I thought I might briefly highlight the recent Cowtan et al. paper Robust comparison of climate models with observations using blended land air and ocean sea surface temperatures. As Ed Hawkins points out, it’s really an attempt to do an apples-to-apples comparison of global temperatures. Dana has already covered this, as has Tamino.The basic issue is illustrated by the figure on the left. Typically the global surface temperature datasets are generated by combining near surface air measurements over land and sea surface temperatures. What’s presented from models, however, tends to be only near surface air temperatures. Another issue is that the change in sea ice cover means that some cells have gone from being near surface air temperatures over the ice to sea surface measurements. All of this means that typically the models are not presenting quite the same thing as is being presented by the observations. The results are shown in the figure on the right. The top panel shows the HadCRUT4 temperature data, the mean model ensemble result without any blending adjustments (red line) and the mean model ensemble result with blending adjustments (blue line). The lower panel shows what happens if you update the model forcings. Essentially, the updating the forcings and making these blending adjustments brings the models and observations into closer agreement.
From what I can see, this is a pretty obvious thing to do. If you’re going to compare models and observations, you really should compare like with like. One might ask why it hasn’t been the norm in the past, but I suspect that that may simply be because it wasn’t expected to make too much of a difference; in most fields you wouldn’t normally have to deal with pedantic nit-picks. As expected, the usual suspects don’t seem to like this new result. I made the mistake of reading some of the Bishop-Hill comments; the least optimal combination of nasty and ignorant. Personally, I think it’s a very interesting piece of work which does something that seems pretty obvious in retrospect and will probably help to improve model-observation comparisons.