The palaeoclimatological literature overflows with purported correlations between proxy records from different sedimentary archives, often claimed on the basis of a couple of wiggles that happen to coincide. These correlations are almost never tested statistically, as the twin problems of uncertain and unevenly-spaced chronologies make it impossible to use standard statistical methods with most archives. Rather than provoking caution, the inability to test correlations has released the fancies and made comparisons as wild and whimsical as Hamlet’s almost obligatory.
Do you see yonder cloud that’s almost in shape of a camel?
By the mass, and ’tis like a camel, indeed.
Methinks it is like a weasel.
It is backed like a weasel.
Or like a whale?
Very like a whale.
William Shakespeare, Hamlet (act 3 scene 2)
Fortunately, a potential antidote has recently been published by Kira Rehfeld and Jürgen Kurths in Climate of the Past Discussion (disclosure: I was one of the reviewers).
The method Rehfeld and Kurths present is summarised by this figure
First, a Monte Carlo age-depth modelling procedure, for example Bacon or Bchron is used to generate an ensemble of possible age-depth models for each core from the dates. Adding the proxy data gives an ensemble of time series. The correlation between each member of this ensemble and the other proxy time series is then calculated using a method that tolerates uneven chronological spacing – the authors introduce several. Rather than having a single estimate for the correlation, this procedure generates a distribution of estimates – some members of the ensemble have better correlations than others.
The significance of the correlation is assessed by comparing the observed correlation with a null distribution derived by using simulated time series with the same autocorrelation structure as the original data.
This method has great potential for banishing camels, weasels and whales from the palaeoclimatological literature. But only if it is used. At the moment, the method is implemented in MATLAB/OCTAVE, but the authors promise to release an R version which should make it more accessible.