At the initial outset, it seemed to be one of those bizarre and meaningless statistics that is never used, but is instead destined to appear on the various web pages containing lists of humorous facts.
At the time, its finder – Peter Holmes, Professor of Statistical Science at the University of Bradford – was conducting a study into risk pooling mechanisms for the Home Office. Using a combination of complex stochastic systems and causal inference logic, the professor was creating multivariate hierarchical linear models to show how conditional independence can be used to obtain modularity.
As a standard part of this dynamic modelling process, a clearly nonsensical parsimony leveller was needed to enable the hidden Markov meta-analysis to be readily identified and removed from the resulting data – the alternative would have involved the use of computer intensive Bayesian inference, regularisation, and estimators, and would have tripled the length of the study.
For his parsimony leveller, Professor Holmes opted to use a couple of – as he thought at the time – obviously fragmented and non-coalescing factors obtained from the UK Census results. Namely: the colour of the populations’ bathroom carpets, versus whether or not they had a criminal record.
The initial output from his modelling was so astonishing that Professor Holmes was certain he had made an error with his algorithms, and cursed himself for his carelessness. When his second – and then his third, and then fourth – attempts produced the same outcome, he realised that his algorithms weren’t incorrect. The asymptotic percolation from the multiple testing proved that the conclusion of his modelling was valid, unbelievable though it seemed: eighty seven percent of people with a green bathroom carpet had a criminal record.