date: Fri Oct 2 09:57:30 2009
from: Phil Jones <p.jonesatXYZxyz.ac.uk>
subject: Re: getting unadjusted temperatures
to: "Don McNeil" <dmcneilatXYZxyz.mq.edu.au>
Not quite clear what you've done - maybe a map of your 86 zones would make it clearer.
Here are a few thoughts - I'd also be happy to look through some of your plots if you want
to send them.
In the papers that we refer to on the page where you've got HadCRUT3 from you should be
able to see that some of the 86 regions will be composed of much more observed data than
others. Your region 83 for example isn't composed of much data as that part of the
Antarctic has hardly any stations. Similarly some parts of the Southern Oceans don't have
much data (45-65S for example).
You might also see this in variance changes through time as some periods have more data
For 1961-2008 you may not see much of this.
The important point is that when doing linear trends for a common period like 1961-2008
you should determine how many years of missing data in each regional average you will
accept. It is your choice, but a thought is that missing data in the earliest and latest
years has a greater influence than missing data in the middle.
As for the absolute temperatures here is what I would do. You said you are using annual
(calendar year) averages from HadCRUT3. This is fine, as you'll have had to deal with the
missing data boxes to get correct results. So rather than add the absolute to HadCRUT3 the
simplest thing for you is to calculate the averages for each of your 86 regions from the
absolute temps. Your same program could be modified for this to give you 12 monthly and
one annual absolute value for each region. Then all you need do is to add the absolute
value for region N to the time series for region N.
It should be as simple as this. Remember that a temperature value is the average (the
absolute) plus the anomaly (from HadCRUT3).
One final thought - when calculating your regions you are using reasonably sized latitude
bands that encompass 20 degrees of latitude. When we do this we use a weighted average as
boxes near the equator are larger than those nearer the poles. Weights are the cosine of
the central latitude band - so for 20-25 north it would be the cos of 22.5. When averaging
the boxes for each month you weight them by this cos and then divide by the sum of the
weights. You're already having to calculate the number of boxes within each area to not use
missing values. I doubt this step will alter things much, but it is technically correct to
One final point - if you look at the Reviews of Geophysics paper from 1999, you'll see
that the anomalies over the ocean are from SST data, but when you add back the absolute
over the ocean you are adding back a marine air temperature absolute.
At 03:17 02/10/2009, you wrote:
My Thai students (in our graduate program in Research Methodology) and I have now done a
statistical analysis of the HadCRUT3 data from your website for full years from 1961 to
2008 inclusive. Since the seasonal effects vary with latitude we've used annual means,
and to simplify the analysis we've also taken means over all grid-boxes within 86 larger
zones of approximately equal areas in 8 latitude bands of widths 25, 20, 20, 25, 25, 20,
20 and 25 degrees, respectively from pole to pole, with these bands containing 4, 9, 12,
18, 18, 12, 9 and 4 regions, respectively.
We've done a simple multivariate linear regression analysis, using the 48 (years) by 86
(regions) data matrix as the response variables and year (ranging from 1961 to 2008
inclusive) as the single independent variable, taking account of correlations between
data in different regions (but not years).
This analysis assumes that the overall trend in each region is linear, but allows the
intercepts and slopes vary with region.
The results show that the temperature anomaly increased in every region except 14 (25-45
degrees N, 180-150 degrees W) and 83 (65-90 degrees S, 180-135 degrees W) where there is
no evidence of a change, and 78 (45-65 degrees S, 20 degrees W to 20 degrees E) where
the temperature anomaly was negative.
Since our model fits a different constant (intercept) term for each region as well as a
slope, we would prefer to use absolute temperatures rather than temperature anomalies,
because these constants will then show how the temperature varies over the Earth's
surface, rather than how they vary with respect to the 1961-1990 baseline average.
If we download the absolute temperatures from your website and simply add them to the
HadCRUT3 data, will that give us absolute temperatures for every grid-box for the whole
I would also appreciate your thoughts on what we have done. We'd like to publish our
method and results in an appropriate journal, because this methodology is easily
extended to include other possibly relevant determinants such as atmospheric CO2,
geologic activity, solar activity, etc.
>>> <P.JonesatXYZxyz.ac.uk> 09/14/09 4:33 PM >>>
In Geneva this week. I can send you some pdfs when I get back, or I
might be able to find one on my laptop.
Basically the analysis is done monthly, so we have averages for all
sites for 1961-90 for each of the 12 months.
Attached one paper - hopefully this says more.
Email me again with papers you'd like - I think I have most as pdfs.
Prof. Phil Jones
Climatic Research Unit Telephone +44 (0) 1603 592090
School of Environmental Sciences Fax +44 (0) 1603 507784
University of East Anglia
Norwich Email p.jonesatXYZxyz.ac.uk