Here is the result of a Gaussian ‘kriging’ of Hadcrut4.5 temperature anomalies for February 2016. This is essentially what Cowtan & Way do.
This clearly shows why the global average of the kriged data gives larger values than Hadcrut4.5 when anomalies at high latitudes are warmer.
However I still prefer Hadcrut4.5 because it makes less assumptions, which is especially important for early years.
One interesting possibility of kriging is that it can also handle irregular grids. This means that you don’t even need to average the data in grid cells first. I might even try this, if I find the time. I suspect this is more or less exactly what Berkeley Earth does.