For years, climate scientists have assured us that NOAA’s homogenized temperature datasets—particularly the Global Historical Climatology Network (GHCN)—are the gold standard for tracking global warming. But what if the “corrections” applied to these datasets are introducing more noise than signal? A recent study published in Atmosphere has uncovered shocking inconsistencies in NOAA’s adjustments, raising serious concerns about the reliability of homogenized temperature records.
The study, conducted by a team of independent climate researchers led by Peter O’Neill, Ronan Connolly, Michael Connolly, and Willie Soon, offers a meticulous examination of NOAA’s homogenization techniques. These researchers, known for their expertise in climate data analysis and critical evaluation of mainstream climate methodologies, gathered an extensive archive of NOAA’s GHCN dataset over more than a decade. Their research involved tracking over 1800 daily updates to analyze how NOAA’s adjustments to historical temperature records changed over time.
Their findings reveal a deeply concerning pattern of inconsistencies and unexplained changes in temperature adjustments, prompting renewed scrutiny of how NOAA processes climate data.
The study analyzed NOAA’s GHCN dataset over a decade and found that:
- The same temperature records were being adjusted differently on different days—sometimes dramatically.
- 64% of the breakpoints identified by NOAA’s Pairwise Homogenization Algorithm (PHA) were highly inconsistent, appearing in less than 25% of NOAA’s dataset runs.
- Only 16% of the adjustments were consistently applied in more than 75% of cases, meaning the majority of “corrections” are shifting unpredictably.
- Less than 20% of NOAA’s breakpoints corresponded to actual documented station changes, suggesting that many adjustments were made without supporting metadata.
In layman’s terms: NOAA is repeatedly changing historical temperature records in ways that are inconsistent, poorly documented, and prone to error.
What Is Homogenization Supposed to Do?
Homogenization is a statistical process meant to remove non-climatic biases from temperature records, such as changes in station location, instrument type, or observation time. NOAA’s PHA algorithm adjusts temperature records based on statistical comparisons with neighboring stations—without needing actual metadata to confirm whether an adjustment is necessary.
This method has been defended by NOAA researchers, who claim it effectively removes bias. However, the new study suggests it might be introducing arbitrary and inconsistent changes that could distort temperature trends.
If NOAA’s adjustments are inconsistent, how can we trust the long-term climate trends derived from them? Here’s why this matters:
- Garbage In, Garbage Out: Climate models and policy decisions rely on adjusted temperature data. If those adjustments are unreliable, the conclusions based on them are questionable.
- Artificial Warming or Cooling? The study did not specifically analyze whether these inconsistencies bias the data towards warming or cooling, but past research has shown that homogenization tends to amplify warming trends.
- Lack of Transparency: NOAA’s daily homogenization updates mean that the past is constantly being rewritten, with little accountability or external validation.
The study’s authors argue that homogenization should not be done blindly without using actual station metadata. Instead, adjustments should be:
- Ground-truthed with station metadata whenever possible—not just assumed based on statistical models.
- Made transparent—users of temperature data should be informed about exactly when and why adjustments are made.
- Re-evaluated for bias—does homogenization systematically increase warming trends?
If NOAA’s temperature records are truly the best we have, they should be robust, reproducible, and verifiable. Instead, this study suggests they are a moving target, adjusted differently depending on the day, and often without a clear reason.
The question we must ask is this: Is the global temperature record a reliable dataset, or just a statistical house of cards?
We need transparency, accountability, and scientific rigor in climate science. Until then, every NOAA temperature dataset should be taken with a grain of salt.
Discover more from Watts Up With That?
Subscribe to get the latest posts sent to your email.
I can’t wait for the usual suspects to step forward and attempt to defend the indefensible…
They never defend it beyond saying -“it doesnt make a difference even when upside down’
You might need more than a grain of salt.
Tony Heller has been quite vehement on temperature “adjustments”.
My personal take is that the procedure is neither blinded or totally open. As the researchers “know” what the temperature trends should be, using adjustments inconsistently is quite plausible. And as they are not documenting just why they made all of the changes, they leave themselves open to people inferring bad intent.
I think it is some combination of sloppiness and Noble Cause Corruption.
I spent some time a couple years ago developing R scripts to examine the monthly USHCN station records Tony was talking about. I wanted to check for myself. Yes, confirmed. The cooling of the early years and warming of the later years by the combination of the TOB and PHA adjustments is very plainly seen by comparison to the “raw” files. To NOAA’s credit for transparency, the raw, tob, and FLs (finalized) files are all still available and are being updated in parallel here for the USHCN stations.
.https://www.ncei.noaa.gov/pub/data/ushcn/v2.5/
“Hiding in plain sight” is a well-recognized tactic for scammers too.
Good point! All the more reason to look behind the curtain.
“they are not documenting just why they made all of the changes”
Of coure they clearly describe the algorithm. And the code is available; you can run it yourself:
You did read the post? Doing so inconsistently?
The adjustments fit theories about CO2 levels better than the raw figures do. Which is why I believe it it Noble Cause Corruption, rather than conscious fraud.
“The adjustments fit theories about CO2 levels better than the raw figures do.”
No, they don’t. ClimDiv adjusted shows a lower rate of rise than USCRN unadjusted.
A point that is consistently ignored on this site.
USCRN is variously referred to here as the “gold standard” or “state-of-the-art” temperature reference data set for US stations.
Over their joint period of measurement (since Jan 2005) USCRN has warmed at +0.42C per decade; a total warming of +0.84C. By contrast, NOAA’s homogenised ClimDiv data warmed at +0.35C per decade; a total warming of +0.71C.
If the adjusted data set is warming more slowly than the pristine one, then the only logical conclusion is that the adjustments are making US temperatures artificially cooler.
The two data sets are provided by NOAA here.
I think the point is even if I agree it is warming by how much is a total guess. Then factor in no-one except the climate zealots gives a toss and the world continues to burn fossil fuels it’s likely going to get warmer and who cares for an inaccurate projection 🙂
Well, you should tell Anthony Watts that then, because he puts great store in USCRN.
Your ignorance is the one on display. If you look at the entire record, you find the attached charts which clearly demonstrate the data is being adjusted to fit the theory, which is the antithesis of “science”.
To spell it out when you plot the adjustments against CO2 concentration you get a straight line. Isn’t that convenient for the doomsayers….. The adjustments have cooled the past and heated the present in the exact relation to CO2 concentration.
REPEAT the Adjustments exactly follow the increase of CO2 for the past century+. Therefore those adjustments are fraudulent and most assuredly non scientific as in adjusting data to fit a theory.
Your argument that they have not been so bad with the last 20 years of “adjustments” belies your willful ignorance of the evidence.
I did look at the “entire record”; in so far as USCRN can be compared with ClimDiv, the ‘entire record’ starts in January 2005.
In the period up to January 2025 the ‘pristine’ USCRN is warming faster than the adjusted ClimDiv.
How many times does this simple fact need to be pointed out?
“USCRN is warming faster than the adjusted ClimDiv.”
Only a mathematical illiterate compares real data with “adjusted” data.
Any difference is from the “adjustment” routine.
There is actually no warming in USCRN except from the effects of the 2016 and 2023 El Ninos.
Data is nearly zero trend for the rest of the time.
So what do we compare ‘real’ data with, imbecile?
This comment illustrates very well what your expectation of data is – simple numbers to be manipulated mathematically! Numbers that can be changed to suit your purpose!
In physical science, data is sacrosanct. If the values are unusable, then they should be declared unfit for purpose and discarded.
LIG thermometers in a Stevenson Screen may provide absolutely accurate data. There would be no measurement bias at all. But, the data from this station may very well differ from a later MMTS station that is also absolutely accurate. The two measurement systems are different and will give different readings.
That is life in physical science.
If I measure the force of gravity on separate days using different stop watches, it is very likely I will get two very different answers.
Can I claim one set of data as “biased” and change the values so they more closely match the other data set?
That’s the whole purpose of “measurement uncertainty” in a measured value. Both readings should fall within the measurement uncertainty interval if you are measuring the same thing. If they don’t then 1. the measurand changed, 2. one of the measuring devices is inaccurate, or 3. the measurement uncertainty interval was improperly evaluated.
In 1. the measurement protocol is not correct. In 2. the measurement protocol is not correct (lack of calibration). In 3. the measurement uncertainty budget needs to be reviewed and modified.
This is why climate science ignoring measurement uncertainty is so “unscientific”. They just ignore measurement uncertainty by using the meme “all measurement uncertainty is random, Gaussian, and cancels”.
How many times does it need to be pointed out that weather follows a cycle of 60-70 years? 2025 – 2005 = 20 years, not even enough time to call it “climate” data.
“REPEAT the Adjustments exactly follow the increase of CO2 for the past century+. Therefore those adjustments are fraudulent and most assuredly non scientific as in adjusting data to fit a theory.”
Exactly right. The Temperature Data Mannipulators have created in their computers, a correlation between CO2 and temperatures that does NOT exist in the real world.
The original, written, historic temperature records from around the world do NOT show a correlation between CO2 and temperatures.
The “Keepers of the Data” are lying to the rest of us. Which amounts to defrauding us because their lies are causing our gullible politicians to spend TRILLIONS of our tax dollars in an effort to stop creating CO2.
Jail time is appropriate for such people, imo.
“The Temperature Data Mannipulators have created in their computers, a correlation between CO2 and temperatures that does NOT exist in the real world.”
To say nothing about correlation not being proof of causation. I haven’t done it but I would bet money that postal rates and the temperature data sets also exhibit a high degree of correlation. Do postal rates drive temperature or does temperature drive postal rates?
After more than 50 years of study climate science has yet to develop a functional relationship between CO2 and temperature. The climate models are nothing more than data matching algorithms used to extrapolate past data into the future, they are *not* functional relationships from which temperatures can be directly calculated, proof being the inability of the models to accurately predict “pauses” in either the short term or in the long term. The typical excuse is “it all averages out in the long term” – a prima facie clue that a “functional relationship” doesn’t exist, only extrapolation of past data trends.
Regardless of whether anything is a gold standard, averaging temperatures from different locations is physically meaningless.
The idea that temperature is a proxy for climate is a joke to start with. If that were the case then Las Vegas and Miami would have the same climate.
Everyone knows that ClimDiv is adjusted to match USCRN. They have just improved their algorithm over time.
Straight from the Haus’s mouth.
Evaluating the impact of U.S. Historical Climatology Network homogenization using the U.S. Climate Reference Network
Difference between USCRN and ClimDiv.. getting closer and closer together.
The difference is becoming more negative, because USCRN shows a larger warming trend than the full bias adjusted station network. If the series had been adjusted to match, the difference across the entire length would be zero.
LOL… That for letting us know that the difference is because of the adjustment routines.
You have to be mathematical empty to draw any comparison between actual data and data that is “adjusted” to math that data.
One is real.. the other isn’t.
Seems they have improved their algorithm parameters over time.
The adjustments are applied across the entire series – there isn’t a different algorithm used at different points in time on the graph you’re showing.
You would make more sense to everyone if you would provide a scientific description that shows the reason for each adjustment. Is there measurement bias which should have been corrected at the time of reading? Are devices not calibrated? Were reading errors made?
I’m actually struggling to simplify this any further. Bnice is suggesting that each month (or day, I guess, for daily values) a new adjustment is applied, specific to that single data point. He is suggesting that gradually, over time, the NOAA is tweaking each month’s adjustment to try to inch the full, bias-adjusted network average closer and closer to the reference network.
But that isn’t what is done. The adjustment algorithm is applied all at once, to the entire time series. The algorithm is blind to the date that it is being applied, so it is physically impossible to observe changes in the time series reflecting changes in the adjustment algorithm over time. There’s just the one algorithm, applied once.
The reason the difference series moves from positive to negative is because the full, bias adjusted network is warming slower than the reference network. It is not because the values are converging to zero (which a quick glance shows is blatantly not the case in the first place) as a result of fraudulently transient methodology.
Is this clearer?
No!
Why are you changing anything to begin with? What is the rationale? What is the need for changes?
Are the past recorded temperatures incorrect?
Do past readings have a systematic measurement bias?
Are the current temperatures measuring the same microclimate?
Is it possible to to just find the slope of each segment that has been measured with the same device and same microclimate and combined them?
Basically what is the bias you are trying to to eliminate? Why does the change need to be made?
These biases are well known and thoroughly described in the literature. They include site moves, changes in instrumentation, changes in the composition of the station network, and changes in observing practices over time.
I don’t understand how you’ve been engaged in these debates for the number of years you have been without understanding these things better.
You didn’t list out a single micro-climate change. Who needs to understand things better?
That is because you still fail to discuss why you call these things a bias.
Not one thing you have listed is a BIAS. To determine a bias, one must know what a measurement SHOULD BE! That is why one uses a measurand whose value is traceable to an international standard to determine what systematic error exists in any given device.
When correcting a bias, you must be able to tell us what the actual value of temperature should be.
Site moves of the same device don’t create a bias. It is measuring a different microclimate and SHOULD HAVE different readings. Unless you can verify that the new microclimate totally duplicates the original microclimate, then a change in values is not only expected but quite likely accurate.
Changes in instrumentation do not create a bias. That normally has a new shelter type and in many cases a new location. It is measuring a different microclimate and SHOULD HAVE different readings.
Here is how NIST defines bias.
https://www.itl.nist.gov/div898/handbook/mpc/section1/mpc113.htm
You have no idea what the true value of any temperature reading, temperature average, or regression is. Therefore, you have idea about what the value of the bias is.
Let’s be honest, what you are calling bias is merely a difference in temperature readings that are expected when repeatable conditions are not used.
Each reading may be 100% accurate for a given device and microclimate over the entire time it is used. Yet when a new device and microclimate is introduced the temperature readings will change. Your position is that bias is the cause. It is not. What has occured is a new record has been created with entirely different, but probably accurate, temperature readings.
You, as a mathematician, say, let’s just change records so we can create a long record that makes regressions look like accurate. That is not science, nor is it allowed in any other scientific endeavor. You treat probability distributions as if they don’t exist. Just ignore them until you can divide by √1000’s.
Which by the way, for an average monthly temperatures, do you have 30 samples of size “n=1”, or 1 sample of size “n=30?
They introduce a trend signal that is not related to the climate. That is what I mean by a “bias” in this context. We want to isolate the climate signal. If we can identify signals that are not the climate signal, we can remove them. We do not need to know what the climate signal is to do this, we merely need to be able to identify what is not the climate signal. The “true” value of the climate signal is unknowable.
Wrong on the first count, correct on the second. A site move can unquestionably introduce a non-climate signal into the trend for a station series.
Wrong on the first count, right on the second. An instrumentation change can unquestionably introduce a non-climate signal into the trend for a station series.
Wrong, even just conceptually. If I can determine what the bias is I can subtract it, whether I know the “true” value of the thing being measured or not.
No, wrong. My position is that bias is the result.
Again, these conversations are wearisome. You never listen, and never learn. Still making the same rudimentary errors you made years ago on this same website. You make no effort to absorb and address what is actually being said to you. You just talk past the people patiently trying to explain it to you.
“We do not need to know what the climate signal is to do this, we merely need to be able to identify what is not the climate signal. The “true” value of the climate signal is unknowable.”
Did you actually read this before you posted it?
Total-Signal = climate-signal + not-climate-signal
If you know the not-climate-signal then you know the climate signal. If you can’t know the climate signal then you can’t know the not-climate signal either!
We know parts of the non-climate-signal, part of it is random, part of it is systematic error we cannot account for. So the true value is forever elusive.
But each part of the not-climate-signal we can remove gets us closer to the climate-signal term we are after.
You wish.
Yet climate science knows the estimated value of the entire global “signal” to something like ± one-thousandths of a degree?
Give me and everyone here a break from your nonsense.
This is nothing more than an argument from incredulity. It’s not compelling.
Malarky! Your answer is the argumentative fallacy of Argument by Dismissal. You didn’t address the question at all.
Can *YOU* know the estimated value of the entire global “signal” to +/-.001 of a degree?
We can know the estimated value to whatever it’s reported to.
Kinda short on any scientific fundamentals! Cat got your tongue?
“But each part of the not-climate-signal we can remove gets us closer to the climate-signal term we are after.”
Again, if you don’t know what the not-climate-signal is then how do you remove even part of it? How do you know you aren’t removing part of the actual climate-signal?
In essence you are trying to say you are attempting to eliminate systematic uncertainty. The problem is that systematic uncertainty is not amenable to statistical analysis. So how do you identify it using statistical analysis?
If you can’t tell that a break-point is caused by someone planting a tree upwind of the station then how do you come up with an adjustment? The station may be 100% accurate in measuring it’s microclimate. It’s the microclimate that changed and not the station. Applying an adjustment to the current readings to make them match past data is misrepresenting what the current environment actually is. Changing past data to match current data is also misrepresenting what the past environment was. In each case you wind up with a record that is not representative of reality.
It’s why when a breakpoint is identified the old record should be ended and a new one begun regardless of whether the breakpoint is caused by station calibration (e.g. a site move) or by microclimate changes. For a global temperature where you have thousands of random variables being combined, dropping one past record and adding one current record should not significantly affect the average. If it does then something is wrong with your data. So spare me the excuse that you need a “long” record to determine what is going on with reality.
That’s exactly what I’m saying.
Yes, you have to understand something about the system(s) producing the systematic bias. Scientists understand these quite well because they’ve been studying them for decades.
Ok, I don’t think this makes any difference. Can you produce the results of your analysis demonstrating that it matters? One long record with the bias removed or two shorter records, split at the breakpoint, are functionally the same thing.
“Yes, you have to understand something about the system(s) producing the systematic bias. Scientists understand these quite well because they’ve been studying them for decades.”
I’ll ask you what the Ghostbusters were asked. Are you God? Are climate scientists God?
Think carefully about your answer.
“Ok, I don’t think this makes any difference. Can you produce the results of your analysis demonstrating that it matters?”
What analysis do you need?
Standard error = SD/sqrt(n). If n is 999 vs 1000 exactly how much does SE change? If n is 1001 vs 1000 how much does SE change?
Assume SD = 2, then 2/999 = .002002002…… and 2/1000 = .002
That’s a difference of .000002/1000 = 2×10-7. An insignificant difference.
Are you *really* that dense? Dropping one data set will *not* make a difference at all. Same for adding a data set.
The real question is where did the measurement uncertainty of the very first measurement disappear to. You can’t find the standard deviation of the final average, divide by √n and assume all the prior uncertainty just doesn’t matter.
Are you saying only god can understand anything about the world? What do you view the point of science as, then?
An actual analysis. Produce a global temperature estimate using your methodology and compare it to existing estimates, demonstrate that it yields a statistically meaningful difference.
No, only God can know the unknowable. Systematic uncertainty can’t be found using statistical analysis. That means that if you know the systematic uncertainty, or even part of it, then you somehow know the unknowable – so you must be God.
I’ve done the analysis. The measurement uncertainty overtakes the anomaly immediately – meaning you can’t know the anomaly’s actual value. Which means that if *you* know it then you must be God.
The temperature data sets don’t even weight the northern and southern hemisphere data to account for their different variances. That alone adds enough uncertainty that it overwhelms any differences.
This even applies to the daily mid-range temperatures. If Tmax has a different standard deviation than Tmin then the average can *NOT* be (Tmax + Tmin)/2. A weighting factor should be applied but climate science never does this. I guess because they are God also and can know the unkowable.
So if we have a documented station move and identify the breakpoint when the station was moved, that makes us god? Interesting theology.
You haven’t shown it here. You’re claiming that treating the series as two separate series at a breakpoint is functionally different than adjusting the breakpoint in-place. You need to demonstrate that. I think it’s the same thing.
If there is a step change, the 2 series are obviously different to a single series derived from adjusting part of the series up or down to remove the step change.
Each series has its own intercept and slope.
Obviously you would normalize the two series before combining them.
I don’t think the GISS algorithm does that.
The GISTEMP algorithm does not spit records at breakpoints, that is a novel approach that old cocky and the Gormans are proposing.
What does that have to do with normalisation?
Isn’t that essentially the BEST approach?
Not in the way the Gormans are suggesting, I think, else they’d simply be endorsing the BEST methodology. But they have yet to describe their method in detail, or how it differs fundamentally from what anyone else is doing.
It’s not all that novel, then 🙂
“But they have yet to describe their method in detail,”
What detail do you need? The detail is to stop trying to create long records by adjusting temperatures before and after a breakpoint.
You don’t *NEED* long records to establish trends of temperatures generated from combining temperatures.
Perhaps you have never encountered a situation where you must do a piecewise integration of a function.
f(x) = {f₁(x), a ≤ x ≤ b
{f₂(x), b < x ≤ c
{fn(x), d < x ≤ e
When you have a piecewise time series, i.e., one with breakpoints in the dependent variable, you need to do regressions on a piecewise basis also. This shows to any reader that a change has occurred at various points in the function and preserves the association between the pieces.
You shouldn’t need to be told this if you have any experience dealing with complex waveforms and functions with nonlinear interactions.
I’m sorry this messes up averages and long term trends, but that is life in the physical sciences.
You have yet to define the term bias with any specificity. It can’t be measurement bias because you have no way to evaluate that specific quantity from stations in the past. Any assumptions other than accurate readings in the past must include evidence to the contrary.
Different readings are indicative of either new devices and/or different microclimates but not inaccracy. If this is what you call bias, then your goal must be to make the readings all the same, which results in your attempting to equalize the microclimates.
Normalization, as you call it, has no use in evaluating sequential data. The breakpoint will still remain and variance of non-Gaussian distributions is problematic. Why don’t you describe the type of normalization being used?
Unfortunately, this doesn’t cut it, because the trend is the thing we are trying to determine.
We want to know how the climate is changing – the trend. Anything that imparts a trend that is not related to the climate is a bias.
So by God, we’re going to get that “trend” even if we have to change data to find it.
You have just defined an end justfies the means.
You would do better asking yourself if the data you have is fit for purpose and if not, why play with it when you don’t have a clue why it isn’t fit for purpose.
You don’t have to do any adjustments at all to see the trend. You just get a more refined estimate if you do them. Here is adjusted versus unadjusted:
What is the trend on this graph, before adjustment and after? Absolute and anomaly.
https://ibb.co/FN4RvfP
I didn’t make the graph, so can’t go back and retroactively calculate the trends for each series. It is obvious on visual inspection that the trends are very similar.
I’ll give the benefit of not seeing what I posted. Here is an online version.
The trend is ~ zero. However there is a breakpoint. Changing one series will have an effect on the calculation of anomalies for this station.
I don’t know what this graph is, you’ll have to offer some more context.
You are being pedantic now.
Here is a graph I have quickly marked up. It has Tmax monthly average for every month between 1900 and 2024. The higher temps are summer temps and the lower temps are winter months.
It doesn’t take a big math algorithm to discern breakpoints. Around 1980 there was probably a change from Stevenson screens to MMTS. Then another change of some kind.
Deal with the issue that a regression from 1900 to 2025 needs to deal with piecewise segments rather than changing data.
It would take some documentation of the station history and/or some comparison with nearby stations to identify breakpoints. But if you identify breakpoints, and can estimate the magnitude of the discontinuity, you can remove it and get a better estimate of the long term climate trend.
“you can remove it”
Which misrepresents what was actually happening at the station. I.e. if falsifies the data.
If the breakpoint is due to a station move then you must be able to verify that the environments at both locations are exactly the same. If they aren’t then you will wind up changing data that accurately represented the environment in the previous microclimate (or in the new location, take your pick).
We aren’t trying to represent what happened at the station, we are trying to represent the regional climate signal.
Identifying a breakpoint means that something not related to the regional climate has changed, and we want to remove that.
I think the is the heart of the deep, fundamental misunderstanding the the contrarians have about what is being done. Each station is a point-level record that we want to use to represent a temperature field surrounding the station. We explicitly do not want features in the record reflecting things that happened to the observational instrument rather than the thing we are trying to observe.
Like it or not, you do not have a physical measurement of a regional temperature.
You have an average of individual physical measurements from individual stations. Modifying a mean results in a change to ALL the values used to calculate the mean.
No wonder you think it’s ok to make changes wily nily with no repercussions.
Right, you don’t have a physical measurement of anything except the exact point-location where the station was sitting at the time the observation was made. But we want to use that point measurement to represent a much wider area via interpolation, so we don’t want to include trends that are specific to the measuring device and not the wider area.
It feels like you’re half-listening to what I’m saying.
Tell how widening the area affects the uncertainty of the value?
Everyone here can give you an assessment of how temperatures change over a short drive.
This is just one more piece of evidence that climate science has no idea what it is doing.
Maybe you can give a paper that discusses the affect on the uncertainty of the interpolation on the basis of physical experimentation.
Here is a picture of a local temperature distribution. Tell what your measurement model calculates the missing value should be along with the uncertainty of the interpolation.
The thing being interpolated is the temperature anomaly. Spatial interpolation is a common technique in geospatial analysis, I’m not sure if you’re trying to say that it’s invalid.
The anomaly is still an intensive property. What is being done in geospatial analysis is usually an extensive property such as mass in a deposit, etc.
No, the fact that the temperature anomaly is an intensive property does not prevent it from being interpolated. What matters is spatial correlation, and temperature anomalies are highly spatially coherent.
You have a hot coal in each hand burning your flesh. What is the “interpolated” temperature in the space between your hands?
As usual, you are pettifogging. What is your definition of spatial? Pikes Peak and Colorado Springs are close based on latitude and longitude but have very different temperature anomalies. Does your “interpolation” include factoring in elevation, humidity, wind, terrain, geography, and cloudiness? Holton, KS snd Topeka, KS are close but have big differences in terrain, elevation, and geography which causes their temp anomalies to be different. How do you “interpolate” those anomaly differences without factoring the other components in as well?
Anomalies have different variances based on all sorts of factors, and those factors are not all “spatial”.
It causes their local climatologies to be different, the anomalies are likely to be similar, because the anomaly is the measured temperature minus the climatology. If it’s a hotter than usual day at Pike’s Peak, it is likely to be a hotter than usual day in Colorado Springs. This isn’t an assumption, peer reviewed research has established this beyond doubt. Anomalies are correlated out to 1000s of km.
There seems to be a deliberate effort on the part of contrarians to misunderstand what is being said.
“climatologies to be different, the anomalies are likely to be similar”
The operative word is “LIKELY”.
Meaning you simply don’t know what their relationship is to the hundredths if a degree. Further meaning that when you GUESS at a value you are ADDING to the uncertainty if anything you the values for.
You just keep digging your hole deepet and deeper. Keep digging!
This is the fundamental nature of geospatial analysis. You take measurements where you can and fill in the missing spaces with the best estimate you can make. It is a better estimate than assuming you have no information about the missing spaces. Because in fact assuming you have no information about the missing space actually is an assumption about the value of the missing space – it’s the assumption that the value is best represented by the worst quality guess you’re able to make.
“You take measurements where you can and fill in the missing spaces with the best estimate you can make”
That works for extensive properties but not for intensive ones.
If I give you a weight at 10C snd a second one at 20C do you have a total temperature of 30C?
It objectively works for a temperature anomaly field where adjacent anomalies are highly correlated.
You NEVER show any references to support your assertions. I post page after page from established universities, from well established blogs, and from textbooks. You NEVER refute them, you just ignore them and keep posting as if you were the world’s expert.
You simply can’t be believed if you never support your arguments. That is a sign of someone with not enough expertise to have a library of supporting information.
My own software storage bibliography has upwards of 2000, papers, textbook notes, blog pages and university online class notes. I have textbooks on metrology and statistics among other engineering texts.
Have you ONE book you can quote from?
See Hansen and Lebedeff, 1987, to start.
It is your assertion. You need to show here the text that you feel justifies your assertion. You can copy and paste here pretty easily.
By doing this there is no misunderstanding about what you are showing as a basis for your assertion.
This what I do with either a url or a book cite. You need to do the same!
The entire paper is easy to follow and extremely relevant. See figure 3 in particular. From the abstract:
Correlation doesn’t mean equal. Height and weight are highly correlated but a change in height doesn’t imply an equal change in weight. It can’t because they have different units!
Just because the temperature at StationA on the east side of a mountain goes up 1deg that doesn’t mean that the temperature at StationB on the west side of the mountain at the same elevation will go up 1deg – because the sun insolation is different. Both stations may only differ by a few kilometers based on longitude and latitude.
Does ANYONE defending climate science actually have any relationship to the physical realities of the Earth and temperature or do they all live in a climate controlled basement somewhere?
The expectation isn’t that the anomalies are exactly equal, but that because they are highly correlated, you can interpolate between pairs of neighboring stations to estimate the anomaly field.
The anomaly already factors in local differences like angle of insolation. It’s a measure of how typical the measured value is relative to the norm for that location.
There is *NO* “temperature anomaly field”. That’s why temperature is classed as an intensive property. How would you interpolate the temperature at the top of a mountain from a station on the east slope and one on the west slope? What does that temperature anomaly field look like?
How would you interpolate the temperature between a station at the bottom of a river valley and one at the top of the river valley?
How do you interpolate the temperature between a station in the transitional hardwood-to-plains area of eastern Kansas and a station in the open plains of western Kansas?
Temperature is a function of a large number of factors including, elevation, pressure, wind, terrain, humidity, and geography. If there is a field it is a multi-dimensional one that simply doesn’t lend itself to a simplistic linear interpolation between physically separated locations.
Everything you’ve offered in this thread is nothing more that unjustified assumptions that are physically incorrect. Their only purpose is to make it easier to say that “temperature determines climate”.
My guess is that you won’t answer a single question above.
Your persistent inability to comprehend the concept of a temperature anomaly is borderline pathological. I’m not sure how I can help you understand at this point.
You didn’t answer a single question.
The only thing pathological here is you making assertions you can’t justify. You have a pathological aversion to admitting that temperature is based on a multitude of factors and is a complex, multidimensional field. You can’t even admit that you can’t interpolate temperature at a mountain top based on the temperature on the mountain’s east slope and west slope based on some phantom correlated temperature gradient field.
Let’s do an experiment. We’ll pick a baseline of 64.2 which is the average of the temperatures and 60 just for fun.
The average is 707 / 11 = 64.2 and the anomaly is “0” when using this value. Rounding using significant digits, the final value is 64.
What about 60 as a baseline? The anomaly is 4.3 and the the absolute temperature 64.3. Rounding using significant digits, the final value is 64.
It really doesn’t matter what anomaly value you use, you will end up with 64.
Here is the photo of the original
64 is a long way from 66.
Even spatial analysis encounters this problem.
At t₁ you are driving at 60 mph at 900 ft. At t₂ you are driving at 60 mph at 800 ft. Did you drive down a 30° slope or did you drive over a 100 ft cliff? How do you know from just the two data points?
You can’t arbitrarily pick a value for the climatology, you need to determine it based on historical values. If you want to interpolate the anomaly between Manhattan and Council Grove in Junction City, you need to know the local climatologies for both sites to determine the anomaly for Junction City, which you will be confident is well estimated by the interpolation because you know the anomalies are highly correlated.
It’s like standing on a hillslope and getting an altitude of 120 meters above sea level at point A and an altitude of 100 meters above sea level at point B 100 meters away. What is the best estimate of the elevation of a C point halfway between them? What if I have a fourth point D 20 meters from point at with a measured elevation of 120 meters?
This is basic stuff. It’s possible that my best estimate based on the measured values is wrong – maybe there’s a boulder at point C 5 meters high. But my estimate is unquestionably better than assuming i have no information about point C whatsoever.
You keep ignoring the fact that a climatology (made up word) is created from individual measurements (observations).
A daily Tavg is a random variable made up of independent observations. Tavg has both a mean and variance. A monthly Tavg is a random variable made up of independent daily averages having means and variance. The monthly Tavg random variable also has a mean and variance.
You can not modify the average value of these random variables without also admitting that the components must change in order to maintain continuity. That defies logic.
To do otherwise simply points out that you have decided that the end justifies the means. You need to show a mathematical proof that would allow this.
That’s irrelevant. I’m saying you can’t make up what the baseline value is, you have to estimate it.
Estimates ALWAYS have uncertainty. Where is your statement of the associated uncertainty?
Refer to the relevant publications for uncertainty estimates (Morice et al., 2020 for HadCRUT, Lenssen et al., 2024 for GISTEMP).
Yes. They can do it, because they don’t have a fixed interval for anomaly base. They use a least squares fit. So the process of anomaly calculation repairs the break. If you had a station which was always 10 before, and always 11 after, the anomaly will be 0 everywhere.
But GISS and others use a fixed interval (GISS 1951-80). Breaks would mess that up.
FWIW my own method TempLS uses least squares, but I don’t slice. That may be why I find that the global average is much the same whether you use adjusted or unadjusted data (and much the same as GISS and BEST).
Is that from the Rohde et al paper, or something later?
“Anomaly” in Section 3 appears to be divergence from the interpolated value.
Yes, it is the 2013 paper. Here is the math:
The red term is the time invariant “climatology”, which does not have a fixed period associated with it. The sum of the green terms is the anomaly.
Incidentally, TempLS was doing this three years earlier, with the same decomposition, as described here.
I wrote a post here showing how the obvious use of an average of whatever data you have doesn’t work, but can be made to work by a little iteration. That is a simple way of doing the least squares fitting for the “climatology”.
I notice your post has no discussion about how this method allows uncertainty to be propagated. It appears you simply assume that the numbers are 100% accurate.
The usual mechanical response. In fact the W is treated as a random term. It includes weather variability, but also measurement uncertainty etc
And what is a typical value and what are the typical component values for them?
Are the values estimated or measured?
They are just residuals. The values are just T-C-ϑ. C is just a slightly modified average local temperature, ϑ is the familiar global average.
Thanks, Nick.
A couple of paragraphs later, it clarifies that the C(x) term is a geographic component – latitude and elevation plus a fudge factor.
The fudge factor is around 5% of the divergence from the interpolated value, so is not insubstantial.
Subtracting a site baseline average has an advantage in this regard, but does require the site to have been operational for a substantial portion of the base period.
Yes, BEST unwisely tries to tie C(x) to lat/lon etc. It should be just derived by fitting ie the average of the data. The TempLS algorithm is:
With ϑ=0, average over t only (fixed x) to get C (ie station means)
Then average over x only to get ϑ
With that ϑ, average again over t to get C
With that C, average again over x to get ϑ
and so on – converges in about 4 iterations.
Length of data is not critical – all the sites in GHCN are long enough.
Thanks.
BEST’s C(x) makes sense because they are attempting to give an absolute temperature (which they show in C, for some reason).
If the only interest is in change over time, empirically deriving the station baseline over the life of the station does seem reasonable.
Does the post-1965 trend skew the results for “new” stations vs “old” stations? Would detrending make a difference?
The iteration I described is detrending. The purpose is to avoid that skew. C is fitted to (W-ϑ). ϑ includes the trend. I showed in this post how the skew arose and how the iteration removes it.
The trouble is that the trend is time-varying prior to about 1965, as per the chart that Alan posted earlier today.
You’re detrending over the life of the record.
As I said, ϑ includes the trend. But it is a general function of time; in fact it is just the widely published global average T. You can derive trends from it at different times.
It’s taken a while to ponder this, so you may not see the reply – perhaps later.
The theta trend isn’t linear, so C will depend on the time period under review. The mild concavity isn’t likely to make a lot of difference, but C will be slightly non-stationary, as is the trend.
The WWII spike is likely to have an effect as well.
Linearity isn’t an issue. The estimates of C are affected by timing, because the T series is not stationary. Estimating and subtracting ϑ makes it as stationary as you can get it.
It must be too long since I’ve done any serious maths, so I’m still missing something 🙁
Agreed, the T series is non-stationary.
Over the 1965 -> period, id does appear to be largely (positive) trend-stationary, but I’m not sure if the variance changes. The trend could possibly be extended back to around 1920 by discarding the 1940-ish outlier.
However, 1880 – 1920 does appear to be stationary (adjusted) or (negative) trend-stationary.
The overall result is a concave up curve.
Unless TempLS’s theta is following that curve, C is going to vary depending on the time period under investigation. I just don’t see how the least squares algorithm as expressed follows the curve rather than providing a linear regression with a slope tangent to the curve..
“Unless TempLS’s theta is following that curve”
ϑ (t) is following that curve. Again, it is just least squares fitting of a statistical model.
.
In the TempLS version,
xsmy=Lsm+Gmy+εsmy
Here L is a local offset for station month (like a climate normal), G is a global temperature function (the desired result) and ε an error term. Suffices (station, month, year) show what each item depends on. G+ε is the anomaly.
It’s the same variables as BEST in the same order, but discretized. s means by station, m by month, and y by year. L has N*12 parameters, for N stations, and G has Y*12, for Y years. These are jointly adjusted to minimise ε. Sounds hard, but the iterative sequence I gave above takes advantage of the weak coupling.
Thanks.
I really should learn R – got the general gist of what you’re doing, but need to know the language better to get into more detail.
I won’t have my usual rant about allowing scientists within 10 miles of a programming language or you’d expect me to rework it “proper” 🙂
You’ve pretty much nailed it. Stokes reply “Estimating and subtracting ϑ makes it as stationary as you can get it.” is pretty revealing. In other words you make the data up that you need through manipulation.
Remember, for linear regression to mean anything physically there has to be a functional relationship between the independent and dependent variables. Temperature is not a function of time, it is a multivariate relationship with a whole host of independent variables.
This is why it doesn’t matter how you manipulate the temperature data, a linear regression line will never predict pauses in warming, step changes at El Nino times, or even climate changes. It’s really nothing more than mathematical masturbation based on the presumption that time drives temperature.
You know the old saying that “it’s good enough for government work”? It’s good enough for climate science work.
Annual averages don’t remove seasonality. Annual averages don’t remove changing variance. Annual averages don’t remove autocorrelations. Converting annual averages to anomalies do not help. All the arithmetic gyrations do is hide things.
Normalizing two data streams means changing their values, in most cases to a common mean.
Problem 1 – they are both time series covering different sequences in time. This is what we’ve been trying to tell you is wrong. You are trying to create a “long record” by modifying measurement values. What you end up with are not measurements. Normalization might be used when both streams cover the same time sequence, but that isn’t the case here.
Problem 2 – if done properly, the resulting stream is not representative of either stream. All you’ve done is found a way to create “long record” and in the process you will lose any value in the original measurements.
Problem 3 – there is no way to directly retrieve the original two data streams without knowing the parameters used in the conversion and those are never shown. Can you provide them for any “corrected” stream so we can determine the original data?
Problem 4 – each stream being normalized has a variance. Tell us how the normalization algorithm determines a combined variance (and subsequent uncertainty). In most cases the streams are ASSUMED to be Gaussian which temperatures are not.
You are basically calling accurate measurements of a given microclimate to be biased. These temperature measurements are not qualitative assessments to be normalized. They are physical data that should be used as recorded. If that means trending a “long record” in segments, that is how it should be done.
How would you combine the two records, then?The thing we care about is retained – the change over time.Yes, NASA has a data tool that lets you see every single station record in its raw and adjusted versions.The variance is the variance. You combine the station records and then determine the variance of the result. No assumptions are made about the variance of the individual records.
No. I’m saying that perfectly accurate measurements of the local microclimate can introduce spurious non-climate trends if these kinds of disparities are not accounted for. We are measuring a field in which the climate isn’t the only variable changing through time – the composition of the instrument network itself is transient. This is what scientists are accounting for with adjustments.
Ok, show us your method and results. If it’s different, it might be worth having a discussion about which approach is better. I don’t think it will be different.
“How would you combine the two records, then?”
You don’t combine the two records! The temperatures from the left side of the break point contribute to the combined data set from the left side of the break point and the temperatures from the right side of the break point contribute to the combined data set on the right side of the break point.
You could conceptually have every station contributing to the data set on the left side of a break point disappear from reality and an exact same number of new stations appear in reality on the right side of the break point and have zero reason to try and adjust individual stations to create “long records” by adjusting the data on the left/right side of the break point to somehow “match”.
You just take the data you have and analyze it. There’s no reason for anything but raw data to appear anywhere in any data set. The only substantial reason given for adjusting the data is that someone thinks it is somehow *WRONG* so they can justify *GUESSING* at how it should be adjusted!
That is not the proper method for conflating two random variables. That is taught in every stats class .
Var (X ± Y) = Var X + Var Y.
The variance is not calculated after combining the terms. It is indictativeof your lack of knowledge about metrology. Variances are uncertainty². Uncertainties are added using RSS (Root Sum Square).
It is no wonder measurement uncertainty gets thrown away at each average that is done. You just accurately described what climate science does. You need to start quoting some metrology texts if you are going to profess how uncertainty is evaluated.
Normalizing changes the mean which is a measurement. You are still avoiding exactly WHY you want to do this when you know a change has been made to a measuring system or to the microclimate being measured.
Give us a measurement model that incorporates how this works to provide a more accurate measurement.
If we are following your method of splitting each record at a breakpoint into two new segments, we need to ensure we are normalizing the before/after segments of the record, otherwise we will be skewing the trend in the exact same way it would have been skewed had we not addressed the breakpoint at all.
“ otherwise we will be skewing the trend”
And exactly how does this “skewing” happen? What is being skewed? The individual long term records that are meaningless when the individual temperatures from multiple stations are combined into an integrated whole?
A station record can’t merely represent the spot(s) where the station is sitting – it needs to represent all the space in between itself and neighboring stations. You don’t want the record (or multiple records if you split them) to retain trends that relate to the specific station history itself. A station move resulting in a temperature discontinuity is a real feature of that station’s history, but it is not a real feature of the climate in the area surrounding the station. So the trend for this region calculated using this station record will be skewed, because it contains both a climate component and a not-climate component.
This “skewing” will happen if you don’t adjust for the discontinuity by whatever method you think is best.
The largest element of this “skew” comes from station records with differing lengths (or missing data), so provided you at least use anomalies, you’ll remove most of the issue without any adjustments at all. The effect of the bias adjustments is pretty small at the global level.
“A station record can’t merely represent the spot(s) where the station is sitting – it needs to represent all the space in between itself and neighboring stations.”
Malarky! The climate is set by the conditions where the station is sitting!
That’s why the climate in Las Vegas is different than the climate in Miami even when they are at the same temperature. You have NEVER addressed that fact even after being asked about it multiple times!
It also shows that you have NO idea of temperature gradients. The temperature in Colorado Springs in NO WAY represents all the space between there and the temperature at Pikes Peak.
“You don’t want the record (or multiple records if you split them) to retain trends that relate to the specific station history itself.”
OF COURSE YOU WANT THE TREND TO RELATE TO THE SPECIFIC STATION HISTORY! What in Pete’s name are you thinking? People around that station experience the microclimate that the station is measuring – including any climate changes or trends. It is *that* which should be represented in any conglomeration of data.
“A station move resulting in a temperature discontinuity is a real feature of that station’s history, but it is not a real feature of the climate in the area surrounding the station.”
I asked you if you are GOD. You never answered. You apparently think you are. How else do *YOU* know that a stations history doesn’t represent the *real* feature of the climate surrounding the station?
“So the trend for this region calculated using this station record will be skewed, because it contains both a climate component and a not-climate component.”
And as I keep pointing out and you continue to ignore, you can’t know the values of either the climate component or the not-climate component so you can’t adjust anything. Unless you *are* GOD, of course. Not-climate components with a systematic uncertainty are not amenable to identifying that systematic uncertainty using statistical analysis. Which, again, you refuse to address. Because you are GOD and *can* know the unknown.
The climate is not set by things happening to the station itself, like being moved from one location to another, or having its instrumentation swapped out.
If I have a station sitting in a grassy meadow, and I move it to a rooftop at an airport, would you say, “the climate in this region has changed” when viewing the resulting discontinuity?
“The climate is not set by things happening to the station itself, like being moved from one location to another, or having its instrumentation swapped out.”
Of course the “climate” can be changed in moving a station from one location to another! If the humidity in one location is greater than in the other then the *climate* will be different. You’ll have different dew points, different freeze dates, etc for each location.
The station measures the climate INSIDE ITS ENCLOSURE. If swapping out a station measuring device includes changing the enclosure, even by sweeping out insect detritus, then the climate being measured has changed!
You have been reduced to just throwing out garbage hoping something will stick. As the old adage says: STOP DIGGING, you are only getting in deeper.
“If I have a station sitting in a grassy meadow, and I move it to a rooftop at an airport, would you say, “the climate in this region has changed” when viewing the resulting discontinuity?”
Of course it has!
And you can’t judge that change in climate based on the temperature readings of the instrument! All kinds of things will change, wind, humidity, the temp gradient between the instrument and the ground, etc.
And you won’t be able to identify the magnitude of any of that based on statistical analysis of the temperature readings!
Two climates being different is not the same as one climate changing. Do you see this distinction? If no, I can simplify it further for you. I am struggling to believe you are engaging in good faith, here.
This is so crazy! What do you think Bill J and Stokes have been discussing? Station moves, housing changes, instrument changes all can cause a discontinuity.
As I’ve told you, different and calibrated measuring devices CAN give different readings. It is a reason the GUM requires these for repeatable measurements.
Climate is the regional average weather, it is not the type of housing a thermometer is in. Changing the housing does not change the climate. But it can impart a non-climatic trend into the station record.
We don’t have a historical station network where no instrumentation changes or station moves have occurred, so there is little point in wishful thinking.
Identifying the break point doesn’t make you God. Thinking that you know the appropriate adjustment to make the data before and after the breakpoint equal is what makes you GOD.
The whole excuse of needing long records is a joke. It implies that you are combining individual station TRENDS into a data set so you need individual long records to create long-term individual TRENDS.
That is NOT what is being done.
Individual station TEMPERATURES are being combined into a data set, be it local/regional/global, in order to determine the trend of the combined individual TEMPERATURES.
In that data set of TEMPERATURES, it simply doesn’t matter if at time t_x, one station stops contributing TEMPERATURES to the data set and another starts contributing TEMPERATURES to the data set. So it’s a fools errand to try and adjust the TEMPERATURES from one station to match the TEMPERATURES from another station.
As Hubbard and Lin showed, an individual station’s data must be adjusted on an individual station basis and not adjusted based on a regional base. That applies to station moves, station change-outs, etc. Each one is a different station and should only be adjusted based on individual characteristics of the different station.
It profoundly matters, unless you’re using anomalies.
At worst, you will encounter a breakpoint in the average of different stations. Breakpoints are not a problem. They are indicative of REAL PHYSICAL change in the readings that were made.
Changing data so you can get a continuous trend with no breaks is a statistical decision to modify data to achieve a desired statistical result.
From: https://statisticsbyjim.com/hypothesis-testing/p-hacking/
(bold by me)
You need to recognize that fudging trends to appear continuous IS NOT objectively displaying what actually occurred. As such, the uncertainty is increased every time you do this.
A real, physical change that is related to the station history, not to the local or regional climate. If the observer started recording temperatures in Fahrenheit instead of Celsius, you could similarly complain that converting the units is data tampering – the change was a real physical change – but it’s obvious that leaving the disparity unaddressed will produce erroneous outcomes for any analysis using the dataset.
You don’t even know enough to understand how the temperature data is collected and recorded!
E.g. for ASOS stations the temperatures are recorded in Fahrenheit (rounded to the units digit) and then converted to Celsius!
The data winds up with more decimal places than the original during the conversion!
You need to exercise good faith in trying to address the argument actually being made instead of looking for an “angle” to attack.
If the observer started recording temperatures in Celsius instead of Fahrenheit, you could similarly complain that converting the units is data tampering – the change was a real physical change – but it’s obvious that leaving the disparity unaddressed will produce erroneous outcomes for any analysis using the dataset.
Malarky! If I have temperature data from 10, 100, 1000 or more stations and just ONE of them stops contributing to the data set – AND it changes the trend of the combined data set then those temperatures should be considered outliers of stupendous proportion and should have been dropped anyway! And that applies to the anomalies from that station as well!
And as I have pointed out, anomalies are garbage unless they are weighted to take care of the different variances associated with the anomalies. And I can’t find where climate science ever even calculates variances let alone use them to generate weighting factors!
A single station dropping out doesn’t matter much to the global trend, it might matter to a regional trend. Many stations dropping out/coming online can definitely matter to the global trend. Especially if they’re in regions of the world that are data sparse to begin with.
It’s odd that you’ve simultaneously adopted the stance of insisting that data quality is paramount while exclaiming that data quality isn’t a big deal and we can just ignore it.
“it might matter to a regional trend.”
Why? If as you have claimed all temperatures are just part of a common “climate” gradient then removing one set of the data shouldn’t change the trend at all.
“Many stations dropping out/coming online can definitely matter to the global trend. “
You are using the argumentative fallacy of Equivocation. You are trying to move to arguing that spatial sampling is being affected. Stations dropping and adding BECAUSE OF SITE MOVES, the primary cause of “break points”, shouldn’t affect the TEMPERATURE trends at either the local, regional, or global scope! There should be no change to the spatial sampling. It might affect the trends of the anomaly associated with that station location but, as I’ve pointed out, if that change is so large that it overwhelms the uncertainty of the anomaly then that station data, both left and right of the break point, should be dropped anyway for being an outlier.
“It’s odd that you’ve simultaneously adopted the stance of insisting that data quality is paramount while exclaiming that data quality isn’t a big deal and we can just ignore it.”
Total and utter bullshite. The data quality is affected by massaging it based on subjective guesses, not by refusing to “adjust” the data.
Because there are regions where there aren’t a lot of stations.
That is not what I’ve claimed. I’ve claimed that the station records contain both a shared (global) and local signal. The local signal contains components that are the result of changes to the observing instruments, not to the thing being observed.
Anomalies are calculated for each month. They are calculated by SUBTRACTING THE MEANS OF TWO RANDOM VARIABLES. When that is done the variances ADD. Var(X-Y) = Var(X) + Var(Y). There can be no argument about this, it is standard basic statistics.
NIST TN 1900 shows the standard deviation for a monthly average of Tmax as 4.1°C which makes the variance “σ²” = 16.8 °C. Even if the variance of the monthly baseline is zero you end up with an anomaly with an expanded combined uncertainty of the mean of 1.8°C.
This should be propagated into further calculations.
Then why is there a change over time?
Because the USCRN is warming faster – the stations in the network have larger warmer trends on average than the full nClimDIV bias-adjusted network.
So the adjustments are made to make it look like temperatures are warming more slowly than they actually are?
I thought you conspiratorially-minded types thought the opposite. Maybe I don’t get rabbit holes.
“Curiouser and curioser, said Alice…”.
No, The graph elsewhere clearly shows that ClimDiv “adjusted” data started higher at the beginning and that they have honed their adjustment parameters to make it match better.
Apart from the 2016 and 2023 El Nino effects, there is no warming
Which is another way of saying that adjustments are making things cooler, not warmer.
What an upside-down world you ‘skeptics’ live in!
Does the code match the outcome in ALL cases? 😉
Try it!
You are the one proclaiming it. Why not try it yourself? All cases. 😉
I have no doubt that Menne and Williams are doing it correctly. If you want to claim they aren’t, then try to provide some backing.
Typical reversal of burden. First, whether you have no doubt is absolutely irrelevant. Second, it is incumbent upon Menne and Williams to prove they are doing it correctly. The onus is on no one else to prove they aren’t.
The onus is always on the claimant. You’re claiming their method is wrong, so the onus is on you to demonstrate that. Menne and Williams provide compelling evidence of the efficacy of their methods in their published work. You’ll need to reference that work in your rebuttal.
Actually, no, I don’t.
Obviously it’s your choice whether you do it or not, it’s just the minimum barrier you need to clear for your objection to be taken seriously. No one is forcing you to be taken seriously.
That is so much hogwash it isn’t funny.
The claim here is being made by Menne and Williams. It is incumbent upon them and/or supporters to show the scientific basis for changing data.
And, before you respond that they are not changing temperature readings, yes they are! You can’t say I am going to change the mean of a set of data without implicitly admitting that the elements that create that mean are also being changed. It is a numbers game with no scientific evidence whatsoever.
Menne and Williams provide the scientific basis for their claims in their research papers, this fulfills their obligation to substantiate. Phil R is claiming they are wrong. Phil R carries the burden of proof for Phil R’s claims. Phil R must actually rebut the scientific basis of Menne and Williams in order for to meet this burden. Phil R has refused. We can summarily dismiss Phil R as an unserious person.
This is childishly simple stuff.
No, they provide no scientific evidence of measurement bias for making changes in the data.
Here is one one their comments.
There is no indication that any investigation was done to physically verify these estimates created accurate values. Everything is done from behind a desk by manipulating numbers. That isn’t physical science.
This is describing a process of generating temperature normals to facilitate comparison between USCRN and USHCN, described in Sun and Peterson, 2005. You need to follow the references to see what was done and why.
They have tried to remove urban bias from urban stations..
Got better at it over time.
You have missed a couple of things in your answer.
“Temperature normals” – this is out and out modifying data for a non-scientific purpose. It is creating metrics and should not claim to be be measured temperatures.
It is not, it’s calculating a value to use for normalization. Normalization is a standard statistical procedure whose use is not remotely controversial in any branch of science.
This is a word salad that says nothing.
Normalization is done on the whole series of data for a several reasons. One is to scale differing magnitudes of values to comparable values. Standardization can be used to match data where both series have a μ = 0 and σ= 1.
None of these have any useful features for knowing when to modify data.
Quit cherry picking stuff from the internet.
Yes. Here it’s being done to facilitate comparison between USCRN and USHCN.
No, its Fake Data fraud, which you help promulgate.
The onus is on the initial claimant, not the person who challenges the claim. If there is a good reason for doubting an assertion, that should be sufficient for the initial claimant to be required to respond and demonstrate that the concern is invalid. Then there is the logical issue of it being difficult to prove a negative.
That isn’t how rational discourse works. If I make I claim, I must provide my supporting evidence. You either accept my supporting evidence or you challenge it. If you challenge it, you are obliged to substantiate your challenge.
If I say Einstein is wrong, and you tell me to prove it, I don’t get to say, “no, you have to prove that Einstein was right.” I have to address all of Einstein’s work that I believe is in error and demonstrate why my claims have substance. Otherwise you can simply dismiss me with a handwave.
You’ve never participated in formal debate have you?
Here is an appropriate reference.
https://essays.williamwcollins.com/2024/09/on-debates-understanding-burden-of.html
Yes, thank you for that quote, which so eloquently articulates the fallacious argument Phil R was engaging in. Also from this essay:
Glad we seem to be in agreement.
No agreement because as the original claimant you have provided no evidence or proof showing the claim is true.
Your evidence is nothing more than opinion – that is, someone has a result and in my opinion it is true.
As pointed out, no description of measurement bias has ever been given. Consequently, no evaluation can really be made concerning measurement bias.
Your essay makes no mention of “original” claimant. It says anyone making a claim is responsible for substantiating it. Menne and Williams unquestionably do provide evidence to support the claims that they make. You are saying they are wrong, the onus is now on you to demonstrate that.
They are not doing it correctly. We must first examine what the term bias means. The GUM says:
Basically “bias” is a systematic error in measurement observations. As such, it is a constant for a given measurement system. One correction need be applied to ALL readings from that device.
Multiple corrections to the same reading means you are not correcting measurement bias. Not correcting ALL readings from a given device for bias means you are not correcting for bias but for something else.
Different measurement systems give different readings. That is a fact of life. The readings from two different devices that use different measurement systems will probably be different. Does that mean that one or the other is “biased” and should have corrections applied? No, it does not. They can both be as accurate as the resolution will allow but be different.
How then does one decide which device should have corrections applied? The older one or the newer one? How does one evaluate a measurement bias decades later?
Hubbard & Lin found in 2002 that regional adjustments cannot be done and this is what “homogenization” is. The problem is that the microclimates are different for each station and this affects the readings.
I do not understand why climate science ignores this physical fact laid out over 20 years sgo.
Climate Science has nothing to do with Physical Science, that’s why.
Nick, the commenter asked you a question. If you don’t want to answer, just don’t answer. Why be a smartarse?
That’s about as stupid as claiming that adding CO2 to air makes the air hotter! Are you that stupid? I hope not.
I’m sure Menne et al are using the homogenisation process to get exactly the result they want to get.
Nick,
Is that a Yes or a No answer to an important question? Geoff S
Can’t its missing in action and last time I did try it had errors.
Nick.
Your link – “We apologize, but the page or resource for which you were searching does not exist.”
No doubt adjusting it?
Here is the actual gz file
More clickbait? Fool me once, shame on you. Fool me twice, shame on me.
A gz file? Really?
A “gz” file, pronounced “jizz”?
It makes sense, I suppose.
gzip, actually
Dear Nick,
The code is in python. It is well documented within the code, but you need access to the Bureau of Meteorology database to make it work (good luck with that!) It is also impossible to get hold of original site instrument files to check that site summary metadata is accurate.
We know for example that since the 1990s most former 230-litre Stevenson screens have been replaced by 60-litre ones, and that the change has resulted in an increase in temperature extremes. However, metadata routinely ignores when the changeover occurred at individual sites.
Having worked on exposing biases and failings in the BoM’s homogenisation methods for more than a decade, problems underlying the methodology are easy to miss. Also, a sound, unified approach to examining the homogenisation issue is lacking. BomWatch protocols have been used for investigating numerous Australian datasets (both raw and homogenised), to my knowledge they have not been used elsewhere.
Even if metadata is available and used to specify changepoints in advance, metadata is often wrong. Thus, if going down this path, there is a need to verify that metadata is ‘accurate’. Metadata may specify station changes that did not happen, or alternatively ignore others that were known.
Changes at Cairns and Townsville are good examples of where station changes were ignored to imply their effect on data was due to the climate. (see: https://www.bomwatch.com.au/climate-of-gbr-cairns/; and https://www.bomwatch.com.au/climate-of-gbr-townsville/).
Alternatively, don’t rely on metadata at all – use other techniques to detect changepoints and use metadata post hoc, as a check. (see: https://www.bomwatch.com.au/data-homogenisation/marble-bar-hottest-place-in-australia/).
It is also important to eliminate the effect of deterministic covariables. This ensures that objectively determined changepoints are attributable to impacts on the variable of interest, not episodic changes in the covariable.
While not questioned by the referenced paper, using comparator data to detect and adjust target site data has no statistical or scientific merit. This is particularly the case using comparators whose first-differenced monthly means are highly correlated with those of the target and are therefore likely to embed parallel faults. Furthermore, in Australia’s case, no attempt is made to test-for or adjust inhomogeneities in comparator data. Thus, the assumption that by adding more comparators, the synthetic series is not likely to be biased, is a fallacy.
While I agree that homogenisation is a dog’s breakfast, it is not for the reasons given by the Authors.
Yours sincerely,
Dr Bill Johnston
http://www.bomwatch.com.au
Like Dr Johnston, I also have studied Australian temperature homogenisation for decades now. Conclusions are relatable with US experience reported here.
The inarguable conclusion we reach is similar. While historic daily temperatures observations are about all we have, they are unfit for reconstructions of plausible past temperatures because of widespread lack of essential metadata.
While I agree with mathematician Nick Stokes that it is possible to create mathematical ways to make good data, we disagree about the meaning of “good”. Nick seems to mean numerically good, while I clearly mean scientifically good.
Homogenisation here involves subjective guesswork. Statistical parameters are then applied to make the numbers look good, but this is clearly in the realm of lipstick on a pig.
I know of no authority to apply statistical procedures to help estimate the uncertainty of guesses.
Scientists concerned with the need for a valid temperature history would be taking different paths, such as comparing past frost observations with thermometer observations for drift and trying to replicate old equipment like screens and thermometers to compare with the electronics of today.
Also, a little humility would help. When Dr Pat Frank reports on LIG thermometer errors, when Dr Johnston reports on rainfall factors and temperatures, proper Scientists do not ignore such findings. They use Science to quantify, validate and correct.
Geoff S
I agree that there seems to be a lot of inconsistency in the way homogenisation is applied to data.
When there is a change in location or type of recording equipment then surely the old and new systems should be run in parallel for at least a year to be able to compare differences in different seasons.
If a guess is made as to conditions in remote areas many miles from official stations, could not a land based version of the Argos buoys be dropped by aircraft in the region to give an actual reading rather than rely on guesswork? Maybe an updated version of the weather forecaster’s radiosonde balloons could do the job?
This still doesn’t address the fact that neither station may have a measurement bias that is considered to be systematic error. It can simply be that different measurement systems will give entirely accurate yet different readings.
‘surely the old and new systems should be run in parallel for at least a year to be able to compare differences in different seasons.”
Won’t help. You can’t know what the reading differences would have been 10 years ago because you don’t know the calibration drift of the old station over those ten years. All you can know is the differences in readings over that year of parallel running. You can’t apply that difference to readings in the past.
The *only* valid way to handle this is to end one set of data and start a new one. In the scheme of “global temperature” losing a few data sets every year and adding a few new data sets won’t change the overall statistical descriptors very much when you are looking at thousands of data points. The whole meme of “we need long records” is typical climate science garbage. If the ended data sets significantly alter the statistical descriptors of the total data set then they are probably “outliers” anyway and should be removed from the overall data set to begin with!
Stokes is by no stretch of the imagination a proper scientist..
That much is obvious what was he a lab technician or a bad mathematician .. I can’t work it out.
Not disagreeing with you or agreeing with Nick, but I do agree with Geoff, Nick is a mathematician. Although math may be the language of science, just because one is good at math doesn’t make one a scientist.
Stokes certainly seems ignorant of the physical realities of measurement and associated uncertainties. He really does believe that a meteorological temperature expressed to three decimal places is meaningful.
It has been my personal experience that mathematicians, unlike engineers and physicists, live in a world where measurements are typically treated as being exact. It was rare that my professors ever paid attention to uncertainty, other than for the one chapter in the calculus text addressing error. It was then promptly forgotten.
I’ll bet no one can show a statistical textbook that deals with uncertain quantities in data and how that affects means, deviations, sampling and the sample means.
That’s where metrology comes in.
Statistics coves the statistical properties, not the measurement properties. The measurement properties can to some extent be analysed statistically – there are a number of learned treatises doing just that.
You know better than this oc! Standard deviation of measurement results is certainly a STATISTICAL property of the measurement data. Look up Type A measurement uncertainty.
That’s why I said “to some extent”.
Metrology utilises statistical analysis, but the underlying field of statistics doesn’t cover the complexities of measurement uncertainties.
Of course the basic field of statistics covers the complexities of measurement uncertainties. From the use of the student-T distribution to the use of Gauss’ formulas to develop how to propagate uncertainties (i.e. root-sum-square) to chi-square tests of measurement distribution it’s all basic statistics. Bevington says in Chapter 4.4: “If we can be fairly confident of the type of parent distribution that describes the spread of the data points (e.g. Gaussian or Poisson distributions) then we can describe the parent distribution in detail and predict the outcome of future experiments from a statistical point of view”.
The issue is not that statistics don’t work the same for measurement data, the issue is that mathematicians, statisticians, and climate scientists just assume that the stated value of a measurement is 100% accurate because all uncertainty cancels.
It’s hard to explain but Bevington goes into some detail on it. If you have a multiplicity of data points and create a histogram for the frequencies of the values then each bin in the histogram will have its own distribution and mean. This distribution is not the same as the uncertainty σ associated with the spread of individual measurements about their mean μ. It is the standard deviation of of the bin values that is the measurement uncertainty of each measurement. It’s from this that Bevington derives the chi-square test.
It’s all basic statistics. But it is complicated to do. That’s why metrology is considered to be an advanced area of expertise. Most mathematicians, statisticians, and climate scientists simply can’t resist the temptation to “simplify” things by assuming stated values are 100% accurate.
It’s t’other way around. The complexities of measurement uncertainties utilise statistics.
Metrology is largely a specialist field of applied statistics, with some other factors involved just for luck (as per your final paragraph)
Econometrics is another specialised sub-field.
Is that Bevington and Robinson, or another work?
Aye, there’s the rub.
“It’s t’other way around. The complexities of measurement uncertainties utilise statistics.”
That’s like saying that basic statistics doesn’t cover experimental physics data. Quantum physics is basically nothing but probability and statistics! Metrology is the same. The real issue is that most metrology texts don’t cover the entire scope of statistics because they *assume* standard simple distributions in most examples, i.e. Gaussian, Poisson, rectangular, etc distributions. In the real world skewed and multi-modal distributions are quite common. E.g. northern hemisphere and southern hemisphere temperatures and anomalies are are multi-modal, their average, medians, and standard distributions don’t really describe what is happening across the globe. It’s like saying the average height of a population consisting of Shetland ponies and Arabians describe the distribution of the heights of “horses”. That “average” really doesn’t tell you much about reality at all.
“Metrology is largely a specialist field of applied statistics”
I see it as a “super”-field of statistics and not a sub-field. It utilizes a wider scope of statistical methods and practices than usual in most disciplines.
“Is that Bevington and Robinson, or another work?”
Yep, my copy is Edition 3.
Your point about “assuming” simple standard deviations has at it core, the assumption of measuring the same measurand multiple times, which may involve multiple measurements of different pieces of the measurand.
The real fun begins when you enter the field of measuring different things. Then you are into the world of quality control and control charts. That is something one can use to monitor changes and how different things react. That allows one to use means and standard deviations of averages of different things to know if something has changed.
Climate science is totally out of their league in analyzing what is happening.
That’s pretty much what I thought I said.
Stats courses and texts don’t cover quantum physics, either, but quantum physics makes extensive use of statistics.
It’s like Kamala’s Venn diagrams. There is quite a lot of overlap, but metrology isn’t entirely within statistics or vice versa.
Thanks. That’s the one I found online.
That should exercise a few old neural pathways. That’s supposed to be good for staving off dementia.
Excellent.
Actually statistics have little to do with the measurements themselves or analyzing measurements using statistics.
Measurements of a certain measurand will form a probability distribution. That distribution is the important thing in metrology.
Statistical parameters are defined in the GUM to provide internationally accepted descriptions of the probability distributions that can occur. They can be uniform, poision, normal, triangular, many others, and even unique.
If you study the GUM and metrology textbooks, identifying the appropriate probability distribution to use is more important than the statistical calculations. From those distributions calculating a standard deviation isn’t much work.
The uncertainty intervals, i.e., the standard deviations, provide internationally accepted descriptions so everyone knows what is being portrayed.
That’s where the mathematicians fall down. They can do the calculations, but don’t know when they are applicable.
I agree.
No, he is a smart person who sucks at making credible counter to most articles and comments criticizing him.
Exactly! Great post.
That Pat Frank paper on LiG Metrology certainly woke me up!
https://www.mdpi.com/1424-8220/23/13/5976
When was the last time that you saw a rigorous propagation-of-error analysis of descriptive statistics? It seems that the unstated assumption in statistics is that all measurements are exact, despite the results often having several significant figures to the right of the decimal point. It seems that statisticians pay little heed to properly rounding off results.
Not only pay little heed but don’t have a clue about why it is important to convey information.
That is an implicit assumption.
Once you get into measurements and measurement uncertainties, you’re into metrology. That utilises statistics, but isn’t entirely a subfield of applied statistics.
It is valid to use additional figures to the same order of magnitude as the sample or population size. For a population of 1,000 members and a total of 2, 128 “things”, a mean of 2.128 is perfectly valid.
Summary statistics summarise the data set. It would be counter-productive to discard information.
NOTE: a mean alone is of very limited use. At least sample size and s.d. should be provided. Median and mode should also be provided if the differ from the mean to any extent.
It’s not proper to round off summary statistics. If measurement uncertainty is involved, it seems more correct to provide an uncertainty range (e.g. (1.28 +/- 0.05)mm, or (1.28 +/- 0.03)mm)
You are using an example of counts instead of measurements. The statistical descriptors serve different purposes for each. For measurements the statistical descriptors are meant to convey information about expected values a subsequent measurement will produce. Giving a statistical descriptor to a higher resolution than the measurements justify is misleading for subsequent measurement results. Thus the use of significant digits.
Even for counts you can get misleading results by basing the “average” resolution on sample size. If you have two different people counting the number of cows herded down a chute to a holding corral what should the resolution of the two different counts give for the average per time period?
Yep. That’s the usual complaint about stats courses and textbooks, but the object of the exercise is to focus on the core of the analysis.
Also yep. Most statistical analysis has to do with agglomerations of classes of “things” such as family size.
That’s for repeated measurements of the same thing, or measurements of things which are supposed to be the same within a specified tolerance.
Which is more misleading?
1,000 “things” measure 0.100m [+/- 0.0005m], and another 100 things measure 0.106m [+/- 0.0005m].
Adding 3 sig figs, the mean is 0.100545m [+/- 0.0003m]
Rounding, the mean is 0.101m [+/- 0.0003m].
Is it valid to add 1 sig fig to make the average 1.005, or do you rely on the range and s.d.?
That’s getting into counting error 🙂
Cows are expensive!
You don’t care about averages, you care about an accurate count.
If you don’t get the number you expected, you either made a mistake, some are missing, or you have stragglers.
In any case, you count them again, more carefully.
If you’re still short, you spend way too much time looking for them.
If you have extra and they don’t have a brand or ear tags, you ask the neighbours if they’re missing any. If they aren’t, you invite them around for a barbecue 🙂
“1,000 “things” measure 0.100m [+/- 0.0005m], and another 100 things measure 0.106m [+/- 0.0005m].”
It depends on the resolution of the measuring device.
“Adding 3 sig figs, the mean is 0.100545m [+/- 0.0003m]
Rounding, the mean is 0.101m [+/- 0.0003m].”
If you publish 0.100545m as the mean when you can only measure out to the thousandths digit then you are fooling people into thinking you have better instrumentation than you actually have.
“you invite them around for a barbecue”
Yep.
PS: How do you count them again when they’ve already been loaded on the freight cars and shipped off?
It’s implicitly the number of sig figs of the measurement.
Adding 3 sig figs for the average of measurements of 1,100 items is a bit pointless given the uncertainty range. On the other hand, rounding does lose information.
btw, I’m assuming above that the Type B uncertainty is the correct approach for the mean.
Cattle are easier to count than sheep, but if I send a truckload of either off to the saleyard I’m going to make damned sure I have the correct count going onto the truck, the bill of lading, and confirmation coming off the truck at the other end.
Not only that, but that the saleyard’s invoice matches what I sent – not only count but weight range.
On Ronald Reagan’s “trust but verify” basis, it’s useful to be at the sale as well.
Sig figs are not determined by the number of items you measure. They are determined solely by the resolution of the measuring device. If I measure 1,100 items to the units digit, I simply have no idea what the decimal digits are for any of the items.
Rounding to the significant digits is used to preserve the information that is known. Going further is creating information that is not known.
The map is not the territory.
A mean maps to the distribution of the items, but does not necessarily correspond to any individual item.
The mode must correspond to an item, and the median either corresponds to an item or is the midpoint between 2 items.
Using the 3 Ms and the s.d. gives a good first pass as to the shape of the underlying distribution of the items.
If the mean lies between 2 items, that information should be preserved.
The information is known. It is information about the distribution, and has been derived from all of the individual items examined. The more items, the more information they provide.
Actually, the example I used is more akin to the Arabians and Shetland Ponies, so binning the measurements is more useful.
The mean, median, mode and s.d. do tell us that the distribution is skewed, so it warrants investigation.
“On the other hand, rounding does lose information.”
Actually it doesn’t when it involves measurements with uncertainty. Again, the expression of a measurement is an estimated value coupled with an uncertainty interval. The uncertainty interval is meant to include the effects of rounding, especially when the rounding is to the resolution of the measuring device. The information you are “losing” is information that you can’t possibly know anyway!
“On Ronald Reagan’s “trust but verify” basis, it’s useful to be at the sale as well.”
I never actually was involved in the sale of livestock but I was in the sale of hay and straw. And yes, you wanted to be at the sale to make sure the quality and quantity of the product was properly considered.
Rounding doesn’t lose information about the measurements of the items, but it does lose information about the distribution of the items which were measured.
If the summary statistics apply to 1,000 items, the distribution of the measurements of those 1,000 items is information which is not available from any individual item.
That’s possibly one of the reasons for stats courses and texts working with counts. It eliminates extraneous factors and ensures that the focus is on the distribution.
You are making the same mistake statisticians, mathematicians, and climate scientists make. The distribution is not defined by the average of the stated values but by the stated values modulated by the measurement uncertainty of the component values. That means the average is a fuzzy number, it is not exact. It *can’t* be exact because that requires assuming the stated values are 100% accurate.
Since the magnitude of the measurement uncertainty is based on the resolution of the measurement device and it modulates the value of the average the average of measurements must be based on the measurement resolution as well.
It’s why I keep telling the statisticians on here trying to support how climate science handles the temperature data, the uncertainty of the average is based on the propagated measurement uncertainty of the component data. The uncertainty of the average is not the average uncertainty and it is not the sampling error. Nor is the average just based on the stated values of the measurements.
Statisticians and climate scientists want to believe that uncertainty doesn’t always grow. They put their faith in the unjustified memes that averaging increases resolution and decreases uncertainty. It doesn’t. *NOTHING* decreases measurement uncertainty, especially when you are working with different measurands and different measurement devices. The measurement uncertainty *ALWAYS* grows. The more uncertain components you add the greater the uncertainty becomes.
It’s like none of them has ever navigated across unmarked territory using just a compass. It doesn’t matter how many readings you make, the uncertainty of where you actually are GROWS, it doesn’t get less. Unless you can “calibrate” your position against an external landmark you can wind up far from your target at the end of journey because of the uncertainties involved i the navigation. No amount of “averaging” your intermediate readings of the compass will decrease the uncertainty of where you are.
“That’s possibly one of the reasons for stats courses and texts working with counts. It eliminates extraneous factors and ensures that the focus is on the distribution.”
Uncertainty is not an extraneous factor. It’s why the sampling error, i.e. the so-called standard deviation of the sample means can’t be based solely on the standard deviation divided by the number of samples. The mean of each sample is based on data elements that are “stated value +/- measurement uncertainty”. That results in the mean of the sample also being given as “stated value +/- propagated measurement uncertainty”. When you have multiple samples, each given as “stated value +/- measurement uncertainty” then the mean of the sample means also becomes “stated value +/- propagated measurement uncertainty.
Students are taught based on 100% accurate counts because it is simpler, not because uncertainty is an extraneous factor. The problem is that over time too many teachers don’t understand uncertainty beacuse they were never taught it and so can’t teach it.
The mean, median, mode and s.d. are defined by the distribution, not the other way around. They are purely summary statistics.
If there are 4 measurements, 1.495, 1.495, 1.494, 1.494, does rounding the mean to 1.495 lose valid information? Does a mean of 1.4945 add spurious information? The median is clearly 1.4945 by definition. Does that add spurious information as well?
Rounding to 1.495:
Using resolution bounds, the mean may lie somewhere between 1.4945 and 1.4955.
Using the Type B uncertainty, the mean may lie within 1.4947 and 1.4953.
Using the additional digit (1.4945):
Using resolution bounds, the mean may lie somewhere between 1.4940 and 1.4950.
Using the Type B uncertainty, the mean may lie within 1.4942 and 1.4948.
What the additional sig fig does is say that the mean sits in the 1.494 to 1.495 bin.
That’s where it is if you use the resolution bounds of the measurements.
Lower bound:
1.4945, 1.4945, 1.4935, 1.4935, mean 1.4940
Upper bound:
1.4955, 1.4955, 1.4945, 1.4845, mean 1.4950
Yes, the mean is fuzzy, as are the median and mode. That’s why the mean (and often the median) need to have explicit uncertainty bounds shown. 1.4945 +/- 0.0003* is not the same thing as 1.4945 [+/- 0.00005]
[*] Using the Type B uncertainty which all you blokes tell me is the correct figure to use.
I’m not doing a very good job of explaining this I guess. Here’s another try.
You want to be able to draw the distribution with a fine point Micron pen with a .3mm width line. The problem is that a distribution, even one that is a count, has uncertainty and must be drawn using a 5yr old’s large crayon.
When you do a histogram each bin has a distribution of its own. If it is a count most people assume a Poisson distribution. It’s a result of doing sampling. Each sample will show different sized counts in each bin thus generating a distribution for that bin.
Now look at the histogram around the central tendency. Pick five of the bins closest to what appears to be the central tendency. Because of the distribution of counts in each of bins it’s impossible to identify which bin is THE actual average. Since the counts in each bin can take on different values based on the distribution you can’t tell if bin1 always has the highest count or if it is bin3 or maybe bin5. The distribution in each of those bins makes the overall distribution “fuzzy”.
With measurements those bins are populated with values determined by the resolution of the instrument. You can’t tell anything about the individual bins because of the resolution limits so you can’t tell anything about overall distribution because of the resolution limits imposed on each of the bins.
If you will this is what is behind the standard error of a distribution. It’s a measure of the uncertainty of the average.
“If there are 4 measurements, 1.495, 1.495, 1.494, 1.494, does rounding the mean to 1.495 lose valid information?”
Actually it doesn’t. Because you can’t know what you can’t know. Trying to add digits is assuming you *can* know what you can *not* know.
“Does a mean of 1.4945 add spurious information?”
Yes. Again, because you can’t know what you don’t know.
That’s a nice analogy.
The larger the sample size, the finer the nib.
I think I see what you’re getting at. It depends on the size of the bin relative to the sample size, or number of samples if you prefer. The larger the sample, or the number of samples, the more closely they will approach the population central tendency.
That’s why I expanded on it. A raw 1.4945 has an implied uncertainty of +/- 0.00005, 1.4945 +/- 0.003 says it’s somewhere between 1.4942 and 1.4948*.
[*] I’m not sure that the Type B uncertainty is more appropriate than the resolution range, but all you blokes seem to insist it is.
“[*] I’m not sure that the Type B uncertainty is more appropriate than the resolution range, but all you blokes seem to insist it is.”
That’s because the resolution doesn’t tell you the uncertainty. My frequency counter can display down to the at least a tenth of a Hz. But it has a uncertainty that overwhelms the resolution because of the accuracy, both short-term and long-term, of the time base used to collect the data samples. It’s barely possible to state a frequency in the units digit accurately. You can use an external time base, say one that is synced using GPS, to increase the accuracy but when used in a standalone measurement you should use the Type B uncertainty specified by the manufacturer plus any other uncertainties in your uncertainty budget.
The freq counter works well to compare two frequencies so if you have a stable standard you are trying to use to calibrate an oscillator the freq counter is an appropriate tool.
Alright, it’s a complicating factor.
Yep. It’s a combination of KISS and “only change 1 thing at a time”.
For purely count-based work, measurement uncertainty isn’t relevant. Statistical uncertainty is relevant. It’s a large part of the curriculum.
Anywhere measurements are involved, measurement uncertainty can certainly be a large factor.
Even purely count-based work which generates a distribution has uncertainty.
Take a six-sided die. Roll it 30 times. Will you get a perfect distribution with equal counts in each bin?
Uncertainty is uncertainty. There isn’t a distinction between metrology uncertainty and statistical uncertainty. They are one and the same. The contributing elements to the uncertainties may be different but uncertainty is still uncertainty.
That reminds me of the old Castrol “oils is oils” ads.
uncoitainty is uncoitainty, but the components which make up the total uncertainty are different and should be treated appropriately. The NASA guide covers this very nicely – see Chapter 5.
Roll it 90 times. 900 times, 9,000 times.
The more rolls, the closer the sample distribution approaches the population distribution.
If the counts in each bin aren’t approximately equal, there’s a good chance you don’t have a fair die.
“Uncertainty is uncertainty. “
I’m not sure about that.
Unless you have a way to conjure a way to read a device beyond its resolution or to know a true value just by eyeballing a measurand, then uncertainty is a real thing, and it grows.
Lighten up, Jim.
Nick told a Dad joke, so it proves he’s human.
All this argy-bargy seems like a diversion (purposeful or not) to avoid Nick having to answer whether he really believes that adding CO2 to air makes the air hotter!
John Tyndall noticed no temperature change at all, when he removed CO2 from an air sample, nor when he dried the sample by removing the water vapour. That was over a hundred years ago. I am unaware of any experiment contradicting Tyndall’s relevant observations, but if Nick can provide results from such experiments, I will offer a most fulsome apology.
I certainly won’t hold my breath waiting for the results of an experiment purporting to show that removing “greenhouse gases” from air results in the temperature dropping.
“John Tyndall noticed no temperature change at all, when he removed CO2 from an air sample, nor when he dried the sample by removing the water vapour. That was over a hundred years ago. I am unaware of any experiment contradicting Tyndall’s relevant observations,”
Crickets from the Climate Alarmist Peanut Gallery. This debunks their whole “CO2 is dangerous” claim, yet not a peep out of them. I think the reason for the silence is they have no rebuttal.
We are wasting TRILLIONS of dollars trying to reduce CO2 based on nothing.
Granted I have not checked a large number, but some have not noted things like ground cover changes, new buildings, tree growth changing both wind and shade, maintenance on screens, calibrations, etc.
This is beyond the fact that bias is in the eye of the beholder. Why are new temps not lowered/raised to match old data if long records is the underlying reason for making changes. The “bias” can easily be contained in the new devices rather than older devices.
“metadata”
Use of that term indicates I should worry that the data was adjusted before anyone started adjusting it.
-Like the now uber-popular “metastudy” of other studies (to relieve student researchers the time consuming slog of actual data collection and experiment replication). Metastudies relieve the newest researcher from the pressure of adulterating data to match expected results. So long as at least a few of the earlier students succumbed to the same pressure, then results will be guided toward the expected result even if most students played honest.
Metadata fo a weather station is a log of changes that might affect readings. Those logs are hard to find and worse, they don’t have a lot of detail.
When one relies on poorly-trained volunteers, they get what they pay for.
I had to look up “a dog’s breakfast”.
dog’s breakfastnoun
Synonyms of dog’s breakfast
chiefly British
: a confused mess or mixture
(I’m an American. 😎
It isn’t a term that I commonly use. It is, however, quite descriptive if one understands that a dog will eat almost anything, including feces from other animals.
Which fact is obscured by reporting weekly/monthly/annual arithmetic means of daily mid-range values and referring to the conflated result as an “average.” The daily mid-range values do not have an associated variance (And are highly sensitive to unrepresentative outliers for either extreme.) Furthermore, Tmax and Tmin are not read at the same time every day, which was an issue with the satellite data as the orbits degraded, and with a cold/warm front moving through, could even be reversed between typical pre-dawn Tmin and usual late-afternoon Tmax. Yet, data manipulators claim high accuracy and precision for the conflation of mid-range temperatures and aggregated arithmetic means of those temperatures, which they routinely report as “Global Mean Temperature.” It is uncommon to see estimated uncertainty envelopes for climatology measurements, and when they are provided, they are typically only for +/-1 sigma, not two as in most other disciplines.
It seems that the sloppy handling of data and its subsequent processing is unique to ‘climate science'( and possibly biology).
+100
Mi-range temperatures are not representative of “climate”. Neither are minimums and maximums. The entire scope of “climate science” is based on a foundation that is not fit for purpose – temperature. If temperature were a useful metric for climate then you could use temperature to distinguish hardiness zones – but that doesn’t work.
The lack of providing the full range of statistical descriptors for the data analyses associated with tenp, including measurement uncertainties, only compounds the lack of meaning of a “global temperature”.
Think the code has been doge’d 🙂
Can’t see it but is it any better than a couple of years ago it had some horror coding errors?
“Available.” Not at our BoM, our temperature gatekeeper- they won’t give reasons! A bunch of liars. Record temperatures once observed and published have been deleted in the last decade. Tony Heller, Jennifer Marohasy, Jo Nova and Senators Malcolm Roberts & Gerrard Rennick Have frequently pointed that out.
Gerrard Rennick trying to get blood from of a stone at Senate inquiry.
**** MEDIA RELEASE: BOM refuses to answer on data records****
“LNP Senator for Queensland Gerard Rennick has called out the Bureau of Meteorology (BOM) during Senate Estimates today with key questions on their Surface Air Temperature dataset.
Despite repeated questions from Senator Rennick about the BOM’s confidence intervals on both its actual and adjusted records, the BOM refused to provide clear answers.
The BOM’s adjusted records, the ACORN-SAT dataset is a time series of daily temperature data from a network of 112 weather stations going back over 100 years in most cases.
Since 2011 the BOM has found it necessary to adjust Australia’s temperature record on two occasions. This process of ‘homogenisation’ is claimed to account for anomalies identified within the raw dataset.
The latest adjustments in the ACORN 2 dataset has increased warming from 0.8 degrees Celsius to 1.23 degrees Celsius from 1910 to 2016.
Senator Rennick asked if the BOM agreed that the homogenisation of Australia’s mean temperature data has effectively increased the rate of warming by more than 50% since 1910. The BOM failed to answer this question and instead sought to defend its adjustment methodology by referencing peer reviews over the past decade.
“If they can’t answer what should be a very straightforward question for them, then I think Australians should be concerned”, said Senator Rennick.
Senator Rennick then asked the BOM to state its accepted margin of error or tolerance level for the original (raw) dataset, ACORN 1 dataset and ACORN 2 dataset. Again the BOM failed to answer the question despite Senator Rennick asking the question more than three times.
The 2011 independent peer review report found that the BOM’s observation practices did not meet international best practice. Furthermore a key recommendation of the report was that the tolerance on temperature sensors be reduced significantly below the present +/-0.5 degrees which is significantly higher than the World Meteorological Organisations standard of +/-0.2 degrees. Another key recommendation of the review was that the BOM should specify statistical uncertainly values associated with calculating Australian national temperature trends.
Senator Rennick said the BOM’s failure to clarify or provide an update on these important recommendations is disappointing given the importance of measuring surface air temperature.
“I intend to submit additional questions on notice on this and related matters that I hope the Bureau of Meteorology will answer expeditiously and in full.”
ENDS.
They never did!
https://gerardrennick.com.au/media-release-bom-refuses-to-answer-on-data-records/
Tony Heller re: BoM’s blatant fraud.
Creating Records By Erasing The Pasthttps://realclimatescience.com/2022/01/creating-records-by-erasing-the-past/#gsc.tab=0
https://realclimatescience.com/2022/01/creating-records-by-erasing-the-past/#gsc.tab=0
Yes, code is available – but it dates from 2012, and is not production code as of October 25 2012 (the date of the code archive) but rather a code snapshot at an earlier date in 2012, and modified for “Benchmarking the performance of pairwise homogenization of surface temperatures in the United States” (https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2011JD016761) and possibly other studies. The snapshot was created at a date earlier in 2012, as indicated by a dated comment in one file, but all the archived files were given a single date when archiving.
52i in the URL provided above indicates that this is the 52i version of the Pairwise Homogenization Algorithm, which was replaced by the 52j version in June 2015. (See the status.txt file in /pub/data/ushcn/v2.5 on the same FTP server). The status.txt file for GHCN-M v3 at /pub/data/ghcn/v3 indicated that GHCN-M version 3.2.2 would be replaced by version 3.3.0 in early June, 2015. GHCN-M version 3 ceased on August 21 2019, GHCN-M version 4.0.0 became operational on October 26 2018. The status.txt files do not indicate any change from PHA version 52j after June 2015, although GHCN-M version was upgraded to version 4.0.1.
The code as distributedis for GHCN v3/USHCN v2.5 and is configured for Peter’s World, not GHCN or USHCN, but it comes with instructions for recompiling for GHCN or for USHCN.
Health warning: This is a compiled version for Peter’s World, based on code frozen in time on some date before October 25th 2012, and I suspect that the modifications for Peter’s World may not all be cleanly removed when it is recompiled for GHCN or USHCN following the instructions supplied. It reproduces the distributed results for Peter’s World exactly, but I have not been satisfied with the reproduction of GHCN results for any GHCN version earlier than the archive distribution date (I would expect one GHCN3 input and output data set earlier than the distribution date should be reproducible, but that other data sets, and in particular later sets, would not reproduce exact output. I have tested with archived data sets for all GHCN-M version 3 subversions)
I have run on both Ubuntu 16.04 and Ubuntu 20.04
20.04 exposes incorrect logical test syntax in BASH shell scripts, which was accepted in 16.04 – my 2004 UNIX book showed that integer comparison and string comparison already had different syntax at that time. So watch out and correct if necessary if using a recent Linux distribution. There is also a redundant test of an undefined variable which makes running the code unmodified a form of Russian Roulette. It barfed on the undefined variable on the first PC I ran it on, but found the undefined variable had an acceptable garbage value on the next two PCs. If the FORTRAN code does not run unmodified. you can either set the compiler switch to zero variable memory before running (unset in the distributed archive, and a practice I would normally not recommend as it disguises other undefined variables which need a value other than zero. But safe here if there is only the one undefined variable and zero is an acceptable value. I hope that this was simply an error introduced in the modified code and not already present in the production code.
Finally, I would remind commenters that our study examined only the PHA performance for a subset of European stations for which we were able to obtain metadata, and not USHCN, USCRN, ClimDiv or any other set of stations. Comparison “between adjusted ClimDiv and unadjusted USCRN” for example does not address our study. I would be happy to correspond with anyone who has actually run the archived code with GHCN-M data, not just with the archived Peter(Thorne)’s World data. leefor asked “You are the one proclaiming it. Why not try it yourself? All cases” – and I suspect the answer is just the Peter’s World case.
Yes, code is available – but it dates from 2012, and is not production code as of October 25 2012 (the date of the code archive) but rather a code snapshot at an earlier date in 2012, and modified for “Benchmarking the performance of pairwise homogenization of surface temperatures in the United States” (https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2011JD016761) and possibly other studies. The snapshot was created at a date earlier in 2012, as indicated by a dated comment in one file, but all the archived files were given a single date when archiving.
52i in the URL provided above indicates that this is the 52i version of the Pairwise Homogenization Algorithm, which was replaced by the 52j version in June 2015. (See the status.txt file in /pub/data/ushcn/v2.5 on the same FTP server). The status.txt file for GHCN-M v3 at /pub/data/ghcn/v3 indicated that GHCN-M version 3.2.2 would be replaced by version 3.3.0 in early June, 2015. GHCN-M version 3 ceased on August 21 2019, GHCN-M version 4.0.0 became operational on October 26 2018. The status.txt files do not indicate any change from PHA version 52j after June 2015, although GHCN-M version was upgraded to version 4.0.1.
The code as distributedis for GHCN v3/USHCN v2.5 and is configured for Peter’s World, not GHCN or USHCN, but it comes with instructions for recompiling for GHCN or for USHCN.
Health warning: This is a compiled version for Peter’s World, based on code frozen in time on some date before October 25th 2012, and I suspect that the modifications for Peter’s World may not all be cleanly removed when it is recompiled for GHCN or USHCN following the instructions supplied. It reproduces the distributed results for Peter’s World exactly, but I have not been satisfied with the reproduction of GHCN results for any GHCN version earlier than the archive distribution date (I would expect one GHCN3 input and output data set earlier than the distribution date should be reproducible, but that other data sets, and in particular later sets, would not reproduce exact output. I have tested with archived data sets for all GHCN-M version 3 subversions)
I have run on both Ubuntu 16.04 and Ubuntu 20.04
20.04 exposes incorrect logical test syntax in BASH shell scripts, which was accepted in 16.04 – my 2004 UNIX book showed that integer comparison and string comparison already had different syntax at that time. So watch out and correct if necessary if using a recent Linux distribution. There is also a redundant test of an undefined variable which makes running the code unmodified a form of Russian Roulette. It barfed on the undefined variable on the first PC I ran it on, but found the undefined variable had an acceptable garbage value on the next two PCs. If the FORTRAN code does not run unmodified. you can either set the compiler switch to zero variable memory before running (unset in the distributed archive, and a practice I would normally not recommend as it disguises other undefined variables which need a value other than zero. But safe here if there is only the one undefined variable and zero is an acceptable value. I hope that this was simply an error introduced in the modified code and not already present in the production code.
Finally, I would remind commenters that our study examined only the PHA performance for a subset of European stations for which we were able to obtain metadata, and not USHCN, USCRN, ClimDiv or any other set of stations. Comparison “between adjusted ClimDiv and unadjusted USCRN” for example does not address our study. I would be happy to correspond with anyone who has actually run the archived code with GHCN-M data, not just with the archived Peter(Thorne)’s World data. leefor asked “You are the one proclaiming it. Why not try it yourself? All cases” – and I suspect the answer is just the Peter’s World case.
Years ago I was able to find station plots comparing unadjusted and adjusted time series of temperature records. These were part of the ghcn v3. datasets. My study focused on surface stations that were rated by Watts et al. as highest quality. Two examples show how the differences arise.
Baker City Oregon:
It is instructive to see the effect of adjustments upon individual stations. A prime example is 350412 Baker City, Oregon.
Over 125 years GHCN v.3 unadjusted shows a trend of -0.0051 C/century. The adjusted data shows +1.48C/century. How does the difference arise? The coverage is about the same, though 7 years of data are dropped in the adjusted file. However, the values are systematically lowered in the adjusted version: Average annual temperature is +6C +/-2C for the adjusted file; +9.4C +/-1.7C unadjusted.
How then is a warming trend produced? In the distant past, prior to 1911, adjusted temperatures decade by decade are cooler by more than -2C each month. That adjustment changes to -1.8C 1912-1935, then changes to -2.2 for 1936 to 1943. The rate ranges from -1.2 to -1.5C 1944-1988, then changes to -1C. From 2002 onward, adjusted values are more than 1C higher than the unadjusted record.
Some apologists for the adjustments have stated that cooling is done as much as warming. Here it is demonstrated that by cooling selectively in the past, a warming trend can be created, even though the adjusted record ends up cooler on average over the 20th Century.
San Antonio Texas:
A different kind of example is provided by 417945 San Antonio, Texas. Here the unadjusted record had a complete 100% coverage, and the adjustments deleted 262 months of data, reducing the coverage to 83%. In addition, the past was cooled, adjustments ranging from -1.2C per month in 1885 gradually coming to -0.2C by 1970. These cooling adjustments were minor, only reducing the average annual temperature by 0.16C. Temperatures since 1997 were increased by about 0.5C each year. Due to deleted years of data along with recent increases, San Antonio went from an unadjusted trend of +0.30C/century to an adjusted trend of +0.92C/century, tripling the warming at that location.
Full analysis:
https://rclutz.com/2015/04/26/temperature-data-review-project-my-submission/
How are adjustments handled for widely spaced stations where a rain storm, or even a cold front, might miss one or more ‘near by’ stations that are used for homogenization?
It’s not just weather that needs to be accounted for. If one station is down in a river valley and the other isn’t there is a difference in microclimate and therefore a difference in temperatures and diurnal ranges. If one is on the east side of a mountain and the other on the west side you will see differences in temps and temp ranges. Homogenization involving the two stations will do nothing but give bad adjustment values and increased measurement uncertainty.
You simply can’t use San Diego temps and Ramona, CA temps together for any kind of homogenization.
Clyde, even more despicable is the erasing of data read at the station to replace it with numbers from somewhere else.
Time to overhaul NOAA!
Defund the NOAA. Isn’t 35 years of this claptrap enough?
As far as temperature goes, the ocean rules, the atmosphere drools…
.
“It will be remembered as the greatest mass delusion in the history of the world – that CO2, the life of plants, was considered for a time to be a deadly poison.” R Lindzen
“I can’t see either of these papers being in the next IPCC report. Kevin [Trenberth] and I will keep them out somehow, even if we have to redefine what the peer-review literature is!”
—Phil Jones, Director of the Climatic Research Unit, disclosed Climategate e-mail, July 8, 2004
.
— “We need to get some broad-based support, to capture the public’s imagination… So we have to offer up scary scenarios, make simplified, dramatic statements and make little mention of any doubts… Each of us has to decide what the right balance is between being effective and being honest.” – Prof. Stephen Schneider, Stanford Professor of Climatology, lead author of many IPCC reports.
“We need transparency, accountability, and scientific rigor in climate science. Until then, every NOAA temperature dataset should be taken with a grain of salt.”
It’s important to note that the study examined monthly GHCN records. The daily station records are not subjected to the PHA algorithm or the TOB (time of observation) adjustment.
An example of the use of these daily records is in Figures 2 and 3 of the most recent NOAA state climate summary report for Wyoming (link below). Read the caption for Figure 2 pasted farther below. NOAA very promptly sent me this list of the 655 GHCN long-term stations they used for the CONUS region (48 states) when I requested it. I have made graphs similar to theirs with updated data through 2024.
https://statesummaries.ncics.org/chapter/wy/
Here is the caption from Figure 2.
“Figure 2: Observed annual number of very hot days (maximum temperature of 95°F or higher) for Wyoming from 1950 to 2020. Dots show annual values. Bars show averages over 5-year periods (last bar is a 6-year average). The horizontal black line shows the long-term (entire period) average of 7.2 days. Values for the contiguous United States (CONUS) from 1900 to 2020 are included to provide a longer and larger context. Long-term stations dating back to 1900 were not available for Wyoming. The highest number of very hot days in Wyoming occurred during the 2000s and early 2010s. Sources: CISESS and NOAA NCEI. Data: GHCN-Daily from 35 (WY) and 655 (CONUS) long-term stations.”
So let’s not implicate the GHCN daily station records. I know of no indication they are being altered over time. Tony Heller uses the global daily station data for his work examining similar questions about the frequency of hot/cold daily values by station, by state, country, etc.
Thank you for listening.
Edit: here is a link to a folder with 6 plots I did recently using the list of 655 GHCN stations.
https://drive.google.com/drive/folders/1fssK0Rijse5t3mM5JX4BFD0rnH7Bp-qn?usp=sharing
“The daily station records are not subjected to the PHA algorithm or the TOB (time of observation) adjustment.”
Nor are the monthly unadjusted temperatures. These are placed in the same location, with the same format and coverage. But of course no-one here is interested in those.
“But of course no-one here is interested in those.”
Spoken to the Almighty again, Nick?
The TOBs adjustments are totally bogus anyway. !
Tony Heller showed that.. no difference in trend in real data.
“he TOBs adjustments are totally bogus anyway. !”
Yes, they are. Adjustments are only useful to Climate Change Liars.
He’s not interested either. He did the adjustments.
“He’s not interested either”.
Mind reading much? I’m interested in why you think you’re right, and everyone else is wrong.
You aren’t just pushing unsupported belief, are you?
That simply is not true. The ISSUE IS the difference between the two records and the reason why.
You need to address the accuracy of the “bias” adjustment from a scientific basis (that is, measurement uncertainty) versus a simple numeric comparison between devices ( that is, homogenization).
Measurement uncertainty *always* has two components, random variation and systematic bias. If you cannot separate the two then any “adjustment” for systematic bias is bound to have its own uncertainty. That uncertainty should ADD to the measurement uncertainty. Since systematic bias is not amenable to identification using statistical analysis the random variation can’t be separated out. Why climate science continues with the meme that all measurement uncertainty is random, Gaussian, and cancels is beyond me.
As is appropriate if a summary mid-range value is reported instead of an actual hourly arithmetic mean. However, it should be made clear when a ‘raw’ mid-range temperature is presented versus the arithmetic mean of many mid-range temperatures. Referring to a “Global Mean Temperature” implies more information than is justified for a summary statistic of mid-range temperatures.
Not only that, but the assumption that changing the mean of a random variable doesn’t affect the individual components is pure sophistry.
How do you change a mean, without admitting you are implicitly changing all the values in a probability distribution along with it?
For temperature, each of these files of daily station data includes only the values of Tmax, Tmin, and Tobs (temperature at the time of observation.)
Example https://www1.ncdc.noaa.gov/pub/data/ghcn/daily/hcn/USC00012813.dly
“Referring to a “Global Mean Temperature” implies more information than is justified for a summary statistic of mid-range temperatures.”
Indeed so.
one of the realities in life is that when your income relies on your outcome, the outcome will be directed to maximse the income. This applies in most walks of life, sometimes with a positive impact, but many times with a negative impact. In science and research, the impact is often negative.
“independent climate researchers led by Peter O’Neill, Ronan Connolly, Michael Connolly, and Willie Soon”
Familiar names, MDPI pay for play journal.
“The same temperature records were being adjusted differently on different days”
Yes. There is more than one set of adjustments that can remove the bias. Same with the next two points.
“Less than 20% of NOAA’s breakpoints corresponded to actual documented station changes,”
Yes. That is the point of the algorithm. The breakpoints are recognised statistically, then tested by comparison with neighbors.
But the main thing is, the algorithm gets it right. Here is the comparison between adjusted ClimDiv and unadjusted USCRN:
The other thing is that, in the end, it doesn’t make much difference. For all the grizzling here about adjustment, no-one wants to show what would happen if you used unadjusted data. The answer is, very little.
OK, I’ll bite –
if “in the end, it doesn’t make much difference”, why bother with adjustments at all?
Because you have to. You know that inhomogeneities exist. They might make a difference. Probaly they will cancel out, as they do. But you don’t know that until you have tried.
Thanks Nick.
So it’s like those people who insist on avoiding stepping on any cracks in the footpath “because you have to”?
(Disclosure – I’ve always avoided engaging with such people.
Looks like I now have to avoid engaging with temperature homogenisers too?)
Nick,
“Because you have to. You know that inhomogeneities exist.” Yes, a chaotic system has “inhomgeneities” (whatever they are).
You can’t deal with chaos, can you? Why not just accept observations as they are? Altering them to fit some preconceived ideas of what “everyone knows” they “should be” is just madness.
Oh Nick,
What happened to your logic?
Of course we know that heterogenieties exist.
We all know that they are places where adjustment should be applied.
…
But, nobody, not you nor me nor anyone else knows how much adjustment is correct for the majority of the cases. In many cases we do not even know the correct sign of an adjustment before it is made. Subjective choices (guesses) grow here.
Geoff S
“But, nobody, not you nor me nor anyone else knows how much adjustment is correct”
Maybe you don’t but others do. The adjustment is whatever is needed to remove the inhomogeneity, aided by the behaviour of nearby stations. I spelt out here the process for adjusting a much discussed inhomogeneity at Amberley. You can pin down well the time and amount of adjustment.
“Maybe you don’t but others do.”
Oh yes, the Almighty told you – he just forgot to mention their names, did he? The mystical “others” appear again.
There is nothing Nick Stokes cannot do.
Except make a compelling argument that isn’t just sophistry.
That like your understanding of generators is complete rubbish.
Even in the comments someone has pinned you for it
“You are adjusting the data using a whim and a fabrication.”
Your response is just laughable and completely unscientific because at the end of the day you aren’t a scientist.
So let me be blunt the Amberley reconstruction might as well be pulled from a ouija board, you presented that crap to me I would fire you on the spot.
Take a logic course FFS. Seems like every post is not an argument or is a logical fallacy. You’re stroking your own ego at best with these nonsense responses.
Can you pin down the systematic error causing the “bias”? Can you pin down the value of that uncertainty? Has that systematic error existed prior to the change? Does the systematic error continue for readings beyond the time when the change was made?
If you can not answer the above, you are simply doing a numeric adjustment to make data look more homogeneous. That is not scientific, it is a guess.
The unstated assumption is that the data should look homogeneous. It is also the unexamined assumption.
Climate science is just chockful of unstated and unjustified assumptions. Including that you can combine random variables having different variances (e.g. northern hemisphere temps with southern hemisphere temps) with no weighting necessary.
HA HA HA, you clearly don’t understand what SheroO1 is saying here since NO one can be sure if the adjustments are truly valid.
Nick,
You merely support what I wrote above, that you are aiming for better math while others like me are aiming for better science.
Example. Pat Frank published a detailed examination of drift in liquid-in-glass thermometers used decades ago. Size of error about 0.2 deg C, compare with guesses of global warming of 1 to 2 deg per century, so a significant error.
Nick, would you please describe how you would incorporate a correction for this error in a homogenisation exercise? So far as I can find, treatment of this error has been to avoid mention of it. Geoff S
As I posted in another reply, homogenization is nothing more than applying regional adjustments to individual stations. Hubbard and Lin found in 2002 that regional adjustments simply don’t work because of the differences in the microclimates of the stations used to come up with the regional adjustment value. Adjustments have to be done on a station-by-station basis and can’t be applied to past measurements but only to future ones because you don’t know the actual drift per time in the past due to station drift and/or microclimate drift (e.g.tree growth, etc).
The whole homogenization idea is garbage, all it can do is 1. spread measurement uncertainty around to other stations and 2. remove natural variation from the data (which you really don’t want to do).
Excellent comment!
If there is a drift then it will be picked up when the thermometers are replaced or re-calibrated, and corrected appropriately.
You can not judge drift by comparing measurements to a “replaced” device. That isn’t how it works. Recalibration can be used to identify drift if it is done regularly. The problem is that the correction is systematic and corrections must be applied at the time of making the reading.
Systematic errors are not resolvable by statistical analysis. It is a common factor that occurs with every reading therefore the only way to identify it is through calibration.
Comparing to a replacement device tells you nothing. As i have mentioned before, different devices will have different readings. This occurs quite frequently when devices with different measurement methods are used. That is why the GUM requires repeatable conditions use the SAME device.
You can’t know the value of the drift in the past. That makes adjusting any past value nothing more than a guess. Guesses ADD to the measurement uncertainty, they don’t reduce measurement uncertainty.
Dear Nick,
Here is what ACORN-SAT said about Amberley:
History
The site has been operating since August 1941. No significant moves are evident in documentation but the data indicate a substantial change of some kind at the site in or around 1980. An automatic weather station was installed on 3 July 1997. Manual observations continued under site number 040910 until September 1998.
My note: Small screen 8/7/1997
Here is what actually happened: https://www.bomwatch.com.au/data-quality/amberley-queensland-australia/
Aerial photographs and RAAF documents show the site moved at least twice when it was in the vicinity of the ATC tower. They also show the site did move around 1980, and that the BoM built a new office and radar facility on the eastern side of the base north of the Old Toowoomba Road guardhouse. They could not have done this without someone knowing and approving expenditure of public funds.
In the post you referenced (https://moyhu.blogspot.com/2014/08/adjusting-amberley-as-it-must-be.html) it is clear that you did not research the site. Knowing what I know now, I suspect the original RAAF Stevenson screen was beside the old (flat-topped) met/operations office.
All the best,
Bill Johnston
Bill
“They also show the site did move around 1980”
“it is clear that you did not research the site”
Yes, I did not. I worked in the conventional way just from T observations there and nearby. Here was my conclusion:
“So there you go. The change happened in August 1980, and it dropped temperatures artificially by 1.4°C.”
It can be done.
Lack of verification leaves your conclusion technically ‘unproved’.
In addition, the actual change in level (degC) is likely confounded with a covariable that you did not accounted for.
Note that I did not investigate Tmin, and also that ACORN-SAT made no adjustment for the effect of the move on Tmax.
Consequently, which gets to the guts of the issue, they blamed the change (which they claim they did not know about – but must have as it was they that moved the site), on the climate.
Cheers,
Bill
“Lack of verification leaves your conclusion technically ‘unproved’.”
But you verified it with photos
But you did not, which is the point!
What point? It is verified.
I showed how you can locate the adjustment time and value by time series analysis. You showed I got it right. Thank you.
Dear Nick,
Just to set the record straight …
Here is what they said in ACORN-SAT V1:
History
The site has been operating since August 1941. No significant moves are evident in documentation but the data indicate a substantial change of some kind at the site in or around 1980. An automatic weather station was installed on 3 July 1997. Manual observations continued under site number 040910
until September 1998.
Here for AcV2:
History
The site has been operating since August 1941. The site moved on 26 August 1980, several hundred metres northeast from its former location near the base buildings. After the site was automated in 1997, manual observations also continued under site number 40910 until September 1998.
(Did you spot the difference!)
Aerial photographs and other records showed the site actually moved in 1975. See Figures 5 & 6 in https://www.bomwatch.com.au/wp-content/uploads/2020/08/Are-AWS-any-good_Part1_FINAL-22August_prt.pdf – in fact why not read the whole report!
The overarching problem is that the the BoM have spent considerable time assuring Australian’s that they have carefully researched site histories, and therefore that adjustments made to data are unbiased. However, that claim is unequivocally wrong!
So how do you explain the 1980 step-change in Tmin, when the site actually moved in 1975, and the change in Tmax was 1991?
Anyway, I’m moving on to something else…
Cheers,
Bill
Wow! Now that is verification!! A reminder that my home-made homogenisation calculation in 2014 concluded
“The change happened in August 1980, and it dropped temperatures artificially by 1.4°C.”
Now you have Acorn V2 in 2018 discovering
“The site moved on 26 August 1980, several hundred metres northeast from its former location near the base buildings”
But somehow you think we’re all wrong, but I can’t see what your evidence is.
No nick,
I just rechecked my report, aerial photographs, and the most recent site summary metadata.
Site summery metadata does not mention the site moved at any time, which is not unusual.
A Queensland government aerial photo (QAP) taken 01 February 1974 (2,500′ ASL) shows the met-radio aids office being built. At that moment in time it did not have a roof.
Foundations for the radar were in place but no radar yet. No activity either at the eventual location of the met-enclosure (Latitude -27.6297, Longitude 152.7111).
A QAP taken 19 August 1975 (10,000′ AMGL (above mean ground level)) shows the office was complete, the radar was in-place, a concrete path had been laid to the met-enclosure, which appeared to contain instruments, and a trench (water, power and telemetry) leading to the MO had been infilled.
As the site was up and working before 1980, ACORN-SAT metadata is wrong again. You have thus mislead yourself into believing something that is not true.
I feel excited for you.
All the best,
Bill
Bill,
The ACORN sttement is specific – the station moved on that day, so I believe they would have a record to match it. Your photos just show what could be preparatory activity in 1975. You might expect the station move to follow shortly, but doesn’t always work like that.
Besides, I pinned it at August 1980 back in 2014 before any of that 🙂
ACORN-SAT is full of irregularities (think Cairns, Townsville, Rockhampton, Sydney Observatory, Marble Bar, Bourke, … the list is long.). The most detailed metadata should be site summary metadata, which is also mostly deficient.
That you allegedly found a change in August 1980 cannot be verified using metadata that is suspect. Furthermore, the Amberley site moved at least two, possibly three times in the vicinity of the operations area, before it moved in 1975. (Did you read the report?)
To get to where it is now from where it was at the Tower, the site had to move 1.21 km directly north. So there was either another site they have not mentioned of they are simply making it up!
And by the way, the BoM produced a report on Amberley (24 September 2014), where they said in relation to Tmin:
2. 1 January 1980—breakpoint detected by statistical methods.• Night-time temperatures started to appear much cooler relative to surrounding stations.• No accessible documentation for Amberley in 1980, but a breakpoint of this size would normallybe associated with a site move.• Min T adjusted by -1.28 °C; no detectable impact on Max T so no adjustments made.
However, they built an office with a balloon tracking radar in 1974, and clearly moved to the new site in 1975 … funds came out of a budget, commonwealth department would have been involved …. Did the office stay empty for 5 years? Were all the APs photoshopped; do you want more?
Why would they do than and then claim in AcV2 (but not AcV.1) they moved it in 1980?? Did the dog eat Blair Trewin’s homework?
It is up to you to explain what happened – it is your job after all.
Cheers,
Bill
Face it Nick,
Your ‘analysis’ could have been responsible for AvV2 deciding the site moved when it did not.
All the best,
Bill
Well, Bill, you are full of conspiracies. How about proving that your photo was actually taken in 1975?
Never seen an aerial photograph Nick?
On the attached, the old Toowoomba road comes in from town on the the right (east). The guardhouse defines the Base boundary.The met office buildings are immediately north of the guardhouse, the balloon tracking radar is the furthest north. Further up the track, beyond the slight bend a concrete path leads off on the left to the instrument enclosure. The filled trench leading to the office can also be seen.
The ribbon along the bottom of the photograph tells you the QAP number, run number, print number, direction north, date, time (read the clock), and altitude (AMGL means above mean ground level).
Someone it telling porkies Nick, and it is not the photograph.
Happy Saturday,
Dr Bill Johnston
http://www.bomwatch.com.au
Bill,
OK, here is your 1975 evidence in detail:
Just a path. No enclosure, no instruments. They came later. How much later? The BoM says 26 August 1980. You say that can’t be. But no evidence.
Remember something that happened in 1975? There was a change of government. Something of an upheaval, in fact. And then followed austerity. Big cuts in government spending. Completing a shift of a met station may well have dropped off the list for a while.
BS Nick,
The new buildings etc. would have been budgeted for in 1972 or earlier – contracts would have to be let by the Department of Public Works etc. you know how it works. My earliest Amberley RAAF AP is 1941, then 1943, 1952, 1955, ’56, ’59, … 1972 and so on.
Here is the picture showing the roofless buildings and the slab for the radar in 1974.
I don’t know precisely when the the BoM took over from the RAAF met-section, but they would have had to provide continuous support for flight operations at arguably, the RAAF’s most important base. The new pre-1968 control tower came to occupy the area where the old met-enclosure was, but a met enclosure could have remained in-place until later, or they may have used a remote analogue instrument wired back to the met office, which was probably located in the tower.
At best these photographs are 300 dpi; a far cry from the 1200 DPI that I used to scan at the National Library. (There appear to be only four images of Amberley available at the NLA.) While it is possible to make-out a Stevenson screen in a B/W photo it is more difficult against a green background in color.
The 1974 picture possibly shows some preliminary works happening at the new enclosure site.
All the best,
Bill
Something else happened around that time Nick.
F-111 entered service with the RAAF at Amberley in 1973. The Canberras had also returned from ‘Nam. This resulted in major changes and reorganisation at the base.
A building plan for the new control tower (which I think was completed in 1968) showed the met-section occupied a few offices in the tower complex, which also had a met observation platform.
I went through all the 1974 APs of the Base (2,500’ ASL), and it seems there was a met enclosure directly in front of the tower complex, which makes sense,although due to shadow, I can’t actually make out a screen. While there is a watered/mown lawn area north of the complex and there is a bit of water lying about, no other shapes look like a fenced enclosure.
I suspect that either the RAAF closed their met-section in 1974, or if the BoM were there, they were booted to the other side of the runway. There are some files in the National Archives with interesting titles, but aside from that and the lack of useful metadata from the BoM (and the photographs), there is too little information that supports that the site moved in 1980. Note that I detected a Tmax step-change in 1991 (the change likely happened the previous year).
All the best,
Bill Johnston
Damn, you’ve done some detailed work here!
Perhaps I should give you an example of what you are doing Nick.
I am studying behaviour of tigers. I study Tiger A recording all it’s behaviour in it’s range and I study Tiger B recording all it’s behaviour in it’s range.
So I now see there is a range between Tiger A and Tiger B range so I homogenize a set of data for Tiger C based on Tiger A and Tiger B.
According to you there is nothing wrong with that.
There is nothing wrong with that theoretically (I think) as long as you then go and try to confirm the actual existence of Tiger C, and the behaviour of Tiger C is consistent with your predictions. If no Tiger C, theory is wrong.
Nick thinks math creates reality. If your homogenization of Ta and Tb creates Tc, then Tc exists and needs no confirmation.
You are exactly correct and in his Blog comments he was pinned on the issue and tried to make out it was scientifically normal.
You have no idea about homogenization.
ROFL that is your comeback to a basic science problem .. good argument. You aren’t should not be allowed anywhere near science you hack.
That is not why homogenization is done. The main reason is to make a continuous dataset, as if changes were not made either to location, equipment or both.
The additional reason is to fill in numbers where data does not exist. A lot of stations have been shut down over time, and homogenization is being used to make an artificial series for such stations.
If global warming is truly global one only needs a single station to show it, as long as it is well maintained, equipped to detect changes relevant to climate, continuously operated, never moved and with a constant and continuous dataset showing changes to its surroundings that make it possible to understand non-climate effects.
So I’ll support Mr. here. If the end result is no or insignificant change, why do it? “Because you have to” is the wrong answer, Nick.
“The main reason is to make a continuous dataset, as if changes were not made either to location, equipment or both.”
Yes. Who said otherwise?
“The additional reason is to fill in numbers where data does not exist. A lot of stations have been shut down over time, and homogenization is being used to make an artificial series for such stations.”
No, they do not do that. In fact, adjusted sets are usually smaller than the original, for each station. The reason is that to apply an adjustment, you have to have enough data before and after to detect a possible break point.
Nick, you’re just being silly now. Saying “The reason is that to apply an adjustment, you have to have enough data before and after to detect a possible break point.”, is just meaningless word salad.
Try harder.
Detect a break point????
You just broke the dam thing in what you are doing the only thing you can detect is artifacts.
Is a break point the same as a tipping point?
Curious minds want to know.
It is a mini-tipping point.
The whole purpose is to detect artefacts. Something that caused a jump/drop in temperature.
Detection is only part of the issue. If you don’t know what caused the artifact then how do you determine an appropriate adjustment? You can’t determine an appropriate adjustment purely using statistical analysis. What if the Corp of Engineers created an impoundment upwind of the measurement station which changed the humidity at the station? That would cause a change in the measured temperatures at that station but that change would accurately portray the current environment at the station. Why would you want to “adjust” the readings if they accurately portray the current environment?
Because climate practitioners are unable to envision anything except straight lines versus time.
AndersV: “The additional reason is to fill in numbers where data does not exist. A lot of stations have been shut down over time, and homogenization is being used to make an artificial series for such stations.”
Nick: No, they do not do that. In fact,
Example of them doing that (start at 5:54 in the video):
https://youtu.be/cF16lDtSVrU?t=354
Valiant defense of Fake Data.
This means your definition of bias is purely based upon a numeric difference and nothing to do whatsoever with measurement bias.
That is not being scientific at all. You have no scientific explanation for a systematic error. You are simply doing a numbers is numbers exercise.
And in so doing, hiding what might be a real physical change of importance.
Any calculated warming trend in USCRN/ClimDiv comes from the bulge at the 2016 El Nino and the 2023 El Nino.
Otherwise, there is nothing.
Its why they homogenise ClimDiv to match USCRN
Evaluating the impact of U.S. Historical Climatology Network homogenization using the U.S. Climate Reference Network
That is not a statement from NOAA, and says nothing about ClimDiv (which didn’t yet exist). It is a paper written by external scientists in 2015 who looked to see if the adjustments to USHCN brought the series closer to USCRN. Well, if the adjustments are working, so they should, since USCRN are higher quality stations. But it doesn’t mean that they adjusted to fit. They adjusted to get it right.
You are ignoring the fact that different measurement systems in different microclimates WILL GIVE DIFFERENT READINGS.
That by itself does not allow the conclusion that there is a measurement bias. Each of two devices could be 100% accurate so any corrections to match them is just playing with numbers and increases the uncertainty interval. In the end, nothing will be accomplished.
I’ve asked before in this thread, what is your scientific definition of bias?
You should have this definition prior to making any decision.
He won’t answer.
It is what I have been trying to get them to admit. If you don’t have a detailed scientific description of what the problem is, then you will never have a scientific explanation of what how to solve it. The more I see folks flounder, the more I believe they are just attempting arithmetic solutions to obtain a number that fits their purpose.
Exactly! Like trying to contort and twist the mean formula into a “measurement model”.
“There is more than one set of adjustments that can remove the bias.” But do they? Why the need for more than one set of adjustments? Perhaps it is like BoM and they need “expert judgment”. 😉
http://www.bom.gov.au/climate/change/acorn-sat/documents/2015_TAF_report.pdf
You don’t need more than one set. You just need one that will do. If there is more than one, applying the algorithm can well switch between adequate soutions.
Ah adequate. So what are these “more than one set”? Which one is correct?
Nick, you wrote “You don’t need more than one set. You just need one that will do.” You commune regularly with the Almighty, do you?
You don’t think that you might be being just a wee bit presumptious? Did you get a note from your mother giving you the power to tell other people what they need to do?
Sorry about the laughter, but I don’t think any rational people want to dance to your odd tune.
“You don’t need more than one set”
There is only one set in the whole world that is remotely representative of that country’s surface temperatures, and that is USCRN (Climdiv is adjusted to match it, as shown elsewhere, so they are, to all intent, the same set.)
Any other data from anywhere on the surface is too tainted by urban warming, site adjustment and bad placement etc, and the products created from that surface data are pure FICTION unless they have a robust pristine system to check against… And they don’t.
No amount of fake random and erratic homogenisation and agenda-driven “adjustment” is going to fix that bad data.. probably just make it even more meaningless.
“The answer is, very little.”
Compared to “very big”, one assumes. In other words, you really have no idea, but want to sound as though you do. Have you found a reproducible experiment which demonstrates that adding CO2 to air makes the air hotter?
Or are you one of the gullible people who just believe that adding a non-reactive gas to another gas causes a temperature increase?
You are spinning stories Nick.
Off-the-shelf homogeneous datasets including the 112 station ACORN-SAT dataset may not be homogeneous, and it is the role of the researcher to provide evidence they are sound and fit for purpose.
All the best,
Dr Bill Johnston
http://www.bomwatch.com.au
The usual noise: “they don’t change anything”.
Then why bother with them?
There is more than one set of adjustments that can remove the bias.
You need to define your use of the term “bias”. Is it considered to be systematic error in measurement observations? If so, how is it determined?
It appears to be more of a numeric change to accomplish something that has nothing to do with measurement accuracy. Is it done to achieve a numerical result instead? Something like creating a “long records” that appears to be made up from cohesive data?
The magnitude of the changes in adjustment is also relevant. And remember that PHA for USHCN uses station metadata whereas PHA for GHCN-M does not. Of course there can be breakpoints recognised statistically which do not correspond to actual documented station changes, but if, for example, parallel measurements at old and new locations after a station location change show a measurable change in temperature we would hope that PHA would detect such a change at at least approximately the correct date. Even before starting our study we had already seen both abrupt changes of adjusted values for the same past date which seemed unreasonably large, and failure to reflect change where parallel measurements were available for a station change. Ideally we would have parallel measurements for all location or instrumentation changes as part of available station metadata, but the only station metadata used by GHCN-M outside the US is latitude and longitude.
“we would hope that PHA would detect such a change at at least approximately the correct date”
It is not very dependent on date. The basic test is that a set of readings on one side of a point in time is statistically different from those on the other side. It is probably a large set of readings. If you shift that test point by a few months, say, it is likely to still be statistically different. So you probably choose the point with max test statistic, but that is a weak constraint. Then you check against neighbors, which will probably sort out the magnitude of change, but doesn’t help constrain the timing .
In the Amberley study I did a further step, in effect minimising second derivative. That worked very well there, giving a very precise time estimate, but won’t always, so I don’t think they do it.
The reason GHCN-M doesn’t use other metadata is that it mostly doesn’t exist in any usable form. End of story. That is why they developed a method that doesn’t depend on it.
Tell me about it!
The US Historical Observing Metadata Repository (HOMR) (Historical Observing Metadata Repository (HOMR) | National Centers for Environmental Information) and the French Informations sur les stations (métadonnées) – METEO FRANCE (https://donneespubliques.meteofrance.fr) are probably the most useful metadata sources for individual countries which I’ve used, although both, in their different ways, are a bit of a pain to use when metadata for many stations is needed. A standardized metadata format is very much needed.
And I pointed to “Benchmarking the performance of pairwise homogenization of surface temperatures in the United States” (https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2011JD016761) in my first comment here above. And remembering that PHA for USHCN uses station metadata whereas PHA for GHCN-M does not, performance in the US is not necessarily a guide to performance in the rest of the world. You might look at our study as a start to addressing benchmarking the performance of pairwise homogenization without available metadata. If I’ve missed a Menne and Williams paper, or work by others, addressing this, do please point it out.
Release of up to date code is also long overdue. A push (more “vigour” than that) might seem needed.
NOAA’s data altering factory is a disgrace to meteorological data integrity — as my video illustrates … https://www.youtube.com/watch?v=cF16lDtSVrU
“Livestock contributions are place between 85-95 Tg (1 Tg = 1 trillion metric tons) each year”
90 trillion tonnes. With a ten year life that means 900 trillion tonnes is Methane. That’s about 38% of the atmosphere….
Livestock only contribute methane? Oh dear.
Oh and each year contributions fall out of the equation. 😉
I’m soooooo over plant food dooming and into methane dooming now-
Discovery of a Massive Methane Leak Escaping Into the Atmosphere—And It’s Worse Than We Feared
Oh no! It’s worse than we thought!
Haven’t I heard that before . . . somewhere . . . about something or other?
Best to ignore it this time.
Which is statistical nonsense, their adjustments can only increase uncertainy.
Their suspect adjustments only increase suspicions.
I will never understand why climate practitioners cannot grasp this simple, inescapable fact.
Good work by O Neil et al. NOAA is not suited for what they are supposed to be doing. One of two things has to happen. All Climate and weather related functions must be removed from this agency and moved to another agency. No employees from NOAA can go to work at the other agency and must be fired. If it is determined that NOAA must do this work all current employees involved must be fired and we will start over with new employees. All new employees will understand that dishonest work, or fudged work or lazy work or half assed work will not be tolerated. Those who don’t produce quality work will immediately be fired.
The research paper reaches no definite conclusion to inspire my following the convoluted analysis to understand it fully. That was my motivation to write the above points.
“Last, but not least, the very slowly increasing temperature anomaly that putatively exists since about 1750 is arbitrarily assigned to anthropogenic causes. “
And of course, even “climate scientists” won’t actually say that they believe that adding CO2 to air makes the air hotter, because that would be laughable.
People are just supposed to assume that “anthropogenic causes” means CO2 added to air!
What nonsense!
A value which a temperature anomaly must inherit—subtraction of a baseline cannot reduce measurement uncertainty.
WAKE ME when the DOGE boys have fed (directed GROK 3 to read all the supposed adjustment ‘code’) all this nonsense into GROK 3 (the latest Musk/Tesla AI ‘engine’) and asked for a summary …
I think that’s what they’ve been doing with the numbers from all these other agencies (to root out the fraud, corruption and financial abuse), rather than calling out the guys with the green eye shades and working them (the bookkeepers, the accountants) into the wee hours of weekday mornings …
Grok is certainly a means to find and establish the chicanery. That is an ‘app’ to which it is well adapted.
The negative: that government agencies mess up like that is appalling.
The positive: Conolly, Soon & al. were able to publish that, which is a step in the right direction.
I hope that many up to now biased scientific journals have realized that “there’s a new sheriff
in town” and realize, that if they want to get government money, playing the climate alarm card
will not be productive for them, and that they stop blocking publications that don’t fit their agenda.
The written, historic, regional temperature records put the lie to the instrument-era Hockey Stick global “temperature” chart The regional charts from all around the world show the temperatures were just as warm in the recent past as they are today. The bogus instrument-era Hockey Stick chart erases the warmth of the recent past and says today is the warmest period in human history.
The UAH satellite chart puts the lie to NOAA and NASA temperature adjustments after 1998. NOAA and NASA claim there were 10 years between the year 1998 and 2015, that they describe as the “hottest year ever!”. The UAH satellite chart shows that NO year after 1998, was warmer than 1998, until the calendar reached 2016. So NOAA and NASA mannipulated the data so they could scare people every year by lying to them and telling them they were living in the “hottest year evah!” year after year after year.
The Temperature Data Mannipulators ARE the problem. They have always been the problem, because they distort reality in an effort to demonize CO2 and keep the Big Bucks rolling in.
Without them, and their temperature lies, ordinary people wouldn’t think there was a problem and wouldn’t be spending TRILLIONS of dollars trying to fix this non-existent problem.
See if you can see any years between 1998 and 2015 that could be described as the “hottest year evah!” using the UAH chart below. As you can see, NO years on the UAH chart were warmer than 1998, until 2016 was reached.
NOAA and NASA are Temperature Data Liars.
UAH has been warming at the same rate as both NASA and NOAA over the past 20-years, +0.29C/dec.
Only because the El Ninos affect the atmosphere more than the surface.
Now, show us the near zero trend periods in GISS from 1980-1997 and from 2001 to 2015.
“Adjusted data” is an oxymoron. Once you “adjust” data, it is no longer “data” at all, just glorified guesswork.
The (actual) raw data should not be “adjusted.” Pepper it with all the footnotes, error bars, etc. needed to reveal its deficiencies, but don’t substitute guesses for instrument readings and call it “data,” because that is simply dishonest.
Of course they will never do as I suggest, because that would reveal the truth – that the data is crap, and is not fit for the purpose of measuring changes to the “climate.”
It becomes Fake Data.
The changes are not scientific at all. Only calibrations can show a measurement bias and determine the corrective adjustments to the current measurement.
Anything else is numbers is numbers gyrations.
The “adjusted data” are the output of a poorly-described model that has not been vetted. It is mostly smoothing and interpolation that can remove potential information. Often times variance is more informative than some variety of average.
And averaging the averages throws most of the variance into the round file.
Climate science assumes you can take two random variables, find the mean of the sum and assume that standard deviations just disappear through cancelation.
They forget that the conflation process is actually adding (or subtracting) the two means and then adding the variances. Dividing by “n” just moves the resulting probability distribution on the x-axis. You do not divide the probabilities by “n” as that drastically modifies the relationship between the data.
Temperature measurements, once “adjusted” are no longer DATA, but merely estimates of what the data might have been. Collections of such estimates are not DATASETS, but merely estimate sets.
homogenised vs raw
Two well matched FAKE graphs.. Very funny !
That’s not “raw” (unmodified) data.
Unmodified historic data does not show a Hockey Stick “hotter and hotter” temperature profile.
You are comparing one set of adjusted data with another set.
corected vs raw
That’s not raw data, either.
As usual, there are no error bars. Thus, it is difficult to asses whether the apparent differences are statistically significant, which depends on the 2-sigma overlap between the two data sets.
sea surface temperatures at record high for most of 2024. 2000metre ocean heat content rising. No effect from UHI or badly located sensors!
Ha, the usual laughable Ocean Heat Graph yet again!
The uncertainties, just in temperature measurement alone, hugely overwhelm any trend.
Not only that, but they had basically no data for HUGE areas of the ocean before 2005 and ARGO.
So the graph is pure FANTASY, probably from a model of some sort.
“So the graph is pure FANTASY,”
Yes.
He needs to show the Ocean Cool Graph, too.
Not all the water in the oceans is heating up. You assume too much.
I calculate the total temperature change of the World’s oceans from 1960 to 2024 to be only about 0.11 degree Celsius.
You really expect me to take your graph seriously?
Oceans are naturally heating since the end of last glaciation. When the answer is known at least qualitatively, it is not difficult to present plausible data.
Natural heating will continue until inevitable descent into the next glaciation. Oceans cannot be heated by atmosphere. They have incomparable by orders of magnitude mass and heat content. Oceans heating is a direct proof that the global warming is natural. The oceans heat the atmosphere, not vice versa.
And the Sun heats the oceans.
But only to about a temperature of 30C (with a very few exceptions), then Mother Nature kicks in and brings the ocean temperatures down through a huge thunderstorm or hurricane, or cyclone, which takes all that ocean heat and dissipates it into space.
So, yes, the oceans are warming but only to a certain extent, and then the warm ocean cools off through creation of powerful storms.
I wonder why these climate change ocean temperature fearmongers never mention that the ocean can only get so warm, and when it does, Mother Nature transforms that ocean heat into radiation going out into the solar system? I guess it wouldn’t sound nearly as scary if a tipping point isn’t included.
The tipping point for ocean heat is 30C. It’s being hit every day, all over the world. And when this tipping point is hit, it dissipates heat into space and cools the surrounding area off. Nothing to fear. It’s been happening since the Earth acquired its oceans.
“And the Sun heats the oceans.”
But only as far as the Sun penetrates – and the warmed water floats to the surface, and as you say, dissipates its heat to space.
The vast bulk of the ocean is of course heated from below. Otherwise, places like the Arctic Ocean would freeze completely during the Arctic winter.
Man-made heat obviously makes the “world” warmer than if man’s energy production and use didn’t exist. Support for this has been shown by comparing the temperature records of two island nations (Britain and Japan), with distinctly different “Industrial Revolution” timings.
In any case, adding CO2 to air doesn’t make the air hotter.
How does the heat in the mixing zone [ https://www.noaa.gov/jetstream/ocean/layers-of-ocean ] get past the thermocline (~200 m) to increase the heat content down to 2,000 m?
Magic, of course!
Comparing inflated estimated current results vs adjusted to lower value previous results is the same tactic used in predicting and reporting markets behavior by businesses, bankers and politicians. Using an estimate of the current quarter results as 100% predict an x% growth. A quarter later quietly revise previous quarter results, as they actually were only, say 95% of the early estimate. Estimate the now current current quarter results at 100%, reporting estimated growth of a few %. Repeat quarterly and yearly as necessary revising the baseline to lower values to always predict a company, a particular market all all economy growth to maintain a rosy picture for voters and investors. They should consider this procedure for Nobel prize in Economics, and perhaps, Ig Nobel prize as well.
Until then, every NOAA temperature dataset should be taken with a grain of salt.
Actually:
Until then, every NOAA temperature dataset should be taken with a high degree of skepticism.
NOAA’s temperature data sets shouldn’t be believed at all. They don’t match the historical record. Not even close. I don’t know how anyone could believe NOAA after seeing ths historical record. And I would say most climate scientists have seen the historic temperature charts, and the difference between the historic temperature profile and NOAA’s temperature profile could not be more different. Polar opposites. Created in a computer. What this means is that a lot of climate scientists are bare-faced liars who know the truth and spout the lies anyway.
The historical temperature record was written down when there was no Human-caused Climate Change Agenda is the Science Community. They had no agenda, they just wrote down the temperatures.
Then along come Climate Alarmists who bastardize the hell out of the written record, and then use this bastardized version of the global temperature profile as “varification” that CO2 is heating up the Earth’s atmosphere a LOT.
Their jobs depend on them continuing to lie about the climate and temperatures.
Somebody ought to ask them how they got a “hotter and hotter and hotter” Hockey Stick chart temperature profile out of data that has no Hockey Stick profile? The original, historic temperature data is the only data available to Hockey Stick creators, and it shows that the temperatures were just as warm in the Early Twentieth Century as they are today,so how do they get a “hotter and hotter” Hockey Stick out of data that doesn’t show a Hockey Stick? Answer: They cheat and lie. See Climategate.
Catastrophic Human-caused Global Warming is a Hoax and so is NOAA’s bastardized temperature record.
Yes, just a few dishonest temperature data mannipulators have ended up screwing millions of people, fooling them into believing CO2 is overheating the world.
I would like to be standing by the Pearly Gates about the time some of those people come up to explain themselves.
“They are lying, God! God: “God Knows”.
I was countering the phraseology of grain of salt with something more applicable, skepticism. You know what skepticism is.
I am aware of all of your points and agree.
That wasn’t meant as a criticism of your comment. I just used your comment to go off on NOAA. 🙂
“But what if the “corrections” applied to these datasets are introducing more noise than signal?”
I’ve been waiting for “them” to reach the limits of correction. Trying to create an exponential with linear data would increase in difficulty every year. The question I’ve always had is how far will “they” go _after_ “they” can’t deny “they’ve” gone too far?