Roy Spencer has fallen in love and become infatuated with regressing population density on temperature anomaly trends in the continental US. Nick Stokes has an interesting discussion but to Eli the real problem is the extremely naive way that Spencer deals with population density.
I computed daily average temperatures at each station which had records extending back at least to 1973, the year of a large increase in the number of global stations included in the ISH database. The daily average temperature was computed from the 4 standard synoptic times (00, 06, 12, 18 UTC) which are the most commonly reported times from stations around the world.
At least 20 days of complete data were required for a monthly average temperature to be computed, and the 1973-2011 period of record had to be at least 80% complete for a station to be included in the analysis.
I then stratified the stations based upon the 2000 census population density at each station; the population dataset I used has a spatial resolution of 1 km.
I then accepted all 5×5 deg lat/lon grid boxes (the same ones that Phil Jones uses in constructing the CRUTem3 dataset) which had all of the following present: a CRUTem3 temperature, and at least 1 station from each of 3 population classes, with class boundaries at 0, 15, 500, and 30,000 persons per sq. km.
By requiring all three population classes to be present for grids to be used in the analysis, we get the best ‘apples-to-apples’ comparison between stations of different population densities. The downside is that there is less geographic coverage than that provided in the Jones dataset, since relatively few grids meet such a requirement.The problem, of course is, that the population has been moving around much more than the temperature, for 1970-2008 there have been huge changes