Steven Levitt made his Freakonomics career with his (and John Donohue’s) abortion and crime study. Now, a few years later, Theodore J. Joyce has a retrospective of the issue in his Abortion and Crime: A Review. Abstract:
Ten years have passed since John Donohue and Steven Levitt initially proposed that legalized abortion played a major role in the dramatic decline in crime during the 1990s. Criminologists largely dismiss the association because simple plots of age-specific crime rates are inconsistent with a large cohort affect following the legalization of abortion. Economists, on the other hand, have corrected mistakes in the original analyses, added new data, offered alternative tests and tried to replicate the association in other countries. Donohue and Levitt have responded to each challenge with more data and additional regressions. Making sense of the dueling econometrics has proven difficult for even the most seasoned empiricists. In this paper I review the evidence. I argue that the most straightforward test given available data involves age-specific arrest and homicide rates regressed on lagged abortion rates in the 1970s or indicators of abortion legalization in 1970 and 1973. Such models provide little support for the Donohue and Levitt hypothesis in either the US or the United Kingdom.
Joyce is a critic of the Levitt-Donohue hypothesis, and so he is not the fairest biographer of the situation, but does include a good number of very insightful and damaging critiques of Levitt-Donohue. Joyce begins by describing the problems in showing causality in this situation. Even if you take their theory as correct (that unwanted children tend to cause more crime and abortion tends to reduce the number of unwanted children), there are complications in regressing against abortion rate as a substitute measure for “unwantedness” of a child. Joyce points out that Levitt-Donohue assume that all of the changes in abortion rates are demand shifts (i.e., that individuals are acting upon their wishes to increase the number of abortions they want) as opposed to supply shifts (i.e., that getting abortions is significantly easier over time).
Joyce also notes several critiques of the theory, starting with his own “identification and specification” argument. Joyce, in previous articles, points out the above supply versus demand shift problem, as well as the relatively sparse data used by Donohue and Levitt. A large number of their data points have 0 abortions, so it is difficult to draw out the effects from a change in abortion numbers as opposed to any other background noise. Finally, using log arrest rates or log homicide rates as opposed to the log arrest or homicide counts results in statistical insignificance. This might be due in part to how Donohue and Levitt did not adjust their standard errors for positive serial correlation. [The two points are Joyce’s; linking them together is my own thought.]
As for other problems, Foote and Goetz pointed out that Donohue and Levitt accidentally left out state-level fixed effects from their regression, and the latter duo admitted that this had a significant effect on their results and decided to revise their models, with the result being that their findings were even stronger than previously. These new models still used log arrest counts rather than log arrest rates; switching to rates made the models insignificant.
Lott and Whitley argued that legalized abortion actually increases rather than decreases homicide rates. It’s an interesting idea, but Joyce points out that the “too many empty cells” problem is even worse in Lott-Whitley than in Donohue-Levitt.
On the empirical front, Joyce notes that there is no evidence of a “cohort effect,” as you would expect if abortion were the driver behind less crime. You would see a staggered decline in crime by age, where the rate for 16-year-olds would drop and then 17-year-olds and then 18-year-olds and so on through the chain. In reality, you can see the violent crime rate increase and decrease as a whole, but you do not see a cohort effect. Meanwhile, studies of other countries seem to corroborate with the anti-DL thesis. In the UK, violent crime has steadily increased since 1984 despite the abortion rate also increasing. In Canada, an article which Levitt and Dubner praised turned out to have faulty assumptions and was unable to explain much. In Romania, where abortion was made illegal in 1967, there was no statistically significant increase in crime that could be tied to the abortion rate. Finally, coming back to the US, we can see in the 10 years after Donohue and Levitt’s initial study period that murder rates, which had fallen during the 1990s, have leveled off, rather than continuing to decline. Given that there’s no reason to expect the rate of “unwanted” children to have increased, we should have seen continual falls (according to Donohue and Levitt) through roughly 2020.