Update: This blog post contains mistakes because of a programming error. The mistakes will be addressed as soon as I am capable of doing so, but in the meantime, I direct anyone who hasn’t been following the comments section to read this comment. My apologies – Brandon Shollenberger
I have a confession to make. I was wrong. Tamino was right.
Yesterday, I happened across a blog post Tamino recently wrote. I had never heard of any of the people or groups being discussed in it, so I only skimmed at first. I was barely paying attention until I reached the second figure:
I immediately did a double take. I was sure there was no way I saw a smoothed line go up while the unsmoothed values went down. Only, I had. Caught off guard, I began reading the post more closely. Immediately, I read:
Lest you object that I’ve applied some fancy smoothing method designed to get what I wanted, let’s apply the simplest “smoothing†method of all. Let’s compute 5-year averages rather than 1-year averages. We get this:
Again, I was shocked. This figure showed an even greater uptick than the last! My mind couldn’t understand how the data was going down, but Tamino’s derived lines were going up. I quickly read the rest of the post, saw no explanation, and made a comment on this site asking people if they could tell me whether those figures made sense to them. A few comments were exchanged, and then I realized I could just ask Tamino:
Tamino, I was wondering if you could explain what kind of smoothing you used for your figures. Your smoothed line extends as far as the data does, implying none is excluded, yet it goes up while at the end, your data goes down. The last four points are equal to or lower than the ten points before them, so I can’t figure out why the smoothed line shows a steady increase.
Incidentally, it’s kind of funny you mention the possibility of someone objecting “that [you]’ve applied some fancy smoothing method designed to get what [you] wanted†as I’m sure some would get the impression you did just that as your smoothed line goes up while the data goes down. The same is true for your five year average graph which has a different visual impact because of the periods you used.
Mind you,none of this has any real relevance to the point you make in your blog post. I’m just curious about how that smoothed line was made.
As I noted in my comment on his site, I didn’t think the issue was hugely important. It was just an oddity that caught my eye. Tamino promptly answered, explaining:
The smoothing method was a modified lowess smooth (I use a different weighting function than the usual tricube) applied to the monthly data (rather than annual averages). Therefore it includes the most recent 4 months, which is astoundingly warmer than any preceding third-of-a-year
I saw he said he had used a (modified) LOWESS smooth, and there were four months of data which hadn’t been displayed in his second figure. Knowing what a LOWESS smooth is like, I assumed the extra data was responsible for what I saw. There was a bit more back-and-forth over other points, then I first expressed that understanding:
I have to disagree with you saying the “intuitive result†was wrong. The reason I was thrown off is I didn’t know different amounts of data were being used in the two lines. Had I known you were including more (and higher) data in one, the curve wouldn’t have seemed unintuitive.
The only way intuition was wrong was it said two lines representing the same data would represent the same data.
Tamino again promptly responded, telling me I was wrong:
Computing the smooth without the first four months of 2012 would not have sensibly changed its result. It contradicts the intuitive result because the intuitive result is wrong, not because data from 2012 contradict it. The fact that you fail to understand this, confirms how important it is to emphasize that noisy data can fool people into drawing the wrong conclusion.
I responded again, telling Tamino I found his response “nearly impossible to believe.” When he responded, he reiterated I was wrong, but this time with a great deal of virulence. I decided his petulant behavior made it not worth posting on his blog, and I planned to stop there.
But I couldn’t get rid of my curiosity. Tamino had seemed very sure of what he said, and I didn’t have any sort of evidence on my side. That bothered me, so I decided to test Tamino’s claim that I’d get “much the same thing” if I tried to replicate his work. A little while later, I had this graph:
You’ll note in my figure, the smoothed line goes down at the end, exactly as I would expect, and exactly as my intuition told me.
As I compared my result to his, I realized the only differences were at the beginning and end of the lines. The middle ~80% was identical. I quickly realized what was going on. Tamino had said he used “a modified lowess smooth” with a different weighting function. That was to blame, not the extra four months of data like I had thought (you can see the difference adding the four months from 2012 makes in blue below):
So as I said, I was wrong. Tamino was right. It was not the addition of four months data that made his end point rise despite the data falling.
But that’s not the end of the story. Why did Tamino get such a radically different end to his series? The data went down. When I used a LOWESS smooth, my line went down. So why did Tamino’s go up?
There’s only one answer. For some reason, Tamino’s modified version of LOWESS gives extremely low weight to the data near the ends of the series. Why? I don’t know. He doesn’t explain it in his post. In fact, he doesn’t even say what he did in his post. It would be completely impossible for anyone reading his blog to know what he did, much less why he did it.
Of course, Tamino offered a second approach in case anyone accused him of using “some fancy smoothing method designed to get what” he wanted. Naturally, I suspected that graph was questionable too, so I decided to test something. What if when calculating the five-year averages, we don’t include 2012, which only had four months of data?
I was speechless. If I took the four months from 2012 as a whole year and averaged them with the data for 2008-2011, I got Tamino’s results. If I took the average from 2007-2011, I got a result nearly half a degree lower. Not only does this have a huge visual impact, it means Tamino used one data set in three different ways:
1) Annual averages, excluding data from 2012.
2) Monthly smoothed, using all data.
3) Five-year averages, using four months from 2012 as a whole year.
Not only is it highly peculiar to use data in so many different ways, Tamino didn’t tell anyone what he had done. Nobody reading his post would realize how different his graphs could look with different choices which were at least as good, if not better than, his undisclosed choices.
Does any of this have any impact on the point of his blog post? No. To be honest, I still know nothing about the Tom Harris his post criticizes. The point of this is people promoting global warming concerns should make it possible to understand their work without having to figure out undisclosed details based on arbitrary and unexplained decisions. In fact, everyone should.
Finally, here is what Tamino’s GISS/USHCN comparison would have looked like if he had used a standard implementation of LOWESS (GISS – blue, USHCN – red:
Update 2 (5/11/2012): After some consideration, I’ve decided to leave this post as is as the basic point of the post is unaffected by the errors in it. As such, the errors will be left for posterity, and anyone interested in the corrected version of the technical details can read the comments section. Moreover, after some more time for discussion, I’ll collect the details into a single, simple explanation, and I’ll update this post to include a link to it.
Brandon,
You have a lot of nerve questioning Tamino on his analysis methods. You should understand that real climate scientists have no obligation to ever explain their methods. 😉
.
But seriously, how could anyone apply an undisclosed method of smoothing, and then object when someone says his conclusions are difficult to understand? IMO, Tamino will never post anything which would add to the uncertainty in the extent of future warming, unless the added uncertainty is only on the side of greater increases; he is consistently a lot less that honest and forthright.
.
This was the part you got right without having to do a bunch of work: “I decided his petulant behavior made it not worth posting on his blog, and I planned to stop there.” The good news is that with this guest post you are almost certainly going to be prohibited from future comments at that echo cham…. err, at that inaptly named blog: “Open Mind”.
I’m sure you remember what happened to lucia:
http://rankexploits.com/musings/2009/two-box-models-the-2nd-law-of-thermodynamics/
“I have a confession to make. I was wrong. Tamino was right.”
You think that will save you?…;-)
You’ve now received the lesson of how Tammy works. Capriciously deceptive and angry when challenged. When pushed he becomes abusive. Why even go there?
Well… I still think you are more right than you are wrong.
1) You were correct that it is very suspicious when the smoothed data goes up at the end when the unsmoothed goes down.
2) Given the temperature data at hand was very suspicious that the 5 year averages end by going up.
In fact, only by making very weird and undisclosed arithmetical choices is it possible to concoct a graph with the 5 year average ending going up. Not withstanding his protestation that the appearance of his graph is not due to a
the appearance does, infact, seem to be the result of some ‘fancy’ smoothing method.
SteveF–
There is little adverse effect in being banned from Taminos. After all, one can always just post the rebuttal elsewhere. Brandon is being too nice. After Taminos first response mentioning that his smooth was computed using data that did not appear in the first figure, Brandon thought that maybe that data made the difference and might explain the bizarre appearance of the graph. After all, Tamino has said the appearance wasn’t due to some fancy smoothing method. So, someone giving Tamino the benefit of the doubt was likely to think “Oh. If what Tamino says is true, it must be the extra data I didn’t know he used.”
In fact, that extra data Tamino used was not the cause of the weird appearance of his graphs. It was the weird smoothing methodology.
steveF:
Nonsense. As my exchange with Tamino shows, plenty of them are happy to explain their methods… in misleading ways.
Niels A Nielsen:
Hardly. I don’t think there’s a thing I could say to make Tamino do anything other than insult me or make deceptive remarks about me. Still, I think it’s important to admit one’s mistakes. Fairness doesn’t stop being a virtue just because other people lack it.
lucia:
To clarify, the extra data isn’t what caused Tamino’s second and fifth graphs. It is, however, what caused his third graph. When you realize it caused the weirdness of one graph, it’s not surprising I’d think it also caused the weirdness of the others. That was just me assuming Tamino didn’t make two extremely weird decisions for one post.
In effect, doesn’t that make my mistake… giving Tamino the benefit of the doubt?
lucia (Comment #95468):
Perhaps an attribution analysis could be undertaken to determine the relative contributions of chance variation and anthropogenic influence. 😉
Perhaps instead of Tamino using a “fancy smoothing method designed to get what I wanted”, used a fancy smoothing method which coincidentally gave him what he wanted. No way to get around it being a non-standard technique of smoothing, maybe something taken from interpolation of data?
Please. No obviously fake emails. @a.xxx is not an email.
max, no not that max
All signs points to this. This means he didn’t set out to deceive people. Of course, I believe he would have changed methodologies if it had caused an inverse effect, so…
No. It’s just a LOWESS smooth with a weighting scheme designed to de-weight points near the ends of a series. I’d assume the reason it was designed is because in some situations, the points near the end of a series are noisier than other points.
Here, I’ll help you out with what Tom Harris was teaching in a University course.
“[Land surface records go back to 1880 and global data from NASA show] “fairly substantial rises, we’re talking about perhaps almost a degree from the zero line up here. Now, if these measurements were real — and that’s a big if, we’ll talk about it later — you would expect this kind of warming to be visible, in let’s say, the United States…. it has by far the best temperature record of anywhere in the world…. You would expect to see that if there was global warming, that there should be some indication of warming in the United States. Well what you see is… there is a very slight overall warming in the United States, but in the 1920s and 30s we were just about as warm as we are today.—
Which is just one of many perversions of logic and evidence in his course.
It occurs to me as an author, I can actually moderate things. That’s going to make it harder to ignore bugs when he makes off-topic posts, like this one where he decides to “help” me do something I have no desire to do.
bugs–
I’m pretty agnostic on what Tom might be teaching in the classroom. Do you have any course notes, handouts etc? I saw the bit you quote comes from an appendix of a report not written by Tom Harris. Page 7 of that report says those quotes come from video recordings. Have those video recordings been released? Put on youtube? It would be nice to hear what various independent 3rd parties have to say about the videos.
I’m not going to wade through the videos– but if you like, I’ll be happy to put Tom Harris on my agenda when I stop into the Heartland conference.
But otherwise, whether Tom is right or wrong is rather orthogonal to the issue here which is that Tamino is presenting some pretty confusing graphs — and it would appear he is meaning analytical choices that are tenuous. Actually, I think indefensible might be the more appropriate adjective.
Tamino’s doing this sort of thing is going to confuse people and mislead them about the trajectory of temperatures in the contintental US. This is true even if this Tom Harris person is, like Tamino, resorting to some analytical stretching of his own.
Brandon,
“That was just me assuming Tamino didn’t make two extremely weird decisions for one post.”
Four months of 2012 ought not be given the same weight as a whole year in calculating a 1 year or a 5 year average, nor should it be simply ignored when calculating an average. Tamino did make more than one weird decision for one post. Oh, and they all happened to generate a rather odd (counter intuitive) upswing at the end of each of his graphs….. talk about staying on message!
Brandon, a suggestion.
Take the monthly anomaly data from the start (1895) to finish (May 2012). Multiply anomaly by yearly fraction then find the mean and subtract mean from each value.
Then do a CUSUM plot. Start at first month, then 1st + 2nd, the 1st+2nd+3rd, etc.
The CUSUM plot shows you rate that things change when you have a low signal to noise ratio, without any averaging or change to the data. True noise is flat, a fall in signal over time is a downward slope and a rise over time is an upward slope.
The CUSUM plot is actually a mathematical chart recorder. CUSUM charts/plots are used for quality control monitoring of tolerances (i.e. the wear on a cutting tool will tend to make things oversize, but less than expected wear will under cut).
These plots are used in most biological measurements, whether people know they are doing it or not.
http://en.wikipedia.org/wiki/CUSUM
Brandon,
“I’d assume the reason it was designed is because in some situations, the points near the end of a series are noisier than other points.”
Yes, almost certainly that is the explanation. No doubt the details are proprietary and so can’t be disclosed publicly. On a related subject, I have recently learned that there exist certain fabrics with the amazing property of being invisible to anyone who is unfit for his position or just hopelessly stupid.
SteveF:
For what it’s worth, Tamino didn’t include those months when making his graph of annual values, and I think that’s perfectly fine. I think it makes perfect sense that if there isn’t a year of data, it doesn’t get graphed. That’s not the only option, but it is a reasonable option.
DocMartyn:
Could you explain why you suggest this? I can see why it might be interesting, but I don’t see how it speaks to the issues at hand.
If that’s the motivation, it’s wiser to either just stop plotting at the point where weighting becomes asymmetric, figure out some way to indicate that the computed endpoints are noisy or to show the a variety of results obtained using different end treatments.
I agree that Tamino should have specified better what he is doing. He’s usually right, but it can be hard to find out why.
We talked about endpoint treatments a while ago with “Mannian” smoothing. Certainly the simplest option is to just not go there. But there is reduced but extra information available about endpoints which you can try to represent on the graphs, recognising that something has to give as end of data approaches.
I think LOWESS lets centering go, using predominantly past points to compute the approximating polynomial. I think that is less desirable – it leads to the effect that Brandon notes, where the result is insensitive to the latest values. Getting the most up-to-date estimate is the point of taking the smooth to the end.
“Mannian” smoothing, which is actually something which has been done in statistics for a long time, retains centering but sacrifices smoothing, so the results become noisier. It has been criticised for taking this to the logical limit, in which the smooth is pinned to the endpoint. No smoothing at all there. But it would certainly track the late downswing.
Nick–
Tamino also didn’t do standard LOWESS. Brandon and I have both created plots of standard LOWESS.
lucia:
I agree with a caveat. If the change in strength of noise is well-quantified, I think it’s a reasonable approach. That’s obviously not the case here, but I can imagine scenarios where that weighting scheme would be useful.
Nick Stokes:
This isn’t what LOWESS does naturally, at all. The only reason it was in this case is Tamino used a non-standard weighting scheme.
I’m not sure why you would think what you think, but if it were true, it’d completely invalidate my post. Given the seriousness of that, and the fact that despite it you didn’t address anything I said in my post, this comment of yours is strange.
“DocMartyn:
Brandon, a suggestion.
Could you explain why you suggest this? I can see why it might be interesting, but I don’t see how it speaks to the issues at hand.”
CUSUM plots are the best way to identify changes in rates of change in serial data-sets. So it allows you to readily identify a change in rate, even if the noise:signal ratio is high.
You do not have to pad either the beginning or the end to observe the end data.
Doc
I don’t think the topic of this post is identifying the change in rate of anything. That might be an interesting post– it’s just not this one.
Sorry Lucia and Brandon, I though this was a critique of smoothing algorithms designed to deal with the ends of series.
I was suggesting a quite different technique that does not rely on smoothing, but really does tell you if a rate is changing.
DocMartyn:
That’s fine. It would be a subject I’d find interesting, and I’m glad you mentioned (I had never seen CUSUM before). I just haven’t really written anything about the subject, so it’s rather off-topic. Also, I get the impression most people wouldn’t be interested in a discussion of that.
By the way, I’d probably try it out and see what the results were on the USCHN data set, but I’m a little tired. This blog post went up only 24 hours after my first comment about Tamino’s blog post (and about half that since I first asked Tamino). Give me a day or a two break, and I’ll be happy to work on a new problem.
We should note that USHCN V2 adjustments are now up to a little over +0.6C.
The trend in the raw Max temperature is 0.0C per decade from 1895 to 2010 while the trend in raw Min Temp is 0.02C per decade according to Williams, Meene and Thorne 2012.
The adjusted temperature trend is 0.06C per decade over that period (and then there was some very large changes implemented in early January this year).
So an increase of over +0.6C as a result of adjustments (basically the same number I came up with using a different method – keeping track of the data over time).
That’s fine Brandon, I would do it myself but am wedded to excel and have two teenagers.
One more point. CUSUM is perfect for showing discontinuities, step changes stick out like sore thumbs, which we see on charts when we inject liquids into chambers (addition artifacts).
DocMartyn:
Two teenagers? Yikes. Sometimes posting on blogs makes me feel so young.
Tamino should have specified better what he is doing.
Well, he has.
I guess a reasonable question would be what is “standard LOWESS” as Lucia and Brandon define it? Tamino talks about “vanilla R lowess” and the parameter value f=0.2 that he used in his smoothing. A parameter value of f=0.15 seems to give the graph that Brandon shows.
Can you verify that using that parameter value generates Tamino’s graph? If it’s true that the difference between “standard LOWESS” and Tamino’s “fancy smoothing method” is a 0.05 difference in one parameter, the dark mutterings about “very weird and undisclosed arithmetical choices” seem a bit overdramatic.
PDA:
It’ll take me a little bit to see if I can figure out what is going on in Tamino’s post, but I can say this much right off the back. I did not use f=0.15 as Tamino portrays. I tested a number of different values for f to see which matched his line best, and I found 0.2 gave a near perfect match. 0.15 gives a terrible match, and it looks dramatically different than the graph I posted. There is no way anyone trying out LOWESS filters would actually think I used f=0.15, and I have no idea why Tamino implies I used it.
Heck, if you just look at the line he comes up with using f=0.15, it’s obviously different than my figure by a large margin. As an example, the last rising slope in my figure pretty much rises steadily, but his supposed replication has a huge flat portion. There’s also the obvious fact the period from 1950-1960 is basically flat in my figure, but in his, it’s rising notably. Given how obvious these differences are, I’m at a loss as to how you can say:
All it takes to see that wrong is to look at the two graphs. That’s it.
PDA (Comment #95499)
Why doesn’t Tamino just specify his weighting function, since he uses “a different weighting function than the usual tricube.” It would make diagnosis of the differences between his and other results much easier.
I tested a number of different values for f to see which matched his line best, and I found 0.2 gave a near perfect match.
OK wait… without a “non-standard weighting scheme,” without “very weird and undisclosed arithmetical choices…” just setting the window width parameter to 0.2 gives a “near perfect match” to Tamino’s graph?
What am I missing here?
I was planning on trying to improve my figures for the next step so I could upload ones which were so obviously amateurish. I thought it’d be a good idea to get the axes right, make a legend, and improve the image quality some. I also thought I’d work on putting my code into a script so it could be run all at once, rather than line-by-line like I’ve been doing. However, given the claim PDA raised just above, I think it’s important I focus my time elsewhere.
I’m going to try to work through Tamino’s new post, but first, I want to point out how obviously wrong it is to say my line was generated by setting f to 0.15. To do so, I’ve generated a graph with my original line (in red), and the line I get by setting f to 0.15 (in blue):
http://rankexploits.com/musings/wp-content/uploads/2012/05/15v20.png
As you can see, I did not do what Tamino portrays. There is no way it could possibly seem like I did. And since I can’t currently release a copy of my code, here’s a picture showing the three commands which generate the graph:
http://rankexploits.com/musings/wp-content/uploads/2012/05/15v20_commands.png
So, unless someone is going to say I lied about what I did, then faked that picture, Tamino’s position on this point is completely and utterly wrong.
Edit: Sorry those images aren’t imbedded. I seem to be having trouble getting that to work. Every time I use the img src tag, the lines just winds up disappearing after I hit Submit.
Re: PDA (Comment #95502)
I don’t know if you’re missing anything. I’m just saying it would be a lot easier for everyone to verify the results if Tamino just posted his procedure. Then we could skip the whole “guessing what actually happened” game and any related discussion about possible “Dunning-Kruger effects.”
Re 95500:
So tell us exactly how you did get your graph! You criticize Tamino for not providing technical details, but won’t say how you produced your graphs! “a little while later, I had this…” and “standard LOWESS” don’t count as providing technical details necessary to replicate the work.
PDA:
You’re missing what I said in my post:
I think it’s fair to say matching someone’s results except for one issue is getting it “nearly perfect.”
Oliver:
In a comment just above, I posted an image (well, a link to an image) which shows the exact commands I used to generate two lines. One of those lines is the line I posted in my blog post. That means anyone can see, and verify, that my code does what I say it does, and not what Tamino says it does.
I guess someone could try to argue I’ve modified the underlying data since I don’t have a script to collect that, but that would be silly. The sheer level of dishonesty that would require is mind-boggling, nevermind the fact it’d be impossible to modify the data in a way which would generate my results. Besides, I could easily post a graph of annual averages to show my data matches Tamino’s (I don’t include it in my graphs right now because it looked cluttered when I did).
So yeah, Tamino and PDA can say what they want, but what I’ve done is easily verifiable.
Oliver:
Because then it wouldn’t give PDA a chance to shill for Tamino.
Simple.
OK, maybe your graphs look different to Tamino’s (using f=0.2) because you’re using different data…he used data from NCDC, you are using USHCN.
I freed a comment from a user named dc from moderation. I suspect this is a reader coming over from Open Mind, and given the contents, I don’t think they’re actually trying to ask me a question. I think it is just a rhetorical ploy to try to smear me. However I may be being too cynical. So, assuming for the moment dc is a serious poster:
dc, what makes you think I “won’t say how [I] produced [my] graphs”? Have you seen someone ask me for an explanation, only to have their request ignored? I haven’t. If I had seen someone ask about what I did, I’d have been happy to explain it to them. As for not providing enough details, the only detail I didn’t provide is the f value. Since I was replicating Tamino’s graph, I assumed he’d know the f value used, and thus, I didn’t think it needed to be mentioned. However, as I said above, it is 0.2, the same as Tamino used. As for any other technical details, you can see the commands I used here
That includes all the information one would need to reproduce my smoothed line, and it contrasts mine to one with a different f value. If you think there is anything else I need to explain, feel free to ask!
I think it’s fair to say matching someone’s results except for one issue is getting it “nearly perfect.â€
Well, I’d say that’s a bit unclear. You called the same graph “radically different” in the OP. I understand now what you mean by “nearly perfect,” and thank you for the clarification.
I definitely encourage both you and Tamino to compare methods to figure out how you got such different results. I have no reason to imagine you “lied” or “faked that picture,” and it certainly seems like there’s something other than the window width parameter in play here.
Out of curiosity, what parameter did you use?
dc:
dc, the NCDC is the agency responsible for USHCN. You can see this by visiting their page on the USHCN.
PDA:
I apologize for being unclear. I probably should have used a different phrase, but in my haste, I didn’t think about the fact people wouldn’t necessarily remember what I had said in my blog post on the issue.
I’d be happy to share whatever information could help, but as I told dc, that image shows you all you need to know to be able replicate my graphs. The only thing not covered by those lines is you’d need to extract the data from this file (you want the lines which start with 110), then do the anomaly calculations.
I unfortunately don’t have a script for that, as I did a number of steps manually, but anyone could do it with Excel in a few minutes. I’d love to be able to give turnkey code for my graphs, but honestly, I’m not that acquainted with R yet.
As I said before, and as that image shows, I used 0.2. I did try out a number of other values back when I started examining the data to see what happened (and because I hadn’t figured out what value Tamino used yet), but once I realized Tamino and I had the same one, I didn’t think about it again.
You’re both smoothing time series data…I’m not sure which smoothed plot is more ” true” but I’m willing to bet that yours, Brandon, is more honest.
The title of Taminos blog is a dead give away for how willing he is to stretch things. Don’t let him discourage you.
Well guys, it looks like I did make another mistake. You’ll remember I said:
I came to this conclusion after testing several different possibilities. I tried calculating five year means covering different periods, and none matched the value Tamino gave. So, on a hunch, I test that, and I got ~0.85, matching the value on Tamino’s graph. This lead me to conclude it was what Tamino had done. He now disputes this, saying:
I’m willing to concede this point. I was wrong when I said Tamino used four months of data as an entire year’s worth. In reality, he simply said, “Let’s compute 5-year averages,” then showed an average based on 28 months of data.
I had assumed Tamino made the innocent mistake I described in my blog post. It turns out, he didn’t. Instead, he knowingly displayed a 28-month average in a graph of 5-year averages. And apparently that’s okay because:
That’s right. Tamino doesn’t deny misrepresenting the data. Instead, Tamino’s position apparently is it was okay for him to misrepresent the data like he did because it didn’t matter for the blog’s main point. And I’m just a jerk because I’m expecting him to:
It couldn’t be that making undisclosed, arbitrary decisions which contradict the text of one’s post and substantially alter its visual impact is bad. No.
I’m just a jerk.
By the way, I decided to check how I reached the conclusion I reached on that point. I took the last five year’s anomalies and averaged them together, getting 0.82.* I then took the last 28 months of data and averaged them together, getting 0.85. As you can see, my result was perfectly in line with what I could see in Tamino’s graph.
I took this as indicating I had found the answer as I could easily imagine how it would happen. All one would have to do was calculate annual averages then average five of those together. If one included 2012 in the process, not thinking about the fact it had less data (since it would appear as a single number just like every other year), they’d get the results Tamino got.
That was the logic I used. I was surprised at the idea of Tamino averaging annual values rather than monthly values, but the difference in approach (in this case) is so negligible I didn’t think much on it. I just figured I had found a simple mistake that would explain Tamino’s results when nothing else would, so that had to be it.
I don’t think I can be faulted for never anticipating the fact what Tamino called a five-year average was really a 2 1/3 year average. If anything, I think the fact I attributed the issue to an honest mistake rather than a misrepresentation speaks well of my character.
Edit: By the way, the fact I omitted the four months of data when calculating the average in that figure shouldn’t be taken to mean much. Shifting the period I used by four months would increase the point I get by… 0.04. That’s not even enough to be visible.
A word about the “principle of charity” in relation to climate blogging: never attribute to malice on the part of others what can be attributed to ignorance on your part.
.
.
Not the standard formulation. And way too many are addicted to the smug feeling of superiority, moral or technical, for me to have any hope of it being more widely adopted.
Are you saying that if Tamino had calculated the last point of the 5-year average graph differently the lowess smoother would head down like yours? Would it have changed the conclusion of Tamino’s original post? If so, please demonstrate.
The main question here is why you’re lowess smooth of monthly anom’s turns abruptly downward when Taminos does not. The only way to compare them properly is for both of you to provide the data you input to lowess() and explain exactly how you got that data (what you did to the raw data first) – you must be inputting different values to lowess(), cos you get very different outputs with f=0.2 (I know Tamino used something different orginally, but he subsequently used lowess in R). If you really want to show that Tamino is wrong, post your input data (“aTemps”) and explain exactly how you got it from the data linked to above.
And this is why I shouldn’t program while lacking sleep. I went back and looked at my data, and at one point, I took data from the wrong variable (annual averages) and put it into aTemps. Since I didn’t have all my commands put in a script, I never caught the mistake. In other words, I’m a huge fool, and I’m unbelievably embarrassed.
I obviously cannot do calculations on annual data and compare it to calculations done on monthly data. Once I fixed that data handling error, I got the exact same graphs Tamino displays in his current post. This means he and I are now in complete agreement about what was done. To be clear, his description of the calculations are completely accurate.
I haven’t slept in 36 hours, so I need to crash. However, when I get up, I’ll edit this blog post to correct the record. If lucia wants to modify it before then (or even take it down), that’s fine by me.
As a final comment for the night, while my mistake is unacceptable, and I am extremely sorry for making it, the basic situation still remains. In fact, Tamino has shown it for us. Changing the f value from 0.2 to 0.15 completely changes the direction his smoothed line takes at the end (it’s actually flat at 0.17 and turns down at 0.16). A reader of his blog post could never have known that.
Beyond that, Tamino has admitted to using a 28-month period in a graph of five-year (60 month) averages, and that caused the graph to curve up rather than down. That decision was also unexplained.
Tamino made two choices which substantially altered the visual impact of his graphs, and he did so without any discussion or justification. One decision was flat-out deceptive, and the other was completely arbitrary. This remains true no matter what you may think about me.
Anyway, I apologize to everyone for making such a stupid mistake. I hope you can forgive me for it and not let my idiocy distract you too much from the real issue being discussed.
Brandon Shollenberger (Comment #95480)
You could show us how you do a better job proving Tom right or wrong, which was the topic you picked up on.
Why do you think the main quesiton is the lowess smooth and not the 5 year averages used to “prove” it’s correct?
If you use a high enough f with lowess, it resembles a straight line. If you use a low enough one, it tracks the data nearly perfectly. Whether or not it is “fancy” and why the end points do what they do depends on POV. But Tamino tried to instill confidence in his LOWESS smooth with the 5 year average turning down. It is only through some weird (or at least inconsistent) choices you can do that.
I can show a variety of 5 year average, ending with different data.
“I can show a variety of 5 year average, ending with different data.” But each choice still shows Tom is wrong. You are just arguing about the end point of the curve.
It appears Tamino has a new post that disputes some of the assertions made here anyway.
Not to mention the logical flaws in Tom’s argument. By selecting the US temperature record only, he is cherry picking a small part of the total global area. He is only choosing the ’48’ states, and leaving out Alaska. There is no evidence provided that the US has the best quality temperature record in the world, nor even that that point is relevant.
Brandon,
@95498 – Mine are 1 and 3. Feel better?
I am really at a loss trying to understand your point Brandon. I don’t think any of the graphs presented show the 1920s and 1930s were hotter in the US mainland compared to more recent data, as Tom Harris claims in his course syllabus:
You would expect to see that if there was global warming, that there should be some indication of warming in the United States. Well what you see is… there is a very slight overall warming in the United States, but in the 1920s and 30s we were just about as warm as we are today.â€
Tom Harris clearly teaching incorrect conclusions based on fallacious fabrications in his course. He deserves strong condemnation by anyone interested in science, math, or statistics; as well as anyone interested in providing good educations for the next generations.
Bugs–
First: I don’t know what Tom said. Second, so what? As far as I can see, this post isn’t about Tom.
Paul_K
Tom may very well be wrong and may need to be condemned. But that doesn’t mean that puzzling confusing graphs should be promulgated by others.
Paul,
I think the point here is an aside relative to the Tom Harris discussion. Tamino seems to be playing fast-and-loose with his data presentation to drive home his own point. This is a bad habit for scientists to get into when they could make much the same points (although perhaps with less visceral impact) without pulling any “fast ones” on the audience.
Brandon,
Do you have a place where we could grab the USHCN monthly data? The NCDC table generator doesn’t seem to be working for me this morning.
Thanks,
Oliver
Bugs–
First: I don’t know what Tom said. Second, so what? As far as I can see, this post isn’t about Tom.
Paul_K
Tom may very well be wrong and may need to be condemned. What does that have to do with this post?
Oliver–
I have various monthly and annual average graphs. To create the graphs, I used data obtained by going to the link Tamino provided and hunting around the page where I found “Please see our ftp site for statewide temperature and precipitation data which are used in making these maps.” The Read Me text suggests I use code “110” to extract data for the (contiguous 48 States).
I thought this must be the data Tamino used. Did he use something different? I’m not getting this
. I’m trying to see if I have a bug or am using different data.
BTW: If I end with April 2012, I get this
But I get a variety of graphs– each somewhat different depending on what I pick for the final month.
Ok– I read… may be using different data. But the misleading endpoint in Tamino’s graph is due to the the final point not being a 5 year average.
PDA–
Tamino himself describes his method as “modified”. Not giving his choice for “f” even after someone read the post asked for details on his smoothing method is certainly undisclosed. Moreover, the parameter choices near f=0.2 gives a smoothing windows of 20 years yet Tamino’s text and graph suggests the reader should somehow not think his method is fancy or selected to show what he wants by comparison to 5 year averaging.
So not withstanding your comment elsewhere that I somehow shouldn’t think this: I absolutely stick by
The facts that Tamino’s choice of f is very near the transition from where higher values show upticks and lower values, that the value results in a smoothing window that is now where near the size he uses in the graph that is supposed to make us feel confident in his undisclosed choice? Those make me think the failure to disclose the choice and it’s effect on the behavior of the end point more dubious not less.
Undisclosed, undisclosed, undisclosed.
Re: lucia (Comment #95532)
Lucia, I get the same graphs you do.
As you’ve probably noticed by now, you get a very similar graph to Tamino if you compute the usual 5-year (60-month) averages but use a mean of the last 24 months of data as the endpoint. It’s fair not to use 30 months, since that would leave a big signature of the annual cycle. It should be noted however that 36 months doesn’t produce the “rising” endpoint, while 12 (which would just pick up the last annual average) produces an uptick which is obviously too big in light of the dip before the rise near the end of the annual values.
I suppose the somewhat larger error bars shown on his last point fit with this explanation as well.
In the future.
When Tamino or anybody posts a graph that looks odd
ask them nicely to produce the the data as used and the code as
run.
Its pretty simple. If people dont post that stuff, I’m free to conclude that they have something to hide, cause they are hiding something.
Bugs
‘There is no evidence provided that the US has the best quality temperature record in the world, nor even that that point is relevant.”
Huh. There is plenty of evidence that the US has the best quality temperature record. stop talking out of your ass.
Given the stated reason for the graph, I don’t think this gives him a pass. The stated reason for that graph is
The smoothing we are supposed to trust by looking at the 5 year averages doesn’t take into account whatever the uncertainties in the measurements are.
Repeating the quote:
Actually, I sort of wondered what the basis for those error bars is. Each years annual average temperature is an observations of what actually happened in specific years. So, for example, the difference in temperature in 2011 vs 2011 is not “sampling error”. Does anyone know how Tamino’s error bars relate to error estimates published by NCDC, NOAA and so on?
I woke up this morning and decided to do some stuff outside while the sun was still low. While I was out there, something occurred to me. In my comment above, I acknowledged Tamino was right about the calculations. However, while I had verified everything he said about them in his latest post, I hadn’t checked what he said that made me start all of this:
When I realized I hadn’t checked that, I realized something else. As Tamino himself has shown, his choice of 0.2 for the f parameter barely creates a graph with an upward end. He showed reducing that parameter to 0.15 is enough to give a downward end. As I commented above, that will actually happen with any value for f below 0.18, and 0.18 will produce a flat end. Given how much of an impact such a small variation in a parameter has, it’s hard to believe omitting four months would have no notable effect. You’d be omitting approximately half of the data which creates the upward end. But you’ll note, in his very first comment to me, Tamino said:
I did. I’ll display the last 12 results from the values it generates:
As you can see, the data turns down at the end. This means if you omit the last four months of data, and you use the same parameter Tamino used, you get a downturning end to the line. This makes one thing clear: I was right. From the very start, I was right.
Now then, I do not believe Tamino’s modified version of LOWESS will get a downturn with f set to 0.2. I can’t prove this since I don’t know what he did, but his version is more responsive to small changes in the data, so it’s the inevitable result. That means if one uses standard LOWESS with the parameter used by Tamino, a downturn is found. If one uses Tamino’s modified LOWESS with the same parameter, that doesn’t happen. So again, I was right. Use the standard implementation of LOWESS, those four months are vital. Use Tamino’s implementation, they aren’t.
My intuition was spot-on this entire time. It spotted the fact a standard implementation of Tamino’s smooth couldn’t produce a similar graph without those last four months of data. It then correctly identified the reason for Tamino’s claim was the difference in methodology.
In fact, my intuition was correct on every point. The only reason I wound up making mistakes is I listened to Tamino instead of my intuition. He said it didn’t matter if you used the last four months or if you used his modified version or the standard version of LOWESS. As it turns out, that’s not true.* He said his graph was of “5-year averages,” but the upturn at the end of it came from him using a 28-month average.
So yes, I messed up in a big way, but that wouldn’t have happened if not for me trusting Tamino’s inaccuracies/deceptions.
*Technically, those each true on their own. It’s only when combined they become false. That’s a lame defense though, just like it was lame when Mann argued tree ring data didn’t matter for Mann 2008 because he could get the same result by using Tiljander.
Re: lucia (Comment #95548)
It absolutely does not give Tamino a pass, especially given the stated reason for the graph. I was only noting that the larger error bars shown seem consistent with forming the last average from fewer data points than the other averages.
I was sort of wondering this as well on the way to lunch. My first hunch, based on some previous appraoches by the author and his sometime-co-authors, is that the signal within each averaging interval is assumed to consist of a “true” signal (mean) + constant-spectrum “noise.” I might try a quick check in a bit.
Paul K2 (Comment #95526),
Next time you should look carefully at what you quote before you try to a attack someone you’re quoting from. He doesn’t claim that the 1920s and 1930s were hotter in the US mainland compared to more recent data. He claims that “there is a very slight overall warming in the United States, but in the 1920s and 30s we were just about as warm as we are today.â€
“Just about as warm” is NOT the same as “were hotter”.
BillC:
Not really. It helps with what I was talking about, but now I just feel bad for you!
Paulk2’s reading comprehension has never been that good
Re 955499
Strange that you don’t you show the actual graph to prove you “were right”. could it be that the “downturn” of 0.00001 C over 4 months (imagine Tamino concealing that by using all of the available data and “fancy” methods) just doesn’t show up on a trend ranging over ca. 1.0 C?
Notice how the data are reported to 2 decimal places – you should probably only report any derived values to the same level (or less). Or just show us the full plot.
or were you joking?
dc–
What graph do you want to see? And what do you think would contradict or confirm that Brandon is right? Because after all– for Brandon to be right, he only needs to be right about what he claimed. He doesn’t need to be right about something that has nothing to do with anything never claimed.
dc:
Why would you just say that is strange? You didn’t ask me to provide the graph, so it can’t be that you want to verify anything. As far as I can see, the only possible reason would be for rhetorical point scoring, and that’s useless.
Nope. Whether or not a downturn can be seen in no way impacts my claim. If a downturn can’t be seen, that just means the end of the graph will appear as flattening out, not decreasing. A flat segment is significantly different than a steadily increasing line, so it proves my point just as well.
To answer your attempt at character assassination, the reason is much simpler. Posting a graph requires I make the graph, ensure its displayed in a way I’m comfortable showing, upload the image, and paste a link. Posting a sequence of numbers requires… posting a sequence of numbers.
Now then, if you’re actually interested in anything other than trying to smear me, you’re welcome to ask me to post a graph showing my results. I have an open offer to share any information involved in what I do, so all you have to do is ask.
But of course, you didn’t ask. Why not?
lucia:
I assume the one involving the numbers I provided. It seems to be the only thing dc referred to, despite the fact I said I was right about several things. Then again, I get the impression he doesn’t actually want to see anything. If he did, he could have just asked.
And I get the impression you just don’t want to see that Tom is teaching garbage in a University course. Do your’s and Tamino’s figures show that Tom is wrong in his claims that use faulty logic to claim there is little warming? The treatment of the endpoints of a time series is an ongoing debate, but the fundamental claims of Tom Harris are just wrong. I see this as being a far more serious matter than the Heartland posters, for example.
Just admit, Tamino is right, Tom Harris is wrong, and I agree with Tamino.
> [H]e could have just asked.
Indeed he could have, abiding by the same token to the principle described by Ron in #95517.
bugs–
I have no idea why you are complaining about Tom Harris. No one has defended Tom Harris. I don’t even know what Tom claims are– so I could hardly defend him. In the main post Brandon said he doesn’t know who Tom Harris is.This post has nothing to do with Tom Harris.
Because you are yammering away about Tom Harris, I previously ask you whether the videos referred to in the appendix of a report Tamino linked which seems to be based on videos. If those videos were available so we could figure out what Tom’s fundamental claims really are rather than relying on 3rd hand messages through the grapevine. You repeat your claim about Tom Harris– but don’t point to underlying references that might help us know what Tom Harris’s fundamental claims are. Are the videos available? If they are and you want people to familiarlize themselves with what Tom Harris claims, maybe people can watch them and judge what Tom really said.
But in the meant time– this post has nothing to do with Tom Harris. I don’t know why you think it does.
bugs actually gets one thing right:
He’s right. I don’t want to see that about Tom. In fact, I don’t want to see anything about Tom. I don’t care about Tom. Tom isn’t in the slightest bit relevant to anything I’ve discussed.
If someone gets to randomly discuss Tom despite him having nothing to do with the discussion, what’s to stop someone else from discussing Sue? How about George? Maybe even Bob? Heck, maybe we can make it interesting and discuss Bob and Tom!
Brandon Shollenberger (Comment #95583)
May 11th, 2012 at 8:08 pm
bugs actually gets one thing right:
And I get the impression you just don’t want to see that Tom is teaching garbage in a University course.
Tom Harris is relevant, because it’s Tom’s claims that Tamino was debunking. If you had read Tamino’s post, it’s not just some random series of graphs he was creating, it was a a series of graphs to show that a Global Warming ‘skeptic’ was teaching garbage in a University course about Climate.
RE: bugs (Comment #95590)
Allow me to clarify:
Tamino believes that Tom Harris is using “fuzzy math” in his class to understate AGW, so Tamino responds with his own “fuzzy math” which overstates AGW. You think Tom Harris is the Devil incarnate and Tamino is the shining knight. Of course you do Bugs. We expect nothing less.
Clue: Maybe we should all dispense with the “fuzzy math” in climate science.
I think bugs should spend his time at Open Mind. He uses a similar style to the people there, and he has just as much interest (or capacity) for accuracy:
He just makes things up, trying to smear me by implying I haven’t read the post I discussed and quoted from. Why?
It’s like he’s channeling Tamino. According to his logic, we couldn’t possibly discuss typos in someone’s post without all the contents of their post being relevant. It’s like saying:
(That phrase is intentionally nonsensical.)
Why so concerned with a possible typo in a post, when Tom Harris is teaching garbage at a University course. You could just say that, having checked the math, you think Tamino has 1% of it wrong, (in your opinion), and 99% of it right.
Did bugs really just take my mocking portrayal of his position as my own position? That’s the only time I ever brought up the subject of a typo, so I don’t see any other interpretation, but…
Whose business is it what Tom Harris is doing besides his own university?
Note that bugs advocates against academic freedom when he doesn’t agree with the conclusions of the professor teaching the course. Bugs appears to think he (or Tamino etc) should be in a position to reprimand professors who don’t toe the line on climate change.
Bugs,
Will Rogers said it best: “When you find yourself in a hole, quit digging.”
On the other hand though, it is quite entertaining to watch. Dig away my friend. 🙂 Teaching garbage is actually quite common at the university as I recall. I believe garbage sifting is required study in many undergrad programs in the USA.
@95543 Mosher
“ask them nicely to produce…the code as run. Its pretty simple. If people dont post that stuff, I’m free to conclude that they have something to hide, cause they are hiding something.”
Totally not a troll or anything*, but WRT this sentiment, can someone link me to the UAH code for their AMSU TLT product? Not the resultant data, or the papers describing the product- just the actual code. Thanks in advance.
*IIRC RS or JC said they were posting it soon a few months back, but I couldn’t find it when I was referring to the ostensible discrepancy between the LT and surface temp obs a few days ago.
thingsbreak.
I will make you the same offer I made Eli.
Write up a nice request to Roy Spencer. I will sign on.
He disappeared. I predict you too will slither off and suck on your thumb.
Or you could go do what magicjava tried.
you probably dont even know what he was doing years ago.
putz.
Hey Mosher
“the data as used and the code”
for the Lukewarmer hypothesis?
I can see that your following statement will apply when you don’t post it?
“Its pretty simple. If people dont post that stuff, I’m free to conclude that they have something to hide, cause they are hiding something.”
Or does this only apply to graphs?
@Nathan Comment #95607
You don’t need data and code to form a hypothesis.
Observations so far seem consistent with a lukewarmer hypothesis.
sorry Brandon, but don t you think that the claim “I was right.” was a little bit too strong?
for example you wrote this above: “I was sure there was no way I saw a smoothed line go up while the unsmoothed values went down. ”
but that is actually exactly what a “smooth” is about: not follow every wiggle in the data!
i didn t do a fancy analysis, but a simple google image search (for lowess smooth). the result are a lot of graphs showing exactly that: the smooth staying up, while data seems to move down (or the other way round).
a nice example can be found here:
http://www.stat.nthu.edu.tw/~swcheng/Teaching/stat5410/lab/08.01_8.gif
Carrick (Comment #95600)
Academics aren’t free to teach falsehoods. There are currently battles being fought to stop the teaching of homeopathy and other garbage at universities, and rightly so. Tom’s logic is fundamentally flawed, as is the claim he makes based on it.
Re “But of course, you didn’t ask. Why not?”
I didn’t ask to see it because I know what it looks like – its not a downturn (your lowess values show that, as do my own graphs based on the data). So I know that these statements are incorrect:
“This means if you omit the last four months of data, and you use the same parameter Tamino used, you get a downturning end to the line.”
“That means if one uses standard LOWESS with the parameter used by Tamino, a downturn is found.”
There is no downturn. Maybe I have been too quick to assume your goal is to deliberately spread doubt and confusion by use of misleading statements and graphs, but if you really want to be taken seriously, don’t claim the data show things they don’t. And how about fixing the graphs in the original post that you know are just wrong (the ones showing a sharp downturn). Whether its your intention or not, some deniers will lap the original post up (esp. the incorrect graphs) as proof temperatures are declining and that climate scientists are covering it up. If you really believe you were right all along – and that the point of the post is still valid – you shouldn’t need to leave graphs with massive errors in them to show it.
Steven Mosher (Comment #95604)
I will make you an offer. A pat on the back if you write the same request to Roy Spencer, and you and McIntyre sign it.
Brandon Shollenberger (Comment #95599)
Are you trying to make yourself so small a target, that you can’t get hit? Tamino was confident enough to take on the whole of Tom’s claim. You are just quibling about 5% of Tamino’s claim.
Brandon Shollenberger (Comment #95583)
May 11th, 2012 at 8:08 pm
He is, it’s his claim that Tamino has debunked. It’s like you are arguing how many angels can dance on a pinhead.
sod:
No. It only seems too strong if you misunderstand what I’m saying I’m right about.
Obviously, I was wrong about this. That’s why I immediately said, “Only, I had”!
That’s true, and I could have been more clear. The problem wasn’t just that the data went down and the smoothed line went up. As you say, we don’t expect it to follow every wiggle. However, we can compare how responsive a smooth is to data in one segment to how responsive it is in another. If you do that, you’ll see the end of the graph is less responsive to the annual averages displayed at the end than it is in the middle.
The reason for that is there is more data used for the smoothed graph than for the annual averages. Not only does it use 12 points of data for every annual average, it also has four extra months worth of data. This causes a disconnect between the smoothed line and the annual averages (which technically is a way of smoothing). That is what I observed.
Using standard LOWESS with the same f paramater as Tamino, we can test to see what effect the extra data has. If we do, we see exactly what my intution said. Take out the four extra months, and the smoothed line flattens out. Include them, and the smoothed line goes up.
So yes, I was right. I just wasn’t always explicit about describing my intuition because… it’s intuition.
Brandon, some points. Months ago I noticed that Tamino’s old posts had disappeared, and thought it was to hide the embarrassing Ian Jolliffe post. If you can find his old posts, there is one called I think On Smoothing, with a description of how he prefers a Lowess smooth. Perhaps that would give you some details. If he has changed the type of the smooth that would suggest deliberate intent.
How many months did he use for the 2011 number, 60 or 40?
Mosher, in response to one of Steve McIntyre’s debunkings after he got a hold of \I think Grinstead’s smoothing algorithm, the denizens at Tamino told me that getting code is unnecessary, and indeed negative. It could lead to the same bad code propagating. This was before RC posted Show Us Your Code.
MikeN:
I found an archive of some of Tamino’s old posts, but none of them have “smooth” in their title. The one you’re thinking of might be in there by a different name though.
He didn’t do a five-year average for 2011. The last five year average was centered on 2007, and it had 60 months of data (of course, the 2012 point was based on 28 months).
@95604 mosher:
“I will make you the same offer I made Eli.”
That’s weird. Because I didn’t ask for “an offer”. I asked for a link to UAH’s code. This isn’t a particularly difficult request- it either exists or it doesn’t.
“Write up a nice request to Roy Spencer.”
I asked for a link to UAH’s code. This isn’t a particularly difficult request- it either exists or it doesn’t.
“I will sign on.”
I don’t care. I asked for a link to UAH’s code. This isn’t a particularly difficult request- it either exists or it doesn’t.
“He disappeared.”
Because instead of providing what he requested, you gave him some irrelevant red herring “offer”? Can’t say I blame him. I asked for a link to UAH’s code. This isn’t a particularly difficult request- it either exists or it doesn’t.
“I predict you too will slither off and suck on your thumb.”
I asked for a link to UAH’s code. This isn’t a particularly difficult request- it either exists or it doesn’t.
“Or you could go do what magicjava tried.
you probably dont even know what he was doing years ago.”
I asked for a link to UAH’s code. This isn’t a particularly difficult request- it either exists or it doesn’t.
“putz…slither…suck on your thumb”
It’s cute when people are so patently helpless that they have nothing to do but unleash their inner five year old.
dc:
No, they show the exact opposite of that. My values show the data goes down. That means there is a downturn. The only reason it isn’t visible in any particular graph is the resolution/zoom factor. In no way does that contradict my statements, nor if it did, would it contradict the point behind them.
You say I shouldn’t “claim the data show things they don’t,” yet I quoted the data showing exactly what I claimed. How could you possibly assume my “goal is to deliberately spread doubt and confusion by use of misleading statements” based on the fact you misrepresented the very data being discussed, something anyone can see with a glance at this post. Something you explicitly acknowledged in this post.
The only basis for you disputing my claims is that with bad resolution/scaling/zooming, it’s possible to hide the downturn in the data, and thus the downturn isn’t found. This means you’re relying on the hiding and misrepresentation of data in order to assume I’m using misleading statements.
The first update to my post explicitly states the post contains errors, and it directs readers to a comment in which I explain them. Then, at the end of the post, a second update explains the reason I’m not changing the contents of the post:
If you truly think “some deniers will lap the original post up (esp. the incorrect graphs) as proof temperatures are declining and that climate scientists are covering it up,” you’re a fool. There is no way anyone could read the current version of my post and do what you claim would happen without intentionally lying to people in a stupidly obvious way.
dc, nothing in your post is valid. Every part of it is based upon misrepresentations. Please stop doing that, or at least do it in a less obviously wrong way.
Incidentally, your posts keep landing in moderation, and this most recent one got labeled as spam. That’s why your posts are getting delayed, if you noticed that. It’s also ironically fitting.
At last, “the embarrassing Ian Jolliffe post” coatrack.
things
I have agreed repeatedly that UAH should release their code. I suggest those who want it ask UAH people for it– and FOIA it. I don’t have it and can’t give it to you.
Most recently, Eli trotted into a comments thread to ask for the UAH code. You two will generally find you get the same answers.
Brandon – you get extra kudos for getting willard to attempt to derail your post. You are doing well.
It appears in 2008, Tamino discussed using f=0.15 when analyzing GISTemp data
http://web.archive.org/web/20081229052117/http://tamino.wordpress.com/2008/12/18/lovehate-relationship/
I don’t know what this means other than in 2008 analyzing GISTemp TAmino used 0.15, but in this ‘snit’ issue he used 0.20 while Brandon used 0.15. I don’t know if Tamino has other posts where he reveals his values of f nor whether he has discussed how he goes about picking htem.
lucia:
I didn’t actually use 0.15, as I was using annual data instead of monthly data. However, if I hadn’t mixed up the two, I could have made my point by using 0.15, and showing the difference between that and 0.2.
It would be interesting to see if he ever did discuss that issue. As he shows in the post you linked to, there are ways of testing one’s choices, so there’s plenty to discuss.
Now I’m curious if there would be any actual justification for picking a particular value with the USHCN data. With how close his parameter was to the changepoint, I doubt you could ever say 0.15 or 0.2 is “better,” but it’d still be interesting to look into.
Things.
You come here asking for a link. It either exists or not.
Go look for it. Dont come here asking for us to do your damn work.
you should, as I suggested, go ask roy. If you do, i will sign on.
I will not go look for a link for you. i would not give you the link if I had it. I will, however, sign on to any request you make of roy.
Understand this is what both mcintyre and I suggest.
1. write to the owner
2. write to the journals
3. FOIA
Note what is missing from that list.
do not ask mosher or lucia to do your damn work.
If you want to see how its done go look at what magicjava did.
note what he did not do
he did not ask mosher to do his work for him.
I dont play fetch. fetch your own code. you will not ask roy for the code. not in a million years.
So, if you want to deal with people in the real world you will have to follow those steps I suggested. i dont suggest anything I am unwilling to do. i dont suggest anything I havent done myself.
You talk as if the data has been revealed by you the way it is really is, when all you have done with the inherently problematic issue of smoothing endpoints in time series in the way that you believe is correct. Tamino offers sound reasons for doing it his way, you offer counter reasons. Trendlines for time series are not raw data, they are derived using methods that vary from case to case.
You still can’t admit that Tamino was correct in his debunking of Tom’s claims that he taught at a University. Why is it so hard for you to do so. Instead you are bickering over a minor technical issue, in this case.
MikeN
You will find all sorts of people saying stupid things about the importance of getting code. They made up all manner of stupid arguments that they dont even remember. from, sharing code is bad, to we already share code. That was epic.
in defense of tamino, he did share his temperature code. it was slick. it was fast. it was well written. I have a couple people who have used it. Saved them lots of work.
bugs:
That’s the rub. Knowing what is true and what is false.
If you are forced to only teach the consensus view on a particular topic, progress becomes impossible.
Anyway, I’d be willing to speculate if his beliefs about what is true didn’t conflict with your own political and ideological goals, you wouldn’t get a rot what he taught.
So the problem here (to me) isn’t that he teaches things that might be wrong, but that he is getting in the way of the activists political goals.
So you still maintain that a drop of 0.00001 C in a line fit to data spanning > 1 C and reported to 2 decimal places constitutes a “downturn” ?
So a tiny difference in rounding anywhere along the chain of data collection and analysis that produced a 0.00001 C increase over that period would constitute an upturn?, and prove your intuition completely wrong? I suspect in that case you’d be calling it a flattening (and you’d be right).
The line flattens if you exclude the last 4 months. It does not turn down. If flattening out proves your were right all along, then say it flattens out! Why insist its a downturn, when your own values show otherwise? This is not an image resolution issue, its a significant digits issue (and common sense). Your insistence that a drop of 0.00001 C constitutes a downturn only adds weight to Tamino’s suggestion that you suffer the Dunning-Kruger effect, and to my suspicion that you trying to mislead naive readers.
Oh? What are the sound reasons for doing things his way and when did he explain what they were? Serious question. Because if you want to discuss the reasons for the choices and/or whether or not they were provided and when, we can do that.
If your methods vary from case to case and your choices result in graphs that look the way you seem to “want” them to look, people will not trust you. They will think you are biased and that you chose your methods based on how much you “liked” the answer. This is not science.
Really bugs, if you didn’t exist, people who hold views that differ from you would have to make you up just to buttress their position.
Carrick–
People rarely get their knickers in a twist because faculty somewhere teach rot. Lots of rot gets taught lots of places. It’s not entirely clear that some rot getting taught always does harm– though it’s certainy true that teaching stuff that is not-rot is better.
dc:
“Downturn” is defined as things like, “a downward turn,” or “a tendency downward.” As such, when I see a downward turn, or a tendency downward, I’m inclined to call it a downturn. Why wouldn’t I? Are you saying I shouldn’t call something what it is because it’s small?
Seeing as “flattening” is “to make flat or flatter,” I wouldn’t call an upturn a flattening. I’d call it an upturn. And I’d call what came before the upturn a flattening, just like I’d call what comes before the downturn I found a flattening. It’s that simple. When a trend changes toward zero, there is a flattening. When it changes in another direction, it is a downturn/upturn.
You can accuse me of bias all you want, just like you’ve been doing since you first posted, but you’re now doing it because I use words in the way they’re defined.
Actually, I am. To get a downturn, one must necessarily first have a flattening (technically, you can also just start with an already flat line). That means any time I say there is a downturn, I’m inherently saying there was also a flattening.
You keep saying it isn’t a downturn because it’s small, but size isn’t part of the definition of “downturn.” You can’t just declare something doesn’t exist because it’s small. And make no mistake, that’s all you’re doing.
If me using words in the way they’re defined makes you immediately think I’m over-confident and dishonest, that says a lot more about you than me.
@lucia
“I have agreed repeatedly that UAH should release their code.”
That’s nice and all, but I was just straight up asking if anyone knew if it had been posted since the intent was (if memory serves) to do so. I looked (I made this quite clear in my initial comment) and couldn’t find it. But that doesn’t mean it’s not public (I could have simply missed it my searches), and I was wondering if anyone else knew the status.
Seriously, it was no more than that.
“I suggest those who want it ask UAH people for it– and FOIA it. I don’t have it and can’t give it to you.”
Again, I am not trying to get RS or UAH to “give it to me”, I was checking to see if it was publicly available yet. I looked and didn’t see anything, but the thought crossed my mind that someone here might be keeping closer tabs on it than I do.
@mosher
“You come here asking for a link. It either exists or not.
Go look for it.”
Of course, I already did that. I already said I did that. I was hoping someone here might be able to point me to it (if I had overlooked it), or confirm that it was still not publicly available. Given the considerable interest of the commentors on this site with various temperature data sets, it seemed a reasonable place to ask.
“Dont come here asking for us to do your damn work.”
“do not ask mosher or lucia to do your damn work.”
Of course, that’s not at all what I did. You seem to think I was asking for more than I was. I was just trying to be thorough.
“you should, as I suggested, go ask roy. If you do, i will sign on”
“I will, however, sign on to any request you make of roy.”
“you will not ask roy for the code. not in a million years.”
I don’t understand why you keep saying this. I am not trying to hassle anyone at UAH about it. Presumably, they will make the code public whenever they feel ready. RS seems to be quite busy lately trying to prove surface warming is spurious, and I would never dream of distracting him from such a valuable scientific enterprise.
“I will not go look for a link for you.”
I wouldn’t expect anyone to.
“i would not give you the link if I had it.”
Why on Earth not? That seems unnecessarily petulant and childish.
“Understand this is what both mcintyre and I suggest.”
That’s nice and all, but you “suggest” a great deal more.
You stated outright that anyone that doesn’t make their code publicly available is “hiding something”, and strongly imply that they’re doing so for nefarious reasons, i.e. “they have something to hide”.
This is all rather beyond the simple request I initially made, but since you insist on expanding the scope of our discussion so much, I suppose it’s on topic. Is it fair to characterize your position on the UAH trop temp product as believing it should not be considered credible, because UAH is “hiding something”? Do you believe that the reason UAH has not made its code public is because “they have something to hide”? If not, is there a reason for this seeming assumption of good faith on behalf of UAH but not other parties?
It seems that the answer to my initial question is that the UAH code is still not publicly available. If anyone else has information to the contrary (am I simply deluded in thinking that either RS or JC said months ago they were going to make it public?), I’d appreciate a comment. Thanks.
> [G]etting in the way of the activists political goals.
.
That’s the rub. Knowing who is an activist and who is not.
The temperature record is a very noisy signal from month to month and year to year. You would expect proper smoothing to look at the long term, not the short term. I would, intuitively, as well.
“You would expect proper smoothing to look at the long term, not the short term.”
bugs, how long is long term?
willard:
I’ve never had any trouble.
My big problem with this discussion is that it neglects uncertainty. DIscussions of upwards or downward trends in the central tendency should include the statistical relevance of any such variability.
Certainly 5-years is not a very long time window for OLS.
See this.
2-years (which is what Tamino ends up with) is totally meaningless.
Lucia:
I’d say it’s a lot more certain they’re causing harm if they are teaching a course on building bridges than if they are teaching a survey course on climate change. (The philosophy of science book that has some rather embarrassing errors in special relativity in it comes to mind, from my college days…getting exposed to that was the result of having a liberal arts school education in physics).
Of course harm is a relative thing. If you’re an activist (somebody that willard at least has problems figuring out who is one and who isn’t apparently, bless him), then promulgating ideas that get in the way of “progress” can be seen as doing harm.
Example: Gleick’s apoplectic response to the title “all models are wrong” of Tasmin Edwards blog. We also learn in watching activists that they have a very different idea of what it means to “communicate” than what a scientist means (and there is a core difference about what it is they are “communicating”). For an activist “communicating” seems to be a code word for “persuading”, for a scientist I would suggest it is “communicating the truth, including its uncertainty, as accurately as possible”.
dc:
Funny… I came away a long time ago with the notion that Tamino suffers from Dunning-Kruger, and his audience Dunning-Kruger by Proxy.
MrE:
If we’re to believe Tamino did the right thing, and that he offered “sound reasons” for it (of course, bugs made that up), then the answer is simple. We should analyze as much data we can while getting an upward trend, but no more. Because that’s exactly what Tamino did.
The value he picked for the f parameter, .2, is practically the lowest value he could have picked and still gotten an upward trend. Any value higher than that would have also gotten an upward trend, but they wouldn’t have tracked the temperatures as well. Any value lower than it (actually lower than ~.18) would have given him a “bad” end to his graph.
In other words, his choice of smoothing gave the best image possible while not showing flattening/decreasing temperatures at the end. A change of just 2% would have dramatically altered his image.
And bugs would have us believe Tamino gave “sound reasons” for that.
@lucia
Will we have breathless threads about the ATI pursuing Spencer and UAH in court? I don’t think so. UAH seems to contain some mysterious kryptonite that affects auditors and traditionalists super powers of skepticism when it comes to seeking emails and source code from there. Mosher is only capable of putting his signature on a piece of paper that someone else writes for him.
This is after trusting in the UAH satellite record in 2005 led McIntyre to make a blunder that relied on him trusting their temperature record, rather than the CRU record. You would have thought that would make him want an ‘audit’.
“If we’re to believe Tamino did the right thing, and that he offered “sound reasons†for it (of course, bugs made that up), then the answer is simple. We should analyze as much data we can while getting an upward trend, but no more. Because that’s exactly what Tamino did.”
Well, I didn’t, Tamino specifically referred to the high noise level present in the signal. I presume making that parameter lower allows it to cope better with a signal that contains a high level of noise. Just read his explanation. I presume you have read it? You have been commenting on the topic at Tamino’s.
Brandon, you are still very confused.
the Tamino graph does not show an “upturn” at then end. Instead it shows a flattening increase.
even if Tamino had used your standard parameter, it wouldn t have given you the “downward” impression that your “intuition” was suggesting. Instead your own data shows, that the result would have been flat. just more flattening out, no visual downturn.
the standard parameter is no guarantee for “intuitive” results, as my examples from above show.
http://blogs.bgsu.edu/math6820kylschw/2011/03/24/lowess/
How wrong your “intuition” was, is most clearly shown by the graph including the last 4 months.
http://tamino.files.wordpress.com/2012/05/usa1yrb1.jpg?w=500&h=325
please be honest: what would your “intuition” have said about the “turn” of that one?
————————————–
ps: i also think you should correct the errors in the original post. your position is like pudding, you can claim whatever you want about which parts were wrong and which were right, when you are called out. i suggest you take a stand and stick to it!
sod, it is cheeky for you to say:
I didn’t say Tamino’s smoothed line shows an upturn, so it would appear you’re the one who is confused. The thing I referred to as an upturn was that at the end of his graph of “five-year averages.” And there’s no denying that is an upturn, not a flattening. Moreover:
I’ve never claimed there was a standard parameter. I assume you’re confused because of my discussion of a standard implementation of LOWESS and Tamino’s modified version. Even if that’s not the explanation, you’re obviously confused. There is no such thing as a standard parameter with LOWESS. In the same way:
My intuition didn’t say there would be a downward slope. It said an upward slope in that graph was weird. This should have been clear in my comments like:
That you are the one confused is especially supported by you saying:
The graph you show does not include the last four months. This fact was explicitly stated by Tamino. Given how wrong everything you’re saying is, when you say:
I can’t take it seriously. If you’re going to complain about someone being unclear, you shouldn’t make things up about what they’ve said, about what other people have said, and about aspects of analysis which have been discussed repeatedly.
Now, since you’ve said I should correct the errors in my post, presumably you should correct basically everything you’ve just said. Will you?
bugs just made me crack up:
bugs says he didn’t make up the fact Tamino gave “sound reasons” for his choice because he presumes setting the
“parameter lower allows it to cope better with a signal that contains a high level of noise.”
If that were the case, why would Tamino’s selected parameter, which was higher than the alternative which showed a different result, be better? If bugs presumes correctly, Tamino’s choice was the worse of the two possibilities he showed. And he tells me to:
He tells me to read an explanation on which he bases his presumption in order to find the “sound reasons” Tamino offered, apparently to justify the choice he didn’t make…
I swear, I have trouble believing bugs actually means anything he says. It’s too hard to believe anyone could actually say such stupid things.
My only consolation, Brandon, is that I’m not as stupid as you.
http://tamino.wordpress.com/2012/05/11/bs-and-dk/
“Why did I choose the degree of smoothing I used? Because it’s the default for my smoothing program. I chose it as the default because considerable experience in its use has shown it to be a good general-purpose value for climate time series. If I notice what looks like real signal that isn’t detected, I’ll try a shorter window, if I notice what looks like response to noise I’ll try a longer time window. If I’m working on a peer-reviewed paper I’ll probably run some statistical tests (like above) to confirm or deny what is signal and what is not. In this case, my intuition told me the default was good. I didn’t try a bunch of values until I got the result I wanted.”
All you are doing Brandon is insisting on including noise in your chart. With climate, the noise is large compared to the signal. You can argue that Tamino’s claim is wrong, but then you would have to test the use of Lowess with climate, and how accurately the setting of the parameter is suscetible to noise.
Either way, you and Tamino both demonstrate that Tom is wrong.
things-
You know that you asked a not on topic question. You know no one here is associated with UAH. It is prefectly fair for people to believe your intention is to be argumentative and thread jack. The answers to your off topic questions are perfectly fair: Go ask the people who run UAH.
Don’t be disingenuous. You can most effectively check by asking people at UAH. You are not too stupid to know this.
If you drop in a post that has nothing to do with getting SSN benefits asked me for a link to about info to help you get your SSN information, I’d tell you to google it or visit SSN. If you ask for links to UAH. The answer to your question is: Go pose your question to the appropriate party. That’s the answer.
Now, desist with this stoooopid thread jack and don’t repeat that sort of thing.
Sound reasons, sound reasons, sound reasons.
bugs
Thanks for quoting so we can read what you think constitutes a “sound reason”. What we see is that “his” program starts with an arbitrary value (which he seems to have landed on by fiddling with images until he got answers “he likes”) but he admits he will fiddles around in individual cases if he doesn’t he doesn’t like the results from his habitual initial value.
This is not what intelligent informed people would call a “sound reason” for selecting a parameter value.
In fact, what Tamino (and you) should learn, is there are sound reasons for stating the value of the smoothing parameters you use rather than leaving them unstated in the first place, or failing that volunteering them when people ask for details about how you did your smoothing. This is better than only revealing them when you absolutely must subsequent to a “snit” and giving the “reason” for your choice to be that you tend to start with a value and fiddle until you get an image “you like”.
@Lucia
R has a default, which is an arbitrary value as well. Climate deals with an unusually noisy signal. You have no idea that you can deduce his ‘motive’, are you a mind reader. Tamino has spent a lot of time working with climate data, perhaps in this case, he has more experience than you. Either way, all this is about is the final few percent of a graph that proves Tom is wrong, either way.
bugs–
The verb “seems” indicates that I am inferring not saying the thing that “seems” to be true must be true. When Tamino provides a discussion of how one determines that particular value of ‘f’ is a “good general-purpose value for … time series” in field ‘x’, I’ll believe his method is something other than “often gives smooths that look the way I like them to look”.
@lucia
“When Tamino provides a discussion of how one determines that particular value of ‘f’ is a “good general-purpose value for … time series†in field ‘x’, I’ll believe his method is something other than “often gives smooths that look the way I like them to lookâ€.”
Guilty until proven innocent.
Seems that “seems” looks like “look the way I like to look”.
What bugs me is that if they don’t agree with data they just call it “noise” and throw out data until their “signal” is supported. Secondly, that they pull out “guilty until proven innocent” but isn’t applied to the guy they were trying to prove wrong on that other blog.
Brandon, I didn’t pay close enough attention to the plots. 5 year averages, but only calculating every 5 years? Is that ever done?
MikeN:
It’s fairly common, though you see it a lot more with annual averages. It doesn’t have any real benefits over using a continuous five-year smooth other than simplicity, but that can be enough of a reason.
Love Hate was not the post I was thinking of. I’ll see if I can find it.
http://web.archive.org/web/20080430110841/tamino.wordpress.com/2007/11/05/analyze-this/
Not the one I remember, but it has the 5 year averages, and analyzing for a step change. I’ll see if temperatures since 1998 argue for a step change according to his analysis methods.
http://web.archive.org/web/20090515212733/http://tamino.wordpress.com/2009/05/11/dangerous-curves/
This is the post I was remembering. He mentions that changing the parameters of the Lowess smooth will change things from increase to flattening.
> I’ve never had any trouble [knowing who is an activist and who is not].
There lies the rub, except perhaps for Humpty Dumpty.
MrE (Comment #95697)
May 13th, 2012 at 8:17 am
The difference is that Lucia is guessing at Tamino’s motivation. A steadily rising signal, such as the temperature record, over 30 years is going to seem strange if it is going to be pointing down to an ice age from just a few years data at the end. She is not actually able to say that Tamino’s logic is wrong for choosing the values he uses.
Tamino is demonstrates that Tom is wrong with his logic and his calculations.
The next post from Tamino about Tim Curtin’s latest effort also involves Tamino pointing out that someone has their analysis very wrong. Perhaps there is another comma to quibble about in Tamino’s analysis of that disaster.
http://tamino.wordpress.com/2012/05/11/tc-and-dw/#more-5082
bug-
No. You said Tamino has sound reasons. Unless you have mental telepathy and know his reasons, the bit you quoted reveal no sound reason. What he wrote is entirely consistent with him simply picking f based on how the results look. Period.
If you have any evidence to support your initial claim that Tamino has a sound reason, please support it.
No one can say it is right or wrong because sound reasons and logic are not revealed not revealed in his explanation of his choice. But you tell us the reasons are “sound”.
If you want to explain where we can find a ‘reason’ in that and why it is sound, have at it!
MikeN, thanks for those links. If nothing else, it’s interesting to see the difference in how Tamino has behaved over the years. It’s also fun to see a time when lucia could still post on his site!
But yeah, as he shows, there are lots of choices you can make when smoothing, and it can often be impossible to say one choice is “better” than another. Given that, it seems Tamino cannot argue his choice of the f parameter was “better” than a different choice, say 0.15. That would mean he’d have to concede he made a choice which significantly impacted the visual impact of his results without disclosing the choice or the effects of the choice.
Which is exactly the point I was trying to make all along, and it is exactly what I say nobody should do.
lucia (Comment #95724)
May 13th, 2012 at 5:44 pm
No mindreading necessary, he explains his reasons, I accept them as being reasonable.
Brandon Shollenberger (Comment #95739)
May 14th, 2012 at 2:22 am
What he said before is entirely consistent with he did now.
Re: bugs (Comment #95741)
Did you go back and read Tamino’s old posts, bugs?
“But yeah, as he shows, there are lots of choices you can make when smoothing, and it can often be impossible to say one choice is “better†than another. Given that, it seems Tamino cannot argue his choice of the f parameter was “better†than a different choice, say 0.15. That would mean he’d have to concede he made a choice which significantly impacted the visual impact of his results without disclosing the choice or the effects of the choice.”
He told you the reason for his choice. The short term noise overwhelms the long term signal if you don’t select the right range of values for f.
Re: bugs (Comment #95744)
bugs,
I think we all get the picture. Either the short term noise overwhelms the long term upward trend, or the long term signal overwhelms the short term uptick. The question that continues to be dodged is whether there any reason, other than wanting to retain both the long term upward trend and the short term uptick on the same graph, to make those choices.
Tamino acknowledged then that the choice of parameter changes the shape of the curve at the end. Is he as revealing currently?
MikeN– Depends what you mean be being revealing. In the original post, he doesn’t mention the existance of parameter, his choice of parameters nor does he talk about how the parameters change the shape of the curve at the end. Mind you– that might be fair enough since his post isn’t really about the end.
Eventually after quite a bit of discussion about end points Tamino does mention the parameters affect the end.
MikeN, my answer is the same as lucia’s.
lucia, I can answer your question from a while ago. I had forgotten about it, so you may have already figured it out, but if not:
Tamino’s error bars do nothing more than show the standard error of the mean of the data. Naturally, he doesn’t say this. Since he doesn’t say that, he also doesn’t he discuss how this would affect their sizes.
I’d be curious what proper error bars would be, but I’m not sure what all would have to be considered for them. Do you know?
Is it possible to find old GISS data? Wayback failed on its crawls. I’m trying to figure out how Tamino gets his trends. I got 1975-2007 as .013 /yr and he reports .018.
Steve Mosher remarks:
“Huh. There is plenty of evidence that the US has the best quality temperature record. stop talking out of your ass.”
Most of the ushcn is cooperative with all of the strengths and weaknesses that implies (see Surface Station). Other countries (Netherlands, UK, Germany, Japan) tend to have nationally operated systems which should be better calibrated.
Add to that the time element (how many states were there in 1860 and what was the population density west of the Mississippi?) and Steve would have a hard time defending his boast.
Finally, Eli has mentioned the problem to Roy on his blog. So far crickets.
I have some.
We’ve got the computer code for the Model E too, but no one really cares about that. Still no code for UAH, but no one cares about that either.
Steve, what trend do you get for 1975-2007, if you use GISS from 2008? Tamino reported .18/decade.