cc: 'Mike Hulme' date: Thu, 13 May 2004 15:47:46 -0700 (PDT) from: Nathan Gillett subject: RE: gu23wld0098.dat dataset to: Mark New Hi Mark, Thanks a lot for your email discussing sources of error in the gridded precip data. I submitted a paper on the issue of detection of volcanic influence on precip to GRL, and have now just received the reviewer's comments. The main point of contention is an apparent disagreement between observed and simulated terrestrial mean precipitation (based on the 2.5x2.5 CRU data). The attached plot shows terrestrial mean precipitation anomalies based on your and Mike's dataset in the solid line (5-yr smoothed), together with a four-member ensemble of PCM simulations (with anthropogenic and natural forcings) (it's similar to the plot shown for HadCM3 by Allen and Ingram (2003)). The reviewer notes that the observations do not lie within the range of intra-ensemble variability and asks us to further investigate the reasons for this disagreement. One reason is that the model appears to underestimate the precip response to volcanoes, but even allowing for this there is still some discrepancy. I'd be very interested to know how large the observational error in this quantity is likely to be - obviously I realise this would take some time to calculate accurately, and is probably not possible within the time in which GRL has given me to revise the paper, but if you have an immediate impression that would be helful. Put another way - how large is the error in five year mean terrestrial mean precipitation precipitation likely to be in percentage terms? Since I'm looking at anomalies, I don't care about a constant systematic error. I'm also sampling the modelled precip anomalies in the grid squares with observations present, so I'm interested only in the terrestrial precipitation over the grid squares with data present. I realise this is a hard question to answer - but any guidance would be very helpful (is it 5% or 50%?)... Thanks very much, Nathan ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ Mailing address: Courier address: Nathan Gillett Nathan Gillett School of Earth and Ocean Sciences School of Earth and Ocean Sciences University of Victoria at Ian Stewart Complex, Room 373 PO Box 3055 University of Victoria Victoria, BC, V8W 3P6 3964 Gordon Head Road Canada Victoria, BC, V8N 3X3, Canada Tel: +1 250 472 4013 Email: gillett@uvic.ca Fax: +1 250 472 4004 http://climate.uvic.ca/people/gillett ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ On Wed, 24 Mar 2004, Mark New wrote: > Hi Nathan, > > These data are really Mike's baby (that is different to the CRU 0.5 > degree data that I was involved in). > > I do however have a comment on using these Mike's gridded data in > detection/attribution work. > > It is important to realise that gridded precipitation data is far more > sensitive to poor station coverage than temperature and precip. Many of > the grid points "with data" in Mike's dataset arise from a single > station; many others from averaging of only two or three stations; and > still others from a gradual increase in the number of stations thru > time. Thus the variance of the grid-point time-series can be biased by > these sampling issues, with gridpoints with low station numbers having > an apparently higher variance. Similarly, gridpoints with few stations > are associated with larger standard errors. > > Ideally one would want to use an estimate of the sampling error, and to > adjust for station density dependent variance biases, in the statistical > model you use for detection/attribution. I did some preliminary work on > error estimates a few years ago, based on the methodology used by Phil > Jones for temperature [Jones, P.D., T.J. Osborn, and K.R. Briffa, > Estimating sampling errors in large scale temperature averages, 10, > 2548-2568, 1997.], but never pursued it beyond some prelimiary analysis. > It seems that is is possible to estimate SE of grid point precip > averages, and you may want to look into this if getting a handle of > observational error if you can inlcude this in the optimal > fingerprinting (assuming you are using OF) statistical model. I seem to > remember that most OF schemes to date do not include observational error > in the statistical model, but concentrate on using climate model control > run data to estimate climate-noise (probably the larger error > component). > > Let me know if you want to pursue this further...this may be the > stimulus I need to continue with the observational error estimates! > > Mark > > > -----Original Message----- > > From: Nathan Gillett [mailto:gillett@ocean.seos.uvic.ca] > > Sent: Wednesday, March 10, 2004 8:09 PM > > To: m.hulme@uea.ac.uk; mark.new@geog.ox.ac.uk > > Subject: gu23wld0098.dat dataset > > > > > > Dear Mike, Mark, > > I obtained a copy of your gu23wld0098.dat precipitation > > dataset from Peter Stott at the Hadley Centre (he had a > > version in pp format, which was easy for me to read). I'm > > currently using this data for a detection and attribution > > study. I've added the appropriate acknowldegement, and I'll > > send you a preprint of the paper once its ready to submit. > > Any comments or suggestions on the use of this data for > > detection and attribution would be welcome. > > > > Cheers, > > > > Nathan > > > > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > > ++++++++++ > > Mailing address: Courier address: > > > > Nathan Gillett Nathan Gillett > > School of Earth and Ocean Sciences School of Earth and > > Ocean Sciences > > University of Victoria at Gordon Head > > Complex, Room 377a > > PO Box 3055 University of Victoria > > Victoria, BC, V8W 3P6 3964 Gordon Head Road > > Canada Victoria, BC, V8N 3X3, Canada > > > > Tel: +1 250 472 4013 Email: gillett@uvic.ca > > Fax: +1 250 472 4004 > http://climate.uvic.ca/people/gillett > > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > Attachment Converted: "c:\eudora\attach\fig1.ps"