 12:52 pm
The trouble is that if the model was soundly based there should not be many outliers of 5 sigma+. It makes no diffference that the outliers are equally spread between warm and cold. The fat tails indicate that the smoothed model is not a good fit for the data. Unless all or nearly all the outliers can be put down to errors in data gathering, the Gaussian distribution is not an appropriate one on which to base the model and another more appropriate one should be sought.
What model are you talking about? Five sigma is an improbable value regardless of whether it’s a Gaussian distribution or not, the actual probability isn’t invoked by CRU it’s just used as a threshold. It seems like a good choice since about 80% of them are correctable errors.
Jim Powell says:
September 5, 2010 at 4:43 pm
The greatly different historical data you cite for Feb 1936 at Huntley MT are but the tip of the iceberg of the massive data revisions evident in UHCN Ver. 2. While some egregious spurious trends were reined in (e.g., Grand Canyon AZ), a great many station records that showed no trend in Ver. 1 suddenly acquired one.
What is unacceptable is that these revisions were made surreptitiously in the so-called “unadjusted” record, downloadable from the GISS site as the default (“from all sources”) option. Many of them simply staircase the yearly data downward from some relatively recent year. Talk about manufacturing spurious trends! Climate data crunchers who want their results to be taken seriously need to look at–and overcome–such fundamental issues of data integrity, rather than contenting themselves with the clerical task of compiling regional or global averages and blindly reporting the “trend.”
I’d urge everyone to treasure whatever downloads they made several years ago. They’re as close to the elusive “raw” data as we”ll come. BTW, if anyone has a Ver. 1 download for Gonzales TX, I’d much appreciate a copy.
Roger Sowell says:
September 5, 2010 at 1:44 pm
Obviously you present the histogram of mean temeprature for all of the months of the year over the entire Abilene TX record. It consequently has a blocky, distinctly bimodal structure due to the nearly sinusoidal seasonal cycle. If you did a similar compilation for each month separately, you’d get something much closer to a bell-shaped histogram. (Interestingly, however, such treatment often produces histograms that are still bimodal, but not extremely so.)
While the goodness-of fit to the gaussian model needs to be established on a case-by-case basis, it is the standard deviation of the unimonthly distribution that is involved in identifying outliers.