Open threads at the Open Thread tab every Sunday and Wednesday

How Likely Are Multifactorial Trends?

Vox recently wrote about 16 Theories For Why Crime Plummeted In The US.

Their story is based on a report by the Brennan Center For Justice, which I haven’t read, so I’m hesitant to critique it too much. The little I got off of Vox I don’t like. For example, if I understand correctly they’re arguing that the lead-crime connection is overblown because although lead was banned in the 1970s (thus affecting people who reached peak crime-committing age in the 1990s), the decline in crime continued even into the 2000s. But lead stays in the environment a long time, there’s still a lot of work to be done eliminating various sources of lead, and so blood lead levels continue to decline. That makes their argument ring a little hollow.

But I want to talk about a more meta-level point.

The analysis ends up concluding that there is no “smoking gun” and crime probably declined because of a bunch of reasons coming together. For example, they say that “up to 12 percent of the drop in property crime during the 1990s was due to the rise in incarceration, but it was probably more like 6 percent”, and “up to ten percent of the drop in crime in the 1990s was caused by hiring more police.” The general picture I get is that there were about ten different factors, each explaining ten percent of the decline.

Imagine two different perspectives on this.

First, a learned professor says “Oh yes, the public always wants to hear about how one big exciting thing caused the decline in crime, but that kind of thinking is unsophisticated. Something as complicated as crime is governed by many factors, and you certainly wouldn’t expect one big knockout change to lower it to this degree. Like everything else, it’s probably a combination of different things that came together, each accounting for a small percent of the variance.”

Second, someone counterargues: “If ten different factors caused the decline in crime, that would require that ten different things suddenly changed direction, all at the same time in 1994. That’s a pretty big coincidence. In fact, let’s reductio ad absurdum this. Imagine it was ten million different factors, each accounting for one ten-millionth of the decline. But that seems stupid. For example, since there are only about ten million criminals in the US, we could structure this as one factor per criminal. Imagine that, in 1994, each of America’s ten million criminals independently and coincidentally had a major life change that made crime seem less attractive. That’s ridiculous. But in that case, any other explanation based on ten million factors should seem ridiculous. And if we give a heavy credibility penalty to a story with ten million factors, we should give some credibility penalty to a story with ten factors.”

The second person seems to me to have a strong argument, which makes me think Vox and the Brennan Center’s model where ten different trends each explain about ten percent of the decline is unlikely.

I feel like somebody has already thought about this and there’s an entire literature I’m missing, but Google is failing me (badly – this was my first search result). Can somebody point me to it? Are there ways to calculate how much less likely a ten-factor explanation is than a one-factor explanation?

[EDIT: Yes, there’s the trivial case where all ten factors are correlated, for example they all have to do with an improving economy. I’m talking about the non-boring version of the question.]

[EDIT2: I might have subconsciously absorbed this thought process from Stefan Schubert]

This entry was posted in Uncategorized and tagged . Bookmark the permalink.

381 Responses to How Likely Are Multifactorial Trends?

  1. Setsize says:

    Well, there’s AIC, BIC and similar stats that penalize for number of predictors.

    • Paul Torek says:

      Wow, I think you just nailed it, giving exactly what Scott is asking for.

      • Scott Alexander says:

        I don’t really understand the Wikipedia page. Anyone want to give me the summary, whether it means I should penalize the ten-factor-of-crime model, and by how much?

        • Setsize says:

          It’s simpler to do than to justify. Take the log of the likelihood of the model fit. Penalize by subtracting the number of parameters. That’s it. (If you want the specific number called “AIC” you multiply by -2 for some reason.)

          So, a model with 10 parameters would need to justify itself by finding a likelihood that’s e^9 times greater than a single-parameter model on the same data.

          • David Manheim says:

            The “some reason” is because of the connection with information theory.

          • Noumenon says:

            e^9 is 8103. I don’t understand the technical definition of likelihood, but it seems like you are ruling out any chance of a ten-parameter model being right, ever.

          • RCF says:


            I don’t see how that follows. If the best you can do with one parameter is to get a likelihood of one in a million, and a ten parameter model gets you a likelihood of 1%, then going by AIC, the ten parameter model would be preferred.

          • RCF says:

            The term “log” is often used to refer to the common logarithm. You do later mention e, but I think you should be explicit in mentioning that it’s the natural logarithm.

          • haishan says:


            In practice likelihood ratios of that size or much bigger crop up all the time.

          • Setsize says:

            Remember that model likelihoods are defined by first calculating the likelihood for each data point and then multiplying them all together. If you have N data points, and you increase the likelihood by 1% per data point, the model likelihood changes by a factor of 1.01^N. This easily covers a factor of e per useful parameter.

          • Noumenon72 says:

            For ten parameters with a likelihood of 1%, AIC is 2 * 10 – 2 (ln .01), which is 29.2. For one parameter with a likelihood of 1/1000000, AIC is 34. So OK, if ten parameters is ten thousand times better at explaining something, it works.

          • Setsize says:

            A likelihood ratio of 10^7 does not mean “ten million times better at explaining.” Thinking of the log of that number would be better, because it can be interpreted as the difference in the amount of information that each model captures about the dataset.

            The connection to information theory goes something like this: imagine you wanted to losslessly transmit the entire dataset as a message, and you wanted to minimize the length of the message. You might be able to shorten the message if you first chose a model, then transmitted the model parameters, followed by the differences between the model’s predictions and the actual data. It only becomes worthwhile to add a parameter to the model if this allows you to shorten the overall message, so you have to save a at least a model parameter’s worth of information from transmitting the residuals. It turns out that this measure of “one parameter’s worth of information” is proportional to the log of the likelihood ratio.

            If under three models the likelihoods of the data are 10^-353, 10^-356, and 10^-359, I would say that Model 1 is as much “better at explaining(*)” than model 2 as model 2 is better than model 3.

            (*) I would prefer “better at saving the phenomena”

        • Troy says:

          The SEP article on Philosophy of Statistics also has a discussion of the AIC:

        • RCF says:

          Suppose you have flip ten coins, and seven of them come up heads. If all of the coins are fair coins, then the probability of that happening is 11%. If the coins are weighted such that they come up heads two thirds of the time, then the probability of seven of them coming up heads is 26%. Now suppose you hypothesize that each coin has a particular side that it always comes up; the one that came up heads were “always comes up heads” coins, and the ones that came up tails are “always comes up tails” coins. Under that hypothesis, the probability of getting seven heads is 100%. So, you have three different hypotheses, and under these hypotheses, the probability of the data is 11%, 26%, or 100%. These numbers are referred to as the likelihood: the likelihood for the first hypothesis is 11%, for the second 26%, and 100% for the third. So if you’re going just by the likelihood, the third hypothesis looks like it’s the best. What that AIC says, though, is that you should penalize the likelihoods by the number of parameters, by dividing the likelihood by e^n, where n is the number of parameters. In the third hypothesis, there are ten parameters (each coin can be either an “always heads” coin or an “always tails” coin), so the penalty is e^10. That makes the adjusted likelihood 0.004%. In the second hypothesis, there is one parameter (we can adjust how much we think the coin is weighted). That makes the adjusted likelihood 9%. If we view the first hypothesis as having no parameters (a fair coin is somehow the “default” condition), then the likelihood stays at 11%. Thus, we should prefer the first hypothesis; although it has a lower likelihood, the other hypotheses are more “ad hoc”.

          Suppose instead we flipped 30 coins and got 20 heads. Now, the likelihood for the first hypothesis would be 2.8%, the likelihood for the second would be 15%, and the adjusted likelihood for the second would be 5.6%. So even taking the penalty for the extra parameters, the second hypothesis is a better fit.

  2. Wrong Species says:

    The idea that 10 different things each contributed to the decline seems pretty plausible to me. Crime started getting really bad in the 80’s and early 90’s and that was probably when people started proposing all of these different ideas to combat it. So many proposals were tried around the same time and this caused the massive drop in crime . But that only makes sense for proposals explicitly used to reduce crime, not the lead and abortion argument.

    • John Schilling says:

      Crime started getting really bad in the ’60s and early ’70s, and yes, people noticed, people were scared, and people were proposing all sorts of different ideas to combat it. Even contradictory ideas like gun control (1968) and Charles Bronson (1974). It is possible but unlikely they all spent twenty years getting it wrong, and then suddenly got it all right at the same time.

      OTOH, the abortion and lead hypotheses do have a fifteen- to twenty-year lead time for producing crime reduction, but those were not on the list of things being proposed for crime-fighting. It is possible but unlikely that a bunch of things that people coincidentally did for unrelated reasons all came together to produce a crime-stopping effect twenty years later.

      OTOOH, if there isn’t a satisfactory answer to be found, falling back on one of the possible-but-unlikely coincidences may be the way to go on this one.

      • Steve Sailer says:

        James Q. Wilson’s 1974 book “Thinking About Crime” put forward the then revolutionary idea that criminals can’t victimize the public while they are locked up in prison, so we ought to lock more criminals up and for longer terms.

        It seems ridiculous now to say that this was considered a new and controversial idea back then rather than a totally obvious one, but it was.

        In general, since liberals dominate the media, the thing they most try to keep covered up is how much they’ve dominated social policy in this country for so long. Liberals pretty much began dominating social policy around, roughly, the JFK assassination more than 50 years ago.

        Where liberalism proved catastrophically wrongheaded, such as over crime — where most of our cities were destroyed and some of them (e.g., Detroit) are still desolate — their policies slowly got rolled back.

        But because they control the media, the juiceboxers at Vox aren’t really very cognizant of the actual history.

        • Scott Alexander says:

          This was the first option Vox started with. They said the statistics suggested it contributed 10% or less of the decline. Other people have mentioned that crime declined to a similar degree in other First World countries, including those that didn’t change their incarceration rate. So that’s not the explanation.

          • Handle says:

            Another way to look at it is that the unifying ‘root cause’ of the decline in crime is .. the intolerably high level of crime itself. Which causes a lot of movers with little bits of authority and skill to react and innovate and try to nudge the situation towards improvement. That means that a lot of factors taking place simultaneously is not in fact increasingly unlikely – as if they were randomly coincidental – but in fact more likely, because it’s exactly what we should expect to see humans do in response to social problems in a system that is decentralized as ours.

          • Steve Sailer says:

            If you look at other first world countries, their crime waves lagged well behind the U.S., which took off with the triumph of Civil Rights in 1964. But crimes like car theft and home invasion were extraordinarily high in Britain by the 1990s compared to America, in part because of all the target-hardening in America that had begun much earlier because crime took off much earlier.

            The British have pursued a more technocratic path (redolent of the one predicted in “A Clockwork Orange” way back in 1962 when home invasion was almost unknown in Britain): cameras everywhere, facial recognition, etc.

            It didn’t keep them from a huge riot four years ago in London, but they managed to lock up a lot of rioters afterwards.

          • Steve Sailer says:

            Handle says:

            “Another way to look at it is that the unifying ‘root cause’ of the decline in crime is .. the intolerably high level of crime itself.”

            Right. The low level of crime in postwar America allowed American elites to stop taking crime seriously. Bill James has some good examples in his recent book on crime of how insouciant elite (e.g., Supreme Court) opinion about crime had become by the early 1960s. Combine this with 1960s elite attitudes about race and you get the 1964-1975 crime bulge.

            Things fall apart faster than they can be put back together again, so it’s hardly surprising that it took longer to teach knuckleheads that society was serious about crime that it took to teach them in the 1960s that we weren’t.

          • Richard Gadsden says:

            From memory of last time I looked at this, you need to be very careful about “home invasion” burglary statistics; UK burglary statistics distinguish between owner-present and owner-absent burglaries, which puts burglaries that involve sneaking into a property while the owners are asleep into the same category as “home invasion” in the usual sense. It also includes deceptive-entry burglary as well as forcible-entry burglary.

            “Home invasion” is not a term used much in Britain, and it’s hard / impossible to get comparable statistics between the UK and US on this.

          • John Schilling says:

            Yes, though I think that’s mostly because it’s hard define “home invasion” in a manner that facilitates easy record-keeping. Whether or not the residents were physically present during the burglary is simple and unambiguous, so that’s usually what gets counted.

            But, while “home invasion” probably adds an emotional charge to the claim that isn’t always appropriate, I do think the categorization is still worth making. Whether hardcore “home invasion” or entry under false pretenses or sneaking in while the residents are asleep, any burglary that takes place while the residents are present, involves at least a risk of confrontation and thus implies a willingness to engage in violence even if that isn’t Plan A.

            American burglars almost always verify that the house is unoccupied before breaking in. This is an easier and more reliable way of avoiding violent confrontation than “Oh, I’ll wait until most people are asleep and then be extra-sneaky as I break in and loot the place”. If we consider violent crime to be worse than property crime, we want to make that distinction in our records and encourage that distinction among burglars.

          • Steve Sailer says:

            Home invasion, especially of the kind in “A Clockwork Orange” in which urban criminals drive out to the countryside, is extremely rare in suburban/exurban America because lots of American homeowners pack heat and racial profiling works pretty well.

            It became a big deal in England in the 1990s, even though it was almost unimaginable when Burgess wrote his novel: he used his wife’s memories of the home invasion she suffered by American GIs stationed in Britain during WWII.

        • HeelBearCub says:

          [sarc]Which is why there is zero crime in England and Australia is just a crime ridden hellhole now.[/sarc]

          Seriously though, don’t try and fit answers to your politics. I don’t get the sense that’s what this blog is supposed to be about.

        • James Picone says:

          For a country that’s had social policy dominated by the left for so long, America has surprisingly little in the way of welfare or universal healthcare. Over a slightly longer time period pretty much every other democracy on the planet has instituted stronger welfare and government-run healthcare policies than America – Medicare in Australia, which is essentially government-owned health insurance everyone gets here, paid for by ~1.5% of income levy – was instituted in 1975.

          France’s was set up in 1945, although I’m not sure how many changes its seen since then, and how close to the current setup it was back then.

          England’s NHS was setup in 1948.

          Not sure when the current unemployment systems of those three countries were set up – Australia’s current unemployment benefit, ‘newstart’, started up in 1991 but it was replacing a previous system. I suspect it’d be a similar timeframe to medicare though, under the Whitlam government, which was kind of a left-wing reform government. Later he was fired by the Governor-General.

          Anyway, the point I’m making here is that if the US’ social policies were dominated by the left, and they couldn’t get all these big complicated bureaucratic things that us liberal types love, then what the hell was going on in the rest of the world?

          • cassander says:

            First, the difference between american and european welfare states is overstated. Second, the welfare state that the US does have has been expanded under every single president since FDR, with the exception of bush the elder. Third, the fact that this expansion has been slightly slower in the US than elsewhere is easily explained by the US’ unique constitutional arrangements which generally act to slow things down and make big changes like the NHS impossible.

          • Steve Sailer says:

            American elites invited Gunnar Myrdal in from Sweden to analyze our problems in the 1940s and then implemented some Swedish-style welfare policies in the early 1960s. Unfortunately, it turned out that large segments of the American population don’t react to welfare the way Swedes do, instead displaying an almost immediate rise in crime.

          • Anthony says:

            Social policy isn’t the same as economic policy. Welfare and healthcare are economic policies; and the left has not been as dominant in the economic sphere as in the social sphere.

          • wysinwyg says:

            Unfortunately, it turned out that large segments of the American population don’t react to welfare the way Swedes do, instead displaying an almost immediate rise in crime.

            Seems a little like begging the question in a thread about what causes increases and decreases in crime.

    • Handle says:

      Right. Except the timing is a little off. The real peak in trouble that caused a multi-faceted reaction was in the early 70’s – I put it at 1972-73, which is when the long-term tide began to turn. A lot of people were under pressure to tackle the problem from a variety of different directions, consistent with their authority and expertise – judges, prosecutors, politicians, detectives, and so forth. Here are some independent and uncorrelated factors that each played a part:

      1. Massive increase in sentencing and incarceration rates.
      2. Dramatic increase in the effectiveness of investigative techniques and forensic technologies. Today we have DNA, video, and records of all your electronic communication and financial activities. IT makes it pretty easy to sort through the kinds of databases that are now feasible to collect, and then hunt you down. All of this was impossible in 1972, when most trials really were dependent mostly on testimony.
      3. Changes in demographics. Always look at the fraction of young males as a fraction of the overall population.

      Even if technology, improvement over time is often the result not of some major breakthrough or advancement among one axis, but a lot of decentralized efforts to improve some aspect by a few percent that contributes to overall performance. When you add them all up, you get a much improved piece of engineering.

      Charles Murray’s favorite graph is the order-of-magnitude decline in the number of fatalities per road-mile over time, the trend of which was seemingly impervious to both Nixon’s NMSL (1974), partial relaxation by the STURA (1987) and repeal by the NHDA (1995). Ask the question, what accounts for the trend.

      Part of it is car safety from the evolution of technology, part from regulatory requirements, and part of it is emergency medicine and part of it having cell phones that ensure that the EMT’s will be contacted and get to you in time, and part of it is …

      See? When there’s an important social objective at stake, lots of players try to tackle it in lots of ways, and a retrospective analysis trying to identify any sole factor is complicated by all these independent simultaneous efforts and phenomena.

      • Steve Sailer says:

        Right. Fighting crime is like Moore’s Law for speeding up computers — it’s not just one thing, it’s a whole bunch of different factors being brought to bear on a common goal.

        • Handle says:

          Also, one thing about lead. You know, we don’t have to go back and infer probably childhood exposures. We have the ability to test violent criminals for the amount of lead still in their nervous systems, even after they’re dead. It won’t be as good as dissolving lead-fed lab rats in acid to test the effectiveness of various chelation thearpies (a former life), but it’s good enough.

          Or, in the alternative, we could test juvenile and very young offenders using the ordinary lead exposure blood tests and see if there is any correlation.

          My take on the available evidence is that, yeah, it matters, but not very much in comparison to the other major crime-reduction factors.

          • Scott Alexander says:

            This has been done and criminals do have way elevated lead.

            Like everything else, this is hopelessly confounded, because poor areas have more lead so if you want you can tell a story about crime coming from the stresses of poverty.

            I think the lead finding remains when you adjust for confounders including poverty, but adjusting for confounders is always risky business.

          • Steve Sailer says:

            If the problem was lead from gasoline being emitted by cars, the worst crime spot in America should have been Sherman Oaks, CA, where I went to high school, because it’s home to the 101-405 freeway interchange, which was the busiest in America for much of the 1970s.

      • ii says:

        Well if we’re talking about technology, 1966 saw the first general purpose credit card and 1990 was when we got American Express. If recent trends about violent crime drops corresponding directly to rise in cyber crime are to be believed…

      • Titanium Dragon says:

        A few objections:

        1) This is plausible. Another possible cause of this is that we got an echo effect – start imprisoning lots of people, and their kids never exist, and thus can’t commit crimes. This might explain the rise in incarceration rates then the decline in crime – we imprisoned more and more people, preventing them from having very many kids, and thus breed out the violent genes from the population.

        Flaws: Other countries saw declines without huge increases in incarceration rates, doesn’t explain upswing in crime rates.

        2) Implausible. While we are better at solving crimes on paper, our murder conviction rates have actually fallen because of lack of compliance with law enforcement in inner-city areas.

        3) Implausible. Demographics are an unreasonable explanation because we have a smaller number of absolute crimes, not just a lower crime rate, with a larger population of said folks.

        However, I do agree that the assumption that one factor is better than multiple factors is very wrong, because it makes a lot of incorrect assumptions as well.

        • Harald K says:

          > start imprisoning lots of people, and their kids never exist, and thus can’t commit crimes.

          There’s plenty of time to have kids before getting arrested, and I’d wager having a parent in prison is less than ideal for the chance of these kids getting into crime. Bear in mind also that for US purposes, a lot of the crimes we are talking about are minor drug offenses.

          • Titanium Dragon says:

            Incorrect. The idea that minor drug offenses are a major driver of imprisonment rates in the US is a myth.

            First off, minor drug offenses don’t lead to much, if any, jail time in many cases. These are mostly “real” drug offenders.

            Secondly, drug offenses are overstated in the system due to plea bargaining; basically, if someone accuses someone of committing multiple crimes, they will frequently plea bargain them down to the lesser offenses, which tend to be drug offenses. Thus, you have a bunch of people in the system who are there for “drug offenses” who are really there for other reasons, and the drug charges are the only ones on the book because of plea bargaining. Likewise, the prevalence of people being jailed for pot is exaggerated for the same reason.

            Thirdly, the idea that lots of people are in jail for drug offenses comes from the federal prison system, which only makes up a small fraction of US prisoners; using the same source, you would think that an enormous percentage of criminals in the US are immigration offenders. Neither is really the case; it is an artifact of the setup of the system, with interstate drug trafficking resulting in federal prison time, along with other offenses, and immigration-related offenses being a federal rather than a state crime.

            In reality, most criminals are imprisoned at the state level, not the federal level, and on the state level, drug crimes make up a much smaller percentage of prisoners.

            This is why anyone who uses federal prison statistics to talk about how the war on crime is the primary driver for imprisonment is a worthless liar who is trying to manipulate you. That’s not to say that there aren’t a bunch of people in jail for drug offenses, but they aren’t the primary prison population.

            Fourthly, a lot of people are in jail on drug charges and other things – they did some crime and they also got nailed for having drugs at the same time. This is pretty common, and some less-scrupulous individuals claim that these folks are in jail due to drugs, when drugs are merely one of many charges laid against them. A friend of mine got run over and killed by a guy who was high on meth while driving a stolen vehicle. The guy got charged with felony murder (because he caused the death of someone while committing a felony, which automatically makes it murder), reckless endangerment, a number of other injuries he caused, grand theft auto, and meth possession. If he goes to jail, is he there because of drugs?

            In reality, the vast majority of prisoners are in jail for other reasons, and the number of people in jail for drug offenses is exaggerated for political reasons.

            And as far as “there’s plenty of time to have kids before being arrested”: somewhat misleading. Peak criminality occurs in young people (25 and younger), and given the delays on reproduction in the population, fewer young people are having children at that age. Indeed, this may be a contributory factor: not only are these people getting locked up in their teens and early 20s, but they’re less likely to have had children prior to that time.

            I suspect that this is unlikely to have a large influence on criminality for other reasons (namely, the question of “why did these people suddenly become so reproductively successful in the 1950s and 1960s?” not being answered, if we assume that genetics were a major driver of the increase in crime) but there is an argument for that as regards people having more children out of wedlock, probably with less suitable parents.

          • Irrelevant says:

            Regrettably, Titanium’s correct here.

            The main debatable point in how many non-violent drug offenders we have locked up is whether you count people who were initially imprisoned for a violent offense, and are now back in for probation violation due to drug possession. But it’s under 5% of the prison population either way.

            All of which I had an incredibly difficult time believing myself, until I dug into several state reports personally and came back with 2%. I’m still a firm believer we need to end the war on drugs, but I no longer consider it anywhere close to an answer to our general crime and imprisonment problem.

          • Steve Sailer says:

            LA Times reporter Jill Leovy’s new book “Ghettoside” makes clear that a lot of people nominally in jail for drug possession are really there for doing very, very bad things for which they terrified witnesses into not testifying against them.

          • haishan says:

            This is why anyone who uses federal prison statistics to talk about how the war on crime is the primary driver for imprisonment is a worthless liar who is trying to manipulate you.

            Or, more charitably, they’re infected by bad memes? People believe lots of false things for many reasons other than “they’re eeevvvilllll.”

          • Nornagest says:

            People believe lots of false things for many reasons other than “they’re eeevvvilllll.”

            That sounds like commie talk to me.

  3. MicaiahC says:

    To point out, it doesn’t have to be ten factors at the same time, which you (rightly) penalize, but it could be ten different factors, all exhibiting different degrees of influence at different times.

    I will add it’s very likely that the additional factors would likely confound the causality by happening at the same time. Example 8 on the the article is one such thing. Sure they can claim that ‘technology keeps people inside’, but additional technologies are being invented all the time; was there a similar drop in crime when books, radio or television came out? Hell, half the causes proposed there seem to be effects (gangs getting less violent? Really?)

    Note: Haven’t had time to read article. Will edit out inconsistencies when I have.

    • Greg D says:

      It doesn’t even have to be just 10 factors.

      It can be 50 factors, 30 of which go one direction, 20 of which go in the other direction. Assume that 20 of the 30 are canceled out, so it seems like 10 factors are causing the difference.

  4. Qiaochu Yuan says:

    Well, here’s a naive remark: it is of course fairly unlikely for ten independent factors to all point in the same direction. But it might be a lot more likely if the factors were moderately correlated, and that wouldn’t be surprising. For example, lots of factors are probably moderately correlated to how well the economy is doing.

    • Troy says:

      Or, as posters above have suggested, tactics designed to fight crime might be correlated with crime getting bad.

    • ryan says:

      There are so many different explanations for the 10 factor correlation. One that seems really obvious to me is there are 90 factors, 50 of which correlated to reduce crime, 40 of which correlated to increase it, and everyone is having fun picking 10 of the 50 factors and ignoring the other 80.

      Completely lost to me is why the data is presented as violent crime victims per 1000 members of the general population. For the love of god normalize that number for the number of men aged 15-40 in the population. I hate to be an ageist and sexist bigot here, but those are the people who commit basically every violent crime.

      That crime rates peaked in the 70’s when the baby boom generation got old enough to start being successfully violent, and then fell off as they didn’t have nearly as many children as their parents, is way, way too easy of an explanation to not at least check to see if it works.

      • Douglas Knight says:

        Yes, it would be better to look at age-adjusted rates. This report has some numbers, but it only goes back to 1980. Figures 3 and 4 on page 4 are the key ones. The crack epidemic was real, not driven by demographics (which is obvious from how fast it came and went). The 25-25 and 35-50 buckets are a lot less violent than in 1980, but the 14-18 and 18-25 buckets aren’t. I guess that’s what you would expect from an incapacitation theory.

        • ryan says:

          That’s great, thank you.

        • Steve Sailer says:

          Thanks. The crack wars were largely fought by young males born after the legalization of abortion in their cities (1969 in LA, 1970 in NYC and DC, 1973 in the rest of the country). This is the absolute opposite of what Steve Levitt’s abortion-cut-crime theory would predict, as I pointed out to him in our debate in Slate in 1999.

          The older men were in prison or dead or older but wiser.

          My best guess is that the combination of crack and gangsta rap fueled delusions of bullet-proofness in quite a few young males for a few years at the end of the 1980s and early 1990s.

          Fortunately, lessons were learned.

          • ryan says:

            One thing I just thought of is that crack in the 80’s was a pretty unique drug in several ways:

            – A very small amount of cocaine once converted to crack can get someone extremely high

            – Converting cocaine to crack is so simple an illiterate 9 year old could do it with ease

            – Cocaine prices in the 80’s were artificially deflated because the wholesalers had protection from law enforcement due to moonlighting as arms traffickers for the CIA in Nicaragua.

            The result is a very cheap drug, so cheap your market could be penniless inner city residents, and one which could be dealt by any random idiot.

            Having Stringer Bell as your gang’s CFO probably didn’t hurt the business, but what really drove the competition would logically have been pure violence.

        • Douglas Knight says:

          Here is the graph Ryan asked for. The black line is the normal rate of homicides per 100k population. The blue and red lines are the number of homicides divided by the youth populations, for two definitions of youth (blue: 15-35; red: 15-45; and the numbers are multiplied by .285 and .405 to make them line in up in 1968).

          It looks like about 20% of the run up and the run down is due to a changing age profile. The normal homicide rate has been cut by 55% since the twin peaks of the cocaine wars and the crack epidemic. But the red line has only been cut by 45%.

          The numbers are from CDC’s Wonder database. It only goes back to 1968, while I’d really like to go back to 1965 or 1960. I’m using it to get demographic data, which ought to be available elsewhere. The database is mainly for cause of death, which suggests other graphs. Here is a graph of homicide death rates for the biggest groups. As in the graph in the other comment, homicide victimization rates for a demographic are similar to, but not quite the same as the rate of committing homicide. Younger groups are being killed at the same rate as in 1968, while older groups are being killed much less often. In particular, in 1968, people aged 35-45 were almost twice as likely to be killed than those aged 15-20, while in 1995 it was the opposite. Today they’re at parity.

  5. Douglas Knight says:

    Steven Levitt holds that both abortion and the end of crack were responsible for 100% of the decline and that the baseline without either would have been doubling crime in the 1990s. But he has no causal explanation for the baseline. So there must be a third cause responsible for 100% of the change, just in the opposite direction.

    • Harald K says:

      Also, “the end of crack” and “the crack epidemic” are in some sense cop-outs. Since we already count crack use/production/traffic etc. as crime. Why did the crack epidemic start, and why did it end?

      Unless Levitt has an explanation for that which stands firmly outside of the other reasons crime rises and falls, it’s not much of an explanation.

      (Aside, it would be nice to have numbers on how much drug use actually causes crime.)

      • Douglas Knight says:

        Sure, we’d like to understand crack, but it’s pretty clear from comparing cities and countries that it’s separate from other causes.

      • ryan says:

        A few thoughts:

        – Drug abuse rates rise and fall in about the same pattern as disease infection rates. Viewing drugs like diseases which initially spread rapidly through a herd, which then develops resistance to them, is probably the most informative way of looking at the situation.

        – Portugal set up free clinics for heroine addicts back in the early 2000’s. Since then rates of heroine abuse have fallen. And most notably the sort of petty thefts typical of heroine addicts seeking to pawn something to get their fix have fallen a lot as well.

        – It’s totally impossible to ascribe causation to the decriminalization laws in Portugal. Heroine addiction rates rose and fell like the rate of infection for a new disease in a herd would. It is entirely possible that no government policy on drugs typical of western countries has any effect whatsoever.

        – The foregoing is not true of a place like Singapore where you get the death penalty for trafficking drugs:

  6. NeshSelg says:

    It seem like the likelihood of multifactor explanations would be much higher then the number of easily identifiable number of factors would suggest due to multiple separate factors having shared causes. All the factor due to changes in policing would likely be casually related and the improvements in the chemical environment would all be part of an overarching trend of better technology and social adaptions to modern (post)industrial living.

    • RCF says:

      But then doesn’t that suggest that the ten factor analysis isn’t complete, and we should try to figure out what the common factor is?

      • Michael Watts says:

        No. Assuming you have some sort of goal and you’re contemplating interventions (for example, we might want to reduce crime even more), it’s counterproductive to just go for the deepest link in the causal chain that you can still see. You’ll actually want to find the shallowest one, because, if things develop stochastically, there’s less opportunity for things to turn out differently the closer you start to your goal.

        The other side of that is that you need to make sure you’ve got a causal relationship, and not, say, two effects of a root cause that happen at different time lags. That argues for going deeper, but you still want to stay as shallow as you can.

  7. Stezinech says:

    Steven Pinker, in The Better Angels of Our Nature, assigns the crime decline mainly to two factors. The Leviathan getting bigger and more effective (larger and more capable police forces), and the decline of 60s counterculture.

    I agree that ten independent factors would be unlikely, but there could be a larger process going on that links many of them. Pinker does a pretty good job of arguing that there was a general re-civilizing trend after the tumultuous 60s.

      • Roger says:


        I read Pinker’s book a few years ago and really enjoyed it. I also really liked your review and commentary on it. Very insightful.

      • Titanium Dragon says:

        The main problem with your article, incidentally, lies here:

        “I was born in an America in which women could walk downtown streets freely at night, where both infanticide and abortion were uncommon, where the prison population was small, and prison rape was not the default punchline as TV detectives handcuffed the bad guys. I have some hopes that, just as with my neighbor’s unlocked car, I might someday live in that America again.”

        The problem here is that you’re conflating multiple different factors here and assuming that they’re all the same.

        Women walking downtown alone at night: safe. Indeed, in most places, it is very safe indeed, and safer than it was back in the 1950s. Why do people believe it is dangerous? Because of mythology. People see reports on how dangerous things are, but things have gotten less dangerous over time; the reason people believe things are getting more dangerous is because of better reporting.

        Infanticide/abortion: What makes people people is not being genetically human, but being intellectually a person. Sperm and egg cells are human, but they’re not people any more than skin flakes or HeLa cells are.

        Pigs are smarter than fetuses, but they aren’t people.

        It is okay to kill and eat pigs, and otherwise destroy them when they are an inconvenience or a threat to our way of life.

        So abortion is no less morally justified, because a fetus is not a person. People matter vastly more than non-people, and unwanted children are much more likely to grow up to be criminals, be subjected to abuse, ect.

        Small prison population: There were fewer criminals back then.

        Prison rape: Just because people didn’t report on it back then doesn’t mean it didn’t exist; it did. People are more aware of it now.

        I lock my door, but it isn’t necessary. It doesn’t cost me anything to do it, though.

        Also, there’s no real indication that Christianity had a positive effect on civilization; we saw declines in crime in non-Christian countries as well, and places like China and Japan have very low crime rates relative to Christian nations. If we rewound time and replaced it with, say, Buddhism, is there any particular reason to believe that the world would have been any worse? Heck, without Christianity, it is possible that Islam would never have come to exist, either.

  8. Vulture says:

    Here’s a model that might be useful: At any given time, there are lots and lots of things going on. Let’s say there are a hundred different factors at play which have some effect on the crime rate. Suppose that 50 of them cause it to go down and 50 cause it to go up, but the total magnitude of the “downward” factors outweighs that of the “upward” factors. Thus the crime rate goes down, with 50 different factors all contributing positively to that trend. In general, then, you’ll see something like this (a very multifactorial trend) any time that there’s something that’s changing which has a lot of factors that effect it.

    • RCF says:

      That would imply that we should expect a Gaussian random walk. Is that model consistent with the data?

      • Alex Richard says:

        Per this, a random walk is better predictor of short-term crime trends than any considered regression, but worse in the long term.

        (This seems consistent with a slightly more complicated model, where the government controls some of the factors (e.g. number of police), but with significant lags.)

      • Wirehead Wannabe says:

        It’s consistent with the data if we throw in the idea that crime begets crime. That would mean that small fluctuations in either direction tend to escalate into big fluctuations.

        There’s also something of a file-drawer problem here. Anything in the middle of the Gaussian curve isn’t going to show up in the media or in academic journals. The end result is the appearance that a bunch of factors all moving to push in the same direction is a common phenomenon, when in reality that might not be the case.

    • Alex Richard says:

      This. Like, upvote, +1.

      To put it another way, I’m pretty sure Scott has to be doing something invalid here. Surely he would agree that, if asked in 1990, the evidence/his prior strongly supports multiple factors having an influence, and that there’s no obvious links between them. (e.g. if I brought Scott back to 1990, he would agree that it seems plausible that most of the mentioned factors would decrease crime, and so at least 5-10 of them actually would, if tried, decrease crime.) Why would observing that, e.g. crime actually did decrease after we increased incarcerations decrease your confidence that incarcerations decreases crime?

      (There’s the obvious counterargument that maybe somebody in the 1990’s should have expected a common cause. And in fact, all cited things do in fact have a common cause: people made changes because they wanted to make things better- keeping criminals off the streets, improving health, improving employment, giving people fun non-dangerous things to do, etc. But altruism seems a constant, not something that changed in the 90’s, and so we’re left wondering why people got better at creating good outcomes between the 70’s and 90’s.)

      • Steve Sailer says:

        The “Sixties” discredited law and order, especially the triumph of Civil Rights. The “Rights Revolution” encouraged people to be criminals because punishment of criminals dropped sharply — Pinker says it dropped 80% relative to crimes committed.

        Society started coming to its senses but it took a long time to convince the criminal classes, who aren’t the sharpest individuals on average, that it was serious about locking them up and throwing away the key.

        • Alex Richard says:

          I am not an expert in this, but per the linked article increased police and incarceration explain only a small fraction of the decline in crime. Certainly, we can’t claim that, e.g. computers or abortion or new drugs or lead regulations were due to the move away from 60’s culture.

        • Titanium Dragon says:

          This is an incorrect view of cause and effect. Civil rights had a net positive effect – the millenials have far more rights than people did in the 1970s, and yet are far less violent.

          The problem was not civil rights but rather a culture of disrespect for centralized authority.

          The reality is that crime is primarily cultural in nature rather than anything else – the Chinese and Japanese are simply expected not to commit crime, and they don’t. People who are simply expected not to commit crime because only garbage would do it are less likely to commit crime because they don’t want to think of themselves as garbage. Indeed, this is true in the US as well amongst low-crime populations – consider the middle-class American response to criminals. “Put them in jail.” Does it matter what they do? No. They just think that throwing them in jail is a good idea unless it is something which they view as a medical issue (drugs), in which case they’re for rehab.

          Thus, you see human garbage in urban areas where crime is seen as acceptable/commonplace who commit crimes because they see nothing wrong with it, and you see human garbage in rural areas (think Waco, the fundamentalist Mormons, that crazy guy who refuses to pay grazing fees, ect.) who live in areas where people disrespect the government who commit crimes, but in places where people respect the government and centralized authority and believe in Civilization, crime rates are much lower.

          If you look at the US, crime rates are highest for blacks and whites in the South. The South has a more violent culture than the rest of the US, and less respect for centralized authority.

          • Irrelevant says:

            the Chinese and Japanese are simply expected not to commit crime

            Those are in no way comparable systems. China posts a pair of police with rifles on every block of Beijing (or did in 2002), while Japan has been so pervasively corrupted by organized criminal interests that they couldn’t pass a RICO law until 2011.

          • Unique Identifier says:

            ‘The reality is that crime is primarily cultural in nature rather than anything else’ – this is a very bold assertion.

            The rest of your post is essentially nothing but a circular argument. Crime is more common in places where crime is common. Who would have thought, indeed?

            The first assertion actually makes the same sort of error. You take culture to be a cause in and of itself, rather than something caused by the people who live in it.

          • Titanium Dragon says:

            @irrelevant: And yet, the Japanese commit crime at a far lower rate than Americans do, despite said corruption.

            Can’t really comment on China – it is a repressive country, but it appears that the repression there is successfully lowering the crime rate. Conversely, Russia is repressive but has a really high crime rate (but not a particularly high homicide rate, IIRC).

            @Unique Identifier: It is by far the most reasonable hypothesis. Do you think that people from inner cities have the same culture as people from a college town with only 50,000 residents? I don’t. The US has a significantly different culture from other countries.

          • Unique Identifier says:

            The primary difference between inner-city and small-town people is of course not their culture, but their very living conditions. You could call it a cultural difference, that in Tinytown everybody knows each other, and when they see a new face on the street they want to figure out who this guy is, whereas everybody in Bigtown treat each other as strangers. But then the world culture quickly loses its meaning, it becomes some sort of all-encompassing synonym for ‘how people live their lives’.

            I think that if you transplanted a group of Bigtowners into their own private Tintytown, and they would adapt into a Tinytown-like culture within three months.

          • Troy says:

            If you look at the US, crime rates are highest for blacks and whites in the South.

            This is not consistent with data I’ve seen on black crime rates.

          • John Schilling says:

            This is an incorrect view of cause and effect. Civil rights had a net positive effect – the millenials have far more rights than people did in the 1970s…

            I question this assertion. The trend has been towards greater civil rights, yes, but I think most of the big-ticket civil rights victories in the United States occurred in the 1960s and early 1970s. Where the millennials are concerned, the major civil rights battles seem to be about gay marriage (or LGBT issues generally), and the “right” to silence people who say offensive things. I find it unlikely that these were major drivers of the US crime rate over the period of interest.

            Please explain what relevant rights you think have been secured since 1979, and why these are more important or more relevant to criminality than the hard-earned victories of the 1960s and early 1970s.

      • Scott Alexander says:

        Someone downthread asks whether I can think of ten factors that should have increased crime.

        Off the top of my head – rising inequality, decreasing % population rural, dysgenics, decline of manufacturing/rise of needing college degree, Bowling-Alone-style social atomization, kids being raised by daycare rather than their parents, rising divorce rate, weaker gun control, welfare reform creates needier poor people, epidemic of crystal meth and bath salts, rising incarceration rates create more hardened criminals.

        Since I can think of ten things that should have increased crime but none of them had an effect, it seems fair to say that even though Vox can think of ten things that should have decreased crime, perhaps none had an effect.

        • Dan says:

          How do you know that the factors you mentioned did not have an effect?

          Perhaps the decline would have been greater without those 10 factors. Perhaps those factors dominated in certain states or countries. That’s why you have to look at exogenous shocks, natural experiments, cross-state analyses, etc.

          But, if the point is about correlation vs causation, this says nothing about the number of factors explaining a phenomenon.

          • Nornagest says:

            If the trend looks monotonic, and the driver isn’t, that’s a pretty good indication that the driver isn’t driving.

          • Fnord says:

            The trend doesn’t look monotonic, even for the nationwide data over the study period, much less if you looked at the smaller resolution that, as Dan suggests, is necessary for discerning the effects of each of the multiple factors.

          • Nornagest says:

            Monotonicity is just a special case that lends itself well to simple explanation. The more general idea is that if a trend is being driven by something, then the two functions should be related in some way — the graphs should look similar. Maybe inverted, maybe derivatives of each other, maybe with a lag, but changes in one should correspond in some way to changes in the other.

            That isn’t the case for most of the things that Scott mentioned. It’d be interesting to collect a bunch of those variables and run a principal component analysis on them, but I don’t particularly fancy doing the huge amount of data entry work that that would take.

          • Fnord says:

            Sure. Unless, of course, the trend is being driven by more than one component, in which case the similarity is still there but confounded by similarities with other variables, requiring (as you say) some more sophisticated analysis than just looking at the graph.

        • Richard says:

          Professional day care seems to be significantly better at producing “good citizens” than a lot of parents. In fact, in the US today it seems to me that the parents who can’t afford daycare are the same parents who should not bring up their kids without a support network.

          Being brought up by one competent parent may be better than being brought up in a dysfunctional family. The alternative to divorce may be being beat up regularly and also watching your mum get beat up regularly. I don’t think the alternative is often an harmonious life in the suburbs with two parents who love both each other and their kids.

          When it comes to gun control, better gun education seems to lower gun crime, banning guns don’t so it depends a bit on your definition of gun control.

          Point is that my intuition seems to be the opposite of yours at times.

          The Lead hypothesis seems to have a good point going for it in that a lot of countries have phased out lead at different times and they all seem to have a drop in crime rates with a lag of about 19 years. The countries that phased out lead in the early 90s seem to get the drop around now, so more data keeps coming in.

          • Irrelevant says:

            Professional day care seems to be significantly better at producing “good citizens” than a lot of parents. In fact, in the US today it seems to me that the parents who can’t afford daycare are the same parents who should not bring up their kids without a support network.

            Alternatively, being the child of parents who can afford professional daycare does a better job producing good citizens, even when said parents delegate much of the responsibility, than a lot of parents.

            Being brought up by one competent parent may be better than being brought up in a dysfunctional family.

            You’re conflating three arguments here. The first and unsurprisingly weakest is the one you’re directly arguing against: that most divorces are mistakes and the result is worse for the children than if the parents had stayed together. I don’t think anyone here is arguing in favor of that position, and I certainly know I’m not.

            The second is that the ability to select a stable and tolerable lifetime partner is an excellent proxy for the sort of long-term planning ability that we want quality parents to have, and by the time we’re weighing whether this divorce or that divorce was the optimal decision under the circumstances, we’re into loss-cutting territory.

            The third cares about the total increase in single-parent households, not particularly whether the parents were ever married.

            I don’t know whether Scott meant the second or the third, but I’m almost certain he didn’t mean the first.

            better gun education seems to lower gun crime

            Data? That’s a pretty surprising claim, unless by lower gun crime you specifically mean less criminal negligence charges resulting from firearm accidents.

            Oh, and I suppose states with strong gun education programs may create gun owners who know better than to admit to brandishing a firearm, which would lower the category. But yeah, really want to see what you’re referring to here.

          • Nornagest says:

            That’s a pretty surprising claim, unless by lower gun crime you specifically mean less criminal negligence charges resulting from firearm accidents.

            I don’t have any data, but I wouldn’t be too surprised to find that a big chunk of crime involving firearms consists of things that involve negligence, even if they don’t hash out to criminal negligence charges.

            A scenario: you’re out at a country bar with an acquaintance. You’re both armed and you’re getting pretty drunk. A petty dispute arises. You draw your weapon to intimidate him. Your acquaintance doesn’t take it seriously. He gets in your face. You get in his face, waving the gun around to make your point. He still isn’t taking it seriously. He shoves you into a table. You startle enough to pull the trigger — a negligent discharge, and a bystander gets hit. Now you’re looking at a long list of charges — I’m no lawyer, but I’m pretty sure we’re talking at least assault and reckless endangerment, up to maybe second-degree murder if enough goes wrong or an ambulance doesn’t get there in time.

            It won’t show up as negligence in any statistics. But it does violate every single rule of gun safety. And if training is at all effective in instilling those rules, making it more widespread among gun owners ought to cut down on similar incidents.

          • Eric Rall says:

            A scenario: you’re out at a country bar with an acquaintance. You’re both armed and you’re getting pretty drunk. A petty dispute arises. You draw your weapon to intimidate him.

            That step right there, brandishing a weapon as a deliberate threat (absent strong justification such as self-defense or defense of another), crosses the line from negligance to malice. I’d expect a reasonable person who doesn’t know the first thing about the four rules of gun safety to know that threatening someone with a gun is not a nice thing to do.

          • Nornagest says:

            That step right there, brandishing a weapon as a deliberate threat (absent strong justification such as self-defense or defense of another), crosses the line from negligance to malice

            Yes. That’s the reason it doesn’t end up looking like an accident from a statistical perspective: because any consequences of negligence end up happening in the context of an assault. (Or whatever your local jurisdiction calls that. It’s pretty universally a crime, though.)

            I’m not saying that firearm safety training makes people less likely to act maliciously. I’m saying that the crime involves both negligence and malice, and could be prevented at the negligence step — three or four steps in this case. The first being that any competent instructor will tell you in no uncertain terms not to draw on somebody unless you’re prepared to shoot them, and I don’t think that’s a message that you’ll clearly get through cultural osmosis.

          • John Schilling says:

            That scenario not only violates the basic rules of gun culture, but also of country bar culture. You don’t bring your gun into a bar. If you do bring a gun into a bar, you don’t get drunk. If you do get drunk, you had better hope your drunken instinct is to leave the gun concealed, because if you even display it you automatically lose. There are bars where one gains status by winning a fistfight, but drawing a gun is an automatic -1000 status points even if everyone decides being obsequiously polite and deferential is the best short-term way of getting the dangerous idiot to go away.

            I live in a town with more than enough bars, plenty of guns, no small number of barfights, and a statistically significant number of shootings. The shootings don’t take place in the bars, the barfights don’t involve guns. The very rare exceptions are, as Eric Rall notes, malicious rather than negligent. The only ones I can recall were outright premeditated.

        • Alex Richard says:

          I can think of ten things that should have increased crime but none of them had an effect

          This claim seems unjustified to me. To look solely at your first example, rising inequality almost certainly has had an impact on crime; a quick google search finds at least 6 studies all supporting this claim, none opposing it.

          I think what you mean is that in the one specific example of the US, we don’t see crime rates as being solely or dominantly determined by the factors you listed? But this is wrong under a multi-factor model as well, since by definition that model is claiming that there are multiple factors responsible, and so it is not necessarily true that any one factor or collection of factors is responsible for the outcome.

        • Titanium Dragon says:

          Inequality is an entirely worthless measurement because it fails to distinguish between the poor getting poorer and the rich getting richer. The poor getting poorer is likely to lead to social unrest; the rich getting richer is not. If inequality grows while the poor get richer and the rich get vastly richer, then you’re likely to see more stability.

          Lower rural population in principle should increase crime rates, but in practice there’s no actual reason to believe this to be so; while cities do have modestly higher crime rates, in reality the extremely high crime rates of cities appears to be a cultural factor. We observed crime rates increasing in same-sized cities during the upswing in crime, and crime rates decreasing during the downswing even as urbanization increased. Additionally, this assumes that historical models for low crime rates in rural areas continue to apply, which may not be the case; if rural areas become poorer, we might expect their crime rates to increase. The other issue lies in the way statistics work; say that there are ten murderers in a population of 10,000 people, or 1,000 murderers in a population of 1,000,000. It is easier to catch all ten murderers than it is to catch all 1,000 murderers, even though their per-capita rate is the same, becaue there are fewer people around to have to check on. In the most extreme case, if you had 1 murderer in your population, you’d only need to catch that one person to get rid of all of them. This artificially lowers the crime rate.

          Dysgenics is a plausible explanation, but it seems more likely that it would create invalids than violent criminals. Indeed, imprisoning people and preventing them from reproducing seems like it would increase human domestication.

          Decline of manufacturing shouldn’t cause any increase in violence – there’s no causual link.

          Rise of college degrees would be expected to decrease violence – educated people tend to be less violent, and to have more to lose. Moreover, the more each human life is worth, the less worthwhile it is to risk that life.

          Social atomization would be expected to decrease violence, because less human interaction means that there are less potential times for an individual to decide to attack another individual.

          Kids being raised by daycare would rely on daycare providing parenting which made children more violent – there’s no reason to expect parents to be competent at raising children. Moreover, one can view education as daycare of sorts which definitely decreases violence.

          Raising divorce rates is only bad if you assume that people staying together in failed marriages is a good thing – if it is more likely to result in domestic abuse or abuse of children (seems likely – stressed people are more likely to commit violent acts) then we might expect increased divorce rates to result in less violence. However, it seems more likely to me that increased divorce rates are a symptom of dysfunction rather than a cause – that is to say, when you have high levels of dysfunction, you’re more likely to get divorced and to abuse your kids or whatever, than if you don’t. Whether or not you actually get divorced is, in this light, irrelevant. And indeed, if you look at divorce statistics, the younger folks are getting divorced less frequently than their parents did – it is the folks of the failed generations which lead to the high crime rates who also get divorced a lot.

          Weaker gun control is irrelevant; there is no link between crime and gun control or the lack thereof. Do a linear regression of homicide rates vs gun ownership and you won’t find any correlation at all; guns neither cause nor prevent crimes, their presence or absence appears to be utterly irrelevant, which is why Wyoming (60% of households own guns) and Hawaii (5% of households own guns) have similar homicide rates. Oregon has a gun ownership rate similar to Louisiana, but Louisiana has the highest murder rate in the US while Oregon has one of the lowest ones.

          Welfare reform creates needier poor people is an interesting idea, but we haven’t seen that in reality; indeed, we saw welfare reform under Clinton around the time that crime rates started falling.

          Drugs may facilitate crime, but the reality is that people who abuse drugs are more likely to commit crimes anyway. Drugs don’t make you commit crimes; they just act as an enabler. The #1 drug of choice for criminals is alcohol by a wide margin, followed by pot.

          Prolonged incarceration is likely to make prisoners less violent simply by releasing them when they’re older; the older you are, the less likely you are to commit a crime. Imprisoning someone during their high-crime years is likely to result in them not committing as many crimes over their lifetime.

          • pensive says:

            broadly agree otherwise but several serious errors need correcting:

            3. criminality correlates with higher fertility in modern world and has a plausible causal mechanism. add in children of immigrants from high-crime societies regressing towards their population means and the frequency of criminal tendencies in general population should be expected to increase

            6. social atomization = lower trust = less empathy for potential victims and less ability to band together against criminal predation

            7. parents might not be more ‘competent’ but they are virtually guaranteed to be more motivated and have more pertinent info on specifics of their child’s needs. the same information/incentive problems which predict command economies should fail apply equally to centralized parenting / education schemes. more evidence of this is seen in how homeschooled children outperform public schooled peers and in the higher incidence of abuse in schools / daycares / churches as compared to within families

            8. ~80% of divorces take place in what experts call ‘low conflict marriages’ which means no abuse no infidelity no serious drug use and about the same level of argument as in otherwise successful marriages. the remaining 1/5th of divorces are in ‘high conflict marriages’ so somewhere less than 20% are potentially in abuse situations. since children of single mothers commit crime at >5x the base rate it would appear divorce is a net cause of crime

            general nitpick: with education / drugs / poverty / race as well as divorce to be fair it could well be they are correlated with crime due to a common cause. iq is a good suspect since it already predicts criminality even among sociopaths and strongly correlates with all of the above. it also provides a causal mechanism for the purported lead/crime and abortion/crime connections to work by

          • Any sources at all for your view on the causality of crime throughout this thread, Titanium?

          • jaimeastorga2000 says:

            Inequality is an entirely worthless measurement because it fails to distinguish between the poor getting poorer and the rich getting richer. The poor getting poorer is likely to lead to social unrest; the rich getting richer is not. If inequality grows while the poor get richer and the rich get vastly richer, then you’re likely to see more stability.

            I keep making this point in the comment section, but again: some of life’s most important things are zero sum, such as land and status, and it doesn’t matter if the increasing inequality is due to the rich getting richer or the poor getting poorer or even if the poor are actually getting richer but the rich are getting so much richer that inequality is still increasing – the bottom line is that the poor will now do worse at zero sum competitions with the rich.

            Decline of manufacturing shouldn’t cause any increase in violence – there’s no causal link.

            Rise of college degrees would be expected to decrease violence – educated people tend to be less violent, and to have more to lose. Moreover, the more each human life is worth, the less worthwhile it is to risk that life.

            The point is that a lack of manufacturing jobs and a requirement that you must have a college degree to get a job make getting a job much more difficult, which leads to crime. Credentialed education is one of the zero-sum competitions I mentioned above.

          • Irrelevant says:

            I disagree that status is zero-sum. There are as many people who are the greatest at something in the world as there are fields of human endeavor, and we keep coming up with more of those fields, and “Best _____ in the World” is a much stronger criteria for satisfactory status than most people actually have.

        • Eric S. Raymond says:

          At least one of those is backwards. “Weaker” gun control suppresses crime.

          Or, to put it in a way that is more generative of insight: crime is suppressed in the presence of higher levels of civilian weapons ownership by the 97% of people who are not in a 3% high-deviant category marked by high levels of crime, substance abuse, domestic volence, and accident-proneness.

          This effect is to some extent masked by regional differences. Criminologically the U.S. consists of Switzerland plus Swaziland – most of it extremely peaceful and low-crime, with scattered hot zones in major urban areas where violent crime (mostly associated with the drug trade) is endemic. Most civilan gun owners live in Switzerland; most criminals live in Swaziland. Statistics collected in one place can seriously mislead about the other.

          The suppressive effect of civilian weapons is especially marked for rape, felony assault, and hot burglary. Upthread someone noted the large difference in hot burglary erates between the U.S. and England. Ever since Gary Kleck’s 1992 study Point Blank: Guns and Violence In America it has pretty well established that hot burglaries are much less common in the U.S. because criminals rationally fear armed homeowners (much more than they fear arrest by cops).

          Other crime-suppressive effects of civilian weapons have been confirmed by longitudinal studies that examined differences in crime rates before and after changes in firearms regulation. Those studies, as much as the 2008 Heller vs. DC decision, are driving the very broad state-level trend away from restrictive firearms regulation.

        • Faze says:

          We should have more appreciation for what it was actually like to experience the transition from the low crime 1950s to the high-crime 1960s. For those who were there, it was stark, sudden and violent in every sense of the word. The most deeply imbedded social norms were overturned in the course of what seemed like a few months, beginning with the Watts riots. Imagine if you woke up tomorrow morning to learn that ISIS was now in charge of your schools, media, public transportation and urban core. That’s how it was for grown ups in the 1960s who awoke into a world that seemed to have been conquered by a tribe of atavistic barbarians. The notion of attributing this catastrophe to a confluence of small causes like the loss of manufacturing jobs and rise of day care, would appear to them as irrelevant as linking the Japanese tsunami of 2011 to the result of the Super Bowl or the Venezuelan economy. It was an earthquake, dammit!

          • Steve Sailer says:

            Right. Unfortunately, the media isn’t interested in recounting what the 1960s rise in crime was like to its victims, especially its white victims. For example, my wife walked a mile to school each day when she was in first grade in 1966 in the Austin neighborhood on the west side of Chicago. It was a walkable urbanist paradise of affordable housing.

            In a few years it turned into a wasteland.

          • Steve Sailer says:

            Here’s a recent Pulitzer Prize winning play by Bruce Norris, “Clybourne Park,” that gingerly pokes at the margins of what happened to so many urban whites in the 1960s, but doesn’t actually touch upon the massive violence that cleansed them from the cities:


  9. Irrelevant says:

    How cohesive is “crime” as a natural category? That is to say, I can easily imagine a ten-factor explanation for the drop in crime if crime is itself five loosely correlated things.

    • RCF says:

      But doesn’t that raise a similar issue, in that we should wonder why all five of them drop at the same time?

    • Steve Sailer says:

      Right, crime consists of a bunch of different things.

      For example, when I was a small child in the 1960s, stealing cars was easy because a lot of people left their keys right in the ignition because car theft was rare in the 1950s. I can recall public awareness articles telling people not to do that. But it was still easy to steal cars because not many people locked their car doors when they parked, and hotwiring cars was not difficult. Then people started locking their cars and manufacturers started armoring the wiring. After awhile, fewer cars were stolen but thieves concentrated upon stealing car stereos. In turn, target-hardening happened with stereos.

      These days, becoming a car thief is a pretty dumb career choice.

  10. Mike Johnson says:

    In some circles throwing lots of semi-independent factors at a modeling problem is referred to as “curve fitting”, and it’s seen as awful because you get a good match with past data but you lose all predictive utility.

    • Lambert says:

      Furthermore, could you find 10 factors that would predict an increase in crime?

      • Mike says:

        Very easily– think of a 10-factor version of this site:

        The probability of finding spurious correlations goes up as you
        (1) include datasets that probably aren’t related to what you’re predicting;
        (2) include more factors.

        It’d be pretty trivial with the right tools to find 10 factors that retrospectively predicted a decrease in crime, closely matching the data, and now predict a huge increase in crime.

    • Texan99 says:

      Couldn’t you just issue new curve-fitting models whenever the predictions didn’t pan out?

      • HeelBearCub says:

        I’m not sure if you are being facetious or you are missing the point.

        What you want is a model that predicts the future fairly well, so that you can use your resources widely. The only way to test a model is to look at the past and see how accurately it “predicts” what has already occurred. If you then tweak the model by adding more and more factors to make it more and more “perfect” at predicting the past, your increase the likelihood that the end model will not predict the future (the opposite of what you intended).

        An example might be trying to predict crowd flow through a doorway and whether people turn left or right once they pass though. You create your model and then test vs. video of people entering a night club. The model does well, but you can make it do better by adding information about the color of shirt/top the person is wearing. Because of the data set you are working with, it makes the model fit the “past” very well, but obviously shirt color doesn’t actually affect traffic flow through doorways in the real world. This is known as overfitting.

        It’s more likely to occur when you don’t have many real world data sets to assess. In the case of “crime rates in the U.S.” you have, depending on how you want to look at it, only one data set. Unlike crowd flows through a doorway, where you can generate thousands of data sets simply by videoing a nightclub entrance or office building or sports venue or …

      • Titanium Dragon says:

        To be fair, if we assume that people who consume organic food are stupid (it is a scam), then we might assume that those who consumed organic food would be more likely to have children with disorders.

        Likewise, we might expect people who are anti-vaccination to have more mentally deficient children due to less intelligence and possibly poorer parenting due to denial of personal responsibility.

        That being said, in neither case would the stupid belief be the cause of the increase.

        Ironically, one idea I came up with a while ago was the idea that vaccines may allow more autistic children to survive – autistic children get sick more often and many of them (30% in some things I’ve read) have varying levels of pica (chewing on/attempting to eat non-food objects). Vaccination would cause fewer of these children to die, and antibiotics could cure them of various bacterial and fungal infections.

        Thus, ironically, vaccination may increase the number of autistic children, but not by causing autism but by preventing autistic children from dying.

        Indeed, I wouldn’t be surprised if we saw more dysfunctional children the better healthcare got.

        • Anthony says:

          This is probably happening with respect to the more serious food allergies. There was no such thing as kids being deathly allergic to peanuts when I was a kid. Now it seems that every classroom has at least one.

    • Harald K says:

      There’s nothing wrong with curve fitting. You can get great predictive utility out of it as long as you are aware of the problem of overfitting and how to avoid it, but any applied statistics/machine learning class will cover that. It really is not rocket science cryptography.

      • Unique Identifier says:

        There’s a lot wrong with curve fitting. There’s a reason it goes terribly wrong all the time. One of these wrongs is that when curve fitting, getting good model-data correspondence is a given, whereas this used to be a great test of a model’s validity. With curve fitting, you have to be very careful and resist temptations, in order to get results which are useful rather than merely pleasant.

        You could rewrite your post, substituting ‘curve fitting’ for ‘guns’, and say that any basic firearms training covers -don’t point a loaded gun at someone you don’t want to kill-, etcetera. There are good and bad things about both guns and curve fitting, and as science is being done today, you would do well to bring your glasses of skepticism whenever you encounter curve fitting. Even though it is a useful tool, in principle.

        • Harald K says:

          One of these wrongs is that when curve fitting, getting good model-data correspondence is a given, whereas this used to be a great test of a model’s validity.

          It is standard practice to divide the data into three: use the first to develop the model, the second to adjust its parameters (and reveal/reduce overfitting), the third to test the final model. Getting a good model-data correspondence on the test set (the only place it counts) is not a given, and if you have it, that’s a good test of the model’s predictive utility.

          • Unique Identifier says:

            Theoretically, this should work. Theoretically, if you never point a loaded gun at someone you don’t want to kill, there won’t be any accidents. Saying these things doesn’t change the reality of things.

            For instance, there is no straightforward way to create training and test data when dealing with crime in the US, and there is no good way of blinding the researcher to the test data (unless he’s born and raised on the Moon, that is).

            [I have churned away at training and test data myself. It takes the tiniest amounts of ‘misconduct’ to get the garbage generator going, and everyone is free to agree or disagree, but -personally- I recommend moderate servings of skepticism when dealing with curve fitting.]

          • John Schilling says:

            Unique has got it right. In theory, under ideal conditions, it is possible to validate a model using pre-existing data to which the modelers were blind during the model’s creation and calibration. In practice, this rarely occurs – and almost never for controversial subjects. The “cheats” that break this sort of validation are things that the scientific community doesn’t intuitively think of as cheating, and too easy to do without deliberate intent.

            Models can be trusted to the extent that they predict the results of experiments that were not conducted until after the model was completed.

          • Tracy W says:

            I have churned away at training and test data myself. It takes the tiniest amounts of ‘misconduct’ to get the garbage generator going,

            Huh! I once recall spending 6 months trying to get some modelling results out for a project and failing utterly.
            Admittedly I only really had 12 data points for the LHS.

  11. Michael Watts says:

    I’d point out that while it’s fair to penalize a complicated explanation for invoking several different factors, it’s also fair to penalize a simple explanation for assigning very large effect sizes to a single effect. How big the penalties are will depend very heavily on your prior understanding of the system you’re trying to describe.

    Everyone else’s point, about multiple factors all being related, is also a good one. Imagine that, oh, europeans start a 250-year trend of getting taller in the 1500s. One model says that this is because of the discovery of the new world (causality is murky, but it’s a simple model and the notional cause occurs at the correct time. A rival model says that this is due to a variety of factors: for example, the influx of new world crops provides better nutrition; quinine provides better malaria treatment; and some population pressure within europe is eased by american settlement.

    (Please don’t take this comment as being historically informed.)

    In this hypothetical, the second, detailed model has a much more direct correspondence with reality, and tells us a lot of useful things that the first model can’t. But it’s hard to say the first model is wrong — all of the particular effects I listed are clearly derived from the new world, and even things like political reforms in europe that end up making it richer and healthier may be knock-ons enabled, or provoked, by the direct effects of the new world.

    On the other hand, it’s also easy to say the first model is wrong — the discovery of the new world can hardly continue for 250 years!

  12. Paul Goodman says:

    If you assume that the ten factors that influence it were all causing the decline equally, that seems pretty unlikely. But if you assume that there are, say, 20 factors, and a couple of them were driving crime upward, another eight or ten were pretty constant, and another ten or so were each moving in a direction that tends to reduce crime by some amount, it seems much more plausible.

    • Scott Alexander says:

      Is that true?

      If there were 20 relevant factors, they all must have been in perfect balance before 1994 in order to create the steady trend.

      For ten of them to suddenly change in 1994, while not enough of the rest of them counter-change to dampen the effect, doesn’t seem more plausible than there only being ten factors that suddenly change.

      • Decius says:

        “Perfect balance” is overstating the trend line before 1994, when specific effort went out to identify and change factors that reduce crime.

        • John Schilling says:

          ” specific effort went out to identify and change factors that reduce crime”, has been ongoing throughout the study period. That’s part of the constant background against which the signal in question is observed.

          Did people suddenly get better at identifying and changing factors that reduce crime in 1994? Because I’m pretty sure they were trying real hard for a long time before then.

          • Titanium Dragon says:

            There’s no particular reason to believe that 1994 is special; while crime began to decrease after that point, there is year-to-year variation, and we saw a long-term decline over a long period of time (1994-present, though the rate of decline has declined).

            You don’t need ten factors to happen in one year; if you have one or two factors with a small influence every year or two, that would cause a long-term decline.

            As others have pointed out, Moore’s Law continuing for a long time was the result of large numbers of discoveries, all of which allowed the improvements to continue.

      • Tom Scharf says:

        It’s not all about humans and social policy. The rise of inexpensive security and surveillance systems could be considered for example. A more recent example is Apple’s move to render stolen phones unusable via a kill switch. This has had a very measurable affect in iPhone theft, it has dropped 40% in SF and 25% in NYC.

      • ryan says:

        Obviously this discussion is starting from the view of “assume this data is accurate, how might we explain it?”

        But we should also keep in mind that this is a department of justice survey program that started in 1973. We don’t know what their survey methods would have shown to be crime rates before 1973. I guess we can be hopeful that the survey methodology didn’t change much from 1973 to the present. And we can also be hopeful that the survey methodology produced accurate information.

        Not the strongest concerns of course, but I have about this same reaction every time I see data plotted without error bars.

  13. CalmCanary says:

    “For example, if I understand correctly they’re arguing that the lead-crime connection is overblown because although lead was banned in the 1970s (thus affecting people who reached peak crime-committing age in the 1990s), the decline in crime continued even into the 2000s. But lead stays in the environment a long time, there’s still a lot of work to be done eliminating various sources of lead, and so blood lead levels continue to decline. That makes their argument ring a little hollow.”

    Also, last I checked 30-year-olds occasionally commit crimes.

    • MicaiahC says:

      Scott is making an argument about the change in amount of crime over time, not about the total amount of crime happening at one instant.

      • CalmCanary says:

        I don’t see how that’s relevant to my point. Suppose that no one commits crimes before age 15 or after age 39, that people exposed to lead as children commit more crimes than people who were not, and that any individual commits fewer crimes per year when they are 30-39 than when they are 15-29. Now imagine all lead is removed from the environment in 1975. From 1990 to 2004, we should see a massive drop in crime as lead-exposed young people mellow out into lead-exposed 30-somethings and are replaced in the 15-29 pool by people who were not exposed to lead. However, we should continue to see a downward trend in the following years, even though by that point none of the people at peak crime-committing age were exposed to lead, because lead-exposed criminals in their 30s will retire and be replaced by non-lead-exposed people.

        (Obviously, the specific numbers here are not important and probably bear only a superficial resemblance to reality.)

        • Steve Sailer says:

          Reading an old Chicago Tribune editorial, I discovered that one reason for building the Cabrini-Green housing project in Chicago was to get poor children out of tenements with lead paint flaking off the walls.

    • Titanium Dragon says:

      The problem is that peak crime happens when you are young (teenager to about 25 or so). This outweighs any residual effects of pollution by a large margin – getting older will decrease your criminality considerably, far too much for the effects of lead to be the cause.

      However, the real problem with the idea is that it is entirely made-up – we have reason to believe that blood-lead levels were very high in the 1950s and 1960s as well, well prior to peak leaded gasoline usage. Indeed, one major flaw is that they’re eliminating another major environmental source of lead – lead paint.

      Why do they not include this?

      Because lead paint peaked way, way earlier – think early 20th century, through the Great Depression.

      According to the obviously wrong lead theory, these folks should have been committing tons of crimes, but they weren’t.

      Likewise, the Chinese have very high blood-lead concentrations and yet commit very few crimes relative to Americans.

      The lead-crime association is spurious and discounts other major sources of lead, and the fact that almost all blood-lead level measurements come from during and after the elimination of leaded gasoline – there are few measurements from prior to that period.

  14. Eric Bruyant says:

    It seems very likely that all of the proposed factors are all correlated with either a general factor of technology level or of economic strength, the two of which are also somewhat correlated, so the credibility penalty should be much lower than if they were each uncorrelated datapoints.

  15. HbNSpcooiae says:

    Couldnt these factors all be the result of officials throwing anything they could at the problem to look strong in the face of rising crime? Explains increased incarceration, more/better police, and maybe lead removal too, if it was known that lead exposure caused violence. Maybe everyone in power decided to do whatever looked like it would work, and we’re just looking at the changes caused the ten things that actually had an effect. These ten factors wouldnt all necessarily have changed independently of each other.

    • Alejandro says:

      It may be the case for better/more policing/incarceration, but not for lead removal, because the strong correlation is between lead removal and crime decline about 17 years later.

      • Titanium Dragon says:

        The problem with lead removal is that there’s no correlation with lead concentration and the rise in crime.

  16. Tim Brownawell says:

    Er, there’s probably a *lot more* than 10 factors that affect crime levels. Which all never stay the same. If several of the more significant ones happen to shift in the same direction at around the same time? It doesn’t seem that unlikely. A *large* fraction (like ten million different things would probably be) all shifting in concert would be unlikely enough to indicate that your model’s a bit off.

    I also wouldn’t expect them to all be the same size. Aren’t there a couple common one-sided distributions that keep popping up for different things (word frequency, scale-free networks, etc)? The factor weights ought to match one of those.

  17. Jacob Schmidt says:

    The second person seems to me to have a strong argument, which makes me think Vox and the Brennan Center’s model where ten different trends each explain about ten percent of the decline is unlikely.

    Unless those 10 factors are related.

    Looking at the list of examples, several things with apparent small effects are linked via the government: greater incarceration rates, larger police forces, police tactics (these three are all directly linked as an attempt at reducing crime), legalized abortion, and environmental laws regarding lead.

    Several other maybe factors are linked via research and development: technology taking over banking so that we use less cash; psychiatric care (prescription of medication specifically); and media technology keeping us indoors.

    It seems to me that we might also model this as 2 or 3 meta-factors that enable progress in many other factors at once. You don’t need to look for ten independent minor reasons for the decline. Look for a couple of reasons that explain lots of minor reasons.

    • Steve Sailer says:

      The general pattern is that liberalism was ascendant in the 1960s into the 1970s, which led to an increase in crime. From the 1980s onward, conservative forces were dominant when it came to the criminal justice system, which eventually led to a decline in crime.

      • Harald K says:

        Again, the rest of the world poses a problem for this “Liberalism causes crime” hypothesis.

        • FacelessCraven says:

          isn’t that fairly slap-bang in the middle of socialism’s heyday? The cusp of the global revolution that everyone was sure was coming any moment? It seems to me it might be an arguable case that the political ferment of the time (and arguably the repressive measures of governments beforehand) did serious damage to the fabric of society worldwide.

        • Steve Sailer says:

          The rest of the advanced world, with the exception of Japan, had a crime bubble which lagged America about as fast as 1960s American ideas (what Pinker calls “the Rights Revolution”) had an impact on their cultures.

          Japan, of course, poses an obvious problem for the lead theory.

      • Titanium Dragon says:

        There’s a few problems with this, the largest being “define liberalism”. The US is more liberal than Europe; Europe is more leftist than the US.

        But, okay. If “liberalism” was the cause in the increase in crime, then we should expect the most conservative parts of the country to have the lowest crime rates.

        Instead, we see the most conservative part of the country (the South) having very high crime rates, while the Left Coast has low crime rates, particularly Cascadia. Indeed, people who are defined as “Liberals” have a low crime rate, whereas the people on the conservative left (blacks) have high crime rates.

        If we look at Europe, the Scandinavian countries, which are probably the furthest to the left and the most liberal part of Europe, have very low crime rates; Switzerland also has a very low crime rate. The UK is more conservative and has a higher crime rate than those places, but a lower crime rate than Eastern Europe.

        The data does not match your hypothesis.

        • Troy says:

          If “liberalism” was the cause in the increase in crime, then we should expect the most conservative parts of the country to have the lowest crime rates.

          Instead, we see the most conservative part of the country (the South) having very high crime rates, while the Left Coast has low crime rates, particularly Cascadia. Indeed, people who are defined as “Liberals” have a low crime rate, whereas the people on the conservative left (blacks) have high crime rates.

          This seems to me a weak argument. First, and most obviously, we need to control for race: it’s compatible with blacks being more conservative than whites and blacks committing more crime than whites that conservative blacks commit less crime than progressive blacks, and that conservative whites commit less crime than progressive whites.

          In addition, there’s a difference between the effects of political attitudes on personal behavior and the effects of political policy on societal behavior. Steve’s hypothesis is that progressive social policies led to increased crime. This is compatible with progressives being personally less likely to commit crimes, whether because progressivism makes people less likely to commit crimes or (more likely) because of common causes (besides race).

        • Cerebral Paul Z. says:

          According to the usual conservative story (which I don’t necessarily vouch for), the “liberal” approach to crime was an elite enthusiasm, which arguably gave it a certain immunity to geography. If judges in the South decide that criminals are misunderstood victims of society, what everyone else in the South thinks of the matter doesn’t really signify.

          • FacelessCraven says:

            my point was less that liberalism and leftism innately cause crime, and more that if we’re looking for a global driver of social change in this particular era, radical revolutionary ideology looks like it might be a decent fit.

            @Titanium Dragon – “But, okay. If “liberalism” was the cause in the increase in crime, then we should expect the most conservative parts of the country to have the lowest crime rates.”

            Less “liberalism”, more radical revolution ideology. Where do you see radical revolution ideology concentrated? Major cities and minority groups, yes? [EDIT] – Also the young, and the poor. So that’s four for four of the groups you’d expect to see the most crime in anyway.

            “If we look at Europe, the Scandinavian countries, which are probably the furthest to the left and the most liberal part of Europe, have very low crime rates; Switzerland also has a very low crime rate. The UK is more conservative and has a higher crime rate than those places, but a lower crime rate than Eastern Europe.”

            Switzerland, low levels of radical revolution ideology. Nordic countries, my guess would be also low levels, and also low conflict within an already-far-left society. England and France, relatively high levels. France had students throwing up barricades at around this point, yes? And most of eastern Europe at this point was ComBloc.

            My point would be less that liberalism obviously destroys everything good, and more that there is, in fact, a globe-spanning political movement ascendant during the period in question, that movement was championed by leftists and liberals, and it exhibited a high degree of hostility toward existing societies and social structures. You can find that hostility in pretty much any sample of liberal/leftist thought of the day.

        • Steve Sailer says:

          In America, crime statistics are driven, overwhelmingly, by differences between races, which overwhelm regional differences among races.

          This is a huge reason why few people are very expert on crime statistics in the U.S.: the racial patterns are so overwhelming that you are quickly confronted with a choice of become a crimethinker, obfuscate, or go think about something else.

        • Steve Sailer says:

          Look at the ratio of imprisonment of blacks to whites across the states: the highest racial inequality is in the most liberal place, Washington D.C. The next highest black to white ratios are in the socially progressive nice white people places like Wisconsin, Minnesota, and Iowa.

          The greatest racial equality in imprisonment rates are found in hard-headed southern and southwestern states like Texas.

        • Steve Sailer says:

          The main reason there isn’t much crime in Cascadia is because there aren’t many blacks there.

          According to an Obama Administration report, blacks made up a majority of homicide offenders over the last three decades.

  18. DanielLC says:

    Imagine there were a hundred factors. Each of them randomly moves in different directions. You would get about fifty in each direction, but not exactly. You’d be off by order of ten. You might have 45 factors increasing it and 55 factors decreasing it, causing a net decrease.

    • Supposing each factor has a variance S=1/N (with N the number of factors), then the variance of the sum of the factors is 1/sqrt(N) (by the central limit theorem). This result should be fairly robust given the various CLT extensions out there.

      Now lets be frequentists and ask ourselves – suppose N factors are involved. What is the probability of seeing an event of size X or larger? The answer is 1 – InverseCdf(X / (1/sqrt(N))) = 1 – InverseCdf(X sqrt(N)) = really fucking small as N grows.

      • Titanium Dragon says:

        This is incorrect. Say that the extinction rate on Earth is a constant, and then is modified by some factors – major volcanic eruptions, major meteoritic impacts, ect.

        Most years, those factors won’t change; in years where those factors do change, though, you see a very significant effect.

        Indeed, biologically speaking, a mutation in one gene can have very profound effects – there are a huge number of factors involved in skin color, for instance, but if you have a mutation in one of them, you can end up an albino regardless of what the other factors say.

        This is a horrible misuse of frequentism.

        • I’m confused. Your albinism example suggests that one factor is a more likely cause than many, which agrees with what I wrote.

          The Bayesian calculation wouldn’t be much different, just plug the frequentist result into Bayes rule. A single cause is far more likely than many.

          • Mr. Breakfast says:

            If I had to guess, the factors are sequential: gene A synthesizes substance a which is the precursor that Gene B uses to produce substance b, etc.

            All steps of the chain can influence the end result, but a complete failure in any one can shut down the whole process independant of the state of the others.

          • Mr. Breakfast, again, the likelihood of one factor in an “AND” statement changing at a given time is far higher than two factors changing simultaneously (assuming independence).

            Now given a causal relation y=x1 AND x2 AND x3, you can certainly say that all three x’s cause y. But if you observe a change in y in, say, 1994, most likely 2 of {x1, x2, x3} were already true and the third factor switched from false to true at that time.

            Pretty sure the arithmetic with poisson processes would support this, whether bayesian or frequentist.

          • Mr. Breakfast says:

            @ C.S. –

            I was responding to the first part of your statement:

            ” Your albinism example suggests that one factor is a more likely cause than many, which agrees with what I wrote. “

            Which I took to be a response to this:

            ” there are a huge number of factors involved in skin color, for instance, but if you have a mutation in one of them, you can end up an albino regardless of what the other factors say. “

            The gist of my response was that the behavior TD was describing is plausible where there is a clearly understood direct causal chain; each factor is a limiting factor in the final outcome. Probablility has nothing to do with the fact that one factor can completely control the outcome by being absent.

    • Baby Beluga says:

      Yeah, this is exactly the right objection, I think. If each of the factors could account for one-tenth of the increase in crime, then this is an entirely plausible model that supports the first professor’s point.

  19. Amisoz says:

    I’m skeptical of the “ten for 10%” argument as well, though maybe for different reasons. Historical causation is a weird thing. Michael Oakeshott wrote that “history is concerned with occasions, not causes”, and while I’m probably not so much of a skeptic to think that causation can’t ever be shown in history, there’s something uncanny about it all, esp. since the past doesn’t ‘exist’ in the way that most things do.

    All this to say that, I’m also interested in any sources people can find on calculating multifactorial explanations … and if that is even the right way to think about this all.

  20. Charlie says:

    When you say “1994,” I get a feeling I should be noticing something. And I think it’s the same feeling I get whenever anyone says “1998” in the context of global warming.

    If you need to explain the decline starting in a specific year, that might need the specificity of a single factor. But if you look at the crime rate curve and think “well, what if the period around 1994 just looks like the elbow because it had some short-term factor making it stand out?” it looks like you could actually start the decline anywhen from 1982 to 1996.

    I have no idea whether short-term bumps like I hypothesized are likely enough to be a good explanation for the data (crime seems even worse than climate!). But I feel like it illustrates the trouble of linear fits to data with complicated, ill-understood dynamics.

  21. Fnord says:

    Imagine that, in 1994, each of America’s ten million criminals independently and coincidentally had a major life change that made crime seem less attractive. That’s ridiculous. But in that case, any other explanation based on ten million factors should seem ridiculous. And if we give a heavy credibility penalty to a story with ten million factors, we should give some credibility penalty to a story with ten factors.”

    That’s not valid reasoning. We would be very surprised to find that ten million people independently and coincidentally had a major life change. It would be very unlikely for those ten million independent things to happen at once. But unlike in the case of ten million personal epiphanies, we already know the things descried in the article actually happened.

    Now, it might be surprising if all those things happened to reduce crime (even if we naively expect each them to do so; conjunction of independent probabilities). The ten for 10% would be an unlikely coincidence. But it would be at least as surprising if none (or only one) of them reduced crime, because that’s also a conjunction of independent probabilities. We’d expect to find some mix of crime-reducing and non-crime-reducing changes (which is basically what they found, except that a bunch of answers are “more data needed”).

  22. Sam says:

    The surprising thing about such an explanation is not that there were a lot of factors, but that they all happened to move in the same direction at the same time.

    Here’s a simple model: We have some number (in this example, 10) of factors which will a priori either increase or decrease their equal effects on crime with equal probability. The chance they all decrease at the same time is (1/2)^10 = .001, exponentially small as you increase the number of factors.

    The problem with this model, of course, is that these 10 factors aren’t the whole picture. What similar-variance factors moved in the opposite direction? If we suppose that these 10 factors decreased crime, but five other factors increased it at the same time, the chances of this happening randomly (in the same model) go up to (15 choose 5)/2^15 = .092. The intuition here is that there are many more ways to pick ten factors to decrease and five factors to increase out of all 15 than requiring that all ten out of ten decrease.

    We could build a more complicated model by adding, e.g. a general bias towards crime decreasing (through these factors), or the possibility that the factors make no (or very small) effect, but the big picture will be the same as in this simple model.

  23. Will says:

    So if you doubt the regression based “how much of the decline does this account for?” type analysis, how willing are you to go down that rabbit hole? This is the core social science methodology, so are you now more skeptical about all those other literature reviews you’ve done?

    If instead of assigning x% declines to many different factors, if they had rolled up all the factors into a “crime quotient” and said “crime declined because the crime quotient went down” would that make it better research?

  24. math_viking says:

    The factors you refer to are almost certainly correlated with each other, and for some are a deliberate response to the crime. For instance, penalizing “more police” and “more incarceration” for being multiple factors contributing to the end of a crime wave seems completely bonkers to me.

    But also, the crime wave of the 70s-90s was way above average going back from now at least to pre-WWII. The crime wave to begin with was highly unlikely, and was likely the result of many factors itself; its end seems like regression to the mean.

  25. NonsignificantName says:

    Maybe there’s an anthropic thing going on, where very rarely a bunch of factors will converge, and these will cause large, hard to explain trends that everyone talks about, so the probability of multiple factors given that the trend is a conversation piece ends up being higher.

    • Troy says:

      This seems plausible. Where causes of trends are fairly obvious (e.g., people are living longer because of better medicine and higher standards of living) they won’t attract our attention so much. But it’s not surprising that some historical trends would have numerous significant causes, and it’s also not surprising that these would be the ones we would find more puzzling and argue about.

  26. Tmick.wtg says:

    Is there a real, functional difference between these two statements?

    “I currently believe each of these ten things are 10% responsible for the drop in crime, but am willing to adjust my proportions as more evidence comes in.”

    “I currently believe one of these ten things is wholly responsible for the drop in crime and have no reason to prefer one over any other, but am willing to adjust my probabilities as more evidence comes in.”

    It seems that your strategies going forward would be the same in either case. It seems policy decisions should be the same in either case.

    That is: if you are 10% certain that broken-window policing is responsible for ~100% of the ‘saved’ lives due to the drop in crime, should you not act as though you are ~100% certain that it ‘saved’ 10%?

    If you’d pay a dollar for a 100% guarantee that you’d save a life, you should pay a dollar for a 10% chance of saving ten, or a 1% chance of of saving 100, right?

    This question smacks of unclear map/territory distinction: crime exists at the individual level and to really ‘get’ it you really would need to model every individual and their environment. That’s the big, capital-T Truth. The model you should prefer is the model most likely to help you achieve your goals, presumably a low crime rate with a minimum of costly or aggressive policing.

  27. Decius says:

    It’s plausible to me that there are about 20 major factors, ten of which account for 20% of the change each, and another 10 that account for -10% of the change each.

    • Steve Sailer says:

      From my review of Pinker’s book:

      “No reductionist, Pinker attributes what he sees as the slow retreat from violence to “six trends” interacting with “five inner demons,” “four better angels,” and “five historical forces.”

      “These 20 factors—ranging from the rise of Leviathan to the expansion of empathy and rationality—aren’t really enough to explain trends in violence, but they’re a start.”

  28. Richard Metzler says:

    Scott, I’m not convinced by your reductio ad absurdum. We can be pretty sure that crime rates are influenced by more than one factor – otherwise we’d have to conclude that, e.g. lead levels in South Africa and Mexico must be outrageously high, and that drug trade, poverty and general corruption have nothing to do with it. It’s also not unreasonable that, among various relevant factors, more than one move in the right direction in a 25-year time span. As you correctly state, the number of factors is plausibly less than 10 million, but that doesn’t rule out 10 as a plausible number 1 < n < 10000000.
    Now, finding out if it's, in our case, 3, or 7, or 10 factors, and which ones… that's the hard part that can't be hand-waved away.

  29. lunatic says:

    I read an article once that charted improvement in all sorts of products (aeroplanes and cars were included, if I recall). It showed that the “exponential” phase of improvement was made up of lots of “mini-improvements” due to new technologies that were daisy chained together. Thus there was no single driver of the overall tend. However, it seems this explanation is easier to swallow when all the factors are driven by some intentional process, which excludes most of the things credited for reducing crime.

    • Steve Sailer says:

      Right. Moore’s Law, for example, hasn’t been driven by one single giant technological breakthrough, but by a steady pitter-patter of smaller breakthroughs.

  30. Peter says:

    The other way of trying to save the multifactorial model is to read the graph differently. You see a flat-ish line then things falling off starting at 1994. That’s my first impression. Alternatively, one might see a gradual fall-off starting around 1982 or so, and superimposed on that, a crime spike about a decade or two wide centred on 1994.

    This suggests a project – randomly generate some crime statistics with causes of changes happening at random times. Try different distributions of cause strength – try all-causes-the-same-strength, a normal distribution, one of those freaky power law distributions. Plot the graphs and see which ones contain things that look vaguely like your crime stats.

    Maybe historical causes are like earthquakes – most of the energy released (and most of the body count) is accounted for by a few big earthquakes, most earthquakes are so small you need a seismometer to detect them, there’s no clear cutoff between big and small. The Gutenberg-Richter Law is a power law distribution.

  31. BD Sixsmith says:

    To add to the above, one or two things can actualise the potential of other things. If John Smith skips work tomorrow he might blame a cold bug without being strictly dishonest but three nights of booze, kebabs and sleeplessness will not have helped.

  32. Steve Sailer says:

    From my review of Steven Pinker’s “The Better Angels of Our Nature:”

    “While we don’t fully understand crime trends—perhaps lead poisoning played a role in the 1960s?—reducing the imprisonment rate while the murder rate was growing was the most characteristic cause of the 1960s disaster. Pinker notes that from 1962 to 1979, “the likelihood that a crime would lead to imprisonment fell … by a factor of five.” That America allowed rape and robbery to get out of control around 1964 reflected a shameful dereliction of duty by elites.

    “We’ve since quelled random violence to some degree, primarily by throwing a vast number of men in jail. The actual outcome of the Rights Revolutions appears to be more freedom for the upper reaches of society and more prison for the bottom. In 1960, only 1 percent of black male high-school dropouts were incarcerated, compared to 25 percent in 2000.”

  33. Steve Sailer says:

    The Los Angeles Times’ crime reporter Jill Leovy has an important new book out, “Ghettoside,” that argues that white people don’t try hard enough to throw black people in prison for murdering other black people.

    Here’s her interview on “Fresh Air” on NPR:

  34. tcmJOE says:

    What about some sort of threshold? I could imagine you could get enough factors working together so that things begin to tip after enough things come into play.

  35. Pete says:

    Many of the factors suggested are based on the USA only, but I’m pretty sure crime is decreasing across the western world. Surely this Suggests that US only policy decisions are only responsible for a small amount of the decrease at most.

    This seems to me to make a 10 factor decline even less likely, because you’d need a corresponding 10 factor decline across all countries showing comparable reduction in crime rates.

    The good news is, because different countries have different policies on things like prison sentencing, drugs, abortion, the environment etc, we should be able to compare results between countries and see if we can correlate different outcomes. I’d be much more sold on the “it’s easy access to abortion” idea, for example, if it could be shown that in western countries where abortions are still difficult to obtain (I believe it’s quite difficult to get an abortion in Ireland for example, at least based on the news that we get here in the UK), crime rates have stayed high. We probably can’t get definitive results, but we can surely get a more complete view.

    • Steve Sailer says:

      Steve “Freakonomics” Levitt’s abortion-cut-crime theory was dependent upon his coding error, as Foote and Goetz pointed out in late 2005:

      I pointed out to Levitt in our debate in Slate in 1999 that he’d forgotten to think about giant national trends like the Crack Wars:

      Levitt replied that, well, maybe I was right about the national trend, but how could I explain the state trends that he had calculated?

      Now we know, however, that Levitt had calculated them wrong by screwing up his programming.

      • More info on the “coding error”:

        Levitt’s conclusion is that “The good news is that the story we put forth in the paper is not materially changed by the coding error.”

        There are further methodological disagreements between Foote & Goetz on the one hand and Levitt on the other. These do change the results. But that’s to be expected. It’s usually the case that our choice of statistical methods affects our conclusions. 🙁

        • Steve Sailer says:

          Sure, Levitt has various excuses, but I pointed out to him way back in 1999 in Slate in our debate that his claims of state level analysis didn’t fit what we could see at the national level just by looking at age groups. He didn’t have any reply other than his state level analysis didn’t agree. Six years later, Foote and Goetz showed he had botched up his state-level coding.

          So, Occam’s Razor suggests than in 1999 I was right and Levitt was wrong. Levitt had six years to fix the obvious problems in his theory, but instead he kept and made the centerpiece of his 2005 Freakonomics bestseller, which made him rich. A half year later Foote and Goetz showed his screw-up, but Levitt was already a celebrity.

    • JB says:

      I wonder if crime in the US represents a major fraction of crime in the Western world. If this is true, then a significant decrease in the US crime rate would make it look like crime was decreasing across all Western countries.

      I haven’t run the numbers, but the US is very populous among Western countries and its crime rate is not that low, I think.

      • Titanium Dragon says:

        Crime decreased independently in all countries – not just as a grouping, but individually.

        Indeed, it decreased across the board virtually everywhere, with very, very few exceptions.

  36. Probably a combination of most of the 16 factors and not just a single one. Maybe add Flynn Effect and obesity to the list.

    • Harald K says:

      Wasn’t there some sort of slogan with the rationalist crowd that in order to not trick yourself into thinking you’ve understood something merely by naming it, you should name it something that reminds you you don’t understand it?

      In that vein, I propose we rename the Flynn effect “the mystery juice geysir”, since it a name for a historical phenomenon about a statistical phenomenon, neither of which we have pinned down very well.

      • Unique Identifier says:

        As far as I have understood, the Flynn effect is well explained by training – both formal schooling as well as everyday activities such as reading, or more intellectually demanding work than manual labor.

        It is my understanding that Flynn effect is (mostly?) limited to the same parts of IQ tests that respond well to tutoring, and is (mostly?) absent in domains like visuospatial and abstract-mathematical reasoning.

        I’m sure there are people here better informed.

        • Titanium Dragon says:

          There’s no evidence that it was caused by that and not, say, better nutrition.

          • Unique Identifier says:

            The evidence is right there in the post. Supposedly [citation needed], the increased performance is limited to the parts of IQ tests, where tutoring works.

            It is of course possible that nutrition and formal/informal education hits the exact same sets of skills, but not particularly likely.

          • Titanium Dragon says:

            Except that assumes that the increase in those is from education, which has never been proven.

            We had better nutrition and better education in the same time period; certain forms of malnutrition which were corrected via supplementation of our diets (iodized salt, ect.) are known to cause neurological issues.

          • Unique Identifier says:

            I will try once more, because this really isn’t that difficult.

            Let’s imagine that the Dystopia was trying to create the perfect basketball team, using all possible means. Assume that two centuries later, Dystopia dominates basketball.

            Everybody else is wondering – why are they so much better than us? Perhaps they have been selectively breeding super athletes? Maybe they start training at six months old? Maybe they have the world’s best coaches and tactics? Maybe they have secret performance enhancing drugs?

            Now, if it turns out that Dystopia’s players are not any taller their competitors, and they don’t jump any higher than their competitors, this tempts us to rule out selective breeding – because height is really useful in basketball, and should be one of the easiest factors to breed for. They probably haven’t been using steroids, because increased muscle mass should help them jump. If they are simply more accurate and cooperate better, the most sensible explanation is that they train more and have better coaches.

            There could of course be other explanations. Maybe they are all on some exotic form of Ritalin. But if they have been breeding for better basketball players, height must be less important than we thought, or very incompatible with other important skills, or very difficult to breed for.

            Nutrition seems a particularly bad explanation, in this case, because we expect nutrition to have a stronger effect on physique (height, strength) than on shooting and passing. It seems unlikely for differences in nutrition to cause this sort of difference in performance – but of course, it could be.

            I contend that a similar argument strongly suggests that the Flynn effect is educational. But of course, there are layers upon layers of conjecture.

        • Harald K says:

          You’re missing the first sin: giving a fancy name to the g-factor has tricked generations into believing they understand it. The g-factor is really short for “anything and everything which we can expect to have an effect on performance on a wide variety of quizzes simultaneously”.

          So whether it’s education, nutrition or anything else which explains the Flynn effect the mystery juice geysir, the answer is “yes”. There’s more of whatever things tend to improve performance on a wide variety of quizzes.

          • Unique Identifier says:

            The g-factor is first and foremost an understanding that there aren’t factors g1, g2, …, g101 – or, insofar as they exist, they are a much weaker signal than the main g. Some would say it is trivially true, but any number of people find it shocking or even repugnant.

            Interestingly, the Flynn effect is supposedly stronger on more weakly g-loaded tasks. This means, for instance, that we don’t expect that talent for chess has improved over the last century, despite the IQ scores rising (and being renormalized).

          • Nita says:

            there aren’t factors g1, g2, …, g101 – or, insofar as they exist, they are a much weaker signal than the main g

            What does that even mean? And is it an observation about intelligence or about mathematics?

          • Irrelevant says:

            It indicates that all intelligence strongly correlates with all intelligence, and they can’t meaningfully separate out single subfactors from within it.

          • Nita says:

            @ Irrelevant

            Thanks, that is better, although you’ve moved the vagueness from “weaker signal” into “meaningfully separate” (possibly losing some of UI’s intended meaning? I can’t tell).

            But, obviously, what factors we can “meaningfully” extract from the data depends on our methods, whereas what factors (in the common, non-statistical sense — “elements contributing to a particular result or situation”) actually exist in the system we’re trying to describe does not.

            The results of factor analysis don’t necessarily justify conclusions about factors in the common sense, so both thinking and talking about this should be done carefully.

          • Irrelevant says:

            “Meaningfully separate” in this case was a reference to the difference between the theory that g is measuring the brain’s processing speed, and the theory that g is measuring the combined effects of a number of cognitive subprocesses which can vary in efficiency independently but which you need to call on a large enough set of in order to solve any given puzzle or perform any real-life task that they can only be measured in combination. The two ideas are equivalent on the macro-scale, but they suggest somewhat different theories of how the brain functions, and figuring out which is true is an interesting challenge for neuroscience.

            I’m unclear on what you mean by “factors in the common sense” though. People don’t have conscious awareness of how their brains’ internal mechanisms function.

          • Nita says:

            The two ideas are equivalent on the macro-scale, but they suggest somewhat different theories of how the brain functions, and figuring out which is true is an interesting challenge for neuroscience.

            Exactly! Furthermore, we will probably have to choose a model on the basis of its usefulness for solving a particular problem, and give up on finding an idea that is both simple and “true”.

            Your “cognitive subprocesses” are factors in the common sense. But the results of factor analysis are factors in the statistical sense, which don’t necessarily correspond to the presence or qualities of particular biological structures or processes.

        • Douglas Knight says:

          It is certainly true that different IQ tests and subtests have Flynn effects of different strengths. Spatial reasoning has been full-strength. But I don’t understand why you’d think spatial reasoning would respond to tutoring less than other kinds of IQ tests.

          Perhaps you are thinking of Armstrong-Woodley (who say that visuospatial memory does respond to training). They claim that the strength of the Flynn effect matches the ability of people to improve by taking the test. They also have a theory, but I’ve never been able to tell what it is. Perhaps that Flynn-heavy tests have too small a bag of tricks? Even if this models the Flynn effect, it does not explain it away. People today are faster at learning the bag of tricks today than in the past.

          • Unique Identifier says:

            I’m mostly going on somewhat foggy recollections here – I cannot even remember who I am trying to paraphrase. Spatial reasoning might very well be the wrong example. [This of course means that I might be entirely wrong, but that goes for mostly everything on the Internet.]

            Part of the idea of an IQ test is that it should be blind to education. That is, the score should reflect underlying cognitive talents, but not education or acquired skills. [We have SATs and grades for the latter.] Ideally, an IQ test should not penalize pre-historic people for not having learned reading, arithmetic, etcetera.

            The reality, of course, is that it is very hard to disentangle -talent- from -training-. It seems sensible that just being used to sitting for hours and taking tests would improve performance.

            All of this is theory, and we cannot get any further than speculation, -unless- it can be shown that subsets of IQ tests respond more strongly to training effects than others, -and- that these same subsets have the strongest Flynn effect.

            Note that you can measure the former independent of the latter, by means such as coaching a control group, or by comparing the performance of westerners with rural peasants from China, etcetera. Hopefully, there are better schemes than these too, as well.

            -If- I remember correctly -and- my source was right, the Flynn effect is either due to -training etc- or some other factor which coincidentally has a very similar footprint.

  37. Very good post on an important and seemingly neglected topic. I wrote about this very problem here:

    My primary example in that post is Jared Diamond’s explanation of why agriculture first arose in the Fertile Crescent, rather than in some other region of the world. Diamond wants to argue that this was down to geography, rather than genetical differences between the inhabitants of the Fertile Crescent and, e.g. Africans or New Guineans.

    Diamond claims that the Fertile Crescent had eight factors that were beneficial for the development of agriculture. No other region of the world seems to have had more than five of them, according to Diamond. Moreover, Diamond does not give any example of a factor that made it *less* likely for the Fertile Crescent to develop agriculture.

    Now the score of pro-agriculture vs anti-agriculture factors in the best region being 8-0, and that of the next best region being 5-3, is a highly surprising distribution of advantages, unless we have reason to believe that the pro-agriculture factors are correlated.

    Do we have reason to believe that they are correlated? Let us forget what we know about agriculture and just think of this abstractly. Under what circumstances would we have reason to believe that causal factors making a certain effect more likely are correlated? I suggest we do that when the effect in turn functions as a *common cause* of these causal factors. Let me give two examples of this: intentional action and natural selection.

    Starting with intentional action, suppose that C1, consumption of nutritious food, C2, exercise and C3, drug-abstention correlate and that they all contribute to health, E. If so, this is presumably due to some people caring more about their health and/or having more self-discipline, and therefore try harder to achieve it. The effect (health) of these actions thus causes people with the right dispositions to take action to achieve the effect. Thus, in a sense E causes C1-C3.

    Similarly, in large African cats, the ability to hunt in packs, C1, and a large size, C2, which both contribute to the ability to kill large prey, E, correlate (lions typically have both, cheetahs and leopards don’t). This is, however, due to the effect causing its causes. Given the nature of African prey, it’s not very useful to have C1 without C2, and vice versa, which means that only the genes of the cats which would have had both of these features, or none, would propagate.

    No doubt there are other similar feedback loops, e.g. in a market economy, where features which contribute to a certain effect are correlated because the effect is a common cause of them. However, my main point is that *in the absence of such a mechanism, we have no reason to believe that the causal factors are correlated*. It seems to me that outside the realm of intentional actions and natural selection, such mechanisms are indeed quite rare. (The philosophical literature on functional explanations, in which feedback loops have been discussed at length, is relevant here.)

    I don’t think there is such a mechanism regarding the decrease in crime or regarding the development of agriculture. Either some of the factors that Vox and Diamond list do not, in fact, point in the direction they would have it (e.g. more permissive gun laws perhaps don’t decrease crime) or (perhaps more plausibly) they have *left out factors that point in the other direction*. It seems likely to me that the Fertile Crescent would have had some (non- negligible) disadvantages regarding agriculture visavi other regions, and that Diamond left them out.

    Given what we know of the extent of confirmation bias and other similar biases, it seems to me that the *standard* explanation of one-sided multi-factor explanations should be that the author in question is being biased. In Diamond’s case, that seems obvious. See also the interesting discussion in the comments of the post above, though.

    Regarding multiple-factor explanations being seen as more nuanced than single-factor explanations: that is also true, and I mention that in my post. One-sided multiple-factor explanations are typically not a sign of nuanced fox-likeness, but of bias. What is truly nuanced are *heterogeneous multiple-factor explanations*: where you say that there were some factors which decreased crime, some which increased it, but on the whole, the former were stronger.

    My guess is that one-sided multi-factor explanations are extremely common especially in the softer sciences, and that they normally signal bias. It seems to me that Pinker, who is discussed in this thread, committed the same mistake in *Better Angels* (see my blog post). This is something that people should look out for and point out whenever they see it.

    As it happens, I’m giving a talk on this very topic tomorrow at Birkbeck University, London. Multiple factor explanations are going to be one example of how we can infer bias from too one-sided or homogeneous belief structures. See here for a blog post on this notion:

    • Steve Sailer says:

      But multifactor causes are quite likely in response to a general desire for a good such as not being a crime victim.

      For example, when crime rises in the 1960s-1970s, people will start locking their doors, they’ll buy elaborate bicycle locks, they’ll move to the suburbs, they’ll rethink their liberalism, they’ll vote for law-and-order candidates, they’ll stop going for walks at night, they’ll buy car alarms, women will stay home and watch TV instead of going to the movies, they’ll start carrying cell phones so they can call 911, their phones will turn into cameras and GPS tracking devices, and so forth and so on.

      • yeah, but but you can’t give a lecture without enough circumlocutions . It’s not about what is most plausible , it’s about what makes you sound the smartest. It’s like it seems obvious, but it cannot be too obvious.

      • Interesting. There could be such common causes, which if so would make this an example of the “intentional action”-common cause I described above. However, it’s surprising, to say the least, that people failed so badly at acheiving this goal that crime actually went up in the 60’s and 70’s, and then suddenly succeeded on *16 different fronts*, while failing at no one.

        Hence even if what you point to could be a common cause of some of the causes of the decrease in crime, the pattern is still much “too good to be true”, especially given that crime actually rose not so long ago. The mechanism you point to was of course operative then, too (i.e. people wanted crime to decrease then too) even if it might have been strengthened after a decade of increasing crime (as you point out).

        • math_viking says:

          It’s fair to say that multifactor causes, with all factors pointing in one direction, are going to be rare. But so are the effects we’re talking about. If the cause of the crime wave and its end were likely, we would have a lot more of them.

        • Troy says:


          Nice points in the grandparent: your explanations are helpful and I’m largely in agreement.

          On the specific question of whether there’s a feedback loop, re: crime, it doesn’t seem too implausible to me that people would — for ideological reasons — make a lot of changes that would increase crime and then, about 20 years later, reverse those changes (or make different crime-reducing changes). It’s only clear that crime really is rising after several years of an increase, and it may only really start to hit home for people after, say, they have children whose safety they’re concerned with, which takes time. Or perhaps the crime wave first hit people with less political power (e.g., poorer and less educated people), and its effects on the upper classes took some time to be felt.

          Also, changes due to government policy are more likely to clump together, since changes in political power are discrete and not continuous — you don’t get a gradual shift from progressive to conservative over 4 years, you get a sudden shift as one administration hands over power to another one.

          • Stefan Schubert says:

            Ever since Kahneman and Tversky and others started to come up with evidence of bias on a large scale, people have come up various ingenious stories to save human rationality. These stories have generally been less than persuasive (Yudkowdsky has a nice post on that -“How they rub it in” or something like that). I think we’re often being over-charitable. In experiments, even stats majors make simple statistical mistakes. Why is it so improbable that Vox can have committed a fallacy in the case Scott discusses? It’s a much more subtle mistake than many fallacies people commit in psychological experiments.

            That said – yes there are some common causes and feedback loops and the discussion on them in this thread is enlightening. But I want to warn against over-charitability. The simplest explanation of why Vox put forward 16 factors pulling in one direction and none in the other is that they thought they were nuanced by listing lots of factors and that they didn’t realize there should be factors pulling in the other direction. Them having had some subtle story about feedback loops or common causes in mind seems implausible.

          • Troy says:

            We’re largely in agreement about cognitive bias. My interest here is entirely in the object-level question of whether the feedback loop mechanism might be operative in the case of crime reduction. I’m not making any claims about what was going in on the head(s) of the Vox author(s); Vox’s favorite causes might not even be the ones that actually go into the above loop.

          • HeelBearCub says:

            I feel that it’s a little disingenuous to put Scott’s conclusion (10 factors each contributing equally to the drop in rate) into the mouth of Vox.

            It seems to me that the Vox is simply listing a number of things that may or may not be causes of the crime rate drop and listing the arguments for and against those factors. They don’t seem to come to any conclusion about what factors actually contributed what amount.

            Essentially it’s a review of possible explanations, rather than a posited model.

          • Anonymous says:

            I understand in what way you think Diamond is biased, but I don’t see how “biased” applies to Vox here.
            They’re listing possible causes of a decrease in crime. That’s the job the article claims to be doing, and it does it. I don’t think they have an intellectual obligation to list all the possible factors that should have caused an increase in crime but didn’t (or maybe did, but no one can tell because of the overall decrease.) I will grant you that the article would have been a lot more interesting if they had. And it would make the given explanations look less post-hoc. But it’s not a form of either cognitive or political bias that they didn’t do it.

      • nydwracu says:

        When was the lead/crime thing discovered?

        (Rediscovered, I should say — some of the Romans knew about it.)

        • Steve Sailer says:

          Worries about lead paint have been around for a very long time — it comes up in the “Studs Lonigan” novels of the early 1930s.

          The problem with the lead theory is that there are a lot of point locations with extreme lead pollution due to lead smelters and the like — Google Superfund site lead pollution — but I haven’t found anybody who has documented unexpectedly high crime rates in those spots.

          For example, I read a lot of articles about a lawsuit filed by parents of lead-poisoned children living near a lead smelter in a small town in Missouri. Their usual complaint was not that their children were out of control but that they were sluggish and lacking in vigor.

          That’s just one place, but it should be fairly straightforward to study the effects of extreme lead pollution on crime, but nobody seems to have done it yet.

    • Scott Alexander says:

      I seem to have upvoted that post, so I must have read it and filed it away in my subconscious. Sorry for not giving you the credit.

    • Dan says:

      “However, my main point is that *in the absence of such a mechanism, we have no reason to believe that the causal factors are correlated*. It seems to me that outside the realm of intentional actions and natural selection, such mechanisms are indeed quite rare.”

      If there is reverse-causality, then it is pointless to talk of ‘causal’ factors to being correlated. The model is misspecified.

      But if you are suggesting that in the absence of endogeneity it is very rare to have correlated causes, that is simply not true.

    • Titanium Dragon says:

      The problem is that you’re looking at it backwards: you’re suggesting that Diamond is leaving out causes, but you have to realize that the Fertile Crescent really DID lead to agriculture and advanced civilization a long time before other regions got the same.

      Or to put it bluntly: if multiple factors need to align for something good to happen, then it is much more likely that somewhere where something good happened had all of those factors align.

      We’re not dealing with a random event here; we’re dealing with something which definitely DID happen. We know the outcome; if we assume we flip a coin eight times in each region, then the odds of any one region coming up with all heads is pretty small, but the odds of one region in the world coming up with all heads isn’t that bad.

      The main problem with Diamond’s hypothesis is that we may well not be identifying the right factors; for instance, Diamond notes that the Americas lack good large domesticable animals, but there is no particular reason to believe this is true. While it is true that llamas are not oxen who can plow fields, we do have caribou, moose, and buffalo – and there’s no particular reason to believe that buffalo could not be domesticated with a similar amount of effort which went into domesticating the aurochs.

      It is true, however, that the North-South orientation of the Americas effectively limited where you could send your animals – a good animal might work well for a range stretching from southern Canada to northern Mexico, but to give it to people on the opposite end of the continent requires you to ship it past a vast equatorial band. Indeed, we saw over in Eurasia that Africa and Indonesia ended up the furthest behind, whereas Japan, while still an island, was at the right latitude to adopt mainland goods and practices.

      However, describing how history went and why it went that way is difficult, and we may be rationalizing.

    • thirqual says:

      I agree with most of what Titanium Dragon wrote (we could discuss how caribou and moose are going to be very difficult as first species to domesticate because the areas where you find them are not going to be favorable for the other positive factors Diamond thinks are key — reindeer husbandry by Sami is not a good counter-argument).

      Moreover in your article you are discussing a multi-component explanation for an event, not for a trend, and Eurasians being superior would just be one additional positive factor. You have build a method which leads to every conclusion from multiple factors being attributable to bias if one squints a little bit, as long as you don’t cite/build a possible negative factor. Which, for the Fertile Crescent, could be the kind of obvious “fragile environment not adapted to trial-and-error methods” to contrast it with, say, many places in China or Eastern Europe. A potential reason why it is not cited is that most if not all Mediterranean climate zones are going to be in this category.

    • Jiro says:

      As someone else pointed out in response to the main question, this could be explained by the fact that unlikely things stand out more and therefore are explained more by scholars. Jared only chose agriculture over, say, the development of a writing system because agriculture is something that developed in only one area; ideas that are preselected to have developed in only one area will have unusually peaked distributions of causes compared to randomly selected ideas.

    • I think in Diamond’s case it does signal bias. He is very explicit in Guns, Germs, and Steel that he wrote the book to counter racialist theories of European dominance. He somehow failed to notice that the book never achieves this – that is, you can walk away very impressed by the evidence he marshals for causes in geography, crop packages, availability of dometicable animals, etc (I did) without thinking that is in itself any evidence against the theory “White people are smarter” (I didn’t).

      Note that I am not endorsing this theory (if only because a cold-eyed look at the psychometric evidence suggests “East Asian people are smarter” would be nearer the mark). What I am trying to point out is that Diamond never engages the possibility of overdetermination: that all the things he writes about are true and important, but that does not mean “White people are smarter” could not also be true and important.

      It’s a large flaw in an otherwise brilliant book, and I can ony ascribe it to blinding biases on the author’s part.

      • thirqual says:

        And his answer would probably be very similar to “Je n’avais pas besoin de cette hypothèse-là”. (“I had no need of that hypothesis.”) — not really by Pierre-Simon Laplace (to Napoléon about why there was no mention of god in his books).

        Then something about the simplest model explaining the data being preferable, and fears of being able to fit anything once you add enough independent variables.

      • Steve Sailer says:

        I was chatting amicably with Jared Diamond in 2002 when I hesitantly brought up the fringe of the key argument against “Guns, Germs, and Steel:” that if the environments on the continents were as different as depicted in his bestseller, they would surely select for different traits in the populations of different continents. His face went blank, he grabbed his papers, and half-jogged out of the Beverly Hilton ballroom.

        In short, Jared Diamond is a very smart guy and he has a guilty conscience about the fast one he pulled in GGS to get rich.

        • Harald K says:

          In fact Jared Diamond explicitly deals with selection for different traits in the populations of different continents. What was the book’s name again? Guns, something something, Steel?

          If I met someone so prone to telling flattering stories of themselves, I think I would have got away in a hurry too.

  38. JK says:

    My impression is that the crack epidemic caused a big upswing in violent crime from the mid-1980s to the mid-1990s. It may be that whatever caused crime to decline started working its magic already in the 1980s, but the crack epidemic masked the decline. The abrupt drop in violent crime after 1994 may just have revealed the true long-term trend that was temporarily disrupted by crack-related crime. Under this scenario, the causes of the drop in crime may be multifactorial and do not have to have started operating at the same moment.

    Another scenario is that the drop was caused by, say, better policing, legal abortion, declining birth cohort sizes, and a rise in incarceration non-additively interacting to effect a much bigger crime decline than any of them alone would have caused. If one of these factors was introduced only in 1994 or so, the interactive effect would not have appeared earlier even if the other factors were present.

    • Steve Sailer says:

      Right. The big historical anomaly that needs to be explained is the increase in crime from roughly 1964-1975.

      After that, crime starts to recede, but it’s interrupted by two drug dealing eras of high crime: the powder cocaine crime wave that peaked in 1980 and the crack cocaine wave that peaked in 1990-1993. The crack wave was so stupidly disastrous (it was concentrated among very young males — crime was already dropping among older people), that it’s not surprising that it receded quickly.

    • Douglas Knight says:

      You can screen off the effect of crack by looking accross countries. Most countries have a single peak, pretty smooth. The best comparison, Canada, peaked in 1975 and slowly drifted down. Most countries peaked later.

  39. Steve Sailer says:

    Why wouldn’t there be multiple causes for something as desirable as avoiding being a crime victim?

    For example, muggings in Central Park at night shot up in the 1960s, but then declined because people stopped going for walks in Central Park at night. Similarly, my wife’s family got mugged a lot in the late 1960s, so they moved way out of Chicago.

    In other words, people made a lot of adjustments to the new, more vicious social reality that emerged in the 1960s.

    • Dan says:

      The flip side of a decline in crime is an increase in crime. Those factors you mention do not explain why the crime rate went up 60’s or if it was high why people did not do the same things to avoid crime. You need something more,like changes: e.g. X made it easier to move to the suburbs, report crime, stop going to the park, etc.

      • Steve Sailer says:

        Crime went up in the 1960s because of the triumph of liberalism — things like more generous welfare payments to single mothers, which liberated lumpenprole males from having to have a job to get sex, Warren Court decisions that encouraged cops to retreat to the donut shop, and black pride movements and the general culture that suggested to blacks that white people owed them.

        There were other factors involving the Baby Boom, but the ideological climate played a huge role.

        • Titanium Dragon says:

          That’s not liberalism but leftism, and it also ignores the fact that leftist countries in Europe and the Far East have lower crime rates than the US does.

          You would have to assume that black people have a uniquely negative reaction to social welfare programs, but we have no particular reason to believe this is so.

          Moreover, leftism surged in the US in the 1930s, and we didn’t see an enormous crime spike then as we saw later on – why not?

          • Troy says:

            You would have to assume that black people have a uniquely negative reaction to social welfare programs, but we have no particular reason to believe this is so.

            Higher time preference and lower IQ would seem to predict worse reaction to being on welfare.

          • Steve Sailer says:

            “You would have to assume that black people have a uniquely negative reaction to social welfare programs, but we have no particular reason to believe this is so.”

            We don’t have to assume that, we _saw_ that in the 1960s when some states, such as New York and Wisconsin, greatly liberalized their welfare programs. The effect on public safety in New York City and Milwaukee, for example, was immediate.

            All of this was documented by scholars by the mid-1970s, but younger people today are ignorant of this social science because the media almost never refers to it.

  40. Steve Sailer says:

    Here’s a graph of the Crime Misery Index showing both the homicide rate (the most trustworthy crime statistic) and the imprisonment rate indexed so that the 1950s = 100.

    • Titanium Dragon says:

      Unfortunately, there’s a problem: homicide, while easily tracked, is also a rare crime. Overall misery would likely track with more common crimes, not the rare homicides.

      The US has a very high homicide rate relative to much of Europe, but our overall crime rates are actually lower than many European countries – the UK, for instance, has vastly more woundings than the US has aggravated assaults, despite a far lower homicide rate.


      No one knows.

      • Unique Identifier says:

        The immediately obvious (and possibly wrong) explanation is that there are parts of the US where crimes go unreported. I assume no one in Detroit calls the police if their bicycle gets stolen – in the UK, they very well might.

        This is why homicide is a very useful number.

        • Titanium Dragon says:

          Sure, but it assumes that homicide correlates with other forms of crime, which is not necessarily the case. The US experiences higher homicide rates than other countries, but lower rates of some other crimes. It is true that people in inner-city Detroit are less likely to report crimes to the police, but there’s no particular reason to believe that is true elsewhere.

          The UK had 7.1 million crimes vs 9.8 million crimes in the US according to the latest data sets I could find, suggesting that that UK has not only a higher crime rate but a vastly higher crime rate – it is very unlikely this is due to a difference in people calling the police. This does not correlate well with our homicide rates, as the UK has a much lower homicide rate than the US.

          Thus, we cannot claim that overall crime correlates well with homicide.

          • Unique Identifier says:

            I would conclude that these numbers don’t correlate well with the reality of things.

            A quick search indicates that US has a 60% higher rate of car theft than the UK.

          • Mr. Breakfast says:

            ” The US has a very high homicide rate relative to much of Europe, but our overall crime rates are actually lower than many European countries – the UK, for instance, has vastly more woundings than the US has aggravated assaults, despite a far lower homicide rate. “

            ” Thus, we cannot claim that overall crime correlates well with homicide. “

            If I had to guess, I would think that high rates of gun ownership and concealed carry by civilians puts a damper on “Get drunk and fight / break shit” as a form of entertainment among young males.

            I remember when I was younger (coming of age in the crack-and-gang era) how older working class men would often complain about how younk kids were crazy/cowards for pulling guns so fast. They would talk about how back-in-the-day a parking lot brawl over a girl or an insult could be had honorably, with well understood norms about the level of violence which was permissable.

            People committing murder are probably in an entirely different state of mind in when assessing risks than people engaging in casual (nonlethal) crimes.

          • Steve Sailer says:

            In recent decades, white people in the UK are more likely to brawl and burgle than their distant white cousins in the US. Here are a number of reasons why:


  41. Steve Sailer says:

    One of the things that’s been going on since 1994 is that the government has been quietly de facto legalizing downer drugs, such as synthetic opiates, while continuing to clampdown upon agitating drugs like crack. If you work the system, you can now legally get all sorts of drugs that will leave you zonked out on the couch, but it’s still hard to legally get drugs like crack that make you want to go out and kill somebody.

    • David Hart says:

      I read your review of Pinker’s book, and I can get behind a lot of what you say. It’s a while since I’ve read the book itself, so I can’t remember to what degree he deals with the hypothesis of the War on Drugs itself being responsible for a fair chunk of the crime increase – by handing over control of a hugely lucrative industry to organised criminals who are incentivised to use violence because they cannot settle their disputes in court, and who are also incentivised to bring to market the most compact and easy-to-smuggle (and therefore most concentrated and dangerous) versions of any given drug – I understand it is much harder to get raw opium or coca tea at the retail level than heroin or crack. Plus the fact that you can’t regulate in favour of low-strength drugs under a prohibition regime like you can with, say, alcohol, where it’s easy enough to apply lower taxes to weaker drinks than stronger ones. Adding on to that the fact that as police budgets are diverted towards arresting drug users, regardless of whether those drug users are posing a significant threat to themselves and others, proportionally less of the police budget is left for going after traditional victimizing property crimes and violent crimes, so a higher proportion of the predatorily-inclined get away with it.

      I’m not sure how much this would explain, but it is certainly seems plausible to me that the answer would be non-negligible. Certainly, given the typical pharmacological effects of cannabis relative to alcohol, it would appear to counterproductive to discourage use of the former in favour of the latter. As we get more data in from the states that have legalized it, the picture should become clearer.

      • Douglas Knight says:

        You can buy raw opium in any big flower shop. Coca leaves are illegal, but easy to get mailorder.

        • David Hart says:

          Well, if you mean in the sense that you can buy lots of poppies, fair enough. Though I’m not sure how much extra effort you’d need to put in towards making it actually smokeable. I’d still suspect that it would be easier to get a given dose of (dia)morphine in the form of street heroin than actually harvested opium.

          And as regards coca tea, I will take your word. Though if so, how long has it been the case that it is easy to get by mail order? We may be touching on a more general claim, that when it is relatively easy to get your illegal consumer goods by mail order, the violence associated with the black market decreases, which if true would be another contributory factor to the crime drop, albeit a marginal and very recent one.

          • Douglas Knight says:

            I have only been aware of the possibility of getting coca leaves by mailorder for a decade, but I think it is much older. I don’t mean buying poppy flowers, but the seedpods. If you want to extract it to smoke, it takes some work, but if you just want to make tea, it’s easy. I’m not claiming that these are popular ways of buying drugs, just that they are not acting like a black market and the government is not treating them like a black market.

          • David Hart says:

            I don’t mean buying poppy flowers, but the seedpods. […] the government is not treating them like a black market.

            Oh, I see. Well I still suspect that a) the amount of work in getting usable opium out of shop-bought seed pods is going to put most users off (whereas being able to buy opium in smokable form would probably absorb some of the current user base for heroin), and that if the government became aware that a shop was deliberately supply seed pods for human consumption, it would crack down fairly quickly. Still surprised about the coca leaves though.

    • Douglas Knight says:

      It’s true that amphetamine usage peaked around 1970, but current usage has probably caught up.

    • math_viking says:

      As far as I’m aware, the evidence is that drug-related crime is almost entirely caused by the drug gangs protecting their turf, rather than drug users bugging out.

    • Titanium Dragon says:

      The problem with this theory is that the #1 and #2 drugs which are found in criminals – alcohol and marijuana – are both depressants, not stimulants.

      • haishan says:

        The problem with this is base rates. Criminals and non-criminals alike use alcohol and marijuana at vastly higher levels than they do, say, bath salts.

  42. Mark Dominus says:

    The Rick Nevin papers about lead decline specifically reject Levitt’s hypothesis that abortion rates are responsible for the decline in crime rates. Nevin points out that the crime rate decline is worldwide, but only the USA had legalized abortion in the 1970s. Other countries had no change, or made it illegal, and saw the same drop in crime.

    Original Nevin lead papers

  43. Stuart Armstrong says:

    In general, even if there are ten factors, you would expect one of them to dominate. It’s unlikely in the extreme that ten factors will have the same magnitude of effect (even if there are ten thousand factors and you choose the ten largest ones, you would expect the largest one to dominate in the real world).

    The exception to that is if some factors are caused by others – or are caused by the decline in crime itself. Like, assume that removing lead caused the initial decline, and then this caused a virtuous circle as attitudes changes, more police became available.

    Actually, I just realised that is an examples of a special case: it’s plausible that you can say “factor 1 was the main cause of decline 1995-1998, factor 2 was the main cause 1999-2003, etc…” So different causes being dominant at different times is plausible, but not different causes sharing the same magnitude over the same period.

  44. Pingback: Probability and Causal Density | askblog

  45. stillnotking says:

    Crime has dropped because people are getting smarter and nicer. Elua is gradually, inexorably kicking Moloch’s hairy ass; interpersonal and interstate violence are simply not as attractive as they once were.

    The rise in crime during the 1970s – 1990s was the historical anomaly that needs to be explained. It was probably down to some combination of unusually dangerous drugs and Baby Boom demographics (a surfeit of crime-prone young men). The drop since then is very much in line with the trend of the last several hundred years.

    Oh, and multifactorial trends are less likely than black people.

  46. Really enjoying the comments on this piece.

    There’s another possible scenario I don’t think is that simple to discount: If nine of the ten factors are prerequisites for the tenth working.

    I’m not saying that’s happening here, mind you. That might also be what you mean when you call the factors correlated – though I took that to mean ‘(appear to) have a common cause, thereby ensuring many or all of the factors that look like data are actually just noise, since they contain no new information’. It’s just something that came to mind.

  47. The 'anging Judge says:

    What is the expected effect of rates of crime reporting?

    Have hardcore criminals just self-segregated into areas where no one calls the cops?

    • Titanium Dragon says:

      To some extent, yes; the worst places in the US have 10x the national homicide rate, and blacks are vastly more likely to commit a number of very nasty crimes (robbery and homicide) than whites are, and are more likely to commit crimes in general across the board.

      However, the decrease in crime does appear to be real, and has affected even most inner-city areas; the worst places are places like New Orleans and Detroit.

  48. I agree that many-factor causal explanations of a single event deserve a penalty. But: 1. it’s not accurate to call the impressive multi-year decline “the decline”. Maybe several things got better over time. You don’t know whether the result is a sum (oversimplifying) of several separate impulses or just one impulse of long duration. It’s probably true that there are several significant causes and during the decline we simply had in any given year over that span more of them leaning toward decline than incline.

  49. JayMan says:

    The simplest explanation: the rise in incarceration:

    Did reducing imprisonment in the 1960s increase crime? – The Unz Review

    For the record, I haven’t looked incredibly deeply at this topic. But at the end of the day, the simple fact of the matter is this: the cause of crime is criminals. This is not as tautological as it sounds.

    • haishan says:

      It seems like most people agree that this is part of the answer. But there are a couple of reasons it seems unlikely to be the whole answer:

      (1) Incarceration rates skyrocketed from the late ’70s to the early ’90s with very little effect on violent crime rates. Either there’s a highly nonlinear relationship between the two variables, or the benefits don’t show up for 15+ years, or there’s something else going on (maybe crack?)

      (2) Admittedly I can’t find data on this, but: If the effect is mediated entirely through keeping criminals off the street, we’d expect to see about the same percentage of convicted felons in the 18-24 age cohort as of 2015 as there were in 1980. (I’m picking ’80 to avoid the crack epidemic.) Maybe this is what the data says, in which case, cool, +1 evidence for the incarceration theory. But if not, there’s something else going on. Maybe it’s mediated through a deterrent effect, but my prior for deterrence working is low. I guess you in particular might claim that it’s mediated through criminals having fewer criminal kids, which is an intriguing hypothesis that I’d have no idea how to test.

      • Steve Sailer says:

        “Incarceration rates skyrocketed from the late ’70s to the early ’90s with very little effect on violent crime rates. ”

        Actually, homicide rates dropped steadily for older men in the 1980s as more of them got locked and/or got the message that society was serious again about punishing criminals.

        The spike in homicide in the early 1990s crack years was largely due to quite young males.

        Personally, I think the emergence of gangsta rap with NWA’s early 1988 album Straight Outta Compton served to glamorize the crack dealer lifestyle and spread it across the country.

        There are a lot of people in the entertainment industry, like new billionaire Jimmy Iovine, who have a lot to answer for.

    • Douglas Knight says:

      Except that it fails to explain the trend in any other country in the world. The best estimate of the effect of incarceration is zero.

      • FacelessCraven says:

        I’m having a very hard time explaining the logic behind that statement. If incarceration has no effect on crime rates, could we just release all prisoners tomorrow and get no increase in crime?

        • Douglas Knight says:

          Of course I mean the marginal effect at the current margin.

          But it is plausible that arrests matter and prison doesn’t.

          • FacelessCraven says:

            Could you elaborate? Or do you have a link that might explain the idea in more detail? It seems really counter-intuitive to me, and it’s got me curious.

          • Douglas Knight says:

            There are two claims. One is empirical: every western country had a crime wave starting in the 60s, all about the same size (doubling). All of these crime waves returned to baseline about the same time, although the peaks were at different times. But only America increased incarceration serveral-fold. That is the basis of the empirical claim that incarceration did nothing.

            Is that plausible? Psychology tells us that certainty and speed of punishment is much more important than quantity of punishment. This makes very similar predictions to the “broken windows” theory. (The pure broken windows theory is that broken windows cause crime. In practice, broken windows policing is about arresting vandals, not just repairing vandalism.)

          • The link you posted doesn’t support your argument. While many countries show modest declines, only two others (Hungary and Italy) show declines that are at all like that in the US. So while it is possible that there is something going on in (almost) every country reducing crime, there are hardly any instances of the giant reduction in crime between 1994 and 2000, and then becoming absolutely level since then.

            Compare it to Canada, for instance. While Canada has had a large decline in murder, based on the chart you’ve posted nearly all of it happened between 1980 and 1995, exactly the period where it didn’t decline the US. Other countries (Korea, Spain, Netherlands, Ireland, Greece, Belgium) show no decline at all. While a cursory look at the data seems to indicate that global first world crime peaked around 1975 and declined since then, it seems likely that there was something going on in the US that wasn’t going on elsewhere.

            (I’ll be happy to revise my opinion if you can provide more/better data, it’s very hard to tell exactly what percent declines were based on a lot of the charts presented and quick google searches aren’t being helpful)

          • Steve Sailer says:

            “every western country had a crime wave starting in the 60s, all about the same size (doubling).”

            The absolute size of America’s crime wave is significant and vastly different. America had great cities devastated by crime, which simply didn’t happen in the rest of the world. Where is the Detroit of Europe that has depopulated itself?

            If in the rest of the world the crime rate went from almost nothing in 1970 to a nuisance in 1980, that’s a lot different from crime going from a major problem in urban America in 1963 to a vast plague driving populations hither and yon within a few years.

          • Harald K says:

            “But it is plausible that arrests matter and prison doesn’t.”

            This is what many criminologists claim: that swiftness and certainty of sanction matters far more than severity of sanction, and you can’t make up deficits in the former by more of the latter. It sounds intuitively reasonable to me, but I note the position also has some detractors, and it’s a partisan mine field.

          • Alexander Stanislaw says:

            I find this highly unintuitive.

            Here is a simple model of how crime works. Suppose that each person has some propensity to crime and the pool of people who have a large propensity to crime is much larger than the pool of incarcerated criminals. Then removing criminals from the population would not significantly affect the rates at which crime occurs (but it would have an effect – criminals have the highest expected propensity to crime). The only way I can see to make this model agree with incarceration having no effect (as compared to just arrests) is that

            1: Being arrested decreases one’s propensity to crime back to the population average.
            2: Incarceration increases other people’s propensity to crime.

            Either that or the effect of incarceration is real but very small (not enough to account for the 50% decrease in crime). Or of course the model is wrong.

          • Irrelevant says:


            The “swiftness/certainty of punishment” model is operating on the assumption that criminals are very bad at weighing long-term consequences. Raising the sentences from 10 years to 20 years is therefore expected to have a very minimal effect on the crime rate, because nobody who was willing to commit the crime at a risk of 10 years in prison is doing the math and saying it’s not worth 20.

          • Douglas Knight says:

            Sure, incarceration decreases crime by incapacitation. I’m just saying that effect is smaller than measurement error. Moreover, smaller than worth talking about.

          • Mr. Breakfast says:

            If crimes committed while in prison still “count” then incapacitation isn’t really a consideration, is it?

      • Steve Sailer says:

        Look at your charts about how much worse crime was in the U.S. than elsewhere.

        • Douglas Knight says:

          Crime was always worse in America. Homicide doubled in most white countries. American whites were already much more violent than other whites and their homicide rate doubled from that high base. American blacks were more violent still, and their homicide rate doubled (and then had an even worse spike with crack). And then it all came back to baseline.

          Maybe your prior is that such ratios are a crazy way to measure crime, but once it happens to so many populations, you have to admit that there is a common thread.

          • Titanium Dragon says:

            The American increase was larger both absolutely and multiplicatively; it went from about 4.5 to 11, whereas other countries were double or less than double (and France actually saw a decrease from its 1960 rate).

          • Alexander Stanislaw says:

            @Titanium Dragon

            I don’t know the numbers, but that would still leave most of the cause of the crime drop unexplained by incarceration. Incarceration would only explain a few percentage points as opposed to the whopping 50% decrease. Which might not be statistically significant.

          • Titanium Dragon says:

            How do you determine what percentage of that drop in crime was caused by incarceration?

  50. gwern says:

    Time-series have trends and are correlated all the time (often spuriously so; I think this point started to be made way back with Yule’s “Why do we Sometimes get Nonsense-Correlations between Time-Series?–A Study in Sampling and the Nature of Time-Series”), so being able to pluck a few time-series out of the ether of thousands of plausible candidates is not improbable. (I’m not an expert on time-series but I think the intuitive argument goes that trends can either go up or down, so it’s pretty likely for there to be an imbalance up or down given a group of trends – it’s unlikely they’ll balance out exactly 50-50.) The question is not so much whether it’s plausible that crime could be declining thanks to ~10 other trends as whether those specific 10 are well-supported.

    There’s also the ‘dense casual networks’ argument: the real-world is really complicated. Why expect a single component to dominate?

    Finally, there’s the economic argument: crime is bad, so various organizations or societies or governments have incentive to reduce it as much as possible, so if there were any silver bullet against crime (like oh say, incarceration and police) then they would invest in that silver bullet until it hits diminishing returns, and then the next silver bullet, and so on until all the remaining crime is caused by a grab-bag of small factors that the organizations don’t know about (lead?) or can’t meaningfully affect (economic growth, whether Halo will be released for PS3); when this remaining grab-bag is investigated, it’ll turn out to be a bunch of little factors (because all the big low-hanging fruit like ‘inventing the idea of a police force’ were already picked).

  51. Maurio says:

    Crime rates in the US seem to be a genuine puzzle to the thoughtful academic community, including comparisons to other developed countries. It seems that the US has latent crime rates comparable to failed states. This paper pretty much sums up why everyone’s personal favorite explanation doesn’t seem to work at all. It also has some interesting statistics on race in the US justice system.

    • John Schilling says:

      Really? Everyone’s favorite explanation refuted? Because there’s one perennial favorite explanation that Spamann carefully tiptoes around.

      To try and be non-inflammatory about it, the United States has for over a hundred years had a giant statue at our front door inviting the residents of failed states around the globe to come here and set up shop. And a hundred years before that, we didn’t so much as invite them as drag them over in chains. The rest of the industrialized world has generally been quite picky about what sort of immigrants it lets in, and only in the past generation have the rank and file of failed states been welcome in any great numbers in e.g. Western Europe.

      • HeelBearCub says:

        If you want to be non-inflammatory about it, you can leave of the “failed state” rhetoric (implying that the immigrants are people who are prone to failure at civilizing) and simply point out that this creates a higher number of in-group vs. out-group problems than this present in more homogenous populations.

        I’ll say it does have a truthy ring to it, but I’m not sure if the data backs it up.

        • John Schilling says:

          I could, but I think it matters where the immigrants come from – mostly in cultural terms, but that ties back into geography, nationalism, and race or ethnicity. Ingroup/outgroup and inhomogeneity doesn’t seem to cause crime in Quebec or Hong Kong or Singapore or Switzerland.

          So, yes, I do mean to imply that the immigrants are representatives of cultures that are prone to failure at civilizing. I also think it is possible to be too civilized, and I want the United States to have a constant influx of immigrants from less-civilized lands for, call it hybrid vigor. But this is a mixed blessing, and I think higher crime rates are an inevitable part of the cost.

          • Has someone actually written something about this? Immigrants bringing along their uncivilized natures and causing crimes and wrecking nations?

          • Mr. Breakfast says:

            @Ahilan – Mein Kampf?

          • Breakfast is nice says:

            My Dad has a crackpot theory along these line (we’re British): all the people that feel socially excluded, and are hungry for money, and are insecure about their social status, have flocked to the same part of the world. Their genes are probably more difficult to work with in terms of civilising a society than other countries. This is probably behind things like disfunctional social policies, crazy republicans, egotistical billionaires, fanatical ideological commitment to capitalist values, and crime rates… personally I have no idea and assume for the meantime that its something to do with a ‘them and us’ mentality that seems to pervade America. But I’m drawing on much less data than US commenters!!

      • Harald K says:

        If that’s your explanation, how about Australia? The immigration profile of that country at best resembles, at worst looks quite a lot worse than that of the USA.

        • Unique Identifier says:

          Counting fairly generously, I get less than 1% of Australia’s population born in what can be called failed states. A large part of it is South Africans, which are a fairly special case.

          I’m not sure if there’s much to be explained, and I would also expect most of their failed state immigration to be very recent.

          [Wikipedia: Demographics of Australia]

        • Jack W says:

          I think the immigration profile of Australia has a large advantage in that it’s under a considerable selection effect – 68% of all Australian immigrants are taken in under skills programs, presumably to be immediately provided jobs.

          The remaining 32% are family of immigrants, of which 79% are partners of other immigrants.

          So roughly ~90% of all Australian immigrants are those who come in with specific skills the country needs, or the partners of those people (assuming each immigrant group has the same number of partner immigrants).

          Conversely, the United States has 66% of its immigrants coming based on family reunification, and 13% admitted based on skills.

          You’d expect the cultural profile of ‘well-skilled and in-demand’ to do better at the abstract notion of ‘civilization’ than the profile of ‘poor, huddled masses’, regardless of country of origin.

      • Titanium Dragon says:

        While a not unreasonable theory in principle, it ignores that most of the population of the US has the same base as European stock, and that Europe, until the 20th century, had a vastly higher rate of violence in the form of warfare.

        • John Schilling says:

          Most of the population in the US does not contribute significantly to the high US crime rates; it’s only the origins of the particularly criminal subcultures that matter. Those mostly aren’t European, and for the one conspicuous counterexample I could argue that Sicily spent the late 19th and early 20th centuries as a de facto failed state.

          Also, we are talking about crime, not warfare. Aside from both involving violence, those are two very different phenomena. Crime is not usually the result of bored/unemployed wannabe soldiers, and wars are not caused by criminals getting together and getting organized. There’s no correlation that I can see.

      • vV_Vv says:

        What is the distribution of intelligence in the US as opposed to Europe?

        IQ scores are artificially normalized, hence they are not a reliable indicator.

        The average g-factor of the US is estimated to be in the same range of Europe, but is it possible that the US distribution has a greater variance or is multimodal?
        If the US has both some subpopulations with unusually high intelligence (let’s call them AJs and EAs, for the sake of the argument) and some subpopulations with unusually low intelligence (let’s call them Bs and Hs) that could explain why the US both excels at intelligence-loaded endeavors (science, technology, world domination, etc.) and has also high crime rates.

        • Douglas Knight says:

          IQ scores are normalized only once. They are not normalized in every country separately. They are normalized so that UK has mean 100 and standard deviation 15. So you can just look up the mean IQ of the US: 98. I’ve never seen published variances, but I have seen it claimed that various populations all have about the same standard deviation, from which you can compute the effect on American variance.

          It is true that the racial balance affects crime rates. Half of homicides in America are by blacks. But that leaves the rest with a homicide rate several times that of Western Europe. It is hard to determine the hispanic homicide rate, but it doesn’t seem to be high enough to drive the aggregate number. Even in 1960, when there weren’t so many hispanics, the white American homicide rate was much higher than than Western European rate.

          • vV_Vv says:

            IQ scores are normalized only once. They are not normalized in every country separately. They are normalized so that UK has mean 100 and standard deviation 15.

            Didn’t know about that, thanks.

            Even in 1960, when there weren’t so many hispanics, the white American homicide rate was much higher than than Western European rate.

            Ok, but perhaps this isn’t just a race effect.

            As others pointed out, even the white population of the US is made of relatively recent immigrants, while the population of Western European countries is mostly made of people whose ancestors have been living in the same lands for thousands of years.
            It is possible that immigration selects for unusual people, with different distribution of genetic and cultural characteristics (e.g. a significant fraction of the US whites come from religious minorities within their ancestral countries).

            The problem with this explanation is that many other countries mostly made of recent European immigrants (specifically, English speaking immigrants) have Europe-level crime rates.
            Argentina and Chile have crime rates similar to the US, but they have a lower HDI than the US and their crime rates are in fact similar (though perhaps a bit higher) to those of Eastern European countries with similar HDI.

            Maybe there was something unusual about the immigration patterns to the US?

          • John Schilling says:

            “We’re Americans, with a capital ‘A’! You know what that means? Do ya? That means that our forefathers were kicked out of every decent country on Earth!”
            – Bill Murray, Stripes, 1981

    • PGD says:

      The Spamann paper contains plenty of evidence that incarceration has been an important factor in reducing crime, underlining that the Brennan Center piece is kind of screwed up.

    • This whole discussion seems to assume that the U.S. has anomalously high crime rates. I don’t believe it’s true. The U.S. has an anomalously high homicide rate. But if you look at the ICVS figures, other crime rates are similar to those in European countries.

      Looking at the 2000 survey figures for total victimization rates, we have:

      “- Above 24% (victim of any crime in 1999): Australia, England and Wales, the Netherlands and Sweden
      – 20%-24%: Canada, Scotland, Denmark, Poland, Belgium, France, and USA
      – Under 20%: Finland, Catalonia (Spain), Switzerland, Portugal, Japan and Northern Ireland. ”

      “The risk of having a car stolen was highest in England and Wales (2.6% of owners had a theft), Australia (2.1%), and France (1.9%). Japan, Switzerland, Catalonia, the USA, Finland, and the Netherlands show risks of 0.5% or less. ”

  52. bellisaurius says:

    I always feel we look at the crime rate wrong. There’s probably a baseline number of crimes that occcur, but the largest chunk of the current number is caused by gang related activities.

    Most people probably model gang activity as a behavioral issue, but what if the better model is politics? Basically, political unrest over a territorial/economic issue. If there was a way of figuring out gang related budget numbers in a couple of cities over that time period (I would think evidence lockers would have a collection of said books), they can see if there was some precipitous drop off in income at that moment.

    • Titanium Dragon says:

      Gangs are only responsible for a small percentage of overall crime – less than 10% of homicides are gang-related, for instance.

  53. Unique Identifier says:

    One could do worse than reading Steven Pinker on lead and crime. You can find it by searching for ‘steven pinker lead crime’. I will post the direct link in a reply to myself, but it might get auto-moderated.

    [Also: Does for instance SAT scores, IQ tests (for instance military data) or similar show the same trend as crime? I would think lead much more directly damages cognitive abilities than it increases criminal behavior.]

    • Troy says:

      Does for instance SAT scores, IQ tests (for instance military data) or similar show the same trend as crime?

      This is a worry I’ve expressed about the lead theory in the past. Data I’ve seen say that IQ scores have been going up for all groups since the 60s, which seems to run contrary to the hypothesis that lead was having a large effect on people.

    • Steve Sailer says:

      Here’s a column I wrote for Taki’s Magazine a few years ago on ways to test the Nevin-Drum lead-crime theory:

      • Unique Identifier says:

        Thanks for the link. Japan is an interesting data point – as always. I really wish more could be said about Pinker’s intervening links. It should be trivial to show that people’s bones contained more traces of lead and that they were cognitively affected, if the effect of lead is strong enough to cause a roughly two-fold (!) increase in crime.

        • Steve Sailer says:

          Right. The lead hypothesis isn’t ridiculous, it just needs a lot more study. I encourage people to carry out the very tests I’ve dreamt up and any others they can come up with.

          • Steve Sailer says:

            One of the problems with the lead hypothesis is that it is often put forward as the silver bullet that will explain the black-white gap in crime. Rick Nevin and Kevin Drum do that a lot but anybody who thinks hard about it realizes that it’s unlikely to work. Nevin and Drum are always talking about how housing projects in Chicago were built near expressways, and they imply that explains why black crime is so much worse than white crime in Chicago. But I’m a white person who lived in highrises full of white people near expressways in Chicago. If you know anything about Chicago, you’ll quickly realize that Nevin and Drum don’t know anything about Chicago.

            I’ve argued that we should study small towns on the EPA Superfund Lead Pollution list to see if crime is higher than expected there. But nobody is very interested in studying what drives crime rates among rural whites.

            In general, a lot of people put a little bit of effort into studying black-white gaps in crime, or in education, or whatever under the assumption that everybody who came before them was a racist idiot so they’ll quickly discover the overlooked proof of racial equality.

            But then they learn more about the actual numbers, get depressed, and move on to something else.

            So, important subjects like what influences crime rates don’t get studied well because anybody who studies them in any detail has to deal with the fact that the Big Factor is that black people commit a lot of crimes. And that just induces crimestop (Orwell’s “protective stupidy”) in most people with a healthy respect for the career prospects.

  54. Douglas Knight says:

    Before you worry about choosing between a 10-factor model and a 1-factor model, you should check whether you actually have a 10-factor model. All that Vox does (well, actually, it does much, much less) is claim that 10 factors each explain 10% of the effect. But do they explain different 10%s or all the same 10%? Do you make better predictions when you combine the factors or just the same?

  55. zslastman says:

    You’re not going to find the easy statistical answer you’re looking for. The answer here depends on a bunch of priors – not just for how likely individual factors are, but for how likely you think they are to be correlated etc. Any out of the box answer, like those involving BIC, will be hiding such priors under the hood somewhere.

  56. Tom Scharf says:

    Yes, the “I have used a COMPUTER using a model based on MATH and have successfully unscrambled the eggs and TRUTH has popped out” claim.


    I almost uniformly dismiss these type of claims, especially when none of the factors being examined can be directly measured. I am unaware of a single instance in which this has occurred that the modeler didn’t end up with an answer that suited his preference.

    Imagine social science stating: “Well I was surprised to learn that increasing incarceration rates and more guns were the main drivers in the reduction in crime”. My model says this is exceedingly unlikely to happen, ever. I would say that if model v1.0 gave them this result, the hard drive would be reformatted and model v2.0 will be attempted within 10 minutes.

    This is a class of problems I would suggest are much more dependent on the model structure and filtering than the raw data fed into it. When you combine this with either a bias or politically pressured subject, science takes a back seat to desired results.

    What I would like to see presented is the results from several independent statistical models that attempt to unscramble the eggs. What you normally get is one model and one set of results. Certainly different assumptions vary the output. I would like to have seen them commit to a specific model before knowing the results. I very much believe there is commonly a feedback process between examining results and “enhancing” the model.

    I assume there are better and worse ways to do this type of thing. I have no faith that any specific instance of this type of modelling is using the better methods, whatever they are. The presentation almost always is that a perfect model was created — massive arm waving exercise — and here is the truth of the matter.

    I’d like to believe this type of thing is done with real expertise. I don’t.

  57. Setsize says:

    Leafing through the Brennan report, it looks like the reason they don’t include lead in their model is merely that they couldn’t find state-level timeseries for it. They also use the phrasing “lead in gasoline” which is slightly worrisome because you’d actually want to be looking at “lead in the environment.”

  58. Anonymous says:

    To what extent would decline in crime from its peak lead to more decline in crime? I’m wondering if we’re looking at a crime “bubble” which popped. I have the impression that in some societies where crime is high the returns on criminality are also very high and the social status of many criminals is also high. Perhaps the US was heading towards that kind of equilibrium at one point.

    Maybe a few of these interventions were enough to fatally injure the infrastructure and social capital of the criminal economy and it’s kept on collapsing ever since because it hasn’t reached equilibrium yet. OK, I’m extremely fuzzy on the details of how this would work; I’m mostly thinking about The Wire and how Stringer Bell is a much more aspirational figure than Marlo. About having to fight harder and meaner for scrappier returns. And I just think in general a lot of processes work in this kind of self-feeding/self-starving way.

    I was going to say the self-starving process might explain declines in career criminality, but doesn’t explain declines in casual violence, but actually it does, doesn’t it? Any kind of violence is basically the epitome of a self-feeding process. Like, maybe the probability of me starting a bar fight is a function of my estimated probability of someone starting a fight with me, which is based on the number of fights I’ve seen in the past few years… You arrest a few persistent fight starters and voila, crime starts a gradual decline.

    Although honestly I would bet on it mostly being the lead thing, because countries outside the US exist.

  59. Phil says:

    Sticking to the non-boring version of the question:

    If there are 1,000 independent and additive possible causes, and they’re all actually just random, the top 10 will almost certainly add up to the amount of the observed effect, won’t they?

    1,000 factors, each with an SD of 1. The SD of the sum of the 1,000 factors will be about 30. But the top ten factors, each of which randomly winds up 3 SD above the mean, will themselves add up to +30.

    In other words, no matter what, the top 10 factors “caused” the observed effect to differ by 1 SD. They’ll be balanced out by the ten factors on the other tail, but you have to remember to look at those.

    As a sports analogy: “If that average team hadn’t had X happen, they would have finished 25th out of 30 instead of 15th.” You can always find an X that makes that sentence true.

    Also: even if you DO consider the other factors, those top 10 really DID “cause” the effect. If they hadn’t randomly come out high, the observed number really WOULD be 3% lower.

    The hard part is, realizing that the “cause” was just random chance, and you can find other balancing factors that were also random.

  60. Pingback: Causes of Many Moving Parts | The Only Winning Move

  61. Ghyl_Tarvoke says:

    The theory that the crime spike in the 60s and 70s was caused by the counterculture is utterly absurd. For that to be true you would have connect countercultural ideas to that of the criminals and except in a few noteworthy cases, this seems very unlikely. Hippies and criminals are, generally speaking, not the same people you know. For large parts of the United States in the 60s the Counterculture was an irrelevance. Did crime spike higher in San Fransisco and New York than in other cities?

    Anyway what this debate seems to be missing here is recognition that these crimes patterns are not exclusive to the United States. Many countries show similar crime rises in the 60s to the 80s and then a colossal drop-off since the mid 90s (Here, for example, is an article from The Guardian on the drop of crime in the UK: and here is an article from the Economist on the global trend: Therefore whatever the cause is it probably isn’t unique to the United States. (This in particular rules out Levitt’s theory imo).

    • Steve Sailer says:

      :For large parts of the United States in the 60s the Counterculture was an irrelevance. Did crime spike higher in San Fransisco and New York than in other cities?”

      Yes. Crime spiked earlier in New York and other liberal cities in the 1960s, just as it did once again in the crack years.

      The Summer of Love of 1967 in San Francisco, for example, was quickly snuffed out by black criminals from the projects preying on high hippie chicks.

      Similarly, the crack wars started in the 1980s in L.A., NYC, and DC (all places that had legalized abortion by 1970, three years ahead of Roe v. Wade), and then spread to Red State America in the 1990s. I’ve looked at homicide trends for scores of cities and the pattern is clear.

      The role of West Coast and East Coast gangsta rappers in spreading the crack dealer ethos to the rest of America is worthy of investigation.

      • ShardPhoenix says:

        >The Summer of Love of 1967 in San Francisco, for example, was quickly snuffed out by black criminals from the projects preying on high hippie chicks.

        I appreciate your posts in general but in the interest of civilized discourse I think it’s best to avoid this kind of assertion without clear evidence.

        • cypherpunks says:

          Steve has written about this before.

        • Steve Sailer says:

          See my old professor Allen Matusow’s chapter on Haight-Ashbury in his history of the Sixties, “The Unraveling of America.”

          • Steve Sailer says:

            Matusow wrote in 1984:

            “Haight-Ashbury was already dying. It’s demise, so similar to the demise of hippie ghettos elsewhere, resulted from official repression, black hostility, and media hype. In San Francisco where city fathers panicked at the prospect of runaway hordes descending upon them, police began routinely roughing up hippies, health officials harassed their communes, and narcotics agents infiltrated the neighborhood. Meanwhile, black hoods from the nearby Fillmore district cruised the streets, threatening rape and violence. Blacks did not like LSD, white kids pretending to be poor, or the fact that Haight-Ashbury was, in the words of a leftover beatnik, “the first segregated Bohemia I’ve ever seen.” Longtime residents began staying home after dark. Finally, the beguiling images of Haight-Ashbury marketed by the media attracted not only an invasion of gawking tourists, but a floating population of the unstable, the psychotic, and the criminal. By the end of the year, reported crime in Haight-Ashbury included 17 murders, 100 rapes, and nearly 3,000 burglaries. In October 1967 community leaders staged a pageant called “Death of Hippie.””

      • Titanium Dragon says:

        Problem with your hypothesis:

        Most crime-ridden part of the US is the South, and there is a strong correlation between immigrants from the South and crime (regardless of whether said immigrants are black or white).

        The South is very conservative and resisted the counterculture movement.

        The South continues to be the worst part of America.

        If liberalism was the cause of criminality, then why is the South so crime-ridden?

        • Jiro says:

          Simpson’s Paradox. The South is more conservative but also has more high-crime subgroups. It may be that conservatives commit less crime than liberals for every subgroup of the South, but the imbalance in subgroups results in a higher crime rate.

          • cypherpunks says:

            It is true that the south has lots of blacks. I don’t know whether they are more conservative or more violent than northern blacks. But southern whites are both more conservative and more violent than northern whites. Do you really propose to break the categories down more? If so, name some categories.

          • Jiro says:

            It isn’t necessarily blacks, it could be poor people. It could be Mexicans, for that matter.

          • cypherpunks says:

            But it isn’t Mexicans. Sure, maybe there’s a Simpson’s paradox hiding there, but just asserting that is not helpful.

        • Dishwasher says:

          The black homicide rate is much lower in the south and the west than in the mid-west and the north-east but the white homicide rate is higher. (There are 2 exceptions that might shed some light Louisiana has a high black homicide rate and NY low.)

          Other crime is hard to measure. (In my experience, not much to go on I know, property crime seemed much higher in the north where I come from than the south where I live now).

        • Steve Sailer says:

          The South has a lot of blacks.

          To a lesser degree of importance, it also has ornerier white people, especially those of Scots-Irish descent, but differences between white people in crime rates are not that important compared to the giant black v. white v. Asian gaps.

          It probably makes more sense to say that the reason the South is more conservative is because it has more blacks and troublemaking whites.

          And, guess what, blacks seem to do better under conservative white rule at the state level: blacks have been moving South on net for about 40 years now. The Atlanta area in Red State Georgia is particularly attractive to black college graduates.

  62. Stefan Schubert says:

    Troy/ Doesn’t seem I can reply to you above so I do it here.

    Fair enough. Yes, that’s a possible mechanism but I think we risk start telling just-so stories here.

    I think that human desire to reduce crime could be a meta-cause that causes different causes of crime reduction, as some people have pointed out above. However, we’re never so much on top of this difficult problem that all relevant causal factors point in one direction. Typically there are side effects of technological development which increase rather than decrease crime.

    • Irrelevant says:

      The normal behavior is you reply to the next level up once the reply depth caps out so that the replies stay threaded. I have no idea what post this is supposed to be in reply to.

  63. NRK says:

    The second person doesn’t have a stronger point in any way, as for a convergence of causes to occur at a given point in time, it isn’t necessary for for each cause to come into play at that exact point. In fact, that is only necessary for the last cause to enter the scheme. This obviously implies that some of the phenomena in question only start having a significant effect on crime once they happen combined with others, which, however, doesn’t strike me as particularly unlikely or uncommon.

  64. Isn’t “how likely are 10 factors to drop crime simultaneously or in succession?” the wrong question? That can be pretty rare, and yet vastly more common than “single factor leads to sustained drop”. Given thousands of factors affecting crime, it doesn’t seem unlikely there would be occasional clustering.

  65. Mr. Breakfast says:

    Every time someone hits the Powerball jackpot, it is because six improbable events occurred simultaneously, 5 at 59:1 and one at 35:1. Yet this happens usually about once a month, and sometimes multiple people win on the same drawing.

    This is not seen as unusual because we know that there are hundreds of millions of trials for each win.

    In the linked article, Jared Diamond’s theory of many favorable geographic factors aligning to favor the establishment of permanent settled agricultural civilization is critisized. But how many bands of nomadic peoples roamed over how many locales for how many generations at approximately that technology level before a few of them “hit the jackpot”?

    Likewise, if the prevalence of crime IS influenced by ten or more factors, then the previous rate it was bopping along at in the middle and late 20th century (4.5% vicimization +-1 or so) WAS ALSO the product of ten or more factors coming together just right and reaching a stable equilibrium trend.

    Am I wrong?

    • Harald K says:

      Jared Diamond doesn’t say that the factors are independent. If you were to build a model out of his theories, you’d put “large east-west landmass” as the most important underlying factor, since it allowed the spread of large-seeded wild grasses, domesticable animals, technology and communicable diseases all along the continent.

  66. Quite Likely says:

    I would think it’s more that there’s a constant churn of different factors making crime go up and down. So if you look at crime in the 1990’s, you’ll see a bunch of these things contributing to a decline, and probably some other factors that worked to increase crime, but were overwhelmed by the declining factors.

    Then you have something like getting lead out of the environment, which can have a long term effect, decreasing crime every year. It seems like just one factor most years, but if it just keeps going, accounting for some portion of the variance every year, pushing crime down, you can get the long term reductions we see.

  67. Psy-Kosh says:

    So, might as well ask here: Did you collude with Eliezer re the timing of this post in particular? There’s some speculation on the relevant hpmor threads on reddit given that this particular post hits a bit too close to home re events in chapter 104.

    Inquiring minds want to know, so I may as well ask: coincidence, or plotting? 🙂

  68. Matthew says:

    Kevin Drum’s posts on lead and on incarceration in response to the Brennan Center report seem worth mentioning.

  69. Timothy Johnson says:

    I don’t find your counterargument convincing. It might seem statistically improbable for ten factors to change at the same time. But I would suggest that that’s only true if you assume that the ten factors given are the only possible factors.

    Suppose that instead of 10 factors we have 100 factors. 10 of these happened to change at around the same time and cause crime to decrease, while the others did not. I’m feeling too lazy to actually do the math, but doesn’t that seem more likely?

    • Unique Identifier says:

      There is barely anything about the counterargument which makes sense. Once you realize that it is really a generalized proof, that either a given variable cannot have multiple inputs, or if it does, many of these cannot change so as to pull in the same direction in a given period of time, it falls very flat.

      Note also that the word ‘sudden’ is used to describe what is really an entire decade, which is not only ample time for coincidences to coincide, but also for complicated feedback mechanisms to matter.

      Other posters have given Moore’s law as an example, where a fairly regular development is actually fueled by a plethora of infinitesimal improvements&inventions.

  70. Unique Identifier says:

    [mis-threaded, redacted]

  71. Titanium Dragon says:

    The lead-crime hypothesis is wrong. Evidence:

    1) It does not explain the upswing in crime in the 1960s. If lead caused said crime, the upswing should have occurred earlier, because massive amounts of lead was used in paint in the 1900-1940s era. We didn’t see an upswing despite lots of lead exposure, and as far as we can tell, lead exposure was high prior to the upswing in crime.

    2) It does not explain the decline in crime.

    3) It does not explain the criminal patterns – blacks commit vastly too many crimes, as do Hispanic immigrants to the US (but not native-born Hispanics, who appear to possess a lower crime rate). People claim higher levels of pollution exposure for them, but historical populations didn’t show same reaction. Very unlikely.

    4) It does not explain China’s extremely low crime rate coupled with their extremely high lead exposure.

    Indeed, the problems with most of the major hypotheses is that they don’t explain the increase in crime; populations with high levels of lead exposure in the 1930s-1950s didn’t commit tons of crime, and abortion and birth control only became widely available after the upswing, so their repression could not have been the cause of the upswing. I am extremely skeptical of any global factor like that which didn’t see a massive change at the time of the upswing. Poverty also can’t be the cause; people were poorer prior to that point in time.

    Moreover, this all assumes that the factors are independent, which is wrong; the factors aren’t independent at all, they were all the result of extremely high crime rates. Indeed, in real life, factors are frequently not independent; you can predict how many people aren’t born in hospitals in some regions by how many swans are there because both are dependent on a third factor, how sparsely populated an area is.

    If we saw ten factors go in the same direction at the same time, then the odds of them all actually being independent factors is pretty low.

    The other problem is that if we assume that all of the factors are having an influence, and only some of them are, we’re probably underestimating the influence of those factors.

    • JK says:

      blacks commit vastly too many crimes, as do Hispanic immigrants to the US (but not native-born Hispanics, who appear to possess a lower crime rate).

      On the contrary, native-born Hispanics have a higher crime rate. The lower crime rates of first-generation immigrants is a general phenomenon that seems to be true of just about all populations.

      • Titanium Dragon says:

        True but misleading; 20% of people jailed in the US are not native-born, but only 15% of the population is not native-born. The children of immigrants are less criminal than the general population, but they’re also less criminal than their parents.

    • Troy says:

      3) It does not explain the criminal patterns – blacks commit vastly too many crimes, as do Hispanic immigrants to the US (but not native-born Hispanics, who appear to possess a lower crime rate). People claim higher levels of pollution exposure for them, but historical populations didn’t show same reaction. Very unlikely.

      Differences between crime rates in different populations might be explained by (say) cultural or genetic factors, while changes in the crime rates over time (both within groups and absolutely) might be explained by lead.

    • thirqual says:

      Using 3) and 4) simultaneously is trying to have your cake and eat it (even if we had not much bigger problems when trying to compare crime statistics between China and the US, or with crime stats in China altogether, especially around the 60s).

    • Steve Sailer says:

      “blacks commit vastly too many crimes, as do Hispanic immigrants to the US (but not native-born Hispanics, who appear to possess a lower crime rate)”

      It’s actually the other way around — Mexican immigrants who arrive as adults try not to attract police attention. It’s their sons growing up on the mean streets of Los Angeles who tended to form showy gangs at young ages.

  72. Robert W says:

    It would be surprising if ten factors flipped from ‘increasing crime’ to ‘decreasing crime’ if there were only ten factors affecting crime in total. But in reality there are dozens, perhaps hundreds, of things that could influence crime rates. Going from 10/10 trends increasing crime to 0/10 trends increasing crime is very unlikely, but it’s not so strange to go from 55/100 factors increasing crime to 45/100 factors increasing crime (with 55/100 now lowering crime).

    But more fundamentally, it’s also surprising for one factor to have a huge effect. So just as it’s odd for ten factors to all push 1 unit in a particular direction, it’s also quite weird for one factor to push 10 units in a particular direction. Huge effects have lower priors.

  73. Ben L says:

    Has anyone already mentioned that one way to explain several things going right is that, as a people, we are constantly looking for ways to reduce crime? What if a lot of things could have or did raise crime slightly, but were drowned out to a net negative trend by intentional attempts at reducing crime / improving well being?

    • Nita says:


      [meta-note: people are now posting comments instead of reading or searching — perhaps it’s time for some technical solution after all]

  74. PGD says:

    People REALLY need to stop trusting kitchen-sink regressions by interested parties as good guides to causation in the real world. In fact I’m astounded that someone like Scott, who is apparently quite numerate and has written some great stuff on the issues that exist in even blinded experiments conducted in a controlled setting, would have this level of credulity when faced with a bunch of observational correlations that have the ‘multiple regression’ fairy dust sprinkled on top. (How about it, Scott, are you an observational regression fan?)

    What this analysis tells you is that year-to-year changes in incarceration rates at the state level are not highly correlated with year-to-year changes in crime rates at the state level, once you control for the overall mean of state crime rates over a long period and the overall mean of crime rates nationwide in the year (a state fixed effect and a year fixed effect). There is a one-year lag in there but that doesn’t move you very far from year-to-year.

    There could be lots of reasons for that to be true even if incarceration rates are highly linked to crime reduction. These include 1) the most dangerous criminals stay locked up for a lot longer than a year, 2) criminals move from state to state a lot, 3) there are long-lasting deterrent effects of increases in incarceration on people who don’t get incarcerated that take years to work through the population, 4) get-tough increases in incarceration are correlated to pre-existing growth in crime, etc. I’m not saying that these correlations are uninteresting or have no evidentiary weight whatsoever, but they are far from dispositive or decisive.

    You can add to this that the Brennan Center is highly liberal and is motivated to find that incarceration does not play the critical role in crime control. That motivation doesn’t have to be very conscious to bias things like regression results which are extremely dependent on which of a vast number of specifications you choose. It’s true that other academics who have looked at this have a hard time finding a very strong short-term correlation between incarceration and crime (but quite a number find a higher correlation than the Brennan Center does — I’ve seen figures like 25-35% of the 90s drop in violent crime caused by incarceration). But I think there’s just as good an argument that this is because the relationship between crime rates and incarceration is a long-term issue and the ‘correct’ lag structure to put in an analysis is unknown.

    • Nita says:

      I’m astounded that someone like Scott, who is apparently quite numerate

      Actually, Scott often claims that he’s bad at mathematics. But he isn’t planning to do anything about it, since his Special Talent is writing.

  75. weareastrangemonkey says:

    Scott, I believe I can give a bit of insight on your question – I think about these kinds of problems a fair bit.

    To begin, there is no way of evaluating whether a multi-factor explanation is more likely than a single factor explanation without specific knowledge about the likely effects of particular factors. This is because in order to calculate the likelihood of an explanation we need priors on the joint distribution of the effects of all the potential factors. But we cannot, without looking at the data of the specific problem, put any priors on the joint distribution of the effect sizes of the potential factors.

    Let’s consider the specific example you use “Is it more likely that crime is all explained by lead levels or by multiple factors?”. The answer depends on what we think the distribution of the effect sizes for the different factors. What is our prior distribution of the effect of lead on crime? What is the distribution of the effect of changes in policing on crime? What is the distribution of the effect of changes in abortion laws on crime?

    Note that we are talking about our prior distributions on the effects of the factors and not the distributions of the factors. This is because we know how the different factors have changed because we can observe them (at least lets assume so); the problem is that we do not know whether any of these factors matter a lot or a little. (If we don’t know the values of the particular factors then instead we instead need priors over the distribution of the product of a factor and its effect size.)

    We can choose fat-tailed distributions for the effect sizes (extreme effects likely) and it will be more likely for a change in a single factor to explain most of the change. However, we can also choose thin tailed distributions (extreme effects unlikely) and then large changes are likely to be explained by many factors. There is really no reason – that I am aware of – for choosing fat or thin tails in general. I am not even sure it even makes sense from a philosophical perspective to talk about an effect coming from a distribution except in that you assign a variety of subjective probabilities to different effect sizes.

    Currently I think there is no rigorous answer to the general question of whether some arbitrary system or variation in outcome is more likely to be explained by k variables or l variables. The best we can do is when we have a specific outcome with data on potential explanatory variables. In this case we can see how well some set of explanatory variables explain variation in that outcome conditional on some set of assumptions. If we keep the set of assumptions the same then we can also rank the likelihood of the different explanations. But these rankings are dependent on our assumptions; we can upset the rankings by changing the assumptions. Of course, we can try to use “plausible assumptions” but there is no rigorous (in the mathematical sense) way of choosing one assumption set over another for a large space of assumption sets.

    We have little choice then at this point to apply judgement in our choice of assumption set. My judgement tells me it is better to apply judgement case by case. It also tells me that we should have some set of general assumptions that we do not change from case to case because that will just lead to politics. My judgement also tells me that it is not clear what “judgement” means. But we have little choice other than to apply it, lest statistics leaves us becalmed in deep epistemological waters.

  76. Dishwasher says:

    Yeah but the crime wave was big news in the 1980’s, so people started working on many ways to reduce crime, perhaps 10 was a small minority of the things people started to try.

  77. Eric Rall says:

    One set of theories I’ve found intriguing is Harcourt’s idea of linking the crime rate to institutionalization rate (prison population + mental hospital committed population) rather than just the incarceration rate.

    Superficially, Harcourt seems to make a pretty strong case. The institutionalization rate correlates with both the rising and falling edges of the American crime rate (the big increase in the 60s corresponds to reforms in treatment of mental illness that lead to a much lower commitment rate, while the big decrease in the 90s corresponds to changes in policing and sentencing that lead to more people being imprisoned for longer), and it seems to hold up in panel studies comparing trends in crime vs institutionalization rate across US states. I can’t find data for whether the correlation holds for countries other than the United States.

    • Irrelevant says:

      Well. The graphs are certainly impressive, though I’d like this not to be true, since I’m not a fan of involuntary commitment either.

  78. Dishwasher says:

    They forgot to mention ER’s are better at saving gun shot victims.

    • Steve Sailer says:


      On the other hand, guns are more common and somewhat easier to shoot a lot of bullets at once.

      • FacelessCraven says:

        firearm lethality hasn’t increased to any appreciable degree since, say, the early 1900s. From what I’ve read, improvements in emergency care, a large-scale transition to handguns as the firearm of choice for most of society, and improved firearms safety training are largely responsible for the decline in firearms fatalities. [EDIT] – I’m pretty sue this has pretty much zero impact on the trends in the murder rate we’re actually discussing, though.

  79. Also Pareto’s principle. Given that factors tend to be distributed in a power law, it’s most likely that 1 or 2 factors had a much higher impact than the rest.

    • Unique Identifier says:

      Does this principle survive contact with Moore’s law? [I am tempted to say it doesn’t instead of asking the rhetorical question, but it does depend a bit on how you choose to group things into factors.]

      • I think I am confused about your objection. Firstly, exponential technological growth seems irrelevant here, or exponential trends of any type. Secondly, if you mean whether Pareto’s principle is valid in the tech world too, I think Moore’s Law is actually a driving principle behind temporal continuity of Pareto’s: because things grow exponentially, instead of some kind of diminishing marginal return regime where competition equalizes, you have the top (of whatever category) still growing exponentially fast and preserving the power law.

        • Unique Identifier says:

          I’m trying to use Moore’s law as a test case, because we know very much about how microprocessors developed, while crime is rather opaque.

          Moore’s law is a case of roughly exponential growth over four decades. The crime decline we are looking at is decays somewhat exponentially over one decade. I don’t see anything particular about graphs themselves, that disqualify one but no the other for the purposes of the Pareto principle.

          Now, we know that there hasn’t been a single or a few driving factors behind Moore’s law. You could, of course, look at narrow time slices, say one or two (five?) years at a time, and the principle might hold locally.

          With this in mind, it seems that the best we can get, applying the Pareto principle to crime, is that ‘locally, a few factors dominate, but over the entire span of the decline, there might have been four or even ten important factors’. Which seems perfectly useless.

          • What kind of driving factors are you looking for? The idea would be to look at some time slice, see who has made the most progress in that period, and see if it follows a power-law distribution and thus was dominated by a few large players (think IBM, Bell Labs). I don’t know the history well but this seems reasonable to me.

          • Unique Identifier says:

            Personally, I would be interested in individual, technological developments, such as listed here, not the names of companies:

            Similarly, I think ‘the government’ is a very unhelpful answer, to the question what made crime decline. Individual policies, new technology changing society etcetera, is more of what I’m interested in.

  80. Pingback: Ensidiga multifaktorförklaringar | DN Debatt-betyg

  81. Douglas Knight says:

    What bothers me most about crime trends is that the American homicide rate went up so smoothly. It doubled from 5 to 10 from 1965 to 1975 perfectly linearly. But the way down was very bumpy. It is hard to imagine anything other than simple demographic trends that would explain the rise. But if the cause of the fall is the same as the cause of the rise, it ought to be equally smooth.

  82. Douglas Knight says:

    You ask about the meta issue of whether you should believe multifactorial explanations, but what about the meta issue of where do multifactorial explanations come from? Do they come from people answered the previous question affirmatively, or do they come from people trying to be inclusive? Do they come from people paid by the word?

    People already have plenty of reasons to reject the individual hypotheses. Giving them a new one to reject to joint hypothesis on the grounds that it is multifactorial fails to address the basic problem that people are failing to reject at all on the grounds that they already have.

  83. Douglas Knight says:

    Some object level responses.

    One thing that most of these analyses do that really bothers me is that they propose an analysis of the crime decline, but not an analysis of the previous rise. It is hard to imagine any useful analysis coming out such historical blindness. I don’t expect much of Vox, but the Brennan Center report acknowledges the history, and then discards it. Vox itself says that #3 broken-windows policing is a myopic theory, but fails to note that all of its analysis is myopic. Also, international comparisons are very important. Canada and much of Europe saw homicide double 1965-1980 and then return to baseline. America is unique in having an extra peak around 1990, the crack epidemic.

    Crack (#12) is not an explanation, but the international and regional comparison shows that it is exogenous. Surely the crack epidemic has a separate explanation, layered on top of the reasons common across countries for the doubling and halving. Hypotheses that try to explain both at once should be looked upon with suspicion. One might make the same argument to ignore the American 1980 cocaine peak. By the metric of homicides, it was just as high and just as abrupt as the crack peak. But it was a uniform crime wave, while the the crack epidemic was concentrated among the young. The rise in homicide by and of youth (say, 15-25) was far faster in the crack epidemic than in the cocaine wars. Anyhow, it hardly matters whether one includes the 1980 peak because that just pulls the peak back five years, which is well within the variation of peak years in Europe.

    The homicide rate today is about the same as it was in 1960-1965. The obvious guess is that the same thing caused the rise and fall. One should be suspicious of explanations that ignore this symmetry. That applies to almost all of Vox’s items, but 1, 2, 3, 7, 8, 9, 11, and 15 are particularly asymmetric. I think #5 and 6, too. #6, the economy, is what the Brennan Center gives the most weight.

    Steve Sailer turns the vague #4 “improved policing” into a symmetric, but still quite vague hypothesis, that the police were hamstrung or demoralized by the Warren Court reforms, but eventually adapted to them, or were just brought out of their funk by public opinion. Maybe even “broken-windows policing” is a return to the old ways. Warren only had direct power in America, but there are vague claims that the general pattern spread throughout Europe.

    But I know that the explanation cannot be perfectly symmetric. The homicide death rates for 15-25 are maybe the same as in 1968 (much higher than 1965), but for 25-35 and 35-45, they are maybe 2/3 the 1968 rates. So the overall average is like 1965, but the distribution is different. That’s pretty weird and I don’t know what it means. Probably we need at least two causes, one increasing crime among the youth (gangs?) and one decreasing overall. Relevant graphs, 1 2 3, discussed above.

    And those demographic trends argue against many explanations. In particular, the lead #16 and abortion #15 theories prediction exactly the opposite of what happened, so I thoroughly reject them.

    I think that #10 Alcohol consumption does have the right symmetry, but the changes are small on the historical scale.

    If all you want to do is explain the crack epidemic, #12 crack and #13 “gangs have gotten less violent” are symmetric. They are definitely the answer, in some sense, but they aren’t much of an answer. If you are interested in a larger time frame, crack is irrelevant, but gangs might be. It would be useful to know if gang violence has declined more or less than other violence from my baseline of 1980. But even if I did know, that would only be a hint, not an answer. Compared to 1960, surely gangs are more violent today. So I endorse the opposite of #13 for the long term.

    Age demographics #14, is the first theory I whole-heartedly endorse. It appears to me to explain 20% of the rise and 20% of the fall. The Brennan Center report calls it 5% during the 90s and nothing during the 00s. Partly that is because of the demographics of the crack epidemic, but even considering that, I do not see how they can get such a small number. It would be useful to do a comparison with Germany, which had a baby boom much later than the winners.

    Going back to #1, incarceration, I reject this theory because it is asymmetric, doesn’t explain any other country, and doesn’t really match the timing. But the one thing going for it is that it that the raw incapacitation version of the theory does match the changing demographics of crime. (Indeed, it is the only theory that does, except for the opposite of #13 gangs). People in prison are systematically older than prisoners, explaining why crime has fallen so much more among the older than the younger. More likely, other policing reforms cause more and better arrests that train people out of crime. But such theories are vague and hard to test.