cc: date: Fri, 30 Jan 2009 15:36:30 -0500 from: "Scott Robeson" subject: RE: More Thoughts to: Phil, Thanks for the helpful comments. Interesting week here -- 30+ cm of snow, which is not typical for southern Indiana. I have picked up the new datasets and run them through 2008. - The v version showed some differences from the previous results -- the difference in the trends between the percentiles actually seems to be larger in the v version. Please see the attachment. Note that these are for the HadCRUT data (previous results were for CRUTEM). I also did a time series of 90th minus 10th percentile time series and the post-war discontinuity in SSTs seems evident here. - Yes, you're right about the high latitude areas (more variable regions) driving the high and low percentiles, but data from all over the globe still contribute somewhat. So, in a sense the trends in the percentiles are most representative of these high-latitude regions. I could start to do some regional analyses, but I'd like to keep the focus on how spatial variability is changing across large spatial scales. Perhaps a hemispheric analysis might be useful along those lines and at least it would ensure that something like having all the 10th percentiles from the SH and 90th percentiles from the NH isn't happening. - Fig. 2 is more erratic since the 1970s as the trends are calculated over increasingly shorter time periods. The last several points on that graph are only for about 30 years while the first ones are for the whole 1881-2008 period. The trend analysis still uses the monthly data, but I just calculated one trend per year (Jan 1881 to Dec 2008, then Jan 1882 to Dec 2008, etc.). So, the original figure caption was misleading in that it didn't mention the months used. - I had thought of a fixed grid analysis too -- then we would know if the changes are due to the inclusion of a larger number of more-variable regions later in the record or to "real" changes in the structure of the thermal anomalies. When you say 90% complete time series, do you mean that a grid point is included if it has 90% data available for a given time period (i.e., excluded if it has more than 10% missing)? Thanks, Scott -----Original Message----- From: P.Jones@uea.ac.uk [mailto:P.Jones@uea.ac.uk] Sent: Wednesday, January 28, 2009 11:28 AM To: srobeson@indiana.edu Cc: willmott@udel.edu Subject: More Thoughts Scott, I picked up a copy of your docs/pics etc in Norwich on Monday. I've now had a read through, so here's some thoughts from Switzerland. I think running the v version through would be worthwhile, as a sensitivity test. If it shows little difference, then you have something that is quite robust. I don't think there are many data issues, just coverage changes. Can you run with a fixed grid - say 90% complete time series over the 1901-2007 period? I'm still wondering where the big increase in 90th percentiles is coming from? I can see how you calculate it, but spatially to my mind this would be dominated by the more variable regions. Maybe if you split into two groups - north of 30N and south of 30N. Seems like your Fig 2 is more erratic since the 1970s. This is an annual whereas Fig 1 is all months. To get annual do you do things annually or average the months. By the way 2008 is complete now. Cheers Phil Attachment Converted: "c:\eudora\attach\hadcrut3_hadcrut3v_comparison.pdf"