The problem is the use of data collected from populations as a whole. A statistical test that would correctly classify 99.99% of the population as being terrorist or non-terrorist sounds like it would just work great. But given that the incidence of terrorists in the US is probably on the order of a hundred or so, that would mean that Law Enforcement would have to intensively surveil 30,000 innocent false positives for every terrorist that might be caught. The cost of the logistics clearly make such testing impossible. It is clear that “old fashioned” police work is the solution, not statistical computer modeling.

The problem comes in with the incidence in the general population. When the incidence is high, identification is easy — but when it is low, detection becomes difficult. This is a well known area of simple algebraic mathematics known as sensitivity, specificity and the predictive value of tests. It is covered on Wikipedia for those who are interested. http://en.wikipedia.org/wiki/Sensitivity_and_specificity