Significance is a new journal published jointly by the American Statistical Association and The Royal Statistical Society. One recent article ( http://www.nr.com/whp/Significance_ToCatchATerrorist.pdf ) discusses why over-sampling people with high-risk profiles, at airport check points for instance, does not work.
I think this article is very controversial, and the statistical explanations are flawed. The metric used to measure the success of "oversampling high-risk profiles" (for terrorism detection) is in my opinion flawed. The message that the authors want to convey is:
- randomly sampling passengers randomly of profile, for further questioning (in hopes to detect as many would-be terrorists as possible), works better than substantially over-sampling people who are 3 times more likely than average to be a terrorist (for instances, people with wires protruding from their suitcase) and under-sampling the other passengers (those that are not a high-risk profile)
- very slightly over-sampling these high-risk profiles (using a square root weight sampling scheme) is the optimum solution, according to the authors of the article
My argument, why this article is flawed, is:
- Let's assume that all terrorists fit with a particular, specific profile, for instance (made up example): they all are very tall, above 6.1 ft tall. Of course, not all very tall people would be a terrorist, only a very small fraction, say 1 out of 10,000. And lets say that 1 out of every 100 passengers is very tall.
- In this example, wouldn't the best sampling scheme be to considerably over-sample very tall people (invite them for further questioning)? Indeed wouldn't the perfect sampling scheme be "sampling ALL very tall people and nobody else"? That is, wouldn't the solution be extreme profiling, the exact opposite of what the authors recommend in their article?
What do you think?
Pdf version of the article: Significance_ToCatchATerrorist.pdf