Climate sensitivity, the change in mean global temperature expected for a doubling of CO2, was estimated by the IPCC AR4 to be 3°C and likely to between 2°C and 4.5°C. This broad uncertainty has scarcely changed for decades, and is the largest source of uncertainty for climate projections beyond a few decades (Knutti and Hegerl, 2008). One way to estimate climate sensitivity is from the difference in temperature from time periods with different climate forcings. The contrast between the Last Glacial Maximum (LGM) and the pre-industrial is a useful target as the change in both the climate forcings and temperature are relatively large. Schmittner et al (2011) use this contrast, a suite of climate models with different climate sensitivity and Bayesian statistics to report a climate sensitivity of 2.3°C, both smaller and, with a likely range of 1.7°C to 2.6°C, more precise than the IPCC AR4 estimate.
RealClimate looked at Schmittner et al (2011) long ago and concluded that climate sensitivity was likely underestimated and too precise. For different reasons, I will come to the same conclusions below.
Schmittner et al’s estimate is dependent on the quality of the proxy data used to reconstruct LGM climate. If the proxy data are biased, Schmittner et al’s results will be biased. If the uncertainty of the proxy data is underestimated, then the uncertainty on the climate forcing may be underestimated.
Schmittner et al’s estimate based on just marine proxies is both smaller and more precise than that based on terrestrial proxies. As the marine proxies dominate the combined result, it makes sense to focus on the marine proxies: fortunately, this is where much of my recent work on proxies has focused.
Most of Schmittner et al’s marine proxies are extracted from the MARGO LGM compilation. This compilation includes alkenone and Mg/Ca temperature estimates, but it is the transfer function-based reconstructions from planktonic foraminifera and, to a lesser extent, dinoflagellate cysts that predominate. Schmittner et al’s climate sensitivity estimate is dependent on these transfer function-based reconstructions being unbiased and their uncertainty being correctly estimated.
The uncertainty on the transfer function-based reconstructions is estimated by cross-validating the modern training set and finding the root mean square error of prediction. Typically each sample in turn is omitted from the modern training set and its temperature estimated from the remainder of the training set. The differences between the predicted and measured temperatures are used to assess transfer function model performance and estimate uncertainty. This procedure assumes that the test observation is independent from the training data during cross-validation. If the data are not independent, for example because of spatial autocorrelation, cross-validation performance estimates will be over-optimistic. Transfer function methods that find the local relationship between species and the environment (for example the Modern Analogue Technique) will be more affected than models that find a global relationship between species and the environment (e.g. weighted averaging). Consequently, one would expect MAT models to have a lower RMSEP that WA models where spatial autocorrelation is strong. Conversely, with data sets with no autocorrelation, performance is similar, or MAT performs worse.
The planktonic foraminifera and dinocyst LGM temperature reconstructions were calculated with MAT or related methods. As sea surface temperature is strongly spatially autocorrelated, we should expect the RMSEP to be over-optimistic, but by how much? One way to estimate this would be to look at the difference between MAT and WA, on the assumption that these should give similar performance when there is no autocorrelation. Another way would be to measure training set performance using a spatially independent test set, for example the South Atlantic for a North Atlantic training set. There are potentially problems with this approach if morphologically similar forams have different ecological niches in the two regions. Both these methods give a similar result, suggesting that the RMSEP for MAT foram reconstructions is about half the true uncertainty. I’ve not tried this analysis on the dinocyst data, but I wouldn’t expect a happy result.
The autocorrelation-induced over-optimism is not the only way for the uncertainty of the LGM reconstructions to be underestimated. A second problem is the lack of good modern analogues for the LGM foram and dinocyst assemblages. If the LGM assemblages lack good modern analogues, it is fairly certain that the uncertainty estimated by cross-validating the training set will be an underestimate, especially with MAT.
The LGM reconstructions may also be biased, if the depth to which the modern assemblages are calibrated is not the one that the assemblages are most sensitive to. Initial analyses suggest that formainifera assemblages are typically more sensitive to sub-surface conditions than the 10m temperature they are usually calibrated against. Reconstructions of sub-surface temperatures, in at least parts of the tropical North Atlantic, can show marked cooling in contrast to surface reconstructions. At the moment it is unclear if the results from the North Atlantic are generally applicable; if they are then the traditional foram-based view that the tropics cooled little may need to be revised (perhaps the coral-based estimates of substantial cooling are not so far-fetched). Even if the bias is small, the uncertainty in the correct calibration set should add some uncertainty to the reconstructions. I’ve not yet tested if the dinocysts have been calibrated against the optimal depth, but as they are mainly used outside of the tropics, I doubt this will be a major issue affecting the validity of these reconstructions – that is, I think there are much more important issues affecting the validity of dinocyst-based reconstructions.
The obvious impact of underestimating the uncertainty on the reconstructions is that the uncertainty on the sensitivity will be underestimated; the sensitivity PDF should be broader. Increasing the uncertainty on the marine data would reduce the weight the analysis puts on the marine data, so the combined land and ocean estimate would be less dominated by the marine data, and more of a compromise. This would result in a higher sensitivity estimate.
The obvious impact of biased reconstructions will be to bias the sensitivity estimate. Schmittner et al’s sensitivity estimate showed that a 0.5°C bias in ocean temperature reconstructions gave a 1°C change in sensitivity estimates. Their sensitivity experiments tested a global bias, but it is possible that the effect of bias is more pronounced in some regions than others.
The impact of the curious, warmer-than-modern, LGM dinocyst-temperature estimates in the Nordic Seas were questioned by Fyke & Eby (2012) in a technical comment. Schmittner et al replied that the impact of these odd reconstructions would have minimal impact as grid-cells with contradicting reconstructions would be given little weight by their procedure. The MARGO paper contrived some explanation for these warmer-than-modern temperatures, I have other explanations.