A few days ago I noted a new paper by Marotzke and Forster which claimed to show that the recent divergence of model predictions and observations was all down to natural variability. The paper was getting considerable hype from Marotzke's employers, the Max Planck Institute:
Sceptics who still doubt anthropogenic climate change have now been stripped of one of their last-ditch arguments...the gap between the calculated and measured warming is not due to systematic errors of the models, as the sceptics had suspected, but because there are always random fluctuations in the Earth's climate.
Marotzke was also quoted as saying: "The claim that climate models systematically overestimate global warming caused by rising greenhouse gas concentrations is wrong" and he went on to get quite a lot of media coverage, including the Mail, the Sydney Morning Herald, Deutsche Welle, and the Washington Post.
Based on media coverage of the paper's contents, I expressed considerable concern over what the authors had apparently done. It seems, however, that my criticisms at the time were understated. It is in fact "worse than we thought".
These uncomfortable facts are revealed in Nic Lewis's analysis of the paper, which has recently appeared at Climate Audit. Nic, assisted by Gordon Hughes and Roman Mureika, has identified some glaring statistical errors in the paper, but these turn out to be just the tip of the iceberg, as I shall now try to explain.
Marotzke's method involved seeing how far inter-model differences in the temperature trends in a group of climate model runs could be explained by differences between models in changes in forcing and two "structural elements": the climate feedback parameter and the ocean heat uptake efficiency. However, the time series for the forcing changes and the estimates of the structural elements came from another paper, by Forster et al. (and hence Forster's appearance as a co-author, or at least so one assumes).
Unfortunately, Marotzke seems not to have understood that Forster had calculated the forcing time series using an equation that expressed them as a function of model temperature. This meant that when the figures were plugged back into Marotzke's regression model he effectively had temperature on both sides of the equation: he was regressing temperature on a function of temperature and the logic was circular!
But in fact, even if the work could be redone without the circularity, Nic reckons the general approach is still "doomed". Readers will have gathered from my introduction that the whole study revolves around model output. It's completely divorced from the real world, apart from a brief and somewhat tenuous claim that "the simulated multimodel ensemble spread accurately characterizes internal variability". But unfortunately, even in model-world, the values Marotzke used for the ocean heat uptake efficiency vary wildly from those estimated from the model runs he uses. For example, in some models a degree of surface temperature rise over 1951–2012 produced twice as much inflow of heat into the oceans as that implied by the model ocean heat uptake efficiency value used by Marotzke. In others it produces only half as much. The same problem may well also arise with climate feedback parameter, but Nic tells me the circularity issue means one can't tell. With the diagnostic data being so poor, you are never going to be able to explain the causes of variation in historical temperature changes between models.
Amusingly, Marotzke declared on the paper's release that "The difference in sensitivity explains nothing really...I only believed that after I had very carefully scrutinised the data on which our graphs are based". It appears that his scrutiny was not careful enough.
I think it's fair to say that this is the last we will be hearing of Marotzke and Forster 2015.