6 ms·

Predictive policing lacks accuracy tests

55 points by danso 7y ago

Barrin92 7y ago
not only does it lack accuracy tests, it has been studied in the past, for example by RAND¹ and there's no evidence that it works at all. more importantly, it should be noted that these are essentially unproven systems that are being beta tested on largely minority groups. That something like this is happening in a country with a rule of law and at least in theory constitutional safeguards is reprehensible. [1]https://link.springer.com/article/10.1007/s11292-016-9272-0?wt_mc=Affiliate.CommissionJunction.3.EPR1089.DeepLink https://link.springer.com/article/10.1007/s11292-016-9272-0?...
- capnrefsmmat 7y ago
  The study you link is of a very different kind of system: creating a list of people considered likely to be victims of homicide or violent crime, and trying to intercede. PredPol, the system in the article, does not do that. PredPol is primarily about mapping crime in space, and trying to predict which regions of space are most likely to have future crime. It doesn't focus on individuals in any way. The people behind PredPol did run a randomized controlled field trial with LAPD and claim to have found reductions in crime by directing patrols to predicted hotspots [1]. Hotspots policing studies generally have found reductions in crime when directing patrols to hotspots, or conducting targeted interventions (like improving street lighting, talking to local businesses, etc.). But I should note that running a good, well-controlled experiment of something like policing is very hard, so the quality of the evidence is not exactly astounding. However, there are indeed constitutional implications, as you note. For example, if a model says an area is a "high crime area", police may be able to use that as part of the justification for stop-and-frisk, reducing the need for any individual suspicion or reason to search you. And, of course, there are fairness problems: high-crime areas tend to be poor and minority areas, and hotspots policing may harm trust in police (if they harass the residents for petty offenses) and either lead to overpolicing or underreporting of offenses. I have a fairly thorough review of the literature on my website [2]. [1] https://www.tandfonline.com/doi/abs/10.1080/01621459.2015.1077710 https://www.tandfonline.com/doi/abs/10.1080/01621459.2015.10... [2] https://www.refsmmat.com/notebooks/predictive-policing.html https://www.refsmmat.com/notebooks/predictive-policing.html
- nroets 7y ago
  How does the current system work ? AFAIK, the police chief looks at various data sources like heat maps and police reports. He will have an incentive organize patrols to protect the politically connected and their property. Then his job is safe.
wry_discontent 7y ago
It's pretty well known that increased policing doesn't decrease crime. These are the same tactics in new clothes used to target PoC. Even if the ML is _working properly_ the biased nature of the existing justice system is going to cause it to target PoC disproportionately.
- bsder 7y ago
  > It's pretty well known that increased policing doesn't decrease crime. I suspect that statement has a LOT of context behind it--citation please? I could believe that increased policing simply pushes the crime around rather than gets rid of it. > These are the same tactics in new clothes used to target PoC. Agreed, unfortunately.
  - capnrefsmmat 7y ago
    There's a lot of existing research here. Generally, the studies show that directing police patrols to high-crime areas does reduce reported crimes. For example, a recent systematic review of studies found that increasing policing to high-crime areas "generates small but noteworthy crime reductions, and these crime control benefits diffuse into areas immediately surrounding targeted crime hot spots." [1] The question of whether policing simply pushes crime to another area is a popular one, and there's been a lot of work to answer it. It appears the answer is usually "no." Crime depends on several factors existing in one place: opportunities for crime, a lack of guardians, and motivated offenders. If you fix one area, the crime will spread to adjacent areas only if those three factors are present there. The fairness and equity problems are, of course, severe. You may reduce crime with hotspots policing, but as the systematic review notes, "only a small number of studies examine the impacts of hot spots policing on police-community relations," or what disproportionate effects they have on any specific group. [1] https://doi.org/10.1080/07418825.2012.673632 https://doi.org/10.1080/07418825.2012.673632
  - bobcostas55 7y ago
    >I suspect that statement has a LOT of context behind it--citation please? It's completely false of course. Not only does policing decrease crime, the US is massively underpoliced under any cost-benefit analysis. http://www.princeton.edu/~smello/papers/Cops.pdf http://www.princeton.edu/~smello/papers/Cops.pdf https://www.mitpressjournals.org/doi/abs/10.1162/REST_a_00694?af=R https://www.mitpressjournals.org/doi/abs/10.1162/REST_a_0069...
freen 7y ago
Unless you are very careful, machine learning is the process of projecting the biases of yesterday into the predictions of today.
thewhitetulip 7y ago
I don't understand why this isn't obvious. Predictive analysis works if you have correct set of data. DOJ is biased towards African Americans, their data is skewed + you can't really predict what someone could do.we are humans and not rule following robots.
nroets 7y ago
The alternative to a rule driven machine is human decision makers whose emotions will play a role, sooner or later. (Prejudice and favouritism is just two emotions that are bad in this case) With a rule driven machine, the rules and be reviewed for any direct form of prejudice and favouritism.
- danso 7y ago
  We've already had rule-driven frameworks: three strikes and mandatory minimums, the latter of which resulted in huge sentences for non-violent drug offenses – particularly those involving crack, a drug more prevalent in black communities, and for which the threshold to trigger a mandatory sentence was 1/100th that of powder cocaine: https://www.aclu.org/other/cracks-system-20-years-unjust-federal-crack-cocaine-law https://www.aclu.org/other/cracks-system-20-years-unjust-fed...
  - AlexTWithBeard 7y ago
    That's something I never managed to understand: if I'm a black person, and I know that current laws will unfairly publish me for drugs possession, and I know that police tenderly loves to search me at any opportunity... Doesn't it mean that I should stay as far as I can from crack and beg all my friends and relatives to do the same?
    - homonculus1 7y ago
      I think there comes a point where you have so little to lose and prison is so normalized in your social environment that the dignity reclaimed by defying the police state and possessing whatever you please outweighs the chance of getting caught. It's not so much a logical defect as it is a different set of premises: if prison looks like a constant threat, it stops being a deterrent and at that point you may as well have some semi-internal locus of control.
    - danso 7y ago
      You could say that for any crime or malady of society, that it's confusing to you why millions of fellow citizens have chosen something you've found easy and obvious to avoid yourself. That doesn't address or excuse why one variation is vastly disproportionately punished.
excalibur 7y ago
Why would they call it PredPol? Don't they get that this could be just as easily interpreted as "Predatory Policing," which is quite likely more accurate?
martingoodson 7y ago
'We have demonstrated that predictive policing of drug crimes results in increasingly disproportionate policing of historically over‐policed communities. Over‐policing imposes real costs on these communities. Increased police scrutiny and surveillance have been linked to worsening mental and physical health;and, in the extreme, additional police contact will create additional opportunities for police violence in over‐policed areas.' https://rss.onlinelibrary.wiley.com/doi/full/10.1111/j.1740-9713.2016.00960.x https://rss.onlinelibrary.wiley.com/doi/full/10.1111/j.1740-...
- AlexTWithBeard 7y ago
  Increased police scrutiny and surveillance have been linked to worsening mental and physical health I find it surprising. I would be happy to see more police in where I live. In my layman's mind police, pretty much by definition, prevents and deters crime. They may not be perfect in that, but in most cases they're fine.
  - martingoodson 7y ago
    They are specifically talking about 'over-policing', which is disproportionate focus of police activity on some sectors of the community.
  - mschuster91 7y ago
    > In my layman's mind police, pretty much by definition, prevents and deters crime. That would be the case if police were adequately trained and equipped. However US police forces on average (!!) are trained only for not even 9 months (per https://www.snopes.com/fact-check/police-training-killings-usa-nordic/ https://www.snopes.com/fact-check/police-training-killings-u...), which means that there are likely cases in which only weeks of training were given. German police, in contrast, train for two and a half / three years (per https://de.wikipedia.org/wiki/Polizeiausbildung_in_Deutschland https://de.wikipedia.org/wiki/Polizeiausbildung_in_Deutschla...). That, combined with a lack of hiring standards (e.g. https://www.portlandmercury.com/blogtown/2019/06/19/26671582/portland-police-lower-officer-hiring-standards-to-address-poor-recruitment-rates https://www.portlandmercury.com/blogtown/2019/06/19/26671582...) leads to the perception of police being nothing more than paid thugs with a uniform. The massive availability of guns as well as former military equipment in police hands doesn't help either, quite to the contrary. When I have an easy sledgehammer (=a tank and a gun), I'll shoot first and ask questions later. When it comes to "overpolicing", you also have to note that it's way easier to arrest someone dealing/smoking cannabis on the street (mostly poor PoC) compared to arresting a rich white banker who is snorting cocaine on his desk.

55 points by danso 7y ago

not only does it lack accuracy tests, it has been studied in the past, for example by RAND¹ and there's no evidence that it works at all. more importantly, it should be noted that these are essentially unproven systems that are being beta tested on largely minority groups. That something like this is happening in a country with a rule of law and at least in theory constitutional safeguards is reprehensible. [1]https://link.springer.com/article/10.1007/s11292-016-9272-0?wt_mc=Affiliate.CommissionJunction.3.EPR1089.DeepLink https://link.springer.com/article/10.1007/s11292-016-9272-0?...

capnrefsmmat 7y ago
The study you link is of a very different kind of system: creating a list of people considered likely to be victims of homicide or violent crime, and trying to intercede. PredPol, the system in the article, does not do that. PredPol is primarily about mapping crime in space, and trying to predict which regions of space are most likely to have future crime. It doesn't focus on individuals in any way. The people behind PredPol did run a randomized controlled field trial with LAPD and claim to have found reductions in crime by directing patrols to predicted hotspots [1]. Hotspots policing studies generally have found reductions in crime when directing patrols to hotspots, or conducting targeted interventions (like improving street lighting, talking to local businesses, etc.). But I should note that running a good, well-controlled experiment of something like policing is very hard, so the quality of the evidence is not exactly astounding. However, there are indeed constitutional implications, as you note. For example, if a model says an area is a "high crime area", police may be able to use that as part of the justification for stop-and-frisk, reducing the need for any individual suspicion or reason to search you. And, of course, there are fairness problems: high-crime areas tend to be poor and minority areas, and hotspots policing may harm trust in police (if they harass the residents for petty offenses) and either lead to overpolicing or underreporting of offenses. I have a fairly thorough review of the literature on my website [2]. [1] https://www.tandfonline.com/doi/abs/10.1080/01621459.2015.1077710 https://www.tandfonline.com/doi/abs/10.1080/01621459.2015.10... [2] https://www.refsmmat.com/notebooks/predictive-policing.html https://www.refsmmat.com/notebooks/predictive-policing.html
nroets 7y ago
How does the current system work ? AFAIK, the police chief looks at various data sources like heat maps and police reports. He will have an incentive organize patrols to protect the politically connected and their property. Then his job is safe.

wry_discontent 7y ago

It's pretty well known that increased policing doesn't decrease crime. These are the same tactics in new clothes used to target PoC. Even if the ML is _working properly_ the biased nature of the existing justice system is going to cause it to target PoC disproportionately.

bsder 7y ago
> It's pretty well known that increased policing doesn't decrease crime. I suspect that statement has a LOT of context behind it--citation please? I could believe that increased policing simply pushes the crime around rather than gets rid of it. > These are the same tactics in new clothes used to target PoC. Agreed, unfortunately.
- capnrefsmmat 7y ago
  There's a lot of existing research here. Generally, the studies show that directing police patrols to high-crime areas does reduce reported crimes. For example, a recent systematic review of studies found that increasing policing to high-crime areas "generates small but noteworthy crime reductions, and these crime control benefits diffuse into areas immediately surrounding targeted crime hot spots." [1] The question of whether policing simply pushes crime to another area is a popular one, and there's been a lot of work to answer it. It appears the answer is usually "no." Crime depends on several factors existing in one place: opportunities for crime, a lack of guardians, and motivated offenders. If you fix one area, the crime will spread to adjacent areas only if those three factors are present there. The fairness and equity problems are, of course, severe. You may reduce crime with hotspots policing, but as the systematic review notes, "only a small number of studies examine the impacts of hot spots policing on police-community relations," or what disproportionate effects they have on any specific group. [1] https://doi.org/10.1080/07418825.2012.673632 https://doi.org/10.1080/07418825.2012.673632
- bobcostas55 7y ago
  >I suspect that statement has a LOT of context behind it--citation please? It's completely false of course. Not only does policing decrease crime, the US is massively underpoliced under any cost-benefit analysis. http://www.princeton.edu/~smello/papers/Cops.pdf http://www.princeton.edu/~smello/papers/Cops.pdf https://www.mitpressjournals.org/doi/abs/10.1162/REST_a_00694?af=R https://www.mitpressjournals.org/doi/abs/10.1162/REST_a_0069...

freen 7y ago

Unless you are very careful, machine learning is the process of projecting the biases of yesterday into the predictions of today.

thewhitetulip 7y ago

I don't understand why this isn't obvious. Predictive analysis works if you have correct set of data. DOJ is biased towards African Americans, their data is skewed + you can't really predict what someone could do.we are humans and not rule following robots.

nroets 7y ago

The alternative to a rule driven machine is human decision makers whose emotions will play a role, sooner or later. (Prejudice and favouritism is just two emotions that are bad in this case) With a rule driven machine, the rules and be reviewed for any direct form of prejudice and favouritism.

danso 7y ago
We've already had rule-driven frameworks: three strikes and mandatory minimums, the latter of which resulted in huge sentences for non-violent drug offenses – particularly those involving crack, a drug more prevalent in black communities, and for which the threshold to trigger a mandatory sentence was 1/100th that of powder cocaine: https://www.aclu.org/other/cracks-system-20-years-unjust-federal-crack-cocaine-law https://www.aclu.org/other/cracks-system-20-years-unjust-fed...
- AlexTWithBeard 7y ago
  That's something I never managed to understand: if I'm a black person, and I know that current laws will unfairly publish me for drugs possession, and I know that police tenderly loves to search me at any opportunity... Doesn't it mean that I should stay as far as I can from crack and beg all my friends and relatives to do the same?
  - homonculus1 7y ago
    I think there comes a point where you have so little to lose and prison is so normalized in your social environment that the dignity reclaimed by defying the police state and possessing whatever you please outweighs the chance of getting caught. It's not so much a logical defect as it is a different set of premises: if prison looks like a constant threat, it stops being a deterrent and at that point you may as well have some semi-internal locus of control.
  - danso 7y ago
    You could say that for any crime or malady of society, that it's confusing to you why millions of fellow citizens have chosen something you've found easy and obvious to avoid yourself. That doesn't address or excuse why one variation is vastly disproportionately punished.

excalibur 7y ago

Why would they call it PredPol? Don't they get that this could be just as easily interpreted as "Predatory Policing," which is quite likely more accurate?

martingoodson 7y ago

'We have demonstrated that predictive policing of drug crimes results in increasingly disproportionate policing of historically over‐policed communities. Over‐policing imposes real costs on these communities. Increased police scrutiny and surveillance have been linked to worsening mental and physical health;and, in the extreme, additional police contact will create additional opportunities for police violence in over‐policed areas.' https://rss.onlinelibrary.wiley.com/doi/full/10.1111/j.1740-9713.2016.00960.x https://rss.onlinelibrary.wiley.com/doi/full/10.1111/j.1740-...

AlexTWithBeard 7y ago
Increased police scrutiny and surveillance have been linked to worsening mental and physical health I find it surprising. I would be happy to see more police in where I live. In my layman's mind police, pretty much by definition, prevents and deters crime. They may not be perfect in that, but in most cases they're fine.
- martingoodson 7y ago
  They are specifically talking about 'over-policing', which is disproportionate focus of police activity on some sectors of the community.
- mschuster91 7y ago
  > In my layman's mind police, pretty much by definition, prevents and deters crime. That would be the case if police were adequately trained and equipped. However US police forces on average (!!) are trained only for not even 9 months (per https://www.snopes.com/fact-check/police-training-killings-usa-nordic/ https://www.snopes.com/fact-check/police-training-killings-u...), which means that there are likely cases in which only weeks of training were given. German police, in contrast, train for two and a half / three years (per https://de.wikipedia.org/wiki/Polizeiausbildung_in_Deutschland https://de.wikipedia.org/wiki/Polizeiausbildung_in_Deutschla...). That, combined with a lack of hiring standards (e.g. https://www.portlandmercury.com/blogtown/2019/06/19/26671582/portland-police-lower-officer-hiring-standards-to-address-poor-recruitment-rates https://www.portlandmercury.com/blogtown/2019/06/19/26671582...) leads to the perception of police being nothing more than paid thugs with a uniform. The massive availability of guns as well as former military equipment in police hands doesn't help either, quite to the contrary. When I have an easy sledgehammer (=a tank and a gun), I'll shoot first and ask questions later. When it comes to "overpolicing", you also have to note that it's way easier to arrest someone dealing/smoking cannabis on the street (mostly poor PoC) compared to arresting a rich white banker who is snorting cocaine on his desk.