3 ms·

Sort of a weak article - no stats. It'd be much more useful if it simply stated things clearly. Something like "90% of reports are for Xs, while only 40% of the

Sort of a weak article - no stats. It'd be much more useful if it simply stated things clearly. Something like "90% of reports are for Xs, while only 40% of the crimes are by Xs". Or even better, "out of n Xs, only m commit a crime". Then show incorrect biases there. (The only stat I saw was % of black population vs % of detainees being black - that's not so relevant without knowing actual crime rates. ) (Edit: I see why they didn't include this stat. According to [1], over 80% of suspects for various crimes were black. So that would explain why the police are pulling over a higher %, right? This still doesn't excuse base rate fallacy but sort of dismisses the one stat the article does provide.) Or more usefully, provide better characteristics off which to classify. If race, gender, and clothing have no predictive power, demonstrate that and show what does. Nextdoor could even build priors into the site, over time and help provide accurate predictions. After all, they have both the suspicious report info plus the actual crime info. Imagine if it would rate your post: "Based on your report of a male walking at 6am, the probability this is a criminal is 0.02%" That'd go a long way towards making people update internal models as well as making them look silly for using non-predictive characteristics. (Though if most crimes are committed by young males, it might be called out as ageist and sexist, but hey if that's reality...) Plus you could grade individual users. So combining the base rate of crime, significant characteristics, and past performance of a user? "MaryX, given your past suspicious activity record, there's a 99% chance this post is incorrect." Without this, all it takes is a busybody to report everything they see (and they can select on anything: race, gender, hair color, shoe size) and eventually get a correct hit. Then ignore the times they were wrong and point out how "see I reported this X and it was a thief!" 1: http://www.sfgate.com/bayarea/johnson/article/Oakland-crime-issue-goes-far-deeper-than-racial-5355633.php http://www.sfgate.com/bayarea/johnson/article/Oakland-crime-...

arthur_pryor 11y ago
I like where you go with this in the second half, because I believe that racial profiling is pragmatically ineffective anyway. But I take issue with this statement: "If race, gender, and clothing have no predictive power, demonstrate that and show what does." I think the burden of proof is on the person using those tools both to show that they have predictive power, and to show that the action being taken on that prediction is reasonable. Put a different way: a couple of principles which I think most Americans probably believe in are: 1) innocent until proven guilty 2) no guilt by association (there are laws that allow restrictions on known felons associating with one another, but at least in that case, the association is based on convictions for actions the individuals have taken, not assumptions about a class to which they happen to belong due to who they were born to). I don't like the idea that someone can look suspicious, without acting suspiciously. I know there's inevitably a lot of overlap and gray area between those two things. But what I'm saying is that even if you have some statistical model which you believe has good predictive power based on inputs of race and dress, and your model says that scruffy looking black men are more likely to commit crimes, you still don't get to investigate someone just because that person's a scruffy looking black guy. (sidenote: I just noticed this reply sitting unposted in a tab I opened a few days ago, so this might not even get read, but it was already written, so whatever.)

arthur_pryor 11y ago

I like where you go with this in the second half, because I believe that racial profiling is pragmatically ineffective anyway. But I take issue with this statement: "If race, gender, and clothing have no predictive power, demonstrate that and show what does." I think the burden of proof is on the person using those tools both to show that they have predictive power, and to show that the action being taken on that prediction is reasonable. Put a different way: a couple of principles which I think most Americans probably believe in are: 1) innocent until proven guilty 2) no guilt by association (there are laws that allow restrictions on known felons associating with one another, but at least in that case, the association is based on convictions for actions the individuals have taken, not assumptions about a class to which they happen to belong due to who they were born to). I don't like the idea that someone can look suspicious, without acting suspiciously. I know there's inevitably a lot of overlap and gray area between those two things. But what I'm saying is that even if you have some statistical model which you believe has good predictive power based on inputs of race and dress, and your model says that scruffy looking black men are more likely to commit crimes, you still don't get to investigate someone just because that person's a scruffy looking black guy. (sidenote: I just noticed this reply sitting unposted in a tab I opened a few days ago, so this might not even get read, but it was already written, so whatever.)