Tuesday, December 25, 2012
Is Gasoline the Root of Crime?
I recall reading about a model linking exposure to lead in infancy with property and violent crime, in 2007 and 2008 (cf my saved links at delicious ). Recently (late November 2012) I saw a group of summary charts from the research paper to illustrate highlights of two recent columns on the subject (Investigating The Crime Drop--The Dish) with a link to the research paper by Rick Nevin, "Understanding international crime trends: the legacy of preschool lead exposure."
I don't recall having looked at the charts or the article itself in 2007 or 2008; looking at these charts now, there seemed to me to be a big problem with France: blood lead levels have hardly diminished, are still as high as U.S. and U.K. levels were at their peaks. I live in France, raised children in France, and so worry about the risk for my children and grandchildren in France. Yet, I also find it improbable that the levels would be so high in a country where over half the cars are diesels (and fuel costs so much). Incidentally, Australian blood lead levels appear also hardly diminished: still over 20 (µg/dL). So, I wanted to review the article and downloaded it about three weeks ago, then forgot about it until last week when someone made reference to the New York Times and Wired write-ups of the 2007 study in the course of a discussion of why American incivility rose so much from 1960 or before until about 1977/8. That question came up, incidentally, contemplating some charts worth studying published by Kieran Healy (cf America is a Violent Country and Assault Death Rates in America: Some Follow-Up ).
A first assessment of the blood lead level blame research reveals two problems. The first is that there is not actually any convincing relationship, consistent across countries. The second is that the model specification does not make sense, and probably should not work; that does leave the possibility that a better specification would reveal a convincing relationship, but that is only a possibility. The problem with the specification is that the "damage done" by lead is accumulated, young offenders persist and may recidivate; the author does try to compensate for this by looking at age group specific crime rates, but without establishing a convincing general model. Also, and this may not be just a detail, very little actual blood lead level data is used; rather, the "lead levels" series are synthesized from gasoline consumption data calibrated to match a few samples of blood lead levels at different times in each country. So what is actually being studied is the lagged relationship between burning leaded fuel and observing crime.
The study was quite ambitious, and involved collection and preparation of a lot of data. A general problem is that available crime statistics vary from country to country, so different set of countries and different aggregate crime indices are used as well; that leads to the "try enough models and you'll find a decent fit" phenomenon. Looking at various sets of countries and measures and reporting the best results does not inspire confidence that the relation is real and not an artefact of the method.
Two of the model relationships will illustrate the first problem, which is that consistent relationships are not observed across countries. For one thing, not all countries have had clear peaks and declines in both the independent variable or 'cause' (preschool blood lead level) and the dependent variable or 'effect' (some measure of crime), so the model is mostly correlating pairs of ramps in the two variables.
One model relationship is charted in Figure 7. "Preschool blood lead versus narrowly defined index crime with a 19 year lag. U.S.A., France, West Germany, Italy, Australia." The U.S.A. is the only one of the five countries in which the dependent variable (index crime) actually decreases after rising. For France, the rise to plateau lag is better modelled at 34 years than at 19, like a hockey stick and a boomerang; index crime rises and plateaus with a couple of subsequent dips. West German index crime similarly rises and plateaus much too early--before blood lead levels stop rising, but also is very high--the highest crime index--whereas the lead level is much lower. The Italian series have the same form as the French and West German ones, but a better timing match. The Australian index crime series does not peak and diminish (or even plateau) after the blood lead level declines. There is no consistent relationship across countries.
Concentrating next on the possible relation with assault, let me describe the tableau in Figure 11. Six countries, three reporting "aggravated assault" and three reporting "violent and sexual assault." In all countries, blood lead levels eventually decline at least modestly, but four of six finish with levels still much higher than in the early years (1930s and 1940s). In only one country does the model look "good" and one other, "plausible": the U.S. and Canada, respectively. (The decline in the dependent variable in Canada is weaker than predicted). In Britain, the aggravated assault rate continues to rise (pronouncedly) well past the reduction in blood lead levels, but the reduction may have been insufficient (more a plateau). The Australian data series is very short, ending before showing any decline in violent and sexual assault. New Zealand's lead levels are half Australia's, the violent and sexual assault rate is much higher than predicted, with a step-change near the middle of the series. Finally, France continues to have high blood lead levels (this is what caught my attention three weeks ago), yet the aggravated assault rate--although rising unabated--is very much below the predicted level (less than half).
The author seems aware of the second problem, and does partially address it in comments and the age group analyses. But the treatment is not systematic, and the linear cause-effect relationship implies that the negative impact of lead on the individual only manifests itself in crime briefly, not recurrently.
In conclusion, I certainly do not deny that lead is bad for health and healthy development. The paper is worth a read if only for the background review and citations of previous studies of these health issues. The project was ambitious, an impressive piece of work.
Yet, as an experienced time series analyst, I am not at all convinced by the "demonstration" that preschool blood lead levels explain the rise and subsequent decline in violent crime rates twenty years later. But it could be, as the author suggests (I think), that the chemically-induced criminality does not go away after it begins to manifest at age 23, so the time to observe a decline rather than a plateau (or flatter ramp) after reducing lead levels could be very long (the rest of a lifetime). Reducing contact with lead is not a bad idea, but the decline in crime rates in the U.S.A. is probably due to something else and the increasing crime rates observed in all the countries included are too monotonic (or hockey-stick) to be reliably attributed to a single monotonic (or hockey-stick) factor.