#302 Steve Fish

Steve, first of all I want to state outright that I have no problem with RC or how they conduct their moderation. If I try to post something and it doesn’t go through then so be it.

You mentioned that some of the comments got through. If you are talking specifically of the “you can’t get there from here” thread I don’t think anyone got through (if I’m missing something point it out to me). Jon P posted a comment that I saw got through initially, but it later disappeared or was removed. I think most of us had links in our comments to the tAV thread which may have been the reason they did not get through, or in Jon’s case – removed.

]]>I haven’t yet partitioned up my grid into the various regions. No real reason for that . . . I just haven’t gotten to it yet.

#68 I just can’t figure out any sensible logic behind the primacy of RE. It really doesn’t make any sense. The more I think about it, the more stunning it is to me that WA and AW got published in the first place, as that argument was essential to both of the papers.

]]>http://www.mmm.ucar.edu/events/antarctic06/presentations/weidner_aws.ppt

]]>RE and CE as terms come out of the tree ring literature – see Fritts for example. The term “Skill Score” in meteorology sometimes matches RE. There are occasional uses of RE under different nomenclature in economics (I think Theil in the 1970s.)

The conceit that RE should rule uber-alles occurs nowhere in any literature prior to the revelation of the MBH failure with other verification statistics – whereupon Wahl and Ammann, writing up Mann’s arguments, advocated RE uber alles.

]]>Figure 8 above shows your all-Ant. recon, with slopes that are positive but insignificant (as adjusted for AR(1)? — as I showed on CA, Steig neglected to adjust for s.c.), even for the full period 1957-2006.

I trust your final product will show similar graphs for each of the 3 regions, including in particular W. Ant., which was the big deal in Steig09. Can you show this now, or are you saving it? ]]>

r^2 is a verification period statistic; R^2 is a calibration period statistic.

I see the difference in notation as you apply it above. I have been too sloppy with my interchangeable use of r^2 and R^2. Thanks for setting me straight.

Correction to Post #62 above: replace R^2 with r^2.

I believe it was Cubasch and Burger who pointed out the history of using RE and CE and its limited to non existent use beyond a couple of fields. Also they pointed that CE and RE should have the same means if they come from the same distribution and implying if the do not something could be wrong with the reconstruction. I suppose that assumes a prior detrending or a constant trend.

]]>I speculated to Jeff a long time ago that the low eigenvector weights of surface station areas were responsible for the “un-natural” pre-1982 Steig PC3 artifact. Spatial grid weighting of station input data constrained this effect and eliminated the un-natural PC3 (I think) by imposing spatial discipline on the station inputs. If this is true, then perhaps an argument which demonstrates the effect of low eigenvector weighting of stations (maybe using Steig’s PC3) could be considered in your publication.

]]>This is a direct measure of the least squares goodness of fit of the regression model that is achieved here by the minimum AICc criterion. However, R2 is known to be a very poor, biased c measure of true goodness of it when the regression model is applied to data not used for calibration (Cramer 1987; Helland 1987), hence the need for regression model verification tests.

Among the verification statistics used here, the CE is the most rigorous. The only difference between the RE and CE lies in the denominator term. However, this

difference generally makes the CE more difficult to pass (i.e., CE . 0). When xy 5 x c, CE 5 RE. But when xy ± x c , RE will be greater then the CE by a factor related

to that difference. This follows by noting that for the CE, the sum of squares in the denominator is fully corrected because xy is the proper mean. However for the

RE, the denominator sum of squares will not be fully corrected unless the calibration period mean is fortuitously identical to the verification period mean. When this is not the case, the denominator sum of squares of the RE will be larger than that of the CE resulting in RE greater than CE

.

]]>The Cook (1999) paper has a mathematical definition of R^2 at the end:

]]>If my memory serves me well, my understanding is that Steig et al. employed the principle components method to solve the problem of rank deficient in the data covariance or correlation matrix when imputing missing values. Isn’t the truncated SVD method is just a principle components method? And an issue with TSVD is the “component-cutting” or “eigenvalue discarding.” A straightforward method in selecting the number of components retained is to minimize the MSE. Well, I cannot tell exactly how the TSVD is applied by reading this post. Anyway, it’s great that you would like improve upon existing research. Good luck!

]]>