This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
Frick, accidentally clicked a link in the sidebar and lost a ton of interesting stuff I had picked out. The highlights, less eloquently put:
Super interesting article and worth a read, thanks! Appendix is also very interesting. They discuss quite a few common questions in detail.
Very interesting setup where they consider victims instead of perpetrators. From a numbers standpoint, it's much more plausible and practical to help victims, plus fewer ethical issues, because few people would say no (presumably) to some kind of intervention to prevent being shot. I like this framing. And presumably you can't target alleged perpetrators without a bad tradeoff: you get ethical issues if you target too narrowly, or you get too many potential perpetrators that it's not practical to do anything substantial, and still might have ethical issues anyways.
The framing of preventing victims is more than just PR though. It's methodologically nicer too. For those who might not know, predictive policing has some issues beyond just pure ethics, though there's that too. The classic example is when I think it was NYC who tried it, they would feed recent crime data back into the algorithm that determined where police should patrol. Might sound like a good idea, but you get a feedback loop! Turns out police "generate" crime to some extent because they naturally uncover crimes on patrol, so if you assign a cop to a neighborhood, yup, they discover more crime, and the algo would send even more cops there, and they'd discover even more crime... etc. You get the idea.
They say it will pay for itself. Two ways of analyzing this. You have direct costs measured (some easy like medical bills, some more hard like productivity decline, social badness, etc.) and then also broader societal willingness to pay (WTP). I looked through at the book and paper referenced the latter, which I think is the source of their claim that halving shooting risk for the top 500 would be worth $67 million. Those WTP numbers came I think from a n= 1200ish survey where people were asked if they'd support an extra 50/100/200 bucks a year for a program that reduced gun violence 30%. I'm not totally sure those two questions are closely enough related that the connection is justified. Though it does raise the question -- if we could, and the model suggests we might, prevent over 100 shootings in Chicago of a specific group of people, how much would that be worth?
I agree their accuracy (well, to be specific, it's precision, since they use the ML algo to finger a subset of people of varying sizes, they try out 500 and 3300 in various spots, and then count how many of those fingered high-risk people actually did get shot) is actually pretty good and plausibly useful in practice.
Note that because of the above, you might notice that their fundamental set-up is a "threshold rule", where the ML algo spits out either "yes" or "no" to the question of if someone will be shot (with "yes" being "high-risk" in practical interpretation). The ML algo as analyzed was not set up to give any other kind of analysis or figure, all of the numbers are relative to this classification paradigm. So 500 as above is the top 500 in risk. When you tell it you want the top 3300 choices, it's less precise (10ish%) but still pretty good. Be Basically a cousin to logistic regression, if that sounds familiar to anyone, which is related to traditional OLS but not quite (IYKYK)
The risk of being shot again if you've already been shot (which makes sense) is about as precise in its risk as the model is. In other words, if the ML says you're high risk, even if you haven't been shot before, it's as if you had been. Which is interesting and could be a wake-up call to someone ID'd as high risk.
Young Black men are correctly identified as the highest risk group, and with slightly better accuracy than other groups, though the model still does pretty well for most races overall.
Obviously if you prevent the algo from considering things it does worse. Withholding demographics hurts, as do own arrests, though the precision drops to only about 10% which is still not so bad. Nice to have that as an ethical option. Dropping felony arrests doesn't hurt at all! Dropping someone's own personal arrest history puts it at 11%. (All top 500 but even out a bit as you increase the number of people you want warned).
What kind of predictors did they use? Notably, network effects (did your name appear along with another person in a police record, are you family, etc). But also, demographics (age, gender, race where availaible), fine-grained arrest data (what type, severity, what beat, how long ago), victimization data (were you the victim of various kinds of crime, fine-grained), gang affiliation indicators. Age actually mattered a lot, and race was never a top predictor, but it sounds like that was partially because the algo to some extent figured out someone's race due to other things (including the network effects). Being arrested yourself sadly was the biggest predictor overall (still a correlation of like .1). Someone you appeared in a record with also being shot and gang affiliations rounded out the top predictors.
They only used people who appeared in the police database at least twice for data quality reasons. How many people did this approach mean they missed? About a third of the victims weren't even in the sample at all. If you finger the top ~3300, you still miss 90% of the victims overall but 10% isn't bad.
No covid issues! They did their cross-validation time-wise so that they could keep their network effects.
Wouldn't you expect a negative effect on crime with police patrolling an area, with any feedback-loop being quite short lived? Else it would be a damning fact for patrolling, showing that it doesn't work. Do you have NY report at hand I would love to read it.
I know what book I read it in but it’s packed I will have to dig it out if you’re still interested - this appears to be an okay overview of a few of the issues and causality concerns in the meantime. The name PredPol sounds familiar?
The core idea is that: no, not necessarily. It heavily depends on the behavior of the police force in question. While some police forces might deter crime by their presence, other forces might by their policies generate crime (reports) by their proactivity in making stops, and yet other forces might lie about their crime data in order to make themselves look better. After all, what actually makes it to a computer-usable report in what is fundamentally a profession which emphasizes discretion is quite variable.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link