More Fun With GISS Temps
Posted by Jeff Id on January 6, 2009
Just another little post to consider the accuracy of our current measurement system. Boris, pointed out how stupid I was for comparing the two measures as they are “not supposed to match” — a valid argument methinks. Well don’t worry Boris, I am oft criticized and I expected it when I made my post with a title like that. Why aren’t the satellites and ground temps. supposed to match? Because of this link graciously provided by Eric Adler: http://en.wikipedia.org/wiki/Satellite_temperature_measurements
Climate models predict that as the surface warms, so should the global troposphere. Globally, the troposphere should warm about 1.2 times more than the surface; in the tropics, the troposphere should warm about 1.5 times more than the surface.
The 1.5 number is in question as John Christy says it should be around 1.3 for the tropics. I used this link because of my basic laziness, I recommend nobody trusts wiki for global warming info but this above statement is easily confirmed through other sources. As an additional confirmation to my previous post, Wiki provided the following at the same link
But what does 1.2 times tropospheric amplification mean you ask. Well it seems to me that it means that the slopes of the UAH and RSS data should be 1.2 times the ground slope measurement so from my graph below the lower UAH red curves should be steeper than GISS.
Without going into explanations as to why the slopes are different, let’s look into the data a bit more. How much is the difference in short term variance between the data?
I used the two year filtered data to calculate the standard deviation of the values below. Unfiltered data produced SD values within a couple percent of each other due to increased GISS noise levels I offer this bit of confusion openly because it is important for proper disclosure of method.
Standard Deviations SD of GISS data = 1.06 SD of UAH data = 1.31 Ratio of SD = 1.23 UAH/GISS.
This value appears to confirm the model data for short term variations in temperature as proposed by John Christy, —Troposphere multiplier of about 1.2 with tropics about 1.3 giving a total of about 1.23. Pretty amazing considering the tropics comprise 40 of 180 degrees or 22% of the signal. 22% x 1.3 + 78% (rest of earth)* 1.2 = 1.22 . Not bad for an aeronautical guy eh?
I plotted the flat trend GISS and UAH data below dividing the UAH data by 1.23 after linear de-trending. 
The difference between the signal in between the two data sources is a quantification of the measurement error level. Standard Deviation calcs – SD_err = square root(SD_giss^2+SD_uah^2). I assumed the whole error for this graph rather than 1/2 the square root. (again in favor of higher noise than actual). The graph is otherwise pretty self explanatory, the residual difference looks to be fairly stationary as expected if I did my de-trending correctly so I applied an ARMA model to the result.
Just to disclose method. Before applying ARMA, I normalized the difference above by SD, fit the model and de-nomrmalized. This was done to insure a better fit for R.
The ARMA values were: AR=0.3974 MA= -0.1788
This graph is a plot of the difference signal as compared to my modeled difference so you can see the quality of the ARMA 1,1 model data. By doing the model this way I have assumed that any signal level under 30 years is stationary and stochastic. I don’t agree with some details of this assumption but the difference in assumption makes my following calc of the standard deviation of slope variation even less. — I am choosing worst case.
I then generated 1000 years of ARMA data based on the above coefficients and plotted the 30 year variations in slope. 
The histogram of the ARMA slope analysis looks like this.
The graph above is forced by the law’s of god and math to be a bell curve so the only interesting thing is the visual appearance of the width and the x axis. The standard deviation (one sigma) of the slope data is 0.001 degC/30 years. WOW! Two sigma 95% significance based on the measurement noise is 0.002 DegC/30 years!!!!!!!!!
Summary, Conclusions and questions:
1 – Model’s say the lower troposphere (UAH) should have a 1.2 times temp over ground measurement as confirmed by short term variation.
2 – UAH slope is substantially less than the 1.2 x GISS 1.83degC/decade or 2.196 C/decade
3 – GISS ground measurement stations have well documented problems in temp quality.
4 – GISS has a very complicated and illegible correction method as demonstrated by people trying to replicate the calcs.
5 – UAH standard deviation is nearly perfect compared to prediction at 1.23 deg C/decade X GISS.
6 – Measurement accuracy analysis determines that 30 year slopes as provided by the different calc methods are minimally affected by short term weather variation.
Questions for the interested-
How can the slope be so far outside of prediction when the short-term (two year) variance falls in line with model predictions?
What are the mechanisms which could account for the difference between the slopes?
And of course with AGW, we have to ask — Is the explanation reasonable?
Finally – does either metric have a problem we should be concerned with?
My thought is that a signal with a proper 1.2 times magnification in short term variance means that the data is properly calibrated. If that is the case this is strong evidence for GISS being over compensated for slope.
Layman Lurker said
Jeff, if we assume for the sake of argument that UAH slope is correct at 0.128 C per decade, then divide by your amplification factor of 1.23 we would expect the surface warming trend to be 0.104 per decade. The gap for GISS would be -0.79 C per decade.
You have already shown the reverse scenario in your post if we assume GISS trend to be correct. The gap for UAH in this scenario is +0.96 C per decade.
The data as it is now and the model cannot both be correct. Your work with SD’s on detrended data makes an interesting argument for the models being correct (unless there is another explanation for the larger UAH SD). Since we are only left with the data it would appear to suggest that one (or both) of the metrics must have a trend bias.
This is interesting in the context of all the discussion about the troposphere temperatures and whether models are valid or not. Perhaps the logic should be reversed.
Many things come to mind when it comes to potential data problems. However these explanations can be narrowed down because not all potential data problems would affect the trend. Some would just cause random noise (if I am interpreting your bell curve properly then there does not seem to be much noise getting in between the two detrended metrics).
Jeff Id said
Dr. Lurker,
I don’t think 1.2 times confirms the trend of the model but it does confirm the amplification effect exists. It does give a bit of credibility to the overall atmospheric structure of the lower troposphere modeling though.
If you search for GISS corrections on the googleverse, I think you’ll find the source of the differences ;).
Even the RSS data is out of range for what is expected. I think I should look at that next.
Layman Lurker said
“I don’t think 1.2 times confirms the trend of the model but it does confirm the amplification effect exists.”
Indeed, it seems to. Furthermore, if the hypothesized tropospheric amplification is confirmed by this analysis, then there must (not might) be a trend bias in one of the metrics. I would suggest pulling the tropical data out and repeating the analysis for further confirmation.
“If you search for GISS corrections on the googleverse, I think you’ll find the source of the differences ;).”
You may be right but IMHO perhaps let further analysis lead you to that conclusion.
Some further thoughts:
1. The apparant trend bias would explain why no amplification effect has been observed so far.
2. Correct me if I’m wrong: The tight SD’s and shape on your bell curve would suggest to me that the bias is “systemic” (such as adjustment algorithms) which would apply equally to all data points. If errors causing the trend bias did not always apply equally to individual data points would not the SD’s be larger? Many types of errors which have been thrown out to explain the differences in the metrics could potentially be ruled out if the error type does not fit with what you have uncovered here.
3. Other types of errors which could affect the trend could be “step” type (perhaps possible with RSS) due to change in data sources etc. and data not properly adjusted. Obviously, with NSIDC still fresh in your mind I don’t have to explain to you about this type of situation. 🙂
4. Is this evidence of the tropospheric “fingerprint”? Similar analysis of the tropics which fit with hypothesized amplification would be very compelling for this. Is the “fingerprint” due to GHG’s or just any warming? Either way it would have profound implications for the models depending on which metric contained the bias.
5. Exploring the other data sets with similar analysis would not only uncover the extent and types of trend discrepencies between individual metrics – it may single one metric out as being more out of line.
Mike Davis said
Jeff id:
If models cannot replicate the past and cannot track 30 years into the future they have been falsified. If data needs to be adjusted it is not fit for use. I read some where that we would have a better idea of global temperature by tracking the data from one site that has been unaffected by its surroundings than the garbage they are using now. That said I feel that GISS and HADCRU are worthless. Maybe some someone could draw some decent unadjusted data from the historic records as they do with paleo data and match it up with UAH/ RSS after they reconcille the differences between those two. However even those data should be used with the realization of error in collection. I feel that anything less than whole numbers is only wishfull thinking! Nature other than Homosapiens only exists and experiences. Our attempts at measuring have no meaning other than a means of communicating an idea with others of our speceis
David Jay said
Mike:
Anthony’s work at Surface Stations clearly shows that using “all the data” by including the bad stations does not improve the quality of the work product.
paminator said
Jeff- Another very interesting post! My initial reaction is that the difference in slopes is attributable to the UHI contamination in GISS. McKitrick et al published an article recently that estimated about half of the surface temp trend is due to uncompensated UHI. If you take this estimate at face value, that would lower the GISS trend to about 0.1 C/decade, which is quite consistent with the LT trend of 0.15 C/decade being 1.2 – 1.3 times higher. It confirms a small contribution from CO2.
Mike Davis said
David Jay:
I agree with Anthony on that point. I was saying that if you use rural data unadjusted and verified you might have a chance to find some value as SPPI has done. I feel that ALL the data is ____ (fill in the blank. I do not to use words that are going through my head)
Eric Adler said
Jeff,
I am not sure what point you are making by looking at the data with the trends removed.
It isn’t clear what the sources of random variation in the data is. In the case of satellite data we could be looking at measurement noise, because we aren’t averaging data over a large number of noisy instruments. Because of this uncertainty, I wouldn’t attach much significance to the agreement between theory and data in this case
There is a discrepancy in ratios of the trends, which is the important unresolved question.
There is also the divergence in the analysis of the satellite data to give the temperatures.
I think you have the ratio of the tropical data to the total data slightly wrong.
area of tropics/total surface=sin(20)=0.34 or 34% not the 22% that you claimed.
DeWitt Payne said
Filtered = autocorrelated. The monthly data has substantial autocorrelation as it is. Are you sure you haven’t lost too many degrees of freedom?
DeWitt Payne said
Why not regress the monthly UAH data against the GISS data and look at the slope rather than the ratio of the SD’s? You would still have a problem with autocorrelation, I think, but maybe not as bad.
Jeff Id said
Eric,
The linear calc was intended as quick check of the magnitude of the 1.3 times. You are correct though (I believe Sin20 is right) which actually pulls the number more into alignment with Christy.
+.34*1.3+.66*1.2 =1.23
DeWitt Payne,
I think of autocorrelation as a multiplicative and filtered as additive versions of the same effect. By modeling the signal noise including both terms, the calcs mean the actual slope variance we would see including AR and MA are accounted for in the long term data. As I see it, the autocorrelation then affects the quality of a single slope fit r value but not the quality of the standard deviation of the slopes we would actually see in a trend fit.
I’m not sure I understand how regressing the data improves the analysis.
—-
I guess the point of my post was just an exploration of the differences in the data. The UAH data has the expected amplification factor over GISS in short term temp variations but the slope is substantially less (very odd). By looking at the ARMA analysis of the accuracy we can actually measure temp it’s pretty easy to see the 30 year trend should be much closer between metrics. What are the mechanisms which drive such a huge difference besides the obvious “errors” in GISS method.
The post also points out just how badly the trend is really known.
When I think about the quality of the GISS stations and the types of ad-hoc corrections applied and the lack of historic stations used pre-1920, this is by far the most suspect.
HadCrut won’t even disclose it’s software corrections so I don’t consider it as anything other than a painting.
—-
I want to look at RSS compared to UAH and GISS when I get time.
Jeff Id said
Eric when you said- –
“In the case of satellite data we could be looking at measurement noise, because we aren’t averaging data over a large number of noisy instruments.”
One of the odd things which showed up in my analysis that I put into italics was that the GISS data is quite a bit noisier as shown by the unfiltered SD values.
Unfiltered data produced SD values within a couple percent of each other due to increased GISS noise levels
That’s why it’s interesting to do these kinds of posts. I think I can learn more from working with the data than I ever would reading articles about it. I get to look at dozens of graphs for a post like this, see things that go against my assumptions (like your reasonable assumption that GISS would be less noisy), plot them any way I want to try and understand. Fun stuff.
The satellites turn out to actually have a better signal quality because of the high degree of area coverage, consistent reading of temp scale and reduced outside influences.
If AGW is real, it sure would be nice to nail down the temp trend measurement a bit better so we can see how bad it really is.
David Jay said
Paminator:
It confirms a small contribution from CO2.
I’m not sure exactly what you mean by “confirmation”. Do you know with certainty which components of the climate system are causing long term temperature changes? And which measured items (such as CO2) are cause and which ones are effect?
Layman Lurker said
Jeff, just to be clear, you should correct your post using 34% for the tropics even though it makes little difference.
Layman Lurker said
I think the key point of the analysis is the ratio of the SD’s. It think it is safe to say that the ratio is consistent with the hypothesized amplification factor . Does it PROVE it? Is this just a remarkable coincidence? The reason this is important is that the question of whether unexplained trend differences between the metrics can co-exist might not be ambiguous any more.
Also, if there are so many data problems with the different metrics, how come the detrended slope differences between the two filtered metrics examined in this post (UAH corrected by SD ratio) gives 2 sigma of the tiny .002 C? Maybe I am not looking at this right but to me this signifies that any of the “micro” level data issues are few, small, random, and cannot even affect the 2 year slope let alone 30. It also confirms an extremely tight relationship between the lower troposphere and the surface so one cannot just toss out “they measure different things” to explain away the discrepency.
Jeff Id said
#14
If it were someone other than Eric, I would leave it the way it is just because it is an engineering check of the quality rather than an exact number. I have to do this type of estimation often in my daily life and it wasn’t intended to claim that the area of a sphere is linear.
Since is is Eric who is somewhat more critical of my work, I will change it but it will have to wait until I return as the change process is time consuming where I’m at and I’d rather work on my next post.
Jeff Id said
I should explain my last comment a bit more. I found that the SD of the raw data had a ratio near to 1:1 as I explained in the post. I viewed the graph and realized there is a higher degree of monthly noise in the GISS data so I ran a two year gaussian weighted sliding filter and took the ratio again. I found the 1.2 ratio was very close and even did a quick check on the 1.3 ratio in the tropics.
It is possible that different filter lengths will result in less fortuitous matches (I haven’t confirmed it) but if I ran a 4 year filter could I get an SD ratio of 1.25:1, I don’t really know. All I do know is the value from a two year filter came out pretty damn close to what Christy and Spencer have determined.
With this level of uncertainty in ratio I didn’t feel a better calc was useful. In engineering you learn to get a feel for how an instrument performs and make quick estimates to confirm your analysis. The two year filter was my first guess as a reasonable way to extract a reliable amplification after looking at the raw data plots. — looks like it was good guess.
paminator said
re David Jay-
Nope to your questions. I’m pointing out that if the model prediction is correct, and if the LT temperature trend is correct, and if the UHI bias in GISS trend is a factor of two, then the surface trend is actually smaller than the LT trend by a factor that is not inconsistent with the model prediction. Whether the trend is all due to CO2 is not at all clear. However, its hard to see how temperature increases much bigger than 1 C/century due to CO2 are possible. I consider this a small effect.
Eric Adler said
The differences between UAH and MSU satellite data has a long history which has not yet been resolved. They have different algorithms for diurnal corrections, combining channels to measure temperature, and different methods for calibration between successive satellites.
Over the years they have come closer together, but large differences still remain.
I don’t think statistical analysis will resolve them.
Tamino at Open Mind looked at the difference between them recently, and one of the things that struck him was a large annual cycle in the difference. The reason for this and he significance was left as a mystery.
http://tamino.wordpress.com/2008/10/21/rss-and-uah/
A more European Satellite program, the CHAMP compared its data with RSS and AMU.
http://www.cosis.net/abstracts/EGU2007/10228/EGU2007-A-10228.pdf?PHPSESSID=6be219b225e17f6d54a32b7f25598920
I don’t get much of value out of abstracts like this.
Here is one quote that stands out.
“In terms of absolute temperature we found that CHAMP TLS temperatures globally
agree better with UAH temperatures outside the summer season and with RSS within
summer, whilst ECMWF temperatures generally agree better with RSS temperatures.”
ECMWF has not been clearly described in the abstract, but it stands for
“European Centre for Medium-Range Weather Forecasts”, which seems to consist of Radiosonde balloon data.
Layman Lurker said
Eric, you sound like a good guy but you seem a little quick to dismiss things (“I don’t think statistical analysis will resolve them.”) and a little premature in your assertions. Perhaps questions instead of statements would be more appropriate.
Tamino’s post you refered to was followed up on by Jeff in two posts:
https://noconsensus.wordpress.com/2008/10/25/an-orbital-heating-signal-from-solar-input/
and
https://noconsensus.wordpress.com/2008/10/26/half-year-cyclic-variaition-in-rssuah-and-giss-anomaly/
Very interesting reading (check out the comment threads and Dr. Svalgard’s comments)
If you eyball the blue line from his latest graph you will see the annual oscilations in the differences between UAH and RSS from about late 90’s onwards. You will also see the step at 1992. These effects actually show up better in the “an orbital heating signal from solar input” where Jeff cuts and pastes the difference graph (unfiltered I think) from Tamino’s post.
You state:
“I don’t think statistical analysis will resolve them.”
That is a curious statement to make. It’s like you are saying that this is all such a mess that we should just move along. Are all of the questions raised by these issues going to be resolved by statistical analysis here? No, and that would be self evident to most of the regular participants of this blog. Are there SOME important questions and insights that can be gained by exploring the data? Yes there are – like for example: whether the annual oscilating cycle in UAH or the step in RSS will affect the 30 year trend….or not. There are others angles which I suspect Jeff will explore, but I don’t want to spoil his fun.
Eric Adler said
I think taking more data from different sources and more work on the physics is needed to sort out this mess. Statistics may show where to look, but I bet that the scientists who are doing the research are doing it. They just don’t have much time to blog.
The analysis is interesting to do, and it is nice that the data is there for the public to mess around with, but it takes a lot of education and time to really figure out what is going on. People who go into science as a profession are generally not stupid. If they can’t resolve these problems working on it full time, I don’t think we can have high expectations for amateurs to make a significant contribution.
Layman Lurker said
Eric states:
“The analysis is interesting to do, and it is nice that the data is there for the public to mess around with, but it takes a lot of education and time to really figure out what is going on. People who go into science as a profession are generally not stupid. If they can’t resolve these problems working on it full time, I don’t think we can have high expectations for amateurs to make a significant contribution.”
Unfortunately it seems I was right in my interpretation of your earlier comment. I don’t think that Jeff or many others here would suggest that scientists are stupid, but your expectation that all issues have been dealt with or understood is naive, as is your suggestion that “amateurs” cannot make a contribution. Again you are being dismissive without really thinking things through. NSIDC made corrections to their sea ice data time series data based on Jeff’s work on this blog. Jeff took Tamino’s (another blogger)observation of the UAH annual temp signal (and temp signals which he pointed out from other metrics) and pointed out the fit with the earth’s distance from the sun to which Dr. Svalgaard responded:
…”when you deal with temperature anomalies, this seasonal variation should disappear, if you do it correctly, i.e. deal with the two hemispheres separately [computing and subtracting the mean for each month [or day, whatever they use]]. If that is not done, or if the coverage is not the same in both hemispheres, or if there are any other little asymmetries, then you very easily get this kind of annual wave. For instance, the geomagnetic Dst-index [that measures the strength of magnetic storms] suffers from being based on 3 Northern and only 1 Southern station. This introduces an artificial annual cycle, see e.g. page 8 of http://www.leif.org/research/AGU%20Fall%202005%20SA12A-04.pdf
I don’t think [don’t know – more precisely] your effect has been noticed before. Good work.”
There are other examples on this blog (and others) which you just dismiss without even attempting to answer for yourself. The whole notion is ad hom.
Scientists themselves interact with bloggers in discussions and even colaborate with them. You just have to go to CA to observe this in action. These are scientists who understand that there are more questions then answers, and they also understand that the answers are not monopolized by scientists.
Roger Pielke Jr., when being interviewed on CE Journal just the other day made the following comment:
“Peer review is simply a cursory check on the plausibility of a study. It is not a rigorous replication and it is certainly not a stamp of correctness of results. Many studies get far more rigorous peer review on blogs after publication than in journals. I use our own blog for the purpose of getting good review before publication for some of my work now, because the review on blogs is often far better and more rigorous than from journals. This is not an indictment of peer review or journals, just an open-eyed recognition of the realities.
“It is hard to say who is outside and who is inside scientific circles anymore. McIntyre now publishes regularly in the peer reviewed literature. [Pielke is speaking of Steve McIntyre, whom I would describe as a climate change gadfly; he publishes a blog called “Climate Audit”] Gavin Schmidt blogs and participates in political debates. [Schmidt is a NASA earth scientist who conducts climate research.] Lucia Liljegren works at Argonne National Lab as an expert in fluid dynamics and blogs quite well on climate predictions for fun. She is preparing a paper for publication based on her work, but she has never done climate work before. I am a political scientist who publishes in the Journal of Climate and Nature Geoscience and blogs. Who is to say who is ‘outside’ and who is ‘inside’? Is participation in IPCC the union card? How about having a PhD? Publishing in the literature? Testifying before Congress?”
link: http://www.cejournal.net/?p=607
What else can I say Eric. I know I am just an “amateur”.
Jeff Id said
“I don’t think we can have high expectations for amateurs to make a significant contribution.”
Nice shot big guy, just how much do I have to get paid before I can analyze a time series. You know I am the first to point out that the statistical sorting methods of CPS don’t only cause distortions in the local data (three papers I’ve read after I figured it out) but also cause distortions in the historic signal. Considering that the historic signal is what the intent is, is that significant? How about the fact that I have found a way to correct for the distortion, is that significant? Perhaps the fact that the NSIDC corrected their website, because of an amateur. How about McIntyre’s efforts to get data available to the public, are you even aware that Mann wouldn’t publish his data in past papers and that his own unpaid, uneducated efforts got it done?
I don’t know Eric, what do you think scientists do? What would happen if a scientist chose to do his work in public on a blog?
Layman Lurker said
Jeff, I noticed something in Eric’s quote from the abstract comparing CHAMP data to UAH:
“In terms of absolute temperature we found that CHAMP TLS temperatures globally agree better with UAH temperatures outside the summer season and with RSS within summer”
Wouldn’t that line up with UAH’s unique annual temp signal?
Jeff Id said
I saw the statement also, it sounds like your right but I haven’t had time yet.
Boris said
I admit I’m part of the conspiracy. But you’ll never catch us!!! Bwahahahahahaha!
Jeff Id said
Tell us, which conspiracy is that? Is it the one to scam money from freddie and fannie, the plot to subjugate the holy Muslim or the one to present weak extremist science as determined fact?
It will help us find you!
Layman Lurker said
Jeff, for the sake of argument, if one was to correct for the UAH annual temp signal by pulling into line with the other metrics, how much would this affect the UAH SD of 1.31?
Boris said
Congrats on uncovering the conspiracy, Jeff.
Jeff Id said
Thanks dude,… consider bakin’ off the weed.