A. FYI: my request for a citation (the specific file is great!) was because “Woodhouse et al 2006″ is not sufficient to tell what paper it is (even if I go to a uni library and look for a hardcopy).

B. Yes, I’m aware that all of the PCs (which are retained) get same weighting. This was my complaint when Ross would show egregious PC1s as examples of data mining and not show the other ones.

C. You may be tired, but tough, Steve. I KNOW that you mentioned it in your EE article. I’m not gigging your knowledge. I’m gigging your fairness in debate. My complaint was on the blog where individual posts and discussions were misleading because of not clarifying the difference. Becuase the impact on a suspect PC is an OVERSTATEMENT of the impact on the overall result! I think one of the John’s (Cross or Hunter) gigged you for actually referring to a PC as a reconstruction and I had to persevere to get the clarification–I did not get an answer to this question when I first asked it a long time ago. So I’m sensative to this and will watch it. And I guess…other things as well.

D. I understand that the post is incidental. My questions are reasonable ones from someone trying to understand what “take-aways”, what inferences to draw. As opposed to number 15 jumping up and down and cackling over a PC4.

E. Jean, I understand that Preisendorfer’s n may be arbitrary. I’d still like to know it in this example. I think what readers need to realize is that “Mannian” PC1 (what Steve and Ross refer to often when asked about mining by readers on the blog) looked like the mean. Steve didn’t show it. Then he looked at PC2…no, still no hockey stick. Then he looked at PC3…still no hockey stick. Then at PC4, he saw a mild hockey stick. And posted that. Just going down to the 4 is interesting (given that there are other times when Ross makes the argument that PC1, while not mathematically special in the reconstruction, is important as the “dominant mode of variation. Similarly, it’s interesting to think just how far Steve went to get his “bad to Mann looking” example. Did he go past Preisendorfer’s n? And then, was it really the “Mannian” nature of the PCA (the off-centering which in the free world is unique to Michael Mann) which led to the restult on PC4 or was it PCA in general. Finally, I think if you look at orbitals (which are decompositions in to parts of the density functional pattern of electrons) you will see some interesting things–symmetric, anti-symmetric modes, etc. So it does not surprise me that individual PCs have some funny patterns. (Note that this also indicts Mann for touting PC1). But I’m about truth. Not about only gigging Mann and never gigging Steve.

Moreover, the Preisendorfer’s rule is IMHO a more or less ad-hoc procedure not much in use outside of the “climate field”.

I didn’t think it was much in use *inside* the climate field either. Didn’t it get mentioned only as a (presumably) post-hoc justification for including the bristlecone PC in Mann’s analyses?

Mark

4a. I get really tired of you “taking credit” for attending to the reconstruction. Our E&E article has a lengthy discussion of the impact of various permutations and combinations on MBH reconstruction. We referred to this in our GRL article. We subdivided the discussion into two papers since, in keeping with much advice to focus academic articles on individual points, we wanted to keep each article to relatively bite-size points.

c. In the regression phase after the PC phase, the regression phase doesn’t care whether the series comes as a PC1 or PC4; both get the same weighting. If you use covariance PCs and 5 PCs (AS WE OBSERVED IN OUR EE ARTICLE), you get a HS shape in the MBH reconstruction because the bristlecones impart a HS shape to the PC4 which now gets in and is overweighted in the reconstruction.

b. Both PCs and the Mannian twist have some odd properties. In this case, I haven’t explored the matter to see whether the Mannian data mining specifically contributes to whatever HS-ness we see in the PC4. That could be thrown up by the PC method seeking series orthogonal to the 3 higher PCs. I’d suspect that the Mannian weights would, in this case, have something to do with it. If I pursue this post past an incidental post, I’ll look at it, and might look at it anyway some time in the future, but I’m chock-ablock with higher priority items right now.

I realize that my comments are not new logical insights to you. My point is that to look at this example fairly, those are the questions that must be answered. This is both for my own sake and for people like #15, who are jumping to conclusions based on an example where you picked the PC4!

1. Could you please answer the citation and tree-line questions from #3?
2. When adding content like the Jean S remark would be useful to clarify with update or better yet do new post.
3. On that material:
a. why does the displayed PC4 have a different start time then the overall average?
b. Is the simple average that you show, that derived by Woodhouse or you? If by you, what does Woodhouse do in their paper (do they have some weighting scheme)?
c. How much do the HS indices differ from your displayed PC4 and simple average. (It does not look visually stunning). What are the HS indices of PC1, 2, 3.
d. What is Preisendorfer’s “n” for this example.

4. Criticism:
a. I’ve had some influence…based on the comment on result in reconstruction versus just the PC.
b. Usually though, you (or Ross) cite purely the PC1 as your example. This does not seem such a dramatic example of data mining. In fact, in this example we are right at the PC4, which usually you scoff at!
c. Why would being in the PC4 drive a reconstruction to be hockey-stickish. We are entering deep into the metaphysical. But, I don’t see how this would do so if this data combined with larger data or even if this were the only data. If you are going to make that hypothetical point, you need to show how the average differs from the average of the PC1/PC2/PC3/PC4/…PCn.
d. You are subtley conflating different issues with the example here of what “Mannian PCA did”. Your readers are used to seeing that term used to indict the thing that was bizzare about it (off-centering). However, in this example you’ve just done, we don’t know to what extent the “Mannian aspect” (off-centering) drives any data mining and to what extent it is variance normalization driving things or even just PCA itself (and PCA itself must “mine” in that it selects some of the series more then others).

From online search :

Washington’s temperature records begin Jan. 1785
Some of his extremely cold readings may indicate that the thermometer was outdoors. (they are not sure of others..my comment)

He wrote on 5 Feb. 1788 of weather so cold that the mercury did not rise out of the bulb of the thermometer all day. But he was writing about one of the coldest days of the century, when near Philadelphia the temperature registered–17° F.