cc: "Andy Wright" <Andy.Wright@umist.ac.uk>, "Geoff Levermore" <geoff.levermore@umist.ac.uk>, "John Turnpenny" <j.turnpenny@uea.ac.uk>
date: Sat Feb 15 16:50:30 2003
from: Mike Hulme <m.hulme@uea.ac.uk>
subject: RE: Using Scaling Factors for HadRM3 data
to: "David Chow" <david.chow@umist.ac.uk>

   David,
   A few answers and comments:
   - the scaling factors (Table 7) are derived from globally-averaged surface air temperatures
   (dry bulb), not from locally derived (e.g. UK) data.  It has to be done this way for the
   pattern-scaling method we employed to make any sense
   - just because you cannot reconstruct values for intermediate periods is no indication of
   how the scaling factors were derived; pattern-scaling is a method that has a long history
   (see attached note for explanation) and can at best only ever be an approximation
   - there are several reasons why you will never get full agreement between pattern-scaled
   estimates and raw/direct model output: - the signal (being scaled is not well-defined
   relative to the noise); climate change is in any case not a linear function; regional
   climate will not behave linearly in relation to global climate.
   - in your exercise of using A2 to reconstruct B2 through pattern-scaling, it matters
   whether you used 30 years or 90 years of GCM data (i.e.., there was an ensemble of three A2
   simulations).  You should use the maximum number of years possible to define your "signal".
   In the end, I am not sure what you are trying to achieve here - proove that pattern-scaling
   is good or bad as a method?  This is a complex subject and Tim Mitchell here in HQ wrote a
   whole PhD thesis about it!  I don't think it is relevant for your research.  You should
   feel confident in using the UKCIP02 scenarios and Hadley data as supplied.
   I hope this helps,
   Mike
   At 09:54 11/02/03 +0000, David Chow wrote:

     Dear Mike,

     In relation to my Tyndall work I'd like to know is how the scaling factors (Table 7 on
     p. 43 of UKCIP02 report) were derived from the
     global data.(Presumably it was just global dry bulb.) An equation and any relevant
     references would be very useful. Were the factors based on 15 min data, hourly, daily or
     monthly? We want to use the factors to derive percntiles for the 2020s and 2050s.

     I have conducted some analysis on spells of data for temperature and solar radiance
     using HadCM3 and HadRM3 data that may be of interest.
     Chart 1 shows the differences between different model runs compared with real observed
     data for 1976-1990. It can be seen that HadCM3 data (dark columns) are significantly
     colder than what was observed in real life, not just with the average values, but also
     for the extremes. So the obvious thing is to use HadRM3 data, which seem to be more
     accurate.
     However, the problem with using HadRM3 data is that there is only data for 1960-1990 and
     2070-2100. Periods in between need to be interpolated, using Table 7 on p. 43 in the
     UKCIP02 Scientific Report (April 2002). I presumed the values from this table are
     obtained from analysis using HadCM3 data. The Report does not specify exactly how they
     were obtained but the obvious thing to use would be average temperature. So I did a
     quick check to see what the scaling factors are for temperature in HadCM3. Chart 2 shows
     the results. It appears that the extreme cold data (dotted line) and extreme hot data
     (solid lines) have significantly different scaling factors. The average data also have a
     trend different to what the Report suggests. (The thick blue line), which suggests that
     the scaling factors are not simply based on average temperature.
     There are 2 ways of obtaining HadRM3 (B2 senario) data for 2080s. One is to use the
     actual data provided by the database (selection), and the other metod is to use results
     from the A2 scenario and apply the pattern-scaling factors. If the scaling factors are
     reliable, the 2 sets of results should be similar. However, Table 1 shows that there is
     a significant difference, with the pattern-scaled data "over-estimating" the increase.
     In particular, one would expect the median (50%) value to be close to the mean and hence
     accurately pattern scaled (with perhaps lower correlation for the extremes), but the
     median difference values are typically about 10% higher thant the data-derived values.
     Thank you very much for helping.
     Regards,
     David Chow

     Research Assistant
     Manchester Centre of Civil and Construction Engineering
     UMIST
     M60 1QD

     Tel. 07879 447760
     e-mail. [1]david.chow@umist.ac.uk

