Dinocysts, the resting stage of dinoflagellates, have been used to reconstruct a range of palaeoceanographic variables, including summer and winter sea surface temperature (SST) and salinity, productivity and sea ice using transfer functions. I became interested in dinocysts as proxies when I saw warmer-than-modern reconstructions of Last Glacial Maximum SSTs in the Nordic Seas, which I thought curious. I was able to demonstrate the dinocyst transfer functions were influenced by spatial autocorrelation, which violates the assumptions of classical statistical methods and makes reconstructions appear more precise that the data warrant, and in extremis, make it appear possible to reconstruct environmental variables that have no ecological relevance.
This week, two papers on reconstructing Arctic sea-ice from dinocyst assemblages were published in Quaternary Science Reviews:
de Vernal, A., Rochon, A., Fréchette, B., Henry, M., Radi, T. & Solignac, S. (2013) Reconstructing past sea ice cover of the Northern Hemisphere from dinocyst assemblages: status of the approach. Quaternary Science Reviews.
de Vernal, A., Hillaire-Marcel, C., Rochon, A., Fréchette, B., Henry, M., Solignac, S. & Bonnet, S. (2013) Dinocyst-based reconstructions of sea ice cover concentration during the Holocene in the Arctic Ocean, the northern North Atlantic Ocean and its adjacent seas. Quaternary Science Reviews.
The first paper, hereafter the methods paper, develops and tests a large (>1400 site) dinocyst calibration set, and transfer function for Arctic sea-ice using the modern analogue technique(MAT).
The second paper, hereafter the application paper, reports reconstructions of Arctic sea-ice from dinocyst stratigraphies in earlier papers.
There are some obvious problems and peculiarities in reconstructing sea ice, for example:
- Sea ice coverage is highly correlated with other environmental variables, obviously SST, but also light availability and nutrient supply, which will make it difficult to determine if sea ice coverage has any direct control on dinocyst assemblages, or if it is the other environmental variables that matter.
- The environmental variables are spatially autocorrelated.
- The environmental gradient is very unevenly sampled: much of the ocean – everything between the Mediterranean Sea and the Norwegian coastal margin – has zero sea ice. This will bias transfer function performance statistics.
- The modern environmental data comes from a warm period, that probably has less sea ice than the proceeding centuries that provided the bulk of the sediment analysed.
The papers rightly discuss this last problem at length. The papers consider that as the error on the sea-ice predictions is the same size as the standard deviation of the interannual variability of the modern sea ice observations that a large fraction of the uncertainty arises from the interannual variability. This seems to me to be a confusion between the uncertainty of the estimate of mean sea ice coverage (ie a standard error) with the standard deviation. The standard error on the modern sea ice coverage will be 1/sqrt(50) =1/7 of the standard deviation – much smaller.
The other problems are ignored or dismissed. Regarding autocorrelation, the methods paper says
Although MAT is an approach widely used in paleoceanography and paleoclimatology, it was criticized because of potential biases linked to spatial autocorrelation yielding to overestimate its accuracy (Telford, 2006). This issue has been examined by Guiot and de Vernal (2011). In order to avoid any bias from spatial autocorrelations, they created a geographically uniform reference database and calculated RMSEP from progressively isolated samples. The results show that strong spatial autocorrelation is inherent to oceanographic parameters and that the spatial structure of data has relatively low effect on the calculation of the error of prediction, which thus provides suitable information on the performance of the approach.
While the application paper writes
It has been argued that MAT underestimates the error of prediction because of spatial autocorrelation ( Telford, 2006 and Telford and Birks, 2009), but tests made from geographically uniform reference databases show that the spatial structure of data has relatively low effect on the calculation of the error of prediction although spatial autocorrelation is inherent to oceanographic parameters (Guiot and de Vernal, 2011).
The idea that spatial autocorrelation has a low effect on the prediction errors when using MAT is fantasy. Both papers rely on Guiot and de Vernal (2011) and ignore the rejoinder to that work from Telford and Birks (2011). Guiot and de Vernal (2011) is an awful paper: it does not address the central predictions of work it criticise; tests a prediction based on a misrepresentation of that work; mis-states the assumptions of the methods they use; and incorrectly interprets the results.
Given the information in the two new dinocyst sea-ice papers, it is not possible to evaluate how good the reconstructions are. All that can be said with certainty is that the prediction errors are underestimated. If the reconstructions are to be used to evaluate climate models and better understand the dynamics of Arctic sea-ice, this is a major weakness and a shame that this opportunity has been wasted.