the Air Vent

Because the world needs another opinion

Roman on How Hockey Sticks are Made

Posted by Jeff Id on September 29, 2009

There is a small discussion hidden in the latest massive thread at CA. TomP made the suggestion that correlation and elimination of non-temperature correlating tree ring data was perfectly ok with him. For those that are a little math savvy and are interested in the problem that creates read my hockey stick temperature distorition posts above. These posts reiterate the conclusioins of VonStorch04 and IMO are much more plainly worded and provide turnkey software code which anyone can run to replicate my results. However, sometimes I read things from other bloggers that just hit the point a heck of a lot better than I’ve written it. This comment by Roman is an excellent example.

If you ever wanted to know how so many papers make hockey sticks, the sorting operation described below is the reason.

The first point quoted is made by TomP.


Rejecting the Schweingruber series [as originally presented] as a good proxy seems reasonable, unless there are doubts about the instrument record. Why it is not a good recent proxy is an important but separate point.

I don’t believe that you have considered the full implication of your argument on the entire reconstruction.

Let’s suppose that you are right and that there are real treemometers which you can identify by comparison to an observed record. Your statement above indicates that there also trees that are not good proxies. Unless, false proxies are a recent phenomenon, the logical conclusion is that there must be a collection of these distributed throughout the entire time period prior to the start of the temperature record since there is no way to identify and exclude those proxies that are not good (unless you know that of course there was no MWP).

So what effect will this have on the reconstruction? Having only good ones in the modern era, we will see that the temperatures have been warming, but we already knew that. The effect the good proxies will have on the early part of the reconstruction will be merely to center it at a particular level of dimensionless chronology units. It will have little or no effect on the quality of the results prior to the actual time at which these known to be good proxies existed.

When we reconstruct the early portion, we will have a mix of good and bad proxies at most time periods. Steve’s sensitivity test shows that the result of including “bad” proxies is to flatten the reconstruction – even in the merged case, the difference is as large as a full unit, of the same order of magnitude as the range of the entire original reconstruction prior to 1800.

The amount of flattening will depend on the relative proportions of good and bad proxies, but the net result will tend towards a hockey stick shape. Any error bars constructed from the fit will seriously underestimate the bias created by the false proxies. Without knowing how prevalent bad proxies are, there is no way to adjust for this bias.

What if the choice of modern proxies is just opportunistic matching to the temperature record? Well, then you get a … hockey stick. But that’s another thread.

My bold.

This is exactly right and the more data you sort the stronger the hockey stick effect. In Mann08 1209 individual series were sorted. The high series count results in a very flat handle as the noisy data average out and a strong pre-sorted blade on the end. When each series is individually scaled for best fit to temp, the handle of the stick is straightened even further. What drives me crazy is smart people like Tom fall for the ruse – after all, why would we include data which didn’t match temp.

These PhD’s are good at math– why can’t they figure out what I did within a minute (probably seconds) after reading it the first time. I remember trashing the CA thread with comments and staying up until 3 in the morning late last August reading and thinking about the huge implications for climate science. Hell, at the time I was just trying to figure out how real this global warming scare was.

Anyway boys and girls, in the stock market you can’t choose your stocks after the trading day ends. Likewise in science, you don’t get to choose which data fits your conclusion after you collect it.

16 Responses to “Roman on How Hockey Sticks are Made”

  1. BarryW said

    This is nothing new. If the data fits my premiss it’s good and if it doesn’t there is something wrong with it even if I don’t know what it is.

    I used to read a lot about Extrasensory Perception (ESP) when I was a kid. One of the most prominent researchers of the time was Dr. Rhine of Duke, who actually attempted to do statistical research on the subject, and showed positive results. At the time his approach seemed quite good, but after having seen the critiques of the tree ring reconstructions you see a similar problem in his approach. He would take a group of volunteers and test them. Those that showed statistically positive results continued in the study and those that didn’t were removed. The argument was that some people had ESP and some didn’t. When one of the subjects stopped showing positive results it was assumed that they had “lost” their ability for some reason, or if they showed an inverse correlation that was also deemed significant. All logical, but statistically and physically wrong.

  2. FrancisT said

    Yes I think I need to add this to my explanation as a sort of background note. I also note that Steve M has a new post up explaining much the same thing

  3. Kenneth Fritsch said

    Jeff ID, I think this thread catches the essence of the misconception of Tom P and other bright people, that includes climate scientists, about the subtle points of selecting data for a reconstruction. Including RomanM’s replies was exactly in line with my views on the best counter for this matter.

    I suspect if some of these authors of reconstruction papers were as forthcoming as Tom P, we could readily point to how this misconception weakens (invalidates) their entire approach.

  4. Layman Lurker said

    If there were complete archives of all samples taken, then it might be possible to shed more light on the nature of the population. For example, in any given year, if the sampled population contained only noise plus the “divergence” signal, then a frequency plot would show a distribution with a population mean centered around that value. If the population containted contained a second “temperature” signal, then the frequency distribution would show a second hump with a mean centered around temperature. The extent to which these poplation distributions overlapped would provide information that could help interpret a level of confidence in the temp signal.

    The stronger the divergence signal relative to the temp signal in a population (assuming that the distributions overlap) – the more likely it is that samples retained on the basis of insturmental correlation are false “noise”.

    Of course it is possible that a tree ring population contains no signal at all (lousy thermometers) – just noise. Or it is possible that it contains only a divergence signal and not temp. In both of these populations there might be noise which correlates with the insturmental record. The point is that without complete archiving of the sample population, no inferences can be made.

  5. MikeN said

    Perhaps you should demonstrate with a look at stock market daily values.

  6. AndyL said

    Here’s a dumb question, but here goes:

    We know that noise plus calibration creates a hockey stick, with a small dip before the blade. We know that calibrating say tree rings in modern period produces a similar shape. So here’s the question: Is there any way we can subtract a hockey stick created from an appropriate set of noise from a hockey stick created from calibrated tree rings, to leave us with just the signal?

    The equation would be something like “random noise with signal minus random noise = signal” – kind of like noise cancelling headphones.

    See – I told you it was a dumb question

  7. Jeff Id said

    #6 that’s my top secret project. I’ve so far been unsuccessful due mostly to lack of time.

  8. Kon Dealer said

    I guess that the same people who support the “statistical methods” that generate hockey sticks- i.e those that remove “bad” data from the analysis would have little time for a pharmaceutical firm that excluded “bad” data from a drugs trial.

    They would righly call it fraud- and that is exactly what it is.

  9. Bill Flastic said

    If there is a good scientific reason for some trees to be good thermometers and others to not be, then sorting the data would be seen as simply selecting the good trees.

    Then it is not just a matter of statistics and correlations.

    Because of their location, or something, some trees will respond more to temperature and others to, say, moisture. Sorting out the trees that match the instrumental record is then seen as finding the right trees. Once you’ve found them, you can use them as far back as they go.

    This would not apply to fossil trees, however.

    Of course, this still leaves the problem of matching the instrumental temperature by chance. And it would be necessary to record the exact location and altitude of each tree, and the instrumental temperatures in that area. Further, the trees should be re-sampled ten years on to see if they still followed the instrumental record.

  10. DeWitt Payne said

    There’s a classic scam involving selection of subjects. You send letters to 8192 (or some larger 2^^n people) and tell half of them that Team A will beat Team B and the other half the opposite. The next week you send letters to the 4096 people that were told the correct winner and tell half that Team C will beat Team D and the other half the opposite. Do that for five consecutive weeks and then ask the remaining 256 people to send you money for your next prediction. After all, you’ve been correct five times in a row.

  11. Layman Lurker said

    #6 & #7

    Jeff, would you create a psuedo scenario with assumptions about noise, temperature, etc to demonstrate?

    Another potential source of error is accuracy of the temp trend which you would calibrate to. If the temp trend is biased or wrong (ie arctic average instead of local Yamal peninsula like Kenneth has alluded to). This effectively would introduce more noise and limit signal during calibration – adding to hockeystickness.

    A cynic would say that the Team WANTS noise captured in the calibration to make sure that they don’t get an MWP in their reconstruction.

  12. Bill Flastic said

    And, of course, if there is “Divergence”, then the tree is not a good treemometer either.

    Nevertheless, there could be a scientific basis for choosing one tree over another.

  13. BDAABAT said

    RE: #9
    Bill Flastic, that’s not correct. You can’t just take samples of a bunch of trees, analyze the data, find the desired signal in some samples and not others, then throw out all of the others that don’t fit your desired outcome. That isn’t science. But, that appears to be exactly what was done in these “studies”.


  14. Jeff Id said

    #9 Actually, I agree with you. You would need to be careful not to over select and choose noise instead of temp trees but stats can help with that. Chuck too much and you’ve crossed the limit of your ability to tell signal in the noise and the data fails the null test for random data.

    The key though for your concept would be to not extend the record beyond known temp data as the historic data is still completely unsorted. The historic extension cannot be attached to the sorted data and becomes completely invalid as an extension of the sorted data.

    #11, I’ve done the scenario of noise and signal in the hockey stick temperature post #2. Link in the bar above. Perhaps I don’t understand your request.

    BTW Lurk, your point about an improved result from more noise in the data is highly valid IMO. The handle of the stick becomes flatter in these sorting schemes with more red noise.

  15. […] Roman on how hockey sticks are made […]

  16. “I bought my childhood home. ”
    #9 Actually, I agree with you. You would need to be careful not to over select and choose noise instead of temp trees but stats can help with that. Chuck too much and you’ve crossed the limit of your ability to tell signal in the noise and the data fails the null test for random data.
    This made me tear up…….

    And that is some serious paneling, my friend.thank u post…..

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: