Stats Made Easy

Practical Tools for Effective Experimentation

Sunday, February 24, 2008

Models snowed by vagaries of winter weather

When warned of winter storms, my wife must pay attention because she teaches preschool, which might be canceled if enough snow falls. The weathercasters like to tell us where their models predict the heaviest bands of precipitation, but often these fall all over the map. The TV meteorologists then say the "the models don't agree" as a hedge against being blamed for a bad forecast. My standard joke to my spouse, who just wants a simple answer on the amount of snow to be expected, is that of course you cannot get models to provide consistent insights on such complicated natural phenomena – they are far too busy primping themselves for their next photographic shoot! (In my mind I always picture at this point the shallow character Ben Stiller played in the movie Zoolander and his inane arguments with fellow models.)

On a more serious note, it came to my attention (a bit belatedly) that the American Statistical Association (ASA) issued a Statement On Climate Change change several months ago. It included this statement: “The design and analysis of computer experiments [DACE] is an area of statistics that is appropriate for aiding the development and use of climate models. Statistically based experimental designs, not currently used in this field, could be more powerful.” I added the acronym DACE as the shorthand for an approach that seems to be getting more-and-more attention as simulations increasingly complex. The objective of DACE is to produce a transfer or surrogate function that provides an adequate approximation of what the simulation actually predicts. Because these computer programs often are very costly to run, a model of its output can be very valuable for taking short cuts to areas of primary interest by researchers. For example, see this DACE done by Canadian hydrologists studying the Smokey-River watershed in the Edmonton, Alberta area.

I also found a website offering Postdoctoral Opportunities in Statistics at the National Center for Atmospheric Research that includes a number of nuggets for further data mining. The abstract for a project on Estimation of Climate Model Parameters, notes that this work is “a novel implementation of new methodology called ‘Design and Analysis of Computer Experiments’ (DACE)” and that “some computer experiments will always be too expensive to run, so one must be judicious in the experiments that are run.”

In any case, I liked all the colored pictures of contour maps, radar images, hurricanes and tornadoes. It looks compelling enough to distract even a real model from their mirror.


Post a Comment

<< Home