May 07, 2005

More on 99 percent solutions

Colby Cosh valiantly comes to his own defense on my snide-isms on a couple recent statistical stories.

The connecting issue to me is that in both cases something was alleged with precisely a precise 99 per cent probability factor: in one case that, as the Telegraph put it, "Dr Benny Peiser... concluded that only one third [of a thousand recent studies containing the words "global climate change"] backed the consensus view [that global warming was primarily a man-made phenomenon], while only one per cent did so explicitly."

What I know about global climate change would definitely not fill a book, but I'm fairly certain that more than 1 per cent of climatologists are going to be convinced of something if it's a "consensus" view. As Tim Lambert points out in the link to my previous post, Peiser listed 34 abstracts of studies out of the thousand that he thought were non-consensus, and found exactly one that explicitly refuted the consensus... and that was a commentary in a petroleum journal. If anything, Pieser's "work" confirms the general thrust of the initial Science piece: that, within the narrow scientific sphere that deals with this issue in daily work, the global warming issue is nearly as settled as evolution is in biology circles. And there is no logical connection from that that I can see to Peiser's claim, apparently also in the paper rejected by Science, that 99 per cent of all the academic work on this is "anti-consensus" or at least agnostic. If Science had been presented with a paper that claimed, on... well, on any basis really... that 99 per cent of biologists are secretly subscribing to the Journal of Intelligent Design, I would understand their not publishing it either.

As to the police Trekkie claims, anyone who does analysis work knows of the numerous studies showing that different listeners assign completely different numerical estimates to the English words meant to indicate probability... different people can assign mental values from 51% to 90% or higher when they are told something is "very likely" or "almost certain" to happen, for instance. Less documented is the similar effect in play when laypeople use numerical estimates in a vernacular sense. Saying something's going to be correct "99 times out of a hundred" is almost certainly not going to be a carefully measured probability estimate by the speaker. Outside of Albanian elections, correlations are never that clean in the best of circumstances... basic high school science is supposed to teach people this key fact, but it rarely seems to work. (Never mind the looseness in terminology involved.... note how 99 per cent "hard core Trekkers" turned in later interviews to "99 per cent" with an undefined interest in fantasy, role-playing, or science fiction. )

I agree with Colby that there's evidence worth investigating here of some nerd-al profiling in the Toronto police department... but that conclusion is almost independent from any need to evaluate what the actual correlation might be, for which we have no more useful evidence now than we did before. In the meantime, I'll be hiding my Lord of the Rings boxed set next time my cop friends come over. Wouldn't want to give the wrong impression.

Posted by BruceR at 11:02 AM