The von Storch and Zorita Comment and the Huybers Comment, together with our replies, were published by GRL this week. I previously posted up on VZ here and on Huybers here, here and here and have nothing to add at this time.
Original copies of VZ is here; our Reply is here; Huybers’ comment is here and our Reply here. All rights reserved to AGU and not to reproduced without permission. VZ issued a press statement here.
Both VZ and Huybers agree that the MBH98 PC method is biased towards finding hockey sticks. VZ did not accurately implement the full hockey stick algorithm and this affects their results, so their affirmation of the "Artificial Hockey Stick" is based on only a partial implementation. VZ pointed out that our GRL simulations were in a red noise situation and argued that, in a situation where there was an actual signal, the mining tendency of the algorithm would not "matter". In such a situation, the real signal in effect competes with the mining tendency for the attention of the PC algorithm. With proxies that have a signal content, the signal will win out and even the biased MBH98 algorithm will find the signal; however, if the proxies have a very high noise level, then the MBH98 algorithm will find a hockey stick instead of the signal. The MBH98 proxies do not have the correlation to gridcell temperature hypothesized by VZ and a high noise situation applies.
If there are some "bad apples" i.e. nonclimatic hockey stick series, even 1-2 bad apples can substantially distort the result. Even 1-2 bristlecones in a network of over 50 other sites suffice to produce a seemingly "significant" PC using the MBH method. So it’s even more biased in an environment with bad apples than in a red noise situation.
As to Huybers, there is no reason to think that the correlation PC1 is a magic bullet for extracting a "signal" from the MBH98 proxy mess. It does give more weight to bristlecones – but surely that’s the issue at point. We do not say that the covariance PC1 is a magic bullet, but it’s what you would do if you were implementing a "conventional" analysis based even on Huybers’ statistical references [Presiendorfer and Rencher]. I’ve commented on this at length.
Huybers does not confront the problem of the failed cross-validation R2 statistic; I think that the referees should have required him to deal with this, but they did not. If you’ve got a failed R2 statistic, it’s only a matter of curiosity as to explaining a spurious RE statistic. We had showed how the flawed PC method was intimately related to spurious RE statistics. Huybers criticized our explanation. We modified this explanation slightly implementing further details on MBH98, including the most up-to-date information on the source code. with only slight modifications, we once again showed that spurious RE statistics resulted from simulations (which included the re-scaling step proposed by Huybers.) So our explanation that the RE statistic was spurious stands.
Neither VZ nor Huybers discuss the peculiar association of MBH98 with the bristlecones. Huybers suggests that bristlecones would be good to research "in the future". Give me a break. This should have been done by MBH98 before they published, not after.
There’s nothing in either of these comments that resurrects MBH98 or refutes any of our findings.
VZ and Huybers Comment and Reply
The von Storch and Zorita Comment and the Huybers Comment, together with our replies, were published by GRL this week. I previously posted up on VZ here and on Huybers here, here and here and have nothing to add at this time.
Original copies of VZ is here; our Reply is here; Huybers’ comment is here and our Reply here. All rights reserved to AGU and not to reproduced without permission. VZ issued a press statement here.
Both VZ and Huybers agree that the MBH98 PC method is biased towards finding hockey sticks. VZ did not accurately implement the full hockey stick algorithm and this affects their results, so their affirmation of the "Artificial Hockey Stick" is based on only a partial implementation. VZ pointed out that our GRL simulations were in a red noise situation and argued that, in a situation where there was an actual signal, the mining tendency of the algorithm would not "matter". In such a situation, the real signal in effect competes with the mining tendency for the attention of the PC algorithm. With proxies that have a signal content, the signal will win out and even the biased MBH98 algorithm will find the signal; however, if the proxies have a very high noise level, then the MBH98 algorithm will find a hockey stick instead of the signal. The MBH98 proxies do not have the correlation to gridcell temperature hypothesized by VZ and a high noise situation applies.
If there are some "bad apples" i.e. nonclimatic hockey stick series, even 1-2 bad apples can substantially distort the result. Even 1-2 bristlecones in a network of over 50 other sites suffice to produce a seemingly "significant" PC using the MBH method. So it’s even more biased in an environment with bad apples than in a red noise situation.
As to Huybers, there is no reason to think that the correlation PC1 is a magic bullet for extracting a "signal" from the MBH98 proxy mess. It does give more weight to bristlecones – but surely that’s the issue at point. We do not say that the covariance PC1 is a magic bullet, but it’s what you would do if you were implementing a "conventional" analysis based even on Huybers’ statistical references [Presiendorfer and Rencher]. I’ve commented on this at length.
Huybers does not confront the problem of the failed cross-validation R2 statistic; I think that the referees should have required him to deal with this, but they did not. If you’ve got a failed R2 statistic, it’s only a matter of curiosity as to explaining a spurious RE statistic. We had showed how the flawed PC method was intimately related to spurious RE statistics. Huybers criticized our explanation. We modified this explanation slightly implementing further details on MBH98, including the most up-to-date information on the source code. with only slight modifications, we once again showed that spurious RE statistics resulted from simulations (which included the re-scaling step proposed by Huybers.) So our explanation that the RE statistic was spurious stands.
Neither VZ nor Huybers discuss the peculiar association of MBH98 with the bristlecones. Huybers suggests that bristlecones would be good to research "in the future". Give me a break. This should have been done by MBH98 before they published, not after.
There’s nothing in either of these comments that resurrects MBH98 or refutes any of our findings.