Crime, punishment, Bayesian informational strategies, and social engineering

Over the past week I’ve been posting some thoughts on crime and punishment at the Volokh Conspiracy.  Reading the comment threads has been in interesting experience.   I was going to put up a final post responding to those comments, but the VC site seems to be down, and it occurred to me that RBC readers might also be interested.

********************

When he invited me to post about crime and punishment, Eugene suggested that I might want to finish up with a response to comments.  So let me offer two general remarks, and some specific responses.

In general:

A.  I can’t read Eugene’s mind, but I suspect that he invited me to guest-blog (despite our basic disagreements) for the same reason that I regularly follow  this blog and other libertarian and conservative information channels.  The basic lesson of Bayesian analysis is that you can learn only from information that disconfirms some part of your current belief set.  But of course the natural tendency of the mind is to minimize cognitive dissonance by accepting confirming evidence and rejecting disconfirming evidence, and that tendency is emphasized when beliefs get to be badges of group membership.

This is a deeply unhealthy tendency, and it tends to defeat one of the basic evolutionary strategies of homo sapiens, which Karl Popper summed up as “letting our beliefs die in our place.”  Unfortunately, it has been on full display in the comment threads to my posts, which consisted (when the comments related to the posts at all rather than merely ranting about unrelated topics) mostly of vigorous attempts to prove that the thoughts offered in the posts were worthless or wicked, and the poster an ill-intentioned idiot.

Eugene no doubt thought he was doing his readers a favor by offering them some reading that might challenge their precoceptions.  There is little evidence in the comment thread that the VC readership shares that view, but it’s possible to hope that the comments are not a representative sample of reader reaction.

B.  The suggestion that various non-punitive programs might control crime, and that doing so was preferable, ceteris paribus, to controlling crime by inflicting damage on offenders, met with an especially furious response, mostly centered on the phrase “liberal social engineering.”  But the project of putting 1% of the adult population behind bars – an incarceration rate five times as high as any other advanced democracy, and five times as high as the U.S. ever had before 1975 – is itself a massive, massively risky, and expensive  social-engineering project, and no less massive, risky, or expensive for never having been thought through.   It also involves a completely unprecedented expansion of the power of the state over the individual.

If all taxation is theft, then the $200 billion required to support the current policing, adjudication, and corrections systems is just as much “stolen” as the much smaller sums that might be usefully expended on improving parental performance by poor young first-time mothers, removing lead from the environment, or improving classroom discipline.  If people who call themselves fiscal conservatives understood that a sentence of life without parole imposed on an 18-year-old represented a present-value expenditure of $1 million, the enthusiasm for “throwing away the key” might be diminished.  (An execution, including the due process required – but not sufficient – to prevent the execution of the innocent, costs more.)

In my view, crime at current levels is such a social problem that even substantial increase in the $200 billion criminal-justice budget would be justified by even modest decreases in crime.  If we can spend an extra $10 billion a year to have reduced crime and reduced incarceration, so much the better.

Now for the specifics:

1.  Evidence about the capacity of nurse-family partnerships to reduce offending by more than 50% (based on a randomized controlled trial) is here.

2.  Evidence about the impact of lead on crime takes two forms:  individual-level studies and econometric analyses. The results are consistent, and the effect sizes are large. Moreover, the biology is understood: lead, even at low levels, damages cognitive function, and lower-level cognitive functioning reduces deterrability, thus increasing crime. Moreover, lead does specific damage to impulse control.

Therefore, lead causes crime, and removing lead reduces crime.  It does so more cost-effectively than increasing incarceration, and it has side-benefits rather than side-costs.

3. I have no doubt that a minimum legal drinking age of 21 reduces drinking among minors, and that relaxing that rule would increase drinking (and drinking-related problems) in that population. It also generates massive disobedience and the mass acquisition of false ID.   Increased alcohol taxes are effective in reducing drinking, and especially in reducing heavy, problem drinking (since an extra tax of a dime a drink wouldn’t much bother someone who averages a drink a day). The biggest impacts are on heavy drinking by minors, whose incomes tend to be limited.  A combination of relaxing the age restriction and raising the price could reduce heavy drinking while avoiding the criminalization of mass behavior.

Author: Mark Kleiman

Professor of Public Policy at the NYU Marron Institute for Urban Management and editor of the Journal of Drug Policy Analysis. Teaches about the methods of policy analysis about drug abuse control and crime control policy, working out the implications of two principles: that swift and certain sanctions don't have to be severe to be effective, and that well-designed threats usually don't have to be carried out. Books: Drugs and Drug Policy: What Everyone Needs to Know (with Jonathan Caulkins and Angela Hawken) When Brute Force Fails: How to Have Less Crime and Less Punishment (Princeton, 2009; named one of the "books of the year" by The Economist Against Excess: Drug Policy for Results (Basic, 1993) Marijuana: Costs of Abuse, Costs of Control (Greenwood, 1989) UCLA Homepage Curriculum Vitae Contact: Markarkleiman-at-gmail.com

21 thoughts on “Crime, punishment, Bayesian informational strategies, and social engineering”

  1. I think part of the objection to that last suggestion, the increase in alcohol taxes, is that most people don't drink excessively, but they'd end up paying the tax anyway. It may rain on the just and the unjust alike, but that doesn't mean the just have to like getting wet.

  2. Brett, follow the link. If we doubled alcohol taxes, a one-drink-a-day social drinker (someone in the highest third of alcohol users by volume) would pay about $3 per month in additional tax. More than half of the tax would be paid by the very heavy drinkers who average four or more drinks a day, because they consume more than half the alcohol. The resulting drop in violence would save a few hundred lives a year.

    Back when Johnnie Walker Black counted as a premium Scotch, their ads said, "If the price difference matters to you, you're drinking too much." Precisely.

  3. Nice post, but I have to question how you can label a typical, evolved human behavior as "unhealthy" and suggest that scientific reasoning is an evolutionary strategy. Occam's razor suggests that any behavior observed in most people is one that has proved itself valuable over the eons of human evolution.

    This is a deeply unhealthy tendency, and it tends to defeat one of the basic evolutionary strategies of homo sapiens, which Karl Popper summed up as “letting our beliefs die in our place.”

  4. @JMG

    Actually, I'm afraid you have it exactly backwards: Occam's razor, combined with an understanding of evolution, suggests that any behavior observed in most people is one that simply has not (yet) proven itself deleterious to reproduction. It may or may not offer any reproductive advantage whatsoever, but it clearly does not harm furthering offspring. Also there's no evidence that I can find that this sort of reasoning is in fact an evolved genetic trait, but may be a carry-along to some other trait, a social trope or something else entirely.

    My question is: Mark says, "I have no doubt that a minimum legal drinking age of 21 reduces drinking among minors, and that relaxing that rule would increase drinking (and drinking-related problems) in that population." Do you have any citations for that? What about 20? or 22? or 30?

  5. Brett, a social drinker might prefer paying $.10 a drink extra to the prices they face otherwise: without it they're enduring more drunkeness and crime, and paying other taxes instead of a booze tax.

  6. One thing about conservatives–even smart conservatives–and crime. Conservatives are deontologists, or at least lean that way. They don't care about consequences of actions as much as the inherent rightness or wrongness of acts. This means that conservatives don't think about crime the way the rest of us do.

    To the conservative mind, it is wrong to condone wrong acts. Failure to punish is condonement. Anybody might go along with these two points as moral principles. But here the conservative mind departs from the liberals' or radicals; mind: condonement is wrong, so we must always punish, no matter the consequences. Hence, the conservative problem with rehabilitation. What kind of punishment is it to help criminals? Or the conservative problem with abortion. A lot of folk feel a bit queasy with the morality of abortion. (And a lot don't.) But only the conservatives stop there, and must criminalize abortion, effects on women and society be damned.

    I don't view this as a masked form of authoritarianism. It is consistent with real sympathy for criminals, and has nothing to do with a police state. The deontological mind cares far more about the law on the books–statements of moral principle–than their enforcement, which is mere consequence. I have a lot of sympathy for it, and there is no way we can do without some of this in our reactions to crime. But it often leads to bad public policy.

  7. Mark,

    I'm glad you entered the fray at VC, despite the disappointing response. Challenging our own ideological positions is an engaging and useful exercise and I would like to see the same done here at RBC.

  8. Mark,

    Given the selection biases involved, it's almost certain that VC commenters are not entirely representative of the readership. Negative emotions are stronger motivators than positive, and you challenged some long-held shibboleths in the libertarian subculture.

    It's like teaching undergraduates: we rarely get to know that we've have positive (or for that matter, negative) influences on them. Let's hope that you got some the readership thinking about the issues.

  9. Mark: "The basic lesson of Bayesian analysis is that you can learn only from information that disconfirms some part of your current belief set. "

    I'll admit that I have a light background in Bayesian decision analysis and statistics, but that is definitely not true. Considering that you then reference Popper, I think that you're confusing Bayesian stuff with Popperian falsifiability.

    JMG says:

    "Nice post, but I have to question how you can label a typical, evolved human behavior as “unhealthy” and suggest that scientific reasoning is an evolutionary strategy. Occam’s razor suggests that any behavior observed in most people is one that has proved itself valuable over the eons of human evolution. "

    Two problems – first, we are not in Ye Olde Sociobiological Baseline Environment; we function in complex societies, dealing with complex technology and problems. Second, a lot of behavior is learned; most of the people in the USA, for example, probably know how to read and write English. I'll bet hard cash that paloelithic peoples didn't know how to read or write English (and you should see how badly they handle hard liquor – and they way that they drive, even when sober will curl your hair!).

  10. JMG says:

    "Nice post, but I have to question how you can label a typical, evolved human behavior as “unhealthy” and suggest that scientific reasoning is an evolutionary strategy."

    I just saw a woman roll by on a powered wheelchair; her legs ended in stubs, not feet. She was quite obese, which led me to believe that she lost her feet to diabetes, caused/aggravated by obesity. That typical, evolved human behavior of 'eat a lot, for tomorrow may hold famine' can cause a lot of trouble when tomorrow always hold more calorie-dense food.

    The idea of scientific reasoning (note, not rigorous) that I believe Mark was getting at was that humans are driven by both instincts, and complex learned behavior, which can be more redily modified than straigh-up instincts. This means that humans can adapt to changing circumstances with strategies which are not simply 'breed a few million slightly different offspring, and let Nature sort them out'.

    Joe S., IMHO there's a big problem about whether people or beliefs are 'conservative' or 'right-wing'. In general, there's more right-wingism than conservatism. As Mark pointed out, most of the commenters are quite happy with radical social engineering, so long as it serves their interests and/or flatters their beliefs. We just saw a radical right-wing administration which seemed to draw massive support from those 'conservatives', no matter what. Perhaps 'conservative' should be seen primarily as a label for right-wingers who want to disavow actual, existing right-wingism.

  11. It is a bit sad, but the vast majority of commenters over there were arguing to win a spat, not to be correct. I could make a joke about lawyers here, but it actually is deeper than that – tribal instincts run deep, even to the point where Eugene himself once penned a defense of supporting "your team" even when you disagree, as a sort of compromise to try to get half of what you want.

    So that's the other side of policy discussions, and Mark, as someone who is more experienced with this than I, I'm sure has processed it to the point that it doesn't depress him as much as it does me. When you can't count on good faith disagreement in a policy discussion, I don't know what to do. Ignoring only goes so far, sinking to matching dishonesty is simply wrong and self-defeating, and playing the comic foil against a Glenn Beck or Karl Rove is adopting the role of a punching bag.

    I hugely respect Mark and others (Jerylyn Merrit, Marcy Wheeler, etc.) who can keep it up, even when I think they're wrong.

    I guess I'm rambling a bit now. I just find it hard to attempt to put forth a what I consider a reasonable argument in the face of people who shriek in response. I'm sure there's a lesson in game theory here somewhere, and the fact that I'm an engineer, not a wonk, by training probably informs my lacking here. (While you will find the occasional blowhard that insists in doing something short-sighted like XORing a pointer in a linked list to save a few bytes and prove they're smart, most engineers recognize baseline reason, at least in their narrow domain. I'm putting aside our tendency to assume we're equally smart in domains where we're not educated.)

  12. Mark,

    1. My subjective impression is that your middle post or posts, where you made more specific (and, to some extent, more value neutral) argument about how a stategy of concentration could reduce crime ceteris paribus got a better quality of comments, compared with your opening summary of premises and your closing listing of unjustified (in the post) policy suggestions. This is not to criticize the opening and closing posts, which had their necessary functions since you were not going to post your entire book at VC. But it suggests what may (often) be the best strategy for meaningful communication accross world-view lines.

    2. I thought your critics at VC touched on (at least) one point where you may not have made yourself clear (whether as a matter of communication or reasoning), or at least not clear to (reasonably liberal) me. You emphasize that what might be labelled "unpleasantness to criminals" should be treated as a cost of crime control methods, to be minimized where we can, even if it is a cost we might often choose to incur to some degree, e.g., to bring about deterrant effects. However, I believe you said you accept retribution as one function of crimal punishment (along with deterrence and incapacitation). In the context of retribution, isn't "unpleasantness to the criminal" in some sense a benefit instead of, or as well as, a cost. (This might also be expressed in terms of inputs and outputs. In a deterrence analysis, unpleasantness to the criminal is one of several inputs, while future reduced criminal behavior is the output. In retribution, isn't unpleasantness to the criminal something close to an output.) One might do a sort of double entry bookkeeping for retribution and put unpleasantness to the criminal on the input side and something like "community feeling of revenge/karmic balance" on the output side. But does putting things this way really add anything? Is this merely like talking about the dormitive property of sleeping pills (see Moliere)?

    Anyway, if you have time and interest, I would be interested to read an elaboration on your view of retribution and how it relates to treating punishment (in the sense of unpleasantness to criminals) as a cost of the system.

  13. Back when Johnnie Walker Black counted as a premium Scotch, their ads said, “If the price difference matters to you, you’re drinking too much.” Precisely.

    That sounds like the converse of a relative of mine's defense of drinking cheap beer: "after 3 of them, you don't taste the difference anyway".

  14. Barry,

    I'm a card-carrying statistician, and Mark's right: taking a Bayesian approach means you are allowing data to modify your beliefs. If the data confirm your preconceptions, you've "learned" nothing, in the sense that the mean of the posterior coincides with the mean of your prior. Certainly the posterior distribution changes, in a way that increases your prejudices. But the amount of weight you allow the data in a classical Bayesian approach is determined by the prior. This is one of the frequentist gripes about Bayesian methods: pick your prior correctly and you don't change your beliefs much, even in the face of strong evidence against them.

    Suppose we have a coin that we believe shows heads 60% of the time, and we model that belief with a Beta(3c, 2c) prior on p. (The parameter c represents your confidence in your beliefs: the bigger c is, the more certain you are that P{H} = 3/5.) If you tossed a coin 20 times and saw no (0) heads, you'd believe something was going on, right? The posterior distribution on p is Beta(3c+S, 2c+20-S). The mean of the posterior is (3c+S)/(5c+20). If S = 12 (exactly confirming our preconceptions) the mean of the posterior is exactly 3/5. What has changed is our belief in our preconception. If we start out with a fairly diffuse prior (c=1), the standard deviation of the posterior decreases to 1/3 of its previous value. If we have a sharp prior (c=500) there is little change.

    Now let's ask if what happens if we collect data that say our preconceptions are wrong, say we observe no heads in 20 tosses of this coin. Now the mean of the posterior is 3c/(5c+20). If we don't have strongly held beliefs (c=1), our best guess at P{H} is 3/25, or 12%. This is much smaller than our prior guess of 60%. If we have strongly held beliefs, though (c=500), our 'new' best guess at P{H} is 59.2%. This is scarcely changed at all.

    So, under a Bayesian analysis you don't modify your beliefs when data confirm them. You only modify them when the data disagree with your prior. And I'll go farther than Mark and add that you only modify them to the extent that you are willing to question your preconceptions.

  15. Dennis says:

    "Barry,

    I’m a card-carrying statistician, and Mark’s right: taking a Bayesian approach means you are allowing data to modify your beliefs. If the data confirm your preconceptions, you’ve “learned” nothing, in the sense that the mean of the posterior coincides with the mean of your prior."

    I think that there's an interpretation question here. I read Mark's statement "information that disconfirms some part of your current belief set. " as data which would lead to rejection of important significant areas of support of a prior (so to speak). For example, any updated posterior represents learning, to me. Even if the prior and posterior distributions might theoretically be non-zero for (-inf, +inf), the practical differences between the posterior and prior might be very large.

    By the definition that Mark is using, one could start with a diffuse/non-informative prior, and update it with lots of data to a very, very tightly concentrated posterior, but there'd be no learning.

  16. @Barry, "I just saw a woman roll by on a powered wheelchair; her legs ended in stubs, not feet. She was quite obese, which led me to believe that she lost her feet to diabetes, caused/aggravated by obesity."

    Or you could just be confusing cause and effect. It is hard to excercise when you have legs ending in stubs. I have a relative confined to a wheelchair after an auto accident, and have seen how hard it is to keep the weight off.

    Basically agree that humans today have mechanisms that become dysfunctional in modern times, and that it becomes harder to overcome when both physiology and culture reinforce them. But I confess I could not quite follow your logic, and how this example was meant to explain your point.

  17. Anonymous,

    I think there is an interpretation question, but I don't interpret "information that disconfirms some part of your current belief set" to mean data that "results in rejection of important areas of support." Certainly changes in the posterior represent a kind of learning, but the change in the posterior depends on the prior. If the prior essentially rejects a large part of the parameter space, data that strongly indicate the parameter lies in a low probability area of the parameter space are essentially rejected.

    I think we need Mark to tell us what he means with that phrase. Also, I took Mark to be using Bayesian updating as a metaphor. People don't generally walk around with priors and update them as data come in. Bayesian inference isn't a perfect model for human reasoning, but it's interesting that it sometimes does reflect some characteristics of human reasoning.

    Another thing to remember is that whether one is Bayesian or frequentist, random sampling is crucial. If you actively seek out information that confirms your prior, and reject information that conflicts with your prior there is no school of statistics that can help you. Far too many people are getting their information under precisely that sort of model.

  18. I was posting under 'Anonymous' – sorry, I gone home, and was posting from another computer.

    In the end, I think that Mark should have used Popper, not Bayes.

    lead says:

    "Basically agree that humans today have mechanisms that become dysfunctional in modern times, and that it becomes harder to overcome when both physiology and culture reinforce them. But I confess I could not quite follow your logic, and how this example was meant to explain your point."

    I'm sorry; I thought that the difficulties posed by a stone-age diet in an environment chock full of cheap, high calorie/fat/sugar foods was pretty much the archetype for why evolution-shaped traits can be unhealthy.

Comments are closed.