The egregious Soon has a paper, Variable solar irradiance as a plausible agent for multidecadal variations in the Arctic-wide surface air temperature record of the past 130 years which is a weasel-worded title if ever I saw one. But anyway: Soon is comparing the “Arctic” temperature record to the solar irradiance (via some proxy or another) and finding a wonderful match, which sends the Durkins of the world into a frenzy. Of course, this means that the solar *doesn’t* match the global record; and of course given almost any shape of solar you could find some region of the world whose temperature trend would match it, but never mind that for now.
What about this Arctic temperature record? Its from Polyakov et al, and you can see their version of it here. You can download their station records (but be careful, there are data bugs in there) and average them yourselves (quite how Polyakov do the averaging I don’t know; I’m just treating each file with weight one. This may not be right: it treats all the buoy data as one point, for example).
My results are shown here (as ever, click for bigger pic; black lines T avg; blue lines station count on the RH axis; all this is the annual mean data); the top left is the one to start with, and fairly well matches P et al. (remember P start their pix in 1880). So things are broadly OK. That pic is done by subtracting the mean of each station before averaging, which is of course necessary. Now look at the top right, where I haven’t done this. There is a precipitious drop in temperature! Why? Because, as time goes by, more and more high-latitude stations come into the mix and the average temperature of a station goes down. Clearly this isn’t a useful figure to deduce a T trend from, but it *is* a useful figure to demonstrate that the P dataset shows substantial shifts in its distribution of stations over time.
You can get rid of that by picking only stations that are always there: the lower plots are the same, but requiring 1900-1980 data to be (nearly always) present. Hence even the non-normalised version of this (bottom right) has no strong trend (after 1900). Sadly, though, the shape of the top and bottom left plots is nearly identical, so the changing data coverage hasn’t made much difference to the overall trend. Ah well. I remain somewhat mistrustful of this dataset, because of the huge changes in coverage.
A much earlier post of mine may well be relevant: using HadCTRU data.