Is spatial autocorrelation introducing biases in the apparent accuracy of paleoclimatic reconstructions?

We address the issue of spatial autocorrelation, an occurrence that may introduce biases in the evaluation of the performance of transfer functions, by using two fundamentally different approaches, one based on calibration (weighted averaging partial least squares regressions; WA-PLS) and the other based on similarities (modern analogue technique; MAT). The tests were made after spatial standardization of 700 North Atlantic surface data points, which include 29 dinocyst taxa and 4 climate parameters. The evaluation of transfer function performance was made by defining a verification dataset that was gradually isolated from the calibration or comparison datasets. Although strong spatial autocorrelation characterizes the original climate parameter distribution, the results show that the spatial structure of data has relatively low effect on the calculation of the error of prediction. They also show that the performances of MAT are generally better than those of WA-PLS, with lower error of prediction. The better performance of MAT in the present study can be explained by the non-modal distribution of salinity and temperature in the studied marine environments, which is not appropriate for the application of WA-PLS. The two methods yield equivalent results about the spatial structure of the residuals based on empirical semi-variograms. The analyses we performed include the comparison of reconstructions based on original raw data and gridded data. Results suggest that the gridding of the reference database may reduce the noise and thus improve the performance of the techniques.