Pit Bulls and Profiling
Malcolm Gladwell· (probably best known as the pop-sociologist author of The Tipping Point·) has an excellent piece in this week’s New Yorker· on what pit bulls can teach us about profiling·. The argument boils down to one about common tendencies in misinterpreting data and drawing the wrong (or “unstable”) generalizations from apparently recurrent phenomena.
Everyone’s favorite security guru Bruce Schneier· has made similar arguments in the past·, but Gladwell’s style in this case is more compelling. I particularly liked this passage about The Godfather:
In July of last year, following the transit bombings in London, the New York City Police Department announced that it would send officers into the subways to conduct random searches of passengers’ bags. On the face of it, doing random searches in the hunt for terrorists — as opposed to being guided by generalizations — seems like a silly idea. As a columnist in New York wrote at the time, “Not just ‘most’ but nearly every jihadi who has attacked a Western European or American target is a young Arab or Pakistani man. In other words, you can predict with a fair degree of certainty what an Al Qaeda terrorist looks like. Just as we have always known what Mafiosi look like — even as we understand that only an infinitesimal fraction of Italian-Americans are members of the mob.”
But wait: do we really know what mafiosi look like? In “The Godfather,” where most of us get our knowledge of the Mafia, the male members of the Corleone family were played by Marlon Brando, who was of Irish and French ancestry, James Caan, who is Jewish, and two Italian-Americans, Al Pacino and John Cazale. To go by “The Godfather,” mafiosi look like white men of European descent, which, as generalizations go, isn’t terribly helpful. Figuring out what an Islamic terrorist looks like isn’t any easier. Muslims are not like the Amish: they don’t come dressed in identifiable costumes. And they don’t look like basketball players; they don’t come in predictable shapes and sizes. Islam is a religion that spans the globe.
Steve Laniel Jan 28
I’m reading through this article now, and I’ve just gotten to the part where the person from New York notes that most every terrorist is a Muslim. Even granting the truth of this premise, they’ve got it backwards: even if all terrorists are Muslims, it doesn’t follow that all Muslims are terrorists. They’ve got their conditional probabilities backwards. They need Bayes’ Theorem to get things right. I’ve written up a little thing (PDF) that goes through the math.
Steve Laniel Jan 28
Reading further, the unstable- versus stable-generalization stuff reminds me of the bits about “ecological validity” in Gigerenzer, Todd, et al.’s “Simple Heuristics That Make Us Smart.” One idea is that it’s often better to use a simple estimator — or several simple estimators — each of which is perhaps no better than chance (e.g., “person looks nervous,” “person gets his story wrong,” etc.), than to use a single complicated estimator (e.g., combining 43 unstable traits).
Good article. I may have to read “Blink” after all.
Steve Laniel Jan 28
Correction: A series of estimators that are no better than chance would not be helpful. What I meant was “each of which is just slightly better than chance.”
We here at Steve Laniel regret the error.