April 26, 2014

"Does Abortion Prevent Crime?" Steven D. Levitt's response to Steve Sailer in 1999 "Slate" debate

Below is the third of four parts of a 1999 debate in Slate between U. of Chicago economist Steven D. Levitt, co-author of the 2005 bestseller Freakonomics: A Rogue Economist Explores the Hidden Side of Everything, and myself, Steve Sailer. We discussed Levitt's most celebrated theory: Did the legalization of abortion in 1969-1973 cause the crime rate to fall? 

I've decided to host this debate on my website because it is of some modest degree of historical importance as the first airing of one of the longer-running social science controversies of the 21st Century, and because Slate deleted our names from their posting of it during a website reorganization. Several years ago, Slate promised to restore our names, but hasn't done so yet. The absence of our names on Slate has made it hard for interested readers to find this using search engines.


E-MAIL DEBATES OF NEWSWORTHY TOPICS.
AUG. 24 1999 9:30 PM

Does Abortion Prevent Crime?


  1. The arrival of crack led to large increases in crime rates between 1985 and the early '90s, particularly for inner-city African-American youths.
  2. The fall of the crack epidemic left many of the bad apples of this cohort dead, imprisoned, or scared straight. Consequently, not only did crime fall back to its original pre-crack level, but actually dropped even further in a "overshoot" effect.
  3. States that had high abortion rates in the '70s were hit harder by the crack epidemic, thus any link between falling crime in the '90s and abortion rates in the '70s is spurious.
If either assumption 1 or 2 is true, then the crack epidemic can explain some of the rise and fall in crime in the '80s and '90s. In order for your crack hypothesis to undermine the "abortion reduces crime" theory, however, all three assumptions must hold true.

So, let's look at the assumptions one by one and see how they fare.
  1. Did the arrival of crack lead to rising youth crime? Yes. No argument from me here.
  2. Did the decline in crack lead to a "boomerang" effect in which crime actually fell by more than it had risen with the arrival of crack? Unfortunately for your story, the empirical evidence overwhelmingly rejects this claim. Using specifications similar to those in our paper, we find that the states with the biggest increases in murder over the rising crack years (1985-91) did see murder rates fall faster between 1991 and 1997. But for every 10 percent that murder rose between 1985 and 1991, it fell by only 2.6 percent between 1991 and 1997. For your story to explain the decline in crime that we attribute to legalized abortion, this estimate would have to be about five times bigger. Moreover, for violent crime and property crime, increases in these crimes over the period 1985-91 are actually associated with increases in the period 1991-97 as well. In other words, for crimes other than murder, the impact of crack is not even in the right direction for your story.
  3. Were high-abortion-rate states in the '70s hit harder by the crack epidemic in the '90s? Given the preceding paragraph, this is a moot point, because all three assumptions must be true to undermine the abortion story, but let's look anyway. A reasonable proxy for how hard the crack epidemic hit a state is the rise in crime in that state over the period 1985-91. Your theory requires a large positive correlation between abortion rates in a state in the '70s and the rise in crime in that state between 1985 and 1991. In fact the actual correlations, depending on the crime category, range between -.32 and +.09 Thus, the claim that high-abortion states are the same states that were hit hardest by crack is not true empirically. While some states with high abortion rates did have a lot of crack (e.g., New York and D.C.), Vermont, Kansas, Hawaii, Massachusetts, and Washington were among the 10 states with the highest abortion rates in the '70s. These were not exactly the epicenters of the crack epidemic.
So, what is the final tally? Two of the key assumptions underlying your alternative hypothesis appear to be false: The retreat of crack has not led to an "overshoot" in crime, causing it to be lower than 1985, and even if it had, the states with high abortion rates in the '70s do not appear to be affected particularly strongly by the crack epidemic. Moreover, when we re-run our analysis controlling for both changes in crime rates from 1985 to 1991 and the level of crime in 1991, the abortion variable comes in just as strongly as in our original analysis.

Crack clearly has affected crime over the last decade, but it cannot explain away our results with respect to legalized abortion.

The best test of any theory is its predictive value. The abortion theory predicts that crime will continue to fall slowly for the next 10 to 15 years. Also, the declines in crime should continue to be greater in high-abortion states than in low-abortion states. What do you predict based on your crack theory? If you are willing to wait 10 years, perhaps we can resolve this debate.
To read more on this topic, see Steve Sailer's 2005 posting after The Economist and the Wall Street Journal revealed that an attempted replication of Levitt's state-level analysis by Boston Fed economists Christopher Foote and Christopher Goetz discovered that Levitt had made a fatal error in his computer code, which explains why Levitt's state-level findings didn't match my national-level analysis in 1999.

Complete debate: Part 1 (Levitt);   Part 2 (Sailer);   Part 3 (Levitt);   Part 4 (Sailer)
   

1 comment:

  1. I think small effects for a multivariable problem should be treated with suspicion. When he makes the comment about a percent per year, than that worries me. If we already know that social factors (e.g. crack) were driving much bigger changes, how can we really adequately factor them out and hone in on the small variable. For one thing, crack is not the only possible social dynamic. If there are some pretty big variable impacts going on, I worry that a small effect (calculated as caused by a regression variable) might just be some change in the larger variables or some other unspecified variable.

    And then something like this (crime rate) could be hugely impacted by social dynamics. Lightbulb effect, religion, general mores, etc. etc.

    ReplyDelete

Comments are moderated, at whim.