All Correlations Are Bastards
July 12, 2024 1:52 AM Subscribe
This paper shows that shootings are predictable enough to be preventable. Using arrest and victimization records for almost 644,000 people from the Chicago Police Department, we train a machine learning model to predict the risk of being shot in the next 18 months. Out-of-sample accuracy is strikingly high: of the 500 people with the highest predicted risk, almost 13 percent are shot within 18 months, a rate 128 times higher than the average Chicagoan. A central concern is that algorithms may “bake in” bias found in police data, overestimating risk for people likelier to interact with police conditional on their behavior. We show that Black male victims more often have enough police contact to generate predictions. But those predictions are not, on average, inflated; the demographic composition of predicted and actual shooting victims is almost identical. There are legal, ethical, and practical barriers to using these predictions to target law enforcement. But using them to target social services could have enormous preventive benefits: predictive accuracy among the top 500 people justifies spending up to $134,400 per person for an intervention that could cut the probability of being shot by half. from Machine Learning Can Predict Shooting Victimization Well Enough to Help Prevent It [NBER; direct link to working paper (PDF)]
Crucially, [a different] approach was developed in consultation with people who live in the affected communities. (page 27)
Bhatt, Monica P., Sara B. Heller, Max Kapustin, Marianne Bertrand, and Christopher Blattman (2023) “Predicting and Preventing Gun Violence: An Experimental Evaluation of READI Chicago,” NBER Working Paper No. 30852, http://www.nber.org/papers/w30852 (pdf)
posted by HearHere at 3:21 AM on July 12 [4 favorites]
Bhatt, Monica P., Sara B. Heller, Max Kapustin, Marianne Bertrand, and Christopher Blattman (2023) “Predicting and Preventing Gun Violence: An Experimental Evaluation of READI Chicago,” NBER Working Paper No. 30852, http://www.nber.org/papers/w30852 (pdf)
posted by HearHere at 3:21 AM on July 12 [4 favorites]
Even if they're good at identifying people at risk, there's nothing (I skimmed the pdf) about what interventions might help. Support of people moving to a different neighborhood, or what?
posted by Nancy Lebovitz at 4:29 AM on July 12 [2 favorites]
posted by Nancy Lebovitz at 4:29 AM on July 12 [2 favorites]
Interventions to help are almost certainly outside the scope of the paper.
posted by MisantropicPainforest at 4:47 AM on July 12 [4 favorites]
posted by MisantropicPainforest at 4:47 AM on July 12 [4 favorites]
You never see a paper called Machine Learning Predicts with Astonishing Accuracy Which Police Officers are Most Likely to Kill a Citizen.
posted by hypnogogue at 4:57 AM on July 12 [40 favorites]
posted by hypnogogue at 4:57 AM on July 12 [40 favorites]
> You never see a paper called Machine Learning Predicts with Astonishing Accuracy Which Police Officers are Most Likely to Kill a Citizen.
Greg Ridgway wrote something like this 15 years ago. Depending on what information you have, it’s just not that hard to predict which cops will plug which people.
This is of a piece with Greg’s piece on LAPD traffic stops 1 hour before dawn and 1 hour after dawn. Care to guess whether the proportion of black people among traffic stops falls or rises once a cop can see what color the person is?
Greg had to move to Philadelphia after that study was published, oddly enough.
posted by apathy at 5:06 AM on July 12 [38 favorites]
Greg Ridgway wrote something like this 15 years ago. Depending on what information you have, it’s just not that hard to predict which cops will plug which people.
This is of a piece with Greg’s piece on LAPD traffic stops 1 hour before dawn and 1 hour after dawn. Care to guess whether the proportion of black people among traffic stops falls or rises once a cop can see what color the person is?
Greg had to move to Philadelphia after that study was published, oddly enough.
posted by apathy at 5:06 AM on July 12 [38 favorites]
I scanned through the paper but I couldn't figure out if they predict being shot by a cop or being shot period.
If it's the former, they are awfully cagey about it.
posted by keep_evolving at 5:07 AM on July 12 [3 favorites]
If it's the former, they are awfully cagey about it.
posted by keep_evolving at 5:07 AM on July 12 [3 favorites]
Support of people moving to a different neighborhood, or what?
"offered an 18-month job alongside cognitive behavioral therapy and other social support" Bhatt et al. supra
posted by HearHere at 5:08 AM on July 12 [5 favorites]
"offered an 18-month job alongside cognitive behavioral therapy and other social support" Bhatt et al. supra
posted by HearHere at 5:08 AM on July 12 [5 favorites]
You never see a paper called Machine Learning Predicts with Astonishing Accuracy Which Police Officers are Most Likely to Kill a Citizen
Where are you going to get the training data for that from?
posted by MisantropicPainforest at 5:12 AM on July 12 [4 favorites]
Where are you going to get the training data for that from?
posted by MisantropicPainforest at 5:12 AM on July 12 [4 favorites]
I thought it was clear that it was about being shot by anyone.
posted by Nancy Lebovitz at 5:13 AM on July 12 [2 favorites]
posted by Nancy Lebovitz at 5:13 AM on July 12 [2 favorites]
Social work Interventions are already happening.
There are already on the ground social service agnecies embedded in these communities who are working with high risk families mostly with known gang involvement. They also come to the hospital and respond to provide social service interventions while there as well.
Not everyone who gets shot is a part of this, there is so much random violence here as well. But there is a subset where police, the social services agency and families would all be aware and honestly not all surprized. But that doesn't mean it isn't an awful thing that happened. And everyone who is shot is a victim regardless of the circumstances.
There is a subset of people who have been shot ,mostly gang involved people but sometimes family members of gang involved people, who will be victims of gun violence more than once.
These people are of course people with complicated motivations and perspectives, often people who have witnessed tragedy after tragedy.
posted by AlexiaSky at 5:32 AM on July 12 [5 favorites]
There are already on the ground social service agnecies embedded in these communities who are working with high risk families mostly with known gang involvement. They also come to the hospital and respond to provide social service interventions while there as well.
Not everyone who gets shot is a part of this, there is so much random violence here as well. But there is a subset where police, the social services agency and families would all be aware and honestly not all surprized. But that doesn't mean it isn't an awful thing that happened. And everyone who is shot is a victim regardless of the circumstances.
There is a subset of people who have been shot ,mostly gang involved people but sometimes family members of gang involved people, who will be victims of gun violence more than once.
These people are of course people with complicated motivations and perspectives, often people who have witnessed tragedy after tragedy.
posted by AlexiaSky at 5:32 AM on July 12 [5 favorites]
They can predict crime with increasing accuracy?
...something something Minority Report something something
posted by AsYouKnow Bob at 5:57 AM on July 12 [2 favorites]
...something something Minority Report something something
posted by AsYouKnow Bob at 5:57 AM on July 12 [2 favorites]
I suspect that since they are working with police data that list of people most accurately predicted to be at risk of being shot comes from data the police collect and are willing to give out in aggregate form. So the data must have involve everyone who was shot in an incident of intimate partner violence who was involved in domestic violence calls, and it must have involved everyone who was shot in gang violence who was earlier involved in a gang violence call, and it is going to involve every black male who interacted with the police, with ultimate significant emphasis on those who drive a vehicle that leaves them vulnerable to traffic stops, driving at unfortunate times of day and who are particularly tall, muscular or dark.
The inclusion of guys with no criminal record other than driving a ganky car has horrified the authors of this study so they are talking more about that than the less loaded data. But the study numbers are large enough they have to also include the mother who provides a home for the young man who is getting involved in gang activity, who called the police twice last month because cars parked in front of her home twice last month and gang members got out and talked to her son and frightened her and she and her son are both at risk of gang violence. And they include the young woman with two toddlers who gets taken into a shelter after her her boyfriend smashes out the windows in her home. The study doesn't highlight the situation of the young gang member and his mother, nor the young woman with the violent boyfriend and her children because naturally they belong in the study; it shows that all five of them died from being shot three months later.
There were 797 homicides in Chicago in 2021. They were not all police shootings. And while one police shooting is obviously one too many, to reduce homicides it might be somewhat easier to work on strategies to reduce shootings in general, and to just assume that both the gangs and the police are locked into a Netanyahu-Hamas type violence spiral where the participants are more concerned with staying in power than anything else, and are using the other side's terror violence as justification for the violence and oppression they themselves inflict. You're not going to get the leaders among the guys with guns on board. You're going to have to work around them and at best can work with low level gang/police members who have not been as fully indoctrinated and do not have enough power to see the violence cycle as essential to their future success and survival.
As for there being no machine learning study as to which cops are most likely to shoot someone, I can tell you what data you need: Which cops interact with the public on traffic stops that are really municipal shakedowns; which cops respond to dispatched calls where they intervene in disputes; which cops are taking steroids; which cops have been give over powered weaponry; and which cops frequent social media or in person groups that promote an us vs. them mentality. Now that I have told you what data it is, I challenge you to go get any police force to release it to you, so that you can compile it. But seriously even the cops know within a few minutes of interacting with another cop, which ones are at risk of shooting someone. They are the cops who are both belligerent and terrified.
You don't need a study, you need to find a way to change the way police interact with the public. Simply changing who interacts with the public by reassigned a few steroid-fueled cops to desk jobs with slight pay increase will result in a new replacement cohort of cops who despise civilians after doing the traffic stops that are really municipal shakedowns (nothing makes you despise someone more than victimizing them), and who are fed up, frightened and disgusted after dealing with civilian disputes, and who then decompress by talking to other cops about what screwed up interactions they are having. It's a short stop from there to them coping with their fear by becoming belligerent and powering up by obtaining over powered weaponry and going on steroids.
The mental health intervention teams staffed by non-cop first responders is one of the ways people are trying to change who interacts with the public. I think everyone who is disgusted with police violence are in support of measures like that one.This study is there to cite, when talking to city hall about getting funding for social services, such as emergency mental health nurses or not closing the emergency domestic violence shelter program.
posted by Jane the Brown at 5:59 AM on July 12 [22 favorites]
The inclusion of guys with no criminal record other than driving a ganky car has horrified the authors of this study so they are talking more about that than the less loaded data. But the study numbers are large enough they have to also include the mother who provides a home for the young man who is getting involved in gang activity, who called the police twice last month because cars parked in front of her home twice last month and gang members got out and talked to her son and frightened her and she and her son are both at risk of gang violence. And they include the young woman with two toddlers who gets taken into a shelter after her her boyfriend smashes out the windows in her home. The study doesn't highlight the situation of the young gang member and his mother, nor the young woman with the violent boyfriend and her children because naturally they belong in the study; it shows that all five of them died from being shot three months later.
There were 797 homicides in Chicago in 2021. They were not all police shootings. And while one police shooting is obviously one too many, to reduce homicides it might be somewhat easier to work on strategies to reduce shootings in general, and to just assume that both the gangs and the police are locked into a Netanyahu-Hamas type violence spiral where the participants are more concerned with staying in power than anything else, and are using the other side's terror violence as justification for the violence and oppression they themselves inflict. You're not going to get the leaders among the guys with guns on board. You're going to have to work around them and at best can work with low level gang/police members who have not been as fully indoctrinated and do not have enough power to see the violence cycle as essential to their future success and survival.
As for there being no machine learning study as to which cops are most likely to shoot someone, I can tell you what data you need: Which cops interact with the public on traffic stops that are really municipal shakedowns; which cops respond to dispatched calls where they intervene in disputes; which cops are taking steroids; which cops have been give over powered weaponry; and which cops frequent social media or in person groups that promote an us vs. them mentality. Now that I have told you what data it is, I challenge you to go get any police force to release it to you, so that you can compile it. But seriously even the cops know within a few minutes of interacting with another cop, which ones are at risk of shooting someone. They are the cops who are both belligerent and terrified.
You don't need a study, you need to find a way to change the way police interact with the public. Simply changing who interacts with the public by reassigned a few steroid-fueled cops to desk jobs with slight pay increase will result in a new replacement cohort of cops who despise civilians after doing the traffic stops that are really municipal shakedowns (nothing makes you despise someone more than victimizing them), and who are fed up, frightened and disgusted after dealing with civilian disputes, and who then decompress by talking to other cops about what screwed up interactions they are having. It's a short stop from there to them coping with their fear by becoming belligerent and powering up by obtaining over powered weaponry and going on steroids.
The mental health intervention teams staffed by non-cop first responders is one of the ways people are trying to change who interacts with the public. I think everyone who is disgusted with police violence are in support of measures like that one.This study is there to cite, when talking to city hall about getting funding for social services, such as emergency mental health nurses or not closing the emergency domestic violence shelter program.
posted by Jane the Brown at 5:59 AM on July 12 [22 favorites]
If we are offering up intervention options, how about getting rid of the guns? Or is that still a bridge too far, even on Metafilter?
posted by Literaryhero at 6:10 AM on July 12 [9 favorites]
posted by Literaryhero at 6:10 AM on July 12 [9 favorites]
Well it looks like almost no one commenting even bothered to read the paper, let alone read the first 5 paragraphs. This paper says nothing about who is committing a crime. It says nothing about arrests. It is entirely focused on predicting who is a VICTIM of a crime, and suggests using these predictions for social work interventions as a way to reduce the probability that these individuals are victims of being fucking shot.
The study doesn't highlight the situation of the young gang member and his mother, nor the young woman with the violent boyfriend and her children because naturally they belong in the study; it shows that all five of them died from being shot three months later.
This is quantitative social science academic paper that will be submitted to a journal with a restrictive word count, and the authors have limited resources and training. What would you like them to do? Train as ethnographers and interview some of the victims? That’s not how this works.
posted by MisantropicPainforest at 6:12 AM on July 12 [20 favorites]
The study doesn't highlight the situation of the young gang member and his mother, nor the young woman with the violent boyfriend and her children because naturally they belong in the study; it shows that all five of them died from being shot three months later.
This is quantitative social science academic paper that will be submitted to a journal with a restrictive word count, and the authors have limited resources and training. What would you like them to do? Train as ethnographers and interview some of the victims? That’s not how this works.
posted by MisantropicPainforest at 6:12 AM on July 12 [20 favorites]
Thanks to Jane the Brown for the insights som of the cycles of violence
This kind of academic study certain doesn’t address all the issues, but if it can help better target effective Inventions, that seems good to me.
For an example of the wrong type of intervention., see
arresting people based on social media posts
posted by CostcoCultist at 6:24 AM on July 12 [2 favorites]
This kind of academic study certain doesn’t address all the issues, but if it can help better target effective Inventions, that seems good to me.
For an example of the wrong type of intervention., see
arresting people based on social media posts
posted by CostcoCultist at 6:24 AM on July 12 [2 favorites]
If we are offering up intervention options, how about getting rid of the guns? Or is that still a bridge too far, even on Metafilter?
On that front, while I think it would help to ban various kinds of gun sales and restrict ownership, we currently have a broken society with a lot of guns floating around, and it's hard to imagine how we'd get rid of the guns in any meaningful sense. We don't have a citizenry who can be reached by "let's not have so many guns, here's a big buy-back". It would be extremely difficult to start enforcing some kind of "guns all live at the gun depot and you pick them up for the day to go hunting" or "everyone has a gun safe which is inspected annually", partly because our society doesn't have shared values around safety, partly because we don't have funding to implement these programs and partly because our police forces are so corrupt that this kind of regulation would be used as an excuse to target, extort and rob people of color*.
We've built a system where there are so many guns and they are so easy to get that reducing gun ownership would need to be a multigeneration project, and we'd need a lot more civic trust to make that work at all.
This is why, regardless of how one feels about guns, it seems like other interventions are the only ones that have even a faint possibility of working in the medium term.
*We had a bike licensing program in my city fifteen years or so ago, and they dropped it because...the police used it to bust Black kids for riding bikes, get them into the system and of course impound the bikes (which could then be resold at cop auctions). It was too racist even for Minneapolis, and it got dropped. Now imagine trying to manage gun law.
posted by Frowner at 6:25 AM on July 12 [22 favorites]
A slight aside Acclivus is one of the social service agencies doing this work that I was taking about above. I linked the information on the hospital response program.
posted by AlexiaSky at 6:54 AM on July 12 [2 favorites]
posted by AlexiaSky at 6:54 AM on July 12 [2 favorites]
If we are offering up intervention options, how about getting rid of the guns? Or is that still a bridge too far, even on Metafilter?
posted by Literaryhero
Ok. How do we do that? Asks this Chicagoan. Yes, please.
posted by tiny frying pan at 7:02 AM on July 12 [2 favorites]
posted by Literaryhero
Ok. How do we do that? Asks this Chicagoan. Yes, please.
posted by tiny frying pan at 7:02 AM on July 12 [2 favorites]
Chicago initiative raises $100M to expand violence intervention programs
I'm pleased our new mayor is trying new things. The police police police approach has gotten us nowhere, and worse.
posted by tiny frying pan at 7:08 AM on July 12 [4 favorites]
I'm pleased our new mayor is trying new things. The police police police approach has gotten us nowhere, and worse.
posted by tiny frying pan at 7:08 AM on July 12 [4 favorites]
Part II: The Body Camera (Alec Karakatsanis's Copaganda Substack)
The public discussion of body cameras fits a pattern in the history of U.S. policing. In the wake of historical police violence against marginalized groups, followed by historical police violence against people protesting the original police violence, there are always calls by people who care about violence to address the underlying root causes. Addressing root causes–such as reducing inequality or making specific investments in housing, health care, education, transportation, toxic cleanup, nutritious food, sustainable ecological practices, the arts, and building strong networks of community members helping themselves through mutual aid and self-empowerment—are mostly ignored by politicians, even though evidence suggests that these strategies lead to short, medium, and long-term violence reduction.posted by box at 7:09 AM on July 12 [7 favorites]
Instead, the segment of the public who cares about improving the lives of marginalized people is distracted and ultimately appeased with the promise of a “reform” like the body camera. At virtually every key moment in U.S. history, those “reforms” have made the government’s punishment bureaucracy bigger, more profitable to private profiteers, more politically powerful, and thus less accountable to local democracy.
Reducing Gun Violence in Chicago
Chicago CRED believes the individuals most at risk are not the problem—they are the solution.
posted by tiny frying pan at 7:11 AM on July 12 [1 favorite]
Chicago CRED believes the individuals most at risk are not the problem—they are the solution.
posted by tiny frying pan at 7:11 AM on July 12 [1 favorite]
ENDING THE CYCLE OF VIOLENCE IN CHICAGO BY ADDRESSING EVERY PHASE.
The Institute for Nonviolence Chicago
posted by tiny frying pan at 7:11 AM on July 12 [1 favorite]
The Institute for Nonviolence Chicago
posted by tiny frying pan at 7:11 AM on July 12 [1 favorite]
Violence Prevention, The Chicago Recovery Plan
posted by tiny frying pan at 7:12 AM on July 12 [1 favorite]
posted by tiny frying pan at 7:12 AM on July 12 [1 favorite]
An experimental evidence from a program in DC showed body cams had no affect on police behavior.
posted by MisantropicPainforest at 7:17 AM on July 12 [1 favorite]
posted by MisantropicPainforest at 7:17 AM on July 12 [1 favorite]
reducing gun ownership would need to be a multigeneration project
Well then we'd better fucking start now.
posted by Mr.Know-it-some at 7:21 AM on July 12 [4 favorites]
Well then we'd better fucking start now.
posted by Mr.Know-it-some at 7:21 AM on July 12 [4 favorites]
A) This is terrifying.
B) Nothing good can come from this.
Only skimmed it.
This doesn't seem scary to me, or particularly impressive apart from the data collection? It's Chicago. Honestly how hard is it to predict who's likely to get shot in Chicago? "People associated with gangs and people associated with those people" will get you some very high percentage of the way there. And, lo, the model *seeeeeeeeems* to be capturing lots of causal and noncausal correlates of first- and second-degree gang association.
They're hiding a lot of what their model is saying inside it being a mysterious machine-learning black box, and they *seem* to be doing the usual garbage machine-learning thing of just throwing a kitchen sink full of variables at the problem and letting the computer sort it out instead of like "having a theory" or "thinking about the operationalization of their theoretical variables." I expect that if they just ran a logit or duration model over a well-thought-out set of IVs they'd have something 90-whatever percent as good, and then we'd be able to see clearly what the model is twigging on.
posted by GCU Sweet and Full of Grace at 7:26 AM on July 12 [1 favorite]
B) Nothing good can come from this.
Only skimmed it.
This doesn't seem scary to me, or particularly impressive apart from the data collection? It's Chicago. Honestly how hard is it to predict who's likely to get shot in Chicago? "People associated with gangs and people associated with those people" will get you some very high percentage of the way there. And, lo, the model *seeeeeeeeems* to be capturing lots of causal and noncausal correlates of first- and second-degree gang association.
They're hiding a lot of what their model is saying inside it being a mysterious machine-learning black box, and they *seem* to be doing the usual garbage machine-learning thing of just throwing a kitchen sink full of variables at the problem and letting the computer sort it out instead of like "having a theory" or "thinking about the operationalization of their theoretical variables." I expect that if they just ran a logit or duration model over a well-thought-out set of IVs they'd have something 90-whatever percent as good, and then we'd be able to see clearly what the model is twigging on.
posted by GCU Sweet and Full of Grace at 7:26 AM on July 12 [1 favorite]
MisantropicPainforest: “An experimental evidence from a program in DC showed body cams had no affect on police behavior.”“The Body Camera,” Alec Karakatsanis, Alec’s Copaganda Newsletter, 11 July 2024
I studied a decade of public discourse about body cameras as well as the internal statements of police, prosecutors, and industry insiders to understand a simple question: how did a coveted tool of repression that police and prosecutors desperately wanted (but couldn’t get funded) come to be seen by the public—especially liberals—as an essential police reform?“Part II: The Body Camera,” Id., 12 July 2024
posted by ob1quixote at 7:28 AM on July 12 [3 favorites]
> An experimental evidence from a program in DC showed body cams had no affect on police behavior.
It does help break the "police did nothing wrong" shield of lies.
Which can build public support for stuff like enforcing the use of body cams (if your cam is off and you engage with a citizen, you get fired. If you fail to store your cam results after an engagement with a citizen, you get fired.)
That helps remove the "presumption of innocence" shield. Police powers are given in exchange for the presumption of guilt; they have to prove their police powers based interaction was done correctly, or they get fired.
Which finally lets you go after qualified immunity and similar; when police powers are a privilege not an entitlement, they don't get to be first-class citizens.
It isn't enough to solve systemic violence, but would it really have no affect? There are plenty of places with a lot less police-on-citizen violence than the USA.
posted by NotAYakk at 7:36 AM on July 12 [2 favorites]
It does help break the "police did nothing wrong" shield of lies.
Which can build public support for stuff like enforcing the use of body cams (if your cam is off and you engage with a citizen, you get fired. If you fail to store your cam results after an engagement with a citizen, you get fired.)
That helps remove the "presumption of innocence" shield. Police powers are given in exchange for the presumption of guilt; they have to prove their police powers based interaction was done correctly, or they get fired.
Which finally lets you go after qualified immunity and similar; when police powers are a privilege not an entitlement, they don't get to be first-class citizens.
It isn't enough to solve systemic violence, but would it really have no affect? There are plenty of places with a lot less police-on-citizen violence than the USA.
posted by NotAYakk at 7:36 AM on July 12 [2 favorites]
Jane the Brown: One more for your list of ways to identify dangerous cops-- the number of complaints against them.
posted by Nancy Lebovitz at 7:41 AM on July 12 [4 favorites]
posted by Nancy Lebovitz at 7:41 AM on July 12 [4 favorites]
this isn't a neural network model. causation has nothing to do with it, the authors are explicit about that they are only measuring one of the potential outcomes (see pages 4 and 5). This would be the exact same thing if they did a normal linear model, just the performacne would likely be worse.
posted by MisantropicPainforest at 8:41 AM on July 12 [1 favorite]
posted by MisantropicPainforest at 8:41 AM on July 12 [1 favorite]
...and we'd have visible coefficients for IVs and interactions thereof. Unless those are buried somewhere I didn't notice.
posted by GCU Sweet and Full of Grace at 8:44 AM on July 12 [1 favorite]
posted by GCU Sweet and Full of Grace at 8:44 AM on July 12 [1 favorite]
pages 58-64 have what you're looking for.
posted by MisantropicPainforest at 8:46 AM on July 12 [3 favorites]
posted by MisantropicPainforest at 8:46 AM on July 12 [3 favorites]
Also without a DAG, who cares? W/o an identified effect the coefficient on any correlated variable could be big/small, positive/negative, significant/not significant in a model and it would make no difference on any intervention.
posted by MisantropicPainforest at 8:48 AM on July 12 [2 favorites]
posted by MisantropicPainforest at 8:48 AM on July 12 [2 favorites]
This is quantitative social science academic paper that will be submitted to a journal with a restrictive word count, and the authors have limited resources and training. What would you like them to do? Train as ethnographers and interview some of the victims? That’s not how this works.
I think you mistake my point about the two families and their importance. I think the study is excellent, and the reason they didn't describe the plight of people caught up in gang violence or domestic violence, is because they were so obviously the people the study was about and who would be the first rank to get supportive social services. They just didn't want young black men who drive home at dusk to be excluded, so they made special mention of them. I made mention of the two families because I thought some of the people reading had missed the fact that they were included and essential, and were presuming the paper was only about police shooting.
It will, of course, be a lot easier to re-home the woman with the son getting involved in gang activities, and find a job for her son to keep him from turning to the gangs for an income, than it will be to find a way to protect the already working young black man from the dangers of his commute. But you solve the problems you can solve, and spend money where it makes improvements, as opposed to not spending money on one problem because it doesn't help another.
posted by Jane the Brown at 9:14 AM on July 12 [4 favorites]
I think you mistake my point about the two families and their importance. I think the study is excellent, and the reason they didn't describe the plight of people caught up in gang violence or domestic violence, is because they were so obviously the people the study was about and who would be the first rank to get supportive social services. They just didn't want young black men who drive home at dusk to be excluded, so they made special mention of them. I made mention of the two families because I thought some of the people reading had missed the fact that they were included and essential, and were presuming the paper was only about police shooting.
It will, of course, be a lot easier to re-home the woman with the son getting involved in gang activities, and find a job for her son to keep him from turning to the gangs for an income, than it will be to find a way to protect the already working young black man from the dangers of his commute. But you solve the problems you can solve, and spend money where it makes improvements, as opposed to not spending money on one problem because it doesn't help another.
posted by Jane the Brown at 9:14 AM on July 12 [4 favorites]
pages 58-64 have what you're looking for.
Thanks. Sort of, sure. If I were a reviewer, I'd ruin their day by asking for the proportional reduction in error over a plain old logit. Or a PRE over the top 500 people, since this is the kind of dataset where it's going to be very hard to improve much over exceedingly naive models.
Also without a DAG, who cares? W/o an identified effect the coefficient on any correlated variable could be big/small, positive/negative, significant/not significant in a model and it would make no difference on any intervention.
A nice thing about policy is that you can just go ahead and try an intervention that seems sensible from this analysis on a small scale and see whether it seems to work, and worry about getting the identification right in the program evaluation afterwards.
posted by GCU Sweet and Full of Grace at 9:32 AM on July 12 [2 favorites]
Thanks. Sort of, sure. If I were a reviewer, I'd ruin their day by asking for the proportional reduction in error over a plain old logit. Or a PRE over the top 500 people, since this is the kind of dataset where it's going to be very hard to improve much over exceedingly naive models.
Also without a DAG, who cares? W/o an identified effect the coefficient on any correlated variable could be big/small, positive/negative, significant/not significant in a model and it would make no difference on any intervention.
A nice thing about policy is that you can just go ahead and try an intervention that seems sensible from this analysis on a small scale and see whether it seems to work, and worry about getting the identification right in the program evaluation afterwards.
posted by GCU Sweet and Full of Grace at 9:32 AM on July 12 [2 favorites]
I'm very happy to see this result, and hope it can be rolled out more broadly.
This methodology worked well in Richmond, CA, and was the cornerstone of a community-centered policing model that substantially reduced violence. It was very controversial because you were spending money to redirect services to people who were at risk of criminal behavior, but it was massively effective in making Richmond much safer and preventing people from.
In those cases they looked at the data by hand without using ML. However, by identifying the people who are most likely to become involved in violence you can save lives and direct resources most effectively to the people who need it, which helps everybody.
posted by a neat little robot at 10:39 AM on July 12 [3 favorites]
This methodology worked well in Richmond, CA, and was the cornerstone of a community-centered policing model that substantially reduced violence. It was very controversial because you were spending money to redirect services to people who were at risk of criminal behavior, but it was massively effective in making Richmond much safer and preventing people from.
In those cases they looked at the data by hand without using ML. However, by identifying the people who are most likely to become involved in violence you can save lives and direct resources most effectively to the people who need it, which helps everybody.
posted by a neat little robot at 10:39 AM on July 12 [3 favorites]
They needed ML to predict that someone who has committed crimes in the past will likely commit crimes in the future? And didn't look at the broader socioeconomic issues that so many face? Sounds very tech-bro-y to me.
posted by tommasz at 10:57 AM on July 12 [1 favorite]
posted by tommasz at 10:57 AM on July 12 [1 favorite]
They needed ML to predict that someone who has committed crimes in the past will likely commit crimes in the future? And didn't look at the broader socioeconomic issues that so many face? Sounds very tech-bro-y to me.
You couldn't read the abstract, paper, or any of the discussion here before a drive by snark?
posted by MisantropicPainforest at 11:02 AM on July 12 [5 favorites]
You couldn't read the abstract, paper, or any of the discussion here before a drive by snark?
posted by MisantropicPainforest at 11:02 AM on July 12 [5 favorites]
They needed ML to predict that someone who has committed crimes in the past will likely commit crimes in the future? And didn't look at the broader socioeconomic issues that so many face? Sounds very tech-bro-y to me.
You couldn't read the abstract, paper, or any of the discussion here before a drive by snark?
Sheesh, I just read the post and knew that wasn't what this paper was about.
posted by oneirodynia at 11:43 AM on July 12 [1 favorite]
You couldn't read the abstract, paper, or any of the discussion here before a drive by snark?
Sheesh, I just read the post and knew that wasn't what this paper was about.
posted by oneirodynia at 11:43 AM on July 12 [1 favorite]
tommasz, the paper was about who was likely to be a *victim* of crimes, not who was likely to be a criminal.
posted by Nancy Lebovitz at 12:33 PM on July 12 [1 favorite]
posted by Nancy Lebovitz at 12:33 PM on July 12 [1 favorite]
Assuming this analysis is good: you would want to be very careful that any planned intervention is not stigmatising and actually works. It's not hard to imagine that when someone looks at your file and sees the intervention team is working with you, they behave differently.
(Real world example: where I live, schools are publicly funded but parent contributions are part of the mix. They are ranked by socioeconomic decile, and the lowest decile schools receive more funding on the premise that the parents in higher ranked schools can do more fundraising and the community behind a lower ranked school needs more help. Despite the fact that deciles are not an indicator of school quality, and in fact this is independently measured and you can find excellent low decile schools, parents typically treat the decile ranking as though it were a school quality measure and try to send their kids to high decile schools, depriving them of student based funding and stigmatising the kids who go to them).
posted by i_am_joe's_spleen at 1:02 PM on July 12 [3 favorites]
(Real world example: where I live, schools are publicly funded but parent contributions are part of the mix. They are ranked by socioeconomic decile, and the lowest decile schools receive more funding on the premise that the parents in higher ranked schools can do more fundraising and the community behind a lower ranked school needs more help. Despite the fact that deciles are not an indicator of school quality, and in fact this is independently measured and you can find excellent low decile schools, parents typically treat the decile ranking as though it were a school quality measure and try to send their kids to high decile schools, depriving them of student based funding and stigmatising the kids who go to them).
posted by i_am_joe's_spleen at 1:02 PM on July 12 [3 favorites]
“Is The US a Police State?” [21:25]—Second Thought, 12 July 2024
posted by ob1quixote at 4:58 PM on July 12 [1 favorite]
posted by ob1quixote at 4:58 PM on July 12 [1 favorite]
“Is The US a Police State?”
A rare counterexample to Betteridge's Law of Headlines!
posted by adrienneleigh at 5:19 PM on July 12 [3 favorites]
A rare counterexample to Betteridge's Law of Headlines!
posted by adrienneleigh at 5:19 PM on July 12 [3 favorites]
tommasz, the paper was about who was likely to be a *victim* of crimes, not who was likely to be a criminal.
My mistake, I skimmed and should have read it more closely.
posted by tommasz at 11:47 AM on July 13
My mistake, I skimmed and should have read it more closely.
posted by tommasz at 11:47 AM on July 13
« Older Reverse Engineering TicketMaster's Rotating... | "The photos will be our treasure" Newer »
B) Nothing good can come from this.
C) This TV show was awesome before Jim Caviezel lost his fucking mind.
posted by Literaryhero at 3:12 AM on July 12 [10 favorites]