Further to my previous comments about reanalysis “data” being nothing but computer model output, I just compared the ERA40 and the NCEP/NCAR reanalysis model output with reality in the form of the HadCRUT3 analysis.
The R^2 of the surface temperatures in both cases is about 0.85, with the errors having a standard deviation of a tenth of a degree, and a maximum of about a quarter of a degree. The R^2 of the first differences, on the other hand, is only about 0.5.
I also looked at the two reanalyses 700mB lower troposphere data vs UAH MSU … the R^2 is worse, only about 0.7.
w.