I haven’t read them properly yet but its good to see Tingley get a mauling in the rejoinder.

It’s a shame the editor felt the need to write:

“Thus, while research on climate

change should continue, now is the time for individuals and governments to

act to limit the consequences of greenhouse gas emissions on the Earth’s

climate over the next century and well beyond.”

**[RomanM: Sounds like a good excuse for a new thread, Steve. I see there is a submission from Ross and yourself in the mix.]**

Tingley claims the LASSO is inappropriate for multi-proxy studies because LASSO is a method for sparse regression (i.e. most appropriate when only a small subset of the possible covariates are expected to have a true effect). Tingley notes that the LASSO is equivalent to putting a double-exponential prior on the regression coefficients and that this does not make sense from a scientific perspective. He is correct that, for a fixed value of the bounding parameter, the LASSO is equivalent to such a Bayesian analysis. However, M&W use k-fold cross-validation to choose the bounding parameter. If, as is claimed, the temperature proxies are all highly informative, the cross-validation would find that very little bounding is needed (i.e. tend towards a standard unconstrained regression with lambda =1). Hence the choice of bounding parameter is highly data driven, meaning the actual procedure is nothing like using a double-exponential prior.

Tingley in the detailed version of his discussion, tries to show how poor M&W’s use of the LASSO is by using the LASSO but he states

“I do not perform the cross-validation procedure used in MW2010 to determine

the LASSO penalization parameter (lambda on page 13 of MW2010). Instead, I use the default setting of the glmnet package, which sets lambda to be 0.05 times the smallest value of lambda for which all coefficients are zero. The LASSO penalization is thus very small.”

Hence instead of doing what M&W have proposed, he is making sure the regression will always be really sparse. Unsurprisingly, he then finds that his version of the LASSO performs badly.

It is a classic straw man argument. However, I suspect Tingley doesn’t understand the importance of the cross-validation step, because what he has done is so easy to rebut and so obviously wrong to anyone with knowledge of the LASSO.

]]>I am not sure of your point. The second paper you reference certainly mentions M&W but does not contain any criticism of the paper. I am also not certain why you would use the adjective “flamboyant” to describe the authors. This paper appears to have a refreshing objective of trying to bring modern statistical methods to climate science. If anything, it is a (mild) criticism of “mainstream” reconstructions for not making use of statistical knowledge available.

]]>http://www.people.fas.harvard.edu/~tingley/

Tingley, Martin P. Spurious predictions with random time series: The Lasso in the context of paleoclimatic reconstructions. A Discussion of “A Statistical Analysis of Multiple Temperature Proxies: Are Reconstructions of Surface Temperatures over the Last 1000 Years Reliable?” by Blakeley B. McShane and Abraham J. Wyner. Submitted to the Annals of Applied Statistics pdf. A more detailed version can be found here.

Tingley, Martin P., Peter F. Craigmile, Murali Haran, Bo Li, Elizabeth Mannshardt-Shamseldin and Bala Rajaratnam. Piecing together the past: Statistical insights into paleoclimatic reconstructions. Manuscript submitted, and currently available as Technical Report No. 2010-09, Department of Statistics, Stanford University. pdf

]]>RE = 1-MSEx/MSEy = (MSEy-MSEx)/MSEy

Here MSEy is the mean squared error from the simple mean, or intercept, model y, and MSEx is from the more complex model x.

Now, if model x was actually a submodel of model y with parameters fitted by least squares, then RE would be like an F-statistic, and necessarily positive (so would not be a sensible threshold for significance).

But model x and model y are apparently measurements from a “holdout”, or “verification” period, using parameters estimated from a distinct calibration period, and hence RE.gt.0 is not mathematically enforced. Even so, given that model y presumably has more parameters and fewer degrees of freedom, one should still expect RE to be bounded away from zero to be significant. One might expect the threshold to depend on the degrees of freedom, as in an F-statistic.

I’d be grateful for any clarification that experts can add to that (vague) assertion.

Rich.

]]>“On the other hand, limiting the

validation exercise to these two blocks is problematic because both blocks

have very dramatic and obvious features: the temperatures in the initial

block are fairly constant and are the coldest in the instrumental record

whereas the temperatures in the final block are rapidly increasing and are

the warmest in the instrumental record. Thus, validation conducted on

these two blocks will prima facie favor procedures which project the local

level and gradient of the temperature near the boundary of the in-sample

period. However, while such procedures perform well on the front and

back blocks, they are not as competitive on interior blocks. Furthermore,

they cannot be used for plausible historical reconstructions!”

This throws the spotlight back on the instrumental temperature record and whether the above statement is correct. There continue to be many reasons to suggest that the temperature record is inaccurate. Unless the relation between a proxy response and a local instrumental temperature is accurately characterised, we can get the instrumental era tail shaking the millennium dog. (Accuracy should not be confused with correlation or precision). There are dangers in accepting the temperature record at face value.

BTW, it would be an interesting analysis if both the initial and final blocks were horizontal lines of similar averages. Many such discrete locations exist as weather stations from 1900 onwards. We might end up with a historic record from books, showing MWP agriculture in cold places, with a statistical reconstruction showing an invariant temperature for over 1000 years. I mention this only in support of the M&W statement that problems arise because the typical proxy response to temperature is weak – and it destructs at this limiting case.

]]>Hi,

I just noticed a little error in your Blogroll, it´s “Klimazwiebel”, not “Klimazweibel”.

(feel free to remove this comment)

greetings from Germany, UL

]]>