Correlation-based interpretations of paleoclimate data - where statistics meet past climates

Correlation analysis is omnipresent in paleoclimatology, and often serves to support the proposed climatic interpretation of a given proxy record. However, this analysis presents several statistical challenges, each of which is sufficient to nullify the interpretation: the loss of degrees of freedom due to serial correlation, the test multiplicity problem in connection with a climate field, and the presence of age uncertainties. While these issues have long been known to statisticians, they are not widely appreciated by the wider paleoclimate community; yet they can have a first-order impact on scientific conclusions. Here we use three examples from the recent paleoclimate literature to highlight how spurious correlations affect the published interpretations of paleoclimate proxies, and suggest that future studies should address these issues to strengthen their conclusions. In some cases, correlations that were previously claimed to be significant are found insignificant, thereby challenging published interpretations. In other cases, minor adjustments can be made to safeguard against these concerns. Because such problems arise so commonly with paleoclimate data, we provide open-source code to address them. Ultimately, we conclude that statistics alone cannot ground-truth a proxy, and recommend establishing a mechanistic understanding of a proxy signal as a sounder basis for interpretation.