The hockey stick: a retrospective - Climate Change: The Facts - page 148

researchers wantingtoobtainhisdataandreplicate hismethodologies, mostofwhichwere
only resolved by the interventions of US Congressional investigators and the editors of
Nature magazine,bothofwhomdemandedfullreleaseofhisdataandmethodologiessome
six years after publication of his original Nature paper.
Mann had re-done his hockey stick graph at some point during its preparation with the
dubious bristlecone records excluded and saw that the result lost the hockey stick shape
altogether, collapsing into a heap of trendless noise. However he never pointed this out to
readers.
He also stated that he had computed test scores called r 2 statistics that he said (or
implied) confirmed the statistical significance of his results, yet when the r 2 scores were
later revealed they showed no such thing, and by then he had taken to denying he had even
calculated them.
Our critique of Mann's method
There are two key parts to the hockey stick-making machine. The first is the principal
components (PC) step, and the second is the least squares (LS) fitting step. The PC step
takes large numbers of temperature proxies and compiles them into a relatively small
number of composite series. The LS step then lines up the final segment of the composites
againstanupward-slopingtemperaturegraphandputsweightontheminproportiontohow
well they correlate. If there are many composites and only one has a hockey stick shape,
the LS step will find it and put most of the weight on it. If none of the composites has a
hockey stick shape, then the LS step will come up blank and the resulting graph will just
look like noise.
Mann's PC step was programmed incorrectly and created two weird effects in how
it handled data. First, if the underlying data set was mostly random noise, but there was
one hockey stick-shaped series in the group, the flawed PC step would isolate it out,
generate a hockey stick composite and call it the dominant pattern, even if it was just
a minor background fluctuation. Second, if the underlying data consisted of a particular
type of randomness called 'red noise'—basically randomness operating on a slow, cyclical
scale—then the PC step would rearrange the red noise into a hockey stick-shaped
composite. Either way, the resulting composites would have a hockey stick shape for the
LS step to glom onto and produce the famous final result.
The use of red noise series is necessary for testing the statistical robustness of the
hockey stick method. This is a procedure called Monte Carlo analysis. For one of our 2005
papers, we generated thousands of series of trendless autocorrelated random numbers and
ran them through the PS and LS steps. 9 This generated thousands of results, each of which
hadanindexofaccuracycalledthereductionoferror(RE)score.Likewisetheactualproxy