Saturday, April 14, 2012

Fight Club

As I predicted a little while ago, this manuscript has attracted a lot of discussion:

It seems that ESD is turning into a bit of a hotbed of debate with plenty of controversial papers. It's interesting to see the review process out in the open, especially in these sort of disputed cases. Too late for anyone to add any more comments now. Unfortunately the second phase of revision and re-review will be hidden from view (this being the standard policy of these journals - I'm not sure why really). So if any commenters have anything else to add, they can do it here instead!


PeteB said...

so can you say who do you think is correct on the r^2 issue ?

James Annan said...

Oh dear does that mean I have to read it all? I was hoping to keep out of this one :-)

Paul S said...

I think Dessler's assertions about r^2 are generally incorrect, but could be reasonable in a particular context.

He's right that a zero slope will always give you 0 (with the exception of slope applied to a dataset with 0 variance, which seems to give me NaN).

It's not generally correct that higher slopes give higher r^2 values. However, if you have two datasets with the same variance properties but different slopes, the one with the larger slope will give a higher r^2 value (with the exception of 0 variance again). I think this is simply because larger slopes carry a more detectable signal.

Carrick said...

I'm not sold on ECMWF or NCEP over the actual surface data. There are difficulties in the physical data sets, but that's from lack of data, and I don't think those can be cured by including a physical model.

(Put another way the only way you can validate the physical model is by having the data to compare against.)

Whilst its technically true (James made the point here) that you're interpolating to the pole, that is only a meaningful distinction if the physical processes are the same in the region where you have data s the regions where you don't, and for the polar region that is substantially not true--as is clear to any causal observer, meteorological phenomena are strongly zonally stratified.

The saying about making silk purses from sows ears comes to mind here.

EliRabett said...

Paul, wrt r2, what it measures is how well you can assign y if you know x. If the slope is 0, any y is equally likely given any x. If the slope is higer, then you have a better ability to assign y given an x.