Comments on: Re-visiting the AWS Recon PCs

By: Geoff Sherrington

Geoff Sherrington — Fri, 20 Feb 2009 23:18:55 +0000

Is it possible to do a sensitivity analysis by artificially changing what you think are the major influences on the PCs, to see if the PCs change in predictable response?

Also, re the use of one site to estimate its effect on another, there are textbooks on methods for handling grades in ore deposits. The enclosure for points to be used in a calculation need not be assumed circular and the weighting need not be assumed inverse square. Assumption can sometimes be improved by derivation. Textural features like the dominant direction of terrain features can have an effect, so a search ellipse can be more appropriate than a circle. That old saw that correlation does not have to be causation enters the arguments.

By: Mark T

Mark T — Fri, 20 Feb 2009 22:18:13 +0000

It probably effects confidence intervals. I’m probably sure this is a sticky wicket with many around here, too, hehe.

Mark

By: bernie

bernie — Fri, 20 Feb 2009 22:05:41 +0000

Re: Ryan O (#50), In that case aren't there rules to determine whether you can actually even use PCA? Normally I would never trust a PCA if the number of cases were not at least twice the number of variables. In this instance if there are 600 monthly records there should be around 1200 real cases -- I am not sure of the impact of infilling of values. But again you guys are the pros on this. I am waiting to retire so that I can catch up on all of this again.

By: Ryan O

Ryan O — Fri, 20 Feb 2009 20:31:11 +0000

Re: bernie (#45), The problem is that as data sets get sparser or shorter (i.e., more missing information), the chances of spurious correlations increase. The idea would be to provide an externally imposed weighting system to discourage spurious correlations between stations that are unlikely to be causally related.

By: Mark T

Mark T — Fri, 20 Feb 2009 20:29:21 +0000

Re: bernie (#45), You'd think so, that is, that there'd be some physical indication in the PCs. I don't know how to specifically address these issues, btw. I haven't given them much thought other than the cursory stuff mentioned here. Mark

By: Mark T

Mark T — Fri, 20 Feb 2009 20:26:00 +0000

Re: David Jay (#44), ? Which comment are you replying to? The atmospheric one? If so, not particularly, though that is a relevant atmospheric phenomena, but a bit different than what I was getting at. I was thinking in particular of the ducting phenomena in which signals bend around in the atmosphere (which is different than ionosphere skipping). The inverse R^2 law doesn't directly apply in such cases, when calculating the path loss from one point to another. Mark

By: Steve McIntyre

Steve McIntyre — Fri, 20 Feb 2009 20:08:24 +0000

Re: Luboš Motl (#43), Luboš, the article says these things but provides no evidence supporting any of these claims. The PC3 as others have observed is astonishingly unphysical and totally at odds with the claim in Nature.

By: UC

UC — Fri, 20 Feb 2009 20:07:06 +0000

Re: Luboš Motl (#43), They have allocated 60 stations for the whole Earth, 2 should be enough for Antarctica.

By: bernie

bernie — Fri, 20 Feb 2009 19:52:51 +0000

Re: Mark T (#42), and Ryan I am not sure I understand where the spatial thing comes in for the current data set and resulting PCs. If, for example, the Antarctica Peninsula had a discrete climate trend pattern then shouldn't that show up as all or a significant feature of one of the PCs - even if a few of the stations were odd-balls - inspection of a rotated factor would show that this bunch of stations all weighted heavily on one of the factors. Don't you have to infer the relationship with some other non-measured variable at least initially by inspection. I am probably approaching this too simplistically.

By: David Jay

David Jay — Fri, 20 Feb 2009 19:47:43 +0000

Mark:

Are you thinking about signal processing for backscatter, for instance?