date: Wed Dec 20 16:44:41 2006
from: Tim Osborn
subject: Re: Choice of model selection for expert elicitation
to: Saffron O'Neill
looks like a suitable choice to me.
Tim
At 14:54 18/12/2006, you wrote:
Mike (and Tim)
I've decided which model to use for the time series component of the elicitation.
Tim and I worked this through and after spotting a few - now corrected! - errors, we
have come up with the following plot (see the second page of the attachment
'seaice_anncyc')). This provides a more thorough method than simply estimating the model
to use 'by eye' - several models (orange, red and pink triangles, light blue and dark
purple circle) appear in the bottom left as showing the least deviation from the
observed and model mean data.
Taking this information and looking at the 'change' graph on page 1, the dotted orange
line appears closest in shape to the model mean change for the seasonal cycle pattern
(especially in the shoulder seasons of melt and freeze up, when the calculations for the
[50%] time series plot to be used will be most affected).
I've checked the timeseries for Hudson Bay, as an region I know a little about now from
the polar bear lit, and the present day data appears reasonable.
Therefore, I will be using the Max Plank echam5 model (unless either of you object to
this). As echam5 has more than one model run time series, I will use the first model run
for each region for consistency. (I've also attached a pdf showing 3 of the regions,
scroll down to the bottom of page 5 for echam5 timeseries).
Saffron
Tim Osborn wrote:
Some further comments:
At 17:43 29/11/2006, Mike Hulme wrote:
My thoughts on this below ....
At 16:20 29/11/2006, Saffron O'Neill wrote:
Tim and Mike
I'm sorry, I did get my wires crossed on that one. To make things clear, I'll put the
maps I need in order as per Mike's email:
(a) model mean spatial map of change in [50%] 'ice-free' length in months
i.e., (max = 12, min = -12) but this means mostly positive values, i.e., general
increases in 'ice-free' length.
yes, -12 to +12, but in fact most values look to be between -1 and +6 months
(b) The time series - one per key region but preserve IAV - Mike, you have suggested I
use the one closest to the model mean here, in which case I need to have figures for the
mean for each of these time series sets and perhaps perform some kind of analysis on
this to pick the closest model to the mean. Or, as I think Tim is suggesting, should I
pick a model by eye that is closest for each regional set to the median value for that
region?
I would have thought choose the model that in its mean seasonal cycle response (when
averaged across the whole Artic) is closest to the model-mean, i.e., choose one model
and stick with it (one could choose different models for different regions, but this
seems too complicated).
I would concur with Mike, that seems simpler to choose one model for all 6 time series.
You now have the data for the annual cycle of changes in total sea-ice area over the
Arctic from me, so you could use that to select a model that is (a) fairly close to the
observed annual cycle in the present-day period; and (b) fairly close to the model-mean
change in the annual cycle. I would just use criterion (b) if it results in choosing a
model that has the target change, but starts from a poor present-day simulation.
(c) Instead of relying on the [50%] value (although this is backed up by the lit. I
think it would be good to provide a little more information here, otherwise as you say
Tim, the same judgment will be given to a 52% to 47% change as an 80% to 15% change). So
instead, a model mean map for Sept and for March of change in sea ice concentration (NOT
sea ice extent).
If I understand it the scale would be from (near) infinity through zero to -100 with
most values on the negative side (i.e., change in concentration in a cell in March from
87% to 63% = -24%, where a change in concentration from 1% to 4% = 400%). Fine, but the
colour scale and legend may need skilful choosing.
Mike -- you are right about choosing the colour scale/levels carefully if the change is
expressed as a percentage of the initial value, i.e.:
%change = 100% * (future - present) / present
because if present is small then %change can be very large. But if this is simply:
change = future - present
then the values will be in a more reasonable range.
Saffron -- when I've had time to get the individual model changes regridded onto a
common grid, ready for averaging into a multi-model mean, perhaps you might want to come
over and we can try various plotting options on screen so you can choose what you think
conveys the necessary information in the simplest/clearest way for your purposes. It'll
probably be Monday or Tuesday next week. Sound ok?
Cheers
Tim
Dr Timothy J Osborn, Academic Fellow
Climatic Research Unit
School of Environmental Sciences
University of East Anglia
Norwich NR4 7TJ, UK
e-mail: t.osborn@uea.ac.uk
phone: +44 1603 592089
fax: +44 1603 507784
web: [1]http://www.cru.uea.ac.uk/~timo/
sunclock: [2]http://www.cru.uea.ac.uk/~timo/sunclock.htm
--
---------------------------------------------------------------
Saffron O'Neill (PhD Researcher)
Tyndall Centre for Climate Change Research
Zuckerman Institute for Connective Environmental Research
School of Environmental Sciences
University of East Anglia
Norwich, NR4 7TJ
United Kingdom
T: +44 (0) 1603 593 911
F: +44 (0) 1603 593 901
E-mail: s.o-neill@uea.ac.uk
Web: [3]http://www.tyndall.ac.uk