American Enterprise Institute

March 28, 2006

[Edited transcript from audio tapes]

12:45 p.m.	Registration
1:00	Introduction:	Jonathan Klick, AEI and Florida State University
1:10		Panel I
	Presenters:	Chris Foote, Boston Federal Reserve
		Ted Joyce, Baruch College, CUNY
		Leo Kahane, California State University
2:10	Discussants:	Phillip Levine, Wellesley College
		Steve Sailer, The American Conservative
	Moderator:	Jonathan Klick, AEI and Florida State University
3:00		Panel II
	Presenters:	John Donohue, Yale Law School
		John R. Lott Jr., AEI
3:40	Discussants:	David Paton, Nottingham University Business School
		Florenz Plassman, SUNY-Binghamton
	Moderator:	Ted Frank, AEI
5:00	Adjournment

Proceedings:

JONATHAN KLICK:  Thank you all for coming today to the American Enterprise Institute.  Today we’re having a four-hour panel on “Abortion Legalization and Crime Rates: Is There a Relationship?”  This panel is going to be broken up into two sections, but the whole panel is basically going to be discussing a provocative bit of research that’s been evolving over the last number of years.  I think I first came in contact with the Donohue and Levitt paper in my first year in grad school down the road at University of Maryland.  I think Levitt was presenting it there.  That was probably back in 1998, I think.  Since then this paper and this research has taken on a life of its own. 

My friend, Paul Rubin, who is also a friend of AEI’s, has a nice point about this paper.  He says, “It’s easy for a guy like me to write a paper that irritates the left, and guys like Ian Ayres and Ted Eisenberg can write papers that irritate the right, but Donohue and Levitt actually irritated everyone.”  That’s actually saying something.  I think that’s pretty apt regarding all of this research, quite frankly.

Why has this research taken on such a contentious tone and why has it generated so much interest beyond a bunch of economists?  Well, it’s a combination of an important policy issue – that is, crime, and specifically why did crime begin to fall in the 1990s? – and a contentious normative issue – that is, basically, abortion access.
 
For those of you who just walked in off the street for the free AEI cookies, you might want to know, “What did Donohue and Levitt actually say?”  They hypothesized that legalizing abortion would affect crime eventually through two channels.  The first channel is a simple cohort-size effect – that is, if there are fewer children being born, then when those children or when that cohort of children reach their prime crime-committing years, there will just be fewer criminals.
 
But then there’s perhaps another, more sophisticated effect, or more subtle effect, when you note that abortion isn’t performed randomly.  That is, it’s not as though a certain group of children who otherwise would have been born are not born, it’s that specific children are more likely not to be born, and who are those children?  Donohue and Levitt hypothesized, on the basis of some existing empirical work, that abortions were going to be more common for children who were in some sense unwanted. 

This is almost definitional.  But what does unwantedness mean?  Well, unwantedness could mean a number of things, but specifically it may mean that the parents of these children (or specifically the mothers of these children) might not have the means or the disposition to take care of these children.  So had they not been aborted, fewer resources would have been invested in these children.  They would have lower human capital.  They would have a harder time growing up and things like this.  All of this very likely would lead to a higher propensity to commit criminal acts eventually. So if we legalize abortion and these children are less likely to be born, we should eventually see some negative effect on the crime rate.

This hypothesis or this idea is not entirely new, although it probably was new to economists, but it’s been pointed out to me that early on in the abortion legalization debates and some of the discussions about abortion, there were a number of people who either implicitly or actually – in the case of Canada – explicitly embraced this reasoning as a reason to legalize abortion or expand access to abortion.  But it’s interesting that Donohue and Levitt really were the first ones to sort of formalize this idea and actually attempt to see if the data bear the conclusions out.

So what in fact did they find when they looked at the empirical evidence?  Their sort of punch-line result was that abortion legalization effectively accounted for about half of the decrease of crime that we witnessed in the early 1990s.  How did they tie this to the abortion legalization?  They have a number of interesting results that just sort of in a nutshell come down to the following issues.

It looks as though the states where abortion was legalized earlier – specifically a handful of states legalized in 1970 – the crime decrease seemed to occur in these sets of states first, then the rest of the states that only legalized in 1973 with Roe v. Wade, who only saw their crime decrease come about a few years later.
 
They also found that the decrease in crime was bigger in states that had higher abortion rates.  So if you go back 18-20 years, 25 years, you find that states that had higher abortion rates actually witnessed or observed more of a crime decrease subsequently. 

In the course of their research and in writing up their results, they make the point that this effect appears to be more important than lots of the other things that people advance as being the cause of the crime decrease in the 1990s, specifically more important than putting more cops on the street, more important than having harsher jail terms, more important than other demographic effects, more important than novel policing strategies – the ‘broken window’ stuff and the Rudy Giuliani stuff in New York – and more important than a good economy.  For a number of these issues, they didn’t say that these things were unimportant, just that in relative terms they were less important than this abortion effect on crime. 

Perhaps most provocatively they argued that this effect hasn’t played its way out entirely, that we should expect because of this effect further crime decreases as years go on for a number of years into the future.

Almost immediately this paper generated interest and criticism – interest outside of the normal channels.  Before this paper even hit the Quarterly Journal of Economics, one of the most prestigious journals in the field, a number of editorials and columns were written about this paper.  I think George Will had a column on this paper.  There were numerous editorials in magazines and newspapers on every side of the political spectrum.  The Washington Post had some stuff on it, the National Review had some stuff on it, as well as lots of others. 

From almost the beginning, as I said, you could sort of feel that both sides were irritated or at least uncomfortable with the results as they came out.  The right, you could have the sense implicitly or it was even made explicit in the handful of folks who commented on this paper, just the fact that there could be these sort of enormous positive benefits of the legalization of abortion probably didn’t sit so well with folks who were normatively predisposed to disliking abortion.  But even on the left there was some discomfort.  It’s not exactly clear where this discomfort comes from, but one could imagine that there are large racial implications of both the hypothesis and the results that Donohue and Levitt laid out.  So one could imagine that these sets of results didn’t sit so well with the left because of this.

But more importantly, and what we’re focusing on here today, are the academic critiques.  We’re lucky to have, quite frankly, essentially everyone who has mustered evidence either suggesting that the Donohue and Levitt results were either wrong or perhaps not quite as clear-cut as Donohue and Levitt had suggested.  We’ve got folks on this panel who have written a number of papers making these criticisms and these critiques and we’re going to hear about quite a few of them today. 

In general terms, what are the criticisms?  Well, a big criticism – or maybe criticism is too harsh a word – but one big question mark is, how is it that one can identify an effect that’s effectively lagged for 20-25 years?  It’s hard enough sometimes – folks who do work in this area – to find sort of immediate effects, much less find something that has a 20 or 25-year lag to it.  Seems as though it’s a difficult task to set before yourself.  Particularly when you recognize there are lots of intervening effects that it’s not quite exactly clear how to control for. 

So the big one you’ll hear, I suspect, a lot about today is the effect of crack on crime rates.  Crack in the time period just before the Donohue and Levitt-identified crime decrease, crack had a huge effect on urban areas, and even more than urban areas.  It looks as though this effect wasn’t uniform across the country.  So the question of how one controls for this crack effect, given that we don’t have particularly good data on crack use and things like that, that’s a very important question here.

There are other, perhaps less sexy problems.  One may be the fact that you’re identifying an effect based on, again, a 25-year lag, and yet people may well be moving during this time period.  So the fact that I happen to be born or have been aborted in a given state doesn’t mean I would have been in that state during my prime crime years. 

On a more pedestrian note, there are fairly large data problems in this area.  We don’t have great abortion data.  We certainly don’t have very good abortion data for the pre-legalization term.  That is, how many abortions were going on before legalization?  That turns out to be a pretty important question to answer in the course of looking at this work.  Then there’s a host of technical, econometric issues that get economists excited but we’ll try to minimize those, I think, to a large extent today.

Then there are also some potential theoretical problems.  The Donohue and Levitt story as told seems very persuasive and convincing until you start thinking of some other effects that may have been going on in this time period.  I’ve done some work suggesting that when abortion was legalized or when access to abortion increases, individuals actually engage in more risky sex.  So it may well be the case that while there is this selective abortion effect getting rid of unwanted children, it may be the case that the number of potential unwanted children grows when abortion is legalized because people have more of an incentive to engage in risky sex or less of an incentive to be careful to avoid pregnancy.  So that may cause some important problems here.

It’s interesting, in terms of Donohue and Levitt’s responses to all of these questions – and  you’ll hear from John Donohue today in terms of his responses to what is said here today –  but on the positive side, Donohue and Levitt have some support from folks who have looked at this potential crime and abortion link in other countries.  Specifically, some folks have looked at the effect of abortion legalization in Romania.  There seems to be some evidence that there is a negative effect on crime of abortion legalization in Romanian data.  Same kind of thing in Canada.  Then there are one or two other papers looking at U.S. data that corroborate the Donohue and Levitt results.

Also, in each round of this argument among Donohue and Levitt versus the folks you see in front of you, in each round of it, Donohue and Levitt have come up with strong defenses of their original results, in fact, each time claiming that in the next iteration the results appear to be stronger than they originally thought.  I’m sure you’ll hear some of that today.

So not to delay this much more, it is interesting to note, I saw on a blog a couple of days ago that was talking about this panel, it predicted that “blood will be drawn.”  Now, I can’t promise that.  Can’t promise you anything that exciting.  We are economists, after all.  We’ll get pretty excited but I don’t know if it will reach that level.  But I do promise you that since we have essentially everyone who’s written with evidence against Donohue and Levitt and we have John Donohue here himself, we’re going to have an interesting discussion.

At the end of the day, though, you might wonder: does this really matter?  This is always an important question that we economists unfortunately don’t ask often enough about our research.  Does this really matter?  Yeah, in fact, I think it does.  Donohue and Levitt, in some of their early iterations on the paper, pointed out that the abortion debate is often framed in normative terms and things like that.  But it’s interesting.  Even though abortion is legal, states can do lots of things to improve or worsen access to abortion.  So there are important policy issues here. 

If you look at the survey data on the normative question, people are pretty much stuck. Either you’re for abortion or against abortion.  So when states go to make these policy decisions, the normative debate probably isn’t going to push them very far.  But it may well be that these policy questions actually do have important effects on questions like, should we have Medicaid funding for abortions?  Should we have parental involvement laws for teen abortions?  Things like this.  In fact, with the new composition of the Court, who knows? This may be more than just an academic question, if the question in general gets thrown back to the states on abortion legality in general.

So these questions, I think, are more than just academic debates.  I think it’s important to come to some sort of understanding of the policy implications and the practical implications of changing abortion access.

Now let me turn it over to the experts.  First we’re going to hear from Chris Foote from the Boston Federal Reserve.

Panel I

CHRISTOPHER FOOTE:  Thanks very much for having me.  I want to point out that what I’m going to present to you today is joint work with my co-author Chris Goetz, also from the Boston Fed.  To start with, there are some important qualifiers, which is that nothing in this seminar represents the official views of the Federal Reserve Board, Federal Reserve Bank of Boston, or anybody whose first name is not Christopher.  I also want to point out that while we are critical of Donohue and Levitt’s 2001 paper in the QJE and the 2006 paper, the response to some of our earlier stuff, I want to acknowledge now how helpful they’ve been to us throughout the process of producing our paper. 

I also want to acknowledge the sensitivity of the topic.  To most people, abortion and crime are not statistical issues but moral issues.  I hope that nothing that I say in the presentation – I’m sure that other people feel this as well –makes light of anybody’s views about abortion.  But it is true, as Donohue and Levitt pointed out in 2001, that whether or not abortion affects crime is an interesting statistical issue.  It’s an interesting empirical question.  That’s really why Chris and I got involved in this.

Today’s version of the stuff I’m going to show you is a substantially expanded version of an original paper that we wrote in November 2005.  So if this stuff doesn’t look familiar, that’s why.

What are a couple of Fed economists doing working on abortion and crime?  It’s not so much the topic of abortion and crime that we were interested in, but the type of data that were used to investigate the topic, specifically, the heavy use of state-level data. 

Donohue and Levitt basically look at six pieces of evidence for an abortion-crime link.  The first two are not really formal statistical tests, they’re sort of back-of-the-envelope calculations to see whether or not, given reasonable parameters on what type of women are likely to have abortions and what are the likely criminal propensities of their offspring, does it make sense that it could have an effect?  The answer is yes. 

They also sort of eyeball some national-level time series data to say when does crime turn down.  Well, about 20 years after legalization.  That makes sense.  That allows them to launch into their four more formal statistical tests of this issue.  All of those tests, in some way or another, involve data from individual states of the country.  For example, five states legalized abortion in 1970, three years before the rest of the country with Roe v. Wade.  Does it look like crime turns down earlier for those states than the rest of the country?  They argue yes.

Number four and five look at states with high abortion rates in the 1970s.  Does it look like those states experienced bigger crime declines in the 1990s?  They argue the answer is yes.

Finally, the sixth test that they use looks at very narrowly defined cohorts of people, based on the state, year, age levels.  So “17-year-olds in Massachusetts in 1990.”  Then going and looking back [to see], was Massachusetts a high or low abortion state when those folks were in utero?  They find that if it was a high-abortion state that was relevant for this particular cohort, it looked like that cohort had smaller amounts of criminal activity, which is a pretty powerful test of their hypothesis.  It’s one that we criticized in our earlier paper.

But you can see that all of those last four, to some way or another, deal with state-level data and use that data in different ways.  I’m the type of economist who likes to use state-level data to look at economic hypotheses.  One, because as the Federal Reserve System, we have to know what’s going on in each of the twelve Federal Reserve districts, the states that make [them] up.  And also, like other economists, we use state-level data to test different hypotheses – not about abortion and crime so much, but what’s the effect of labor demand on labor force participation?  You could use state-level data to test that stuff too.  That’s the type of things that we were interested in when we started getting involved in the abortion and crime [debate].

Anytime you use state-level data to test some sort of economic hypothesis, you always have to worry about the problem that economists call endogeneity.  That’s basically a fancy word that involves the following question: is the abortion rate of a state related to other factors?  Is it endogenously determined by the other factors that might determine crime?  So for example, New York is a high-abortion state.  Utah is a low-abortion state.  If I see New York crime going like this [up] and Utah crime going like this [down], the first thing that I think of to explain that may be – it must be because New York had a lot of abortions in the past.  There may be other things that make New York crime qualitatively different than Utah crime, so that any differences between those states in changes in crime over a particular time period for those two states may not have to do with abortion.  That’s called the endogeneity problem.

Donohue and Levitt are obviously well aware of this problem.  They use a very innovative way to check for likely endogeneity of crime with respect to the abortion rate.  Namely, they look at how the crime rates of the different states behaved between 1973 and 1985, when abortion should have no effect on a state’s crime rate because the cohorts of young people being born are still too young to really have an effect on the crime rate.  Not a lot of seven-year-olds knocking over liquor stores.  So if you don’t see very much difference between New York crime and Utah crime in that earlier period, you’re on more solid ground saying, yeah, it is reasonable to use state-level data to test an abortion-crime relationship.

Specifically, what they find is that between 1973 and 1985, in high-abortion states, property crime is falling relative to low-abortion states.  So New York property crime [was] going down 1973-85, relative to the changes going on in Utah’s property crime.  Murder, however, is rising and trends in violent crime in general – of which murder is one – are the same. 
So they interpret those data, those patterns in the changes in crime in the individual states over time, in the following way.  They say there should be no effect of abortion on crime between 1973 and 1985.  To the extent that high and low-abortion states systematically differ in crime rates in this earlier period, questions are raised about the exogeneity – opposite of endogeneity – of the abortion rate.  It is reassuring that the data reveal no clear differences in crime rates across states between 1973 and 1985 as a function of the abortion rate.  In some instances, crime was rising more quickly in high-abortion states; in some cases, the opposite is true.

What Chris and I have done is go back and ask – let’s ask a different question about abortion in the different states in this earlier period.  Instead of asking about the changes in state-level crime rates – is New York crime changing differently than Utah crime is changing – let’s just look at the average or the levels of crime in the two periods.  When you do this, you find some interesting relationships.

I want you to focus first on that top row, or the top panels.  Each column in that slide there graphs the abortion rate on a horizontal access and a particular type of crime on the vertical axis.  So farthest away from me is property crime, in the middle is violent crime, and over here is murder.  The top row of panels includes data from all 51 states – D.C. we count as a state.  You look at that and you can see – huh.  Maybe there’s an upwards sloping relationship between the average crime rate in a state between 1970 and 1984, and the average abortion rate in a state.  But you might think maybe that’s the eye being deluded, because D.C. is driving a lot of that relationship.  You can see that the District of Columbia has both high crime rates and high abortion rates.  So if you look down at the bottom level, it’s the exact same data except we’ve just dropped D.C.  So you can see what’s going on among the remaining 50 states a little bit more easily.  You can see there that yeah, it does kind of seem to see, especially for property crime and violent crime over in the far two panels, that there is a relationship between abortion and crime in this earlier period.  Not maybe so much in the changes in crime but in the actual just averages of the crime rates.

One thing you might think is, well, this is a pretty raw calculation.  We’re just sort of graphing abortion against crime.  Maybe what’s going on is that there’s sort of different regional effects that are driving this relationship, so that there are high-abortion, high-crime regions of the country and low-abortion, low-crime regions of the country.  What you can do, though, is strip out geographic effects in this graph by basically taking out the variation in both abortion and crime that can be explained by geographic factors. 

The statistical way to do that is just to run a regression of each state’s crime rate or each state’s abortion rate on a series of nine dummy variables corresponding to the nine census divisions of the country: New England, Middle Atlantic, whatever.  And then you look at the residuals.  The residuals there are the portions of abortion and crime means that can’t be explained by geography. 

What do those look like?  Far from driving the relationship that we see here, if you take out the geographic effects, the relationship gets even stronger.  So what this is saying is if you sort of look at states that are very close to one another, like Nevada and Utah, Nevada has more crime than Utah does.  Nevada has also more abortions than Utah does.  So the correlation between abortion and crime, this positive correlation in the earlier period, gets stronger when you take out the geographic effects.

This is sort of interesting, but what does it really show?  Why does it matter that pre-1985 levels of crime are higher in high-abortion states?  One issue is, remember, what we’re trying to explain here.  We’re trying to explain why crime is falling in high-abortion states after 1985.  What if there’s a tendency for convergence in state-level crime rates over time?  What if over time what seems to be going on is that the states that have high crime rates are sort of moving down to be closer to the rest of the states in the country?  Then you might infer high-abortion state is falling relative to everybody else – maybe what’s going on is just high-crime states are falling.  So there’s a potential issue there.

Is it reasonable that we see convergence in the data?  We would argue yes.  This is a figure that appears in the revised version of our paper, which will be out shortly.  What these graphs are, these lines are basically measures of dispersion in state-level crime rates over time.  One of them takes out the geographic effects – that’s the dotted line.  The solid line is just basically the raw standard deviations.  You can see that there’s a general pattern here.  I have to ask you to ignore the two vertical lines for now.  But there’s a general pattern, that these standard deviations are getting smaller, that there is convergence over time.  So it’s realistic to at least think that over time you can see the high-abortion states decline but it may not have anything to do with abortion.  It may be that just high-crime states see their things declining.
The second thing is, remember what we saw when we accounted for the geographic effects.  We saw that if we accounted for geographic effects, the positive correlation between abortion and crime in the 1970-84 period actually got stronger.  That’s going to make our lives even harder because there are other things that we might want to control for when we start talking about what’s going on with state-level crime rates over time.  One thing we might want to control for is, does it look like there’s sort of time-varying geographic effects?  Like crime in the Northeast seems to be moving around differently than crime in the South, or crime in the West.

So in our data, typically what we do – and Professor Donohue and Levitt did this in a robustness check on their data – they included some controls that sort of accounted for that in one of the tests of the paper that I’m going to show you in a minute.  When you do that, typically what the statistical procedure does is basically say, okay, I’m not going to compare all of the states together at the same time, I’m going to compare states that are close to one another and see how they do, and then average all those different correlations together, basically looking within census divisions to see if there’s a relationship.  But, because the problematic correlation between abortion and crime gets stronger when you look within the census division, that means that accounting for these geographical effects may make problem number one even worse.

Specifically, what’s going on?  Remember that fourth test that I talked to you about: states with high abortion rates in the 1970s saw bigger crime declines in the 1990s.  This is the effective abortion rate regressions in the Donohue and Levitt 2001 paper.  What these regressions are, basically statistical procedures to try to explain crime – that’s the left-hand-side variable – in state i at time t. 

There are three things you can use to try to explain that.  The top line is the effective abortion rate.  That’s basically a prediction of, given past abortions, how much did those past abortions affect crime right now? 

The second thing is something we put in this new paper, which tries to account for this potential for convergence. That is, we take the mean of the crime rate in state i from 1970 to 1984 – that’s one number for each state – and then we multiply that times a trend, which equals 1 in 1985, 2 in 1986, et cetera.  What that term is going to allow the data to do is to say no, it’s not the high abortion states that see their crime decline, it’s the high crime states that are seeing their crime rates decline.

That last piece is the region [multiplied by] year interactions.  Those are things that can account for time varying geographic effects, allowing there to be a New England 1985 effect on crime, a New England 1986 effect on crime, et cetera.  Then we account for serial correlation in the error terms, which would be exceptionally boring if I were to discuss how we did that.

So what do we get when we re-run Donohue and Levitt’s effective abortion rate regressions with some of these new variables?  Let’s start in the far column over here, by looking at just the regression that basically they ran, looking at just allowing abortion as the only interesting variable here to affect crime.  They put in some other variables as well that we’re not going to put in, but these are the most important ones.

So you find a big negative effect.  That -.114 is a lot bigger than the standard error of .026.  Update the data, instead of stopping in 1997, go to 2003, gets even stronger.  Now we’re up to -.133 and still a low standard error.

Now I’m going to put the geographic interactions in, but I’m not going to put in anything about convergence.  You can see the geographic controls are listed on the bottom row.  I’m going to put region [multiplied by] year interactions.  There are four census regions in the country: the Northeast, the South, the Midwest, and the West.  I’m going to allow each of those things to have an individual effect on crime that sort of varies with each year.  When you do that, the abortion variable still comes in significant.  This is exactly what they found when they ran the same robustness check in their 2001 paper.  Accounting for these geographic effects doesn’t affect the strength of the abortion coefficient very much.
But remember, what we argued could be going on here is basically we’ve got a nasty correlation in our data that gets even nastier when we throw in controls for geography.  So now I’m going to put the controls for geography in, in that fourth column, and also put in our convergence term, to allow the data to say, it is convergence, or the high-crime-rate states that are declining, and bang: abortion runs out of the regression.  It’s no longer significant.  The crime [multiplied by the] trend variable comes in very significant.  The regression says, yes, it’s the high-crime states that are seeing their crime rates decline, not the high-abortion states.

Same story if you use more precise divisional controls.  There are nine divisions in the country and there you go.  Same story for violent crime and murder going on.

[The] other thing they do is they look at the five states that legalized abortion in 1970.  For various reasons, the convergence problem still exists.  That’s because those five states that legalized early are also high-crime states.  California and New York are two of those five states.  When you use population weighting to come up with a mean, you are more likely to find that, yeah, California and New York have a lot to say about how that mean for the early legalizers goes, and they’re also high-crime states.  I don’t have time to talk to you about how that matters.  I’ll have to refer you to the paper.

The main theme of what I’ve said so far in the last fifteen minutes is that using very state-level variation to test this relationship can be very dangerous.  That’s because there can be very strong differences between crime in high-abortion states versus low-abortion states that have nothing to do with abortion.

So what you’d want to do is you’d want to use the state-level data in a special way, in a way that didn’t have the contaminating variation of New York versus Utah variation sort of identifying your effect.  What you’d want to do is sort of look within New York in a specific year – say, 1990 – and instead of having a data set that has just one observation for New York in 1990, have several observations for New York in 1990 that vary on the basis of age cohorts.  So I’ll have 15-year-olds in New York in 1990, 16-year-olds in New York in 1990, all the way to 24-year-olds.  So for New York in 1990 in my data set, I will have ten observations.  What I’ll do is each one of those cohorts is going to have a different level of abortion exposure, because abortion rates change over time.
 
So what I can do is I can look within that cell, that state-year cell, get at an effect of abortion and crime, [and] see if the high-abortion cohorts are committing more crimes.  Then sort of average together all the correlations that I find for New York 1990, New York 1991, Utah 1991, and sort of define things that way.  That’s what they supposedly did in their last test.
The problem was – and this is what we wrote about in November – there was a small programming error that left out some crucial controls, some crucial variables in that regression, that forced the comparison to be within the state and year.  So what they thought they were running was a comparison or a test that only looked within a state and year and averaged things together, but instead their test had contaminating variation.  When you fix that programming error, you get their coefficient goes from -.028 to -.013.  It drops about in half.

Another thing they did in that test was they switched from looking at per capita data to data that was just total arrests.  So it was unable to distinguish between the first channel, just lowering the number of people, versus the second channel, the controversial selection effect.  Does it look like per person crime rate is falling if you have a high abortion rate?  When you add those data, then the population data, there’s no per capita effect.

Since then they’ve made some adjustments to their data.  They’ve defined abortions by the state of residence of the mother, not by occurrence, and this data was [made] available fairly recently.  They’ve accounted for inter-state migration and the timing of births.  By the way, all the stuff that I showed you earlier, except for that last table, all that data was abortions based on the state of residence of the mother.  That last table it wasn’t.

If you do their corrections, though, for violent crime you can go from something like .000 to -.021.  So you’re getting close to what they got before but for various other reasons, which other people will probably talk more about, that’s still not a significant effect because it’s making heroic assumptions on how much correlation there is elsewhere in the data. But with property crime, which is the biggest part of crime, their corrections actually go the other way.  They now become slightly positive.

So I’m about out of time. But this is sort of where we were coming from when we were looking at their stuff.  We’re very interested in how to use state-level data to test economic hypotheses.  We found an endogeneity problem here.  When you look at the data the right way, we would argue that the evidence for a selection effect on abortion is pretty weak.

MR. KLICK:  Thank you, Chris.  Next we’re going to hear from Ted Joyce, who is a professor in the department of economics at City University of New York, Baruch College.  In addition to his work challenging Donohue and Levitt’s work on crime and abortion, Ted is also one of the leading economists to look at effects of abortion in general, having dozens of articles in the economics journals.  So we’re really happy to have Ted speak today.

TED JOYCE:  Thank you.  I want to echo Chris’ comments about the openness of John and Steve regarding data and responses.  They’ve been collegial to a fault in many ways, and we’ve all benefited.

I’ve got a pretty aggressive title.  I really think the consequences of abortion on crime rates – I underline “rates” – is really inconsequential.  That’s going to be my point today.  So let me just structure these comments.  I want to talk a little bit about what we call these total crime regressions.  I’m going to discuss the limitations of age-specific crime rate regressions, and if time permits try to outline an approach I thought would test this in a different context.

This [graph shown on slide] is what all the noise is about.  This is the homicide rate in the U.S.  These are homicides per 100,000 population.  So they’re peaking in 1991 at 9.8 and they’ve fallen to 5.5 by 2001 - clearly a dramatic and very important decline.  We’re at levels we haven’t experienced since the mid-1960s.  If you go to where I live, in New York City, the decline is actually spectacular.  We’re talking 30.7 in 1991, down to 7.3.  A huge, huge fall in homicides.

Some of the criminologists have been on this.  Their take is an important place to start.  Alfred Blumstein, for example, writes, “A dramatic rise in homicide in the latter half of the 1980s, peaked during the 1990s and declined at an equally dramatic rate.  Such trends in homicide rates can be understood only by examining rates in specific age, sex and racial groups.  The increase primarily involved young males, especially black males, occurred first in the big cities, and was related to the sudden appearance of crack cocaine in the drug markets of the big cities around 1985.”  I think it’s a nice way to structure the discussion. 

Here’s what he’s talking about.  If you look at homicide rates by age, for example, you can see where all the story is.  The top line is 15 to 19-year-olds.  That’s the homicide rate, 20 to 24-year-olds, 25-29, 30-34.  In other words, the decline I showed you previously – this decline – is being driven entirely by kids or adults less than 25 years of age.  That’s where all the action is.  You can see it here. 

If you go by race, these changes are actually spectacular.  If you go to 1984, for example, the homicide rate among blacks 15-19 was around 25.  It rises to over 100 in a matter of seven years.  So these are staggering increases.  They go up and they come dramatically back down again.  Look at 20 to 24-year-olds, you get the same story.  If you go to older adults, there’s not much going on.

So [there are] a couple of points that are really important here.  One, whatever happened, happened to a specific age group and it happened at a specific time, about 1984-85.  Whatever you want to call it – crack, whatever it is – it hit those groups.  It didn’t seem to hit all other groups. 

Second of all, they peak in the same time.  In other words, the story that John and Steve were trying to get at is what we call a cohort effect.  They’re trying to find the effect of something that happened near the time of birth and trace its impact all the way out 15-25 years later.  If you look at this, cohort effects really present very differently than this.  John Lott is going to say more about this and I won’t speak too much about it.  But I needed to motivate why I think their total regressions are not the place to go looking for a relationship.

Anyway, this is a period effect.  Whatever drove crime seemed to drive it for all kids 15-24 years old.  Whether you were exposed to legalized abortion or not, you were affected.  They peak in the same year and start to come down.  That’s kind of an important point to stay with.  If you go to whites, for example, you get some of the pattern but it’s not nearly as dramatic.  So it really is about age, race and crime rates, is really my point here.

I don’t want to beat anyone up with regressions, for example, but the devil is in the details, so let me start here.  This is really what got all the play in their analysis.  This is the total crime rate.  This is the homicide rate of all people per 100,000, or the violent crime rate, or the property crime rate.  That’s the left-hand side variable.  What they’re going to explain it with is what we call this effective abortion rate.  It’s a pretty crude measure of exposure, I’m not going to spend much time on it today.  But the point I really want to make is that you can’t identify age or cohorts with this regression.  I think the most important features of the rise and decline in crime is being driven by age and race, and neither of those factors are identifiable really in this model at all.  It’s not age-specific.  They can’t identify cohorts at all.  Their abortion rate tries to get at it through a back door, very clever, but it’s really quite crude.
When you look at their effects, I think their effects are huge.  These coefficients are coming from the Journal of Human Resources, in which they responded to my criticism.  Let me just tell you what this means.  They’ve scaled them nicely so that .153 means that an increase of 100 abortions per 1,000 live births – which is one standard deviation – should lower crime about 15.3 percent – total crime.  Property crime should fall about 11 percent and homicide should fall about 16.6 percent. 

Remember, the left-hand side is the total crime rate.  As they, again, appropriately do, they say the percent of arrests accounted for by those under 25.  So let me quickly turn – I’ll come back to this.  Here’s their quote.  “On average, about half of those arrested are under the age of 25.  Thus, to generate the crime reduction in Table 4” – the one that I just showed you – “requires coefficients on young arrests that are twice as large as the coefficients on overall crime.”  In other words, these effects in that top row underestimate the effect on specific cohorts by about 50 percent.  They give you those numbers, so I then inflate them.
In other words, what we’re looking for in the cohort analysis, we’re looking for declines in crime of about 30; 17 low; 35 percent high.  These are huge decreases in crime.  But that’s what their model, this model, implies we should be finding in the arrest rates.  Now, arrests [are] different from crime, we can make arguments about that, and that’s fair enough.  But still, we’re unmistakably looking for very large effects.  If you look at the change in crime from its peak in the 1990s to 1997, when the data ended, these declines are the huge declines I showed you – 19 percent, 16 percent, 30 percent.

So really, what they’re arguing is that the change in crime among the cohorts that have been exposed to legalized abortion are gigantic.  I think they’re too big to be real.

So this is, I think, a much more credible approach to the problem.  This is what I give them great credit for.  Now they’re going to look at age-specific arrest rates.  So the left-hand side is the arrest rate of age group a, state j, in year t.  So, for example, if this were 1990 and these are 20-year-olds, that would be the arrest rate, let’s say, for violent crime among 20-year-olds in New York State in 1990.  They’re going to regress it on this AB, this abortion rate.  Again, I think a much more credible approach to the problem.  The abortion rate, but lagged A-1 years. 

In other words, take that 20-year-old in 1990 – they’re going to regress the arrest rate of 20-year-olds in 1990 on the abortion rate in 1969.  That makes sense.  That’s when these kids most likely were in utero and their mother had access or didn’t have access to legalized abortion.  So this is a very credible regression in my mind.  You can now get at the age effects, the cohort effects.  This is where I think we should be looking.

But, what are the issues?  [One is] the underestimation of the standard errors.  I won’t push this too hard, but it’s actually a very important point.  I think the effects are small if they’re properly scaled.  The effects of cohort size, to my mind, are questionable.  And the exogeneity of the instrument, I think I’ll pass.

Here’s what their data looks like.  The top row are crime years.  So this is 1985-1998.  For each year, they have arrest rates by single year of age.  So they have arrest rates of 24-year-olds, 23, 22, et cetera.  So every year, they have ten age groups, which is very nice.  If you then look at the vertical column there, I’m linking the age groups to their year of birth.  Again, that’s important, because what they’re going to do is they’re going to say, okay, the 24-year-old arrest rate in 1985 – that top left corner there – should be associated with the abortion rate in approximately 1960.  That’s what they’re going to argue.  The problem is they have no abortion data.  The only time they have abortion data that’s relevant for the cohorts is in the yellow area.  So [for] that entire other non-yellow area, they have no data on abortion.  You couldn’t estimate their model in that period.  You can really only estimate the model with an abortion rate linked to it for the cohorts that are in the yellow.  So in other words, 60 percent of their data cells have no abortion rate.  Was abortion zero in those days?  No.  Do we know what it was?  Not exactly.  We know it’s not zero.

[Some] other points.  I want to make one other point.  Go to 1974 and go across on the horizontal line.  They have the abortion rate in 1973, which they’re going to associate with, let’s say, 15-year-olds in 1989.  So here are 15-year-olds in 1989 and they’re going to associate this with the abortion rate in 1973.  Fine.  The problem is, what they’re assuming here is a couple of things.  One, the abortion rate in 1973 is going to be used with this group, this group, this group, this one – in other words, they don’t have independent effects of abortion.  The abortion rate is the same for every group born here.  It just repeats over and over again.  So they adjust for that. 

We call that clustering.  That’s an appropriate thing to do.  But they pose a very strong restriction on the data.  They say there is no relationship or no correlation between 15-year-olds in 1989 – and their arrest rates in New York City, for example – with 15-year-olds in 1988 in their arrest rates in New York City.  And yet we know that crime has strong serial correlation – we call it strong persistence over time.

So they make some very strong restrictions on their statistics that impact on the statistical significance.  So this is really Chris [Foote]’s insight.  This is now a coefficient which shows you, for example, that if you increase the abortion rate by one standard deviation, which is 135 abortions per 1,000 live births, you’d expect violent crime arrest rates to fall by 3.5 percent.  I think that’s small.  This is the biggest result we’re really going to find in their data – 3.5 percent is much smaller from the 30 percent I was showing you before, which is what you would expect, all else constant, given their first regressions analysis.

But then when I correct for the standard errors, and when you cluster I think appropriately, what you find is a result that’s no longer statistically significant.  In other words, this is statistically not different from zero, and it’s a pretty small effect, 3.5 percent.
I then said, look, why don’t we just run your data from 1974 to 1981?  In other words, let’s just use the cohorts – let’s just use these people.  These cohorts have abortion rates, real abortion rates.  Abortion is growing dramatically over this period.  So if your story is correct, it really should hold for this stuff, or actually it should be strongest.  The response to me in the Journal of Human Resources, they said even if these are zeros, it’s just going to make our results smaller than they really are.  Okay, well, we can kind of look at that.

So I then ran their analysis from 1974 to 1981, using those cohorts, and this does go up.  It’s no longer statistically significant, marginally, whatever.  The effect is slightly larger, 4.6 percent.  But it’s not a big decline here.  And this is where everything should be happening.  We have data on abortions, abortions growing rapidly.  There’s not much going on.  If you go to property crime, there’s nothing going on whatsoever.  Nothing improves when you go to periods in which you have data versus periods where you have zeros for abortion rates, for 60 percent of your data.

When they responded to me, they did use murder arrest rates and they did use homicide rates.  So let’s go look at it.  When you look at it and adjust the standard errors correctly, or look at a time period in which we really think the effects should be strongest, there’s nothing here.  Not one of these coefficients is significant.  There is no association between abortion and crime across violent crime arrest rates, property crime, murder arrest rates, and homicide rates. 
One can argue arrest rates are a little funky because they reflect not just violence but police behavior.  But the homicide rate is a little cleaner in that sense.  It does reflect violence.  I think John Lott is going to have more to say about that later.

So there’s nothing going on here.  This is, again, probably for the economists involved here.  But if you look at what they did and how they changed their model – so from the QJE to the QJE and JHR, this was how they specified the model.  They controlled for state, year and age effects, or state-age effects, year-age effects.  We’ve got 549 parameters or variables.  It’s a pretty heavily parameterized model.  This would say positively related to crime, statistically significant.  This is negatively related to crime, statistically significant.  This says positively related to crime, statistically significant.  Positive—negative.  The results are very, very sensitive to how you specify the models.
 
At the end of the day, you might want to conclude this is the best model.  But remember, you’ve got 1,100 variables in this model.  Your abortion data only varies across 560 cells.  I think it’s a fragile kind of model and it doesn’t really stand up.  Again, if you stopped right here, you’d conclude that abortion and crime were positively related.  In their initial articles, the two articles they did, they stopped right here, and you get different results if you use rates as opposed to logged arrests. 

I can do this for all the others.  Again, the story doesn’t change.  There’s not much going on.  It seems sensitive.  I’m using their standard errors, which are really I think incorrectly estimated.  I can show you things but I won’t drive you through all that.

But the point being here is that their model is sensitive.  It’s not robust, as we say.  Anytime you change something, the coefficients change dramatically.  If you thought this was the best model, you’d infer, again, positive relationship to crime; you thought this was the best model, you’d take a negative relationship to crime.  It’s unsatisfying that the model seems so fragile.

So they would back up and say, wait a minute.  Really it’s more important to look at the total effect, because they’re going to argue abortion does two things.  It lowers the size of a cohort and at the same time affects the crime rates.  So what they did in their regressions, they looked at arrests – not arrest rates, arrests.  Their effects are substantially larger when you do that.  Chris brought out the November paper that when you go from arrest to arrest rates, the coefficients fall dramatically.  In some cases, become totally insignificant.

So now they have to argue that everything is really working through the size of a cohort.  That’s a strong statement, because if you look at this, what does it mean, the size of a cohort?  What I’m showing you here are birth rates in the U.S. from 1961 to 1985, and abortion rates in the U.S.  These are the Roe states, the states that legalized abortions after Roe v. Wade.  These are the states that legalized abortion earlier.  They follow the same pattern.  They break off right here. 

This is Phil Levine’s work, in which 1971 to 1973 we see about a 5 percent decrease in birth rates among those in these early states relative to these states, but both trending.  They converge around 1976, 1975.  So what?  That’s important because the abortion rate never converges.  In other words, there’s a huge difference in abortion rates between states that legalized earlier and those that legalized later.  This difference in abortion is not reflected in smaller cohorts at all.  So the cohort effect, if it exists, only exists in this period.  Yet their analysis goes from 1961 to 1983.  The cohort effect is really limited to this little window right here.  After that, there really is no cohort effect at all.  They acknowledge that in the response to me in the JHR.  They said as long as the number of unwanted births fall, even if total births do not decline, one would expect to see better life outcomes on average for the resulting cohort. 

What are they saying?  Abortion affects timing – not cohort size, but timing.  That a teen delays a birth or a woman of twenty delays a birth until after college, and the total number of births is not really changing, because you can’t find a relationship between abortion and birth rates, but the timing is being improved.  It’s a little softer argument, because there’s unwanted births and there’s unintended births.  Timing is unintended births.  The evidence that unintended births are kind of adversely affected is pretty weak.  So cohort size is not a big deal here.  They, I think, acknowledged it in the JHR.

One last statement.  They say, “Because the dependent variables are denominated by a population under age 25, the abortion coefficients only reflect changes in arrest rates per person.  If the impact of abortion was solely through changes in cohort size, then the per capita specifications we run would yield zero coefficients on the abortion variable.”
What are they saying?  If it’s really operating through cohort size, then anytime you look at a rate, the abortion coefficient would be zero.  What’s the flip of that?  The flip of that is this.  If the impact of abortion was solely through selection on crime rates, then the per capita and the level specifications would yield similar coefficients.  That’s the implication.

So I did that.  This is from their Journal of Human Resources paper.  This is showing the change in violent crime arrest rates for people less than 25 years of age, again, for property crime and homicide.  This is just the rates.  Again, what they said was this.  If there’s really a strong cohort effect, then when I don’t use rates, I use levels, this should be a much larger coefficient.  It isn’t.  It’s actually smaller.  So I don’t think there’s strong evidence for the cohort effect. 

I think it’s time for me to wrap up, which I will.  I think the most credible test of abortion-crime involves age-specific crime rates.  I don’t think you can test cohort effects unless you can identify cohorts and know the age of the individuals involved.  We also saw that the crime change is being driven almost totally by those less than 25 years of age. 

The effects, I think, are relatively small.  Given their results for total crime, we’re looking for really big effects out there.  I think they’re small effects.  I’m not sure the cohort effects are identified credibly and there are other ways to test this.  That’s for another day.  Thank you.

MR. KLICK:  Thank you very much, Ted.  Now we’re going to hear a slightly different set of results or comments by Leo Kahane, who’s a professor of economics at California State University-East Bay.  He’s going to discuss some of the same types of analyses on British data.

LEO KAHANE:  Thanks to John Lott and AEI for putting this together.  It’s a pretty exciting chance to have all of these guys together to talk about a very narrow topic.  You rarely get these opportunities.  This paper is written with two co-authors: David Paton, who’s here – neither one of them is named Chris, by the way.  Two co-authors, David Paton, who’s here as a discussant and will help me out when I run into trouble with the British data, and Rob Simmons, who’s not here today. 

Before I begin, just to satisfy my own curiosity, I want to pose two questions to everyone in the room, including the panel.  The first one is, setting aside the econometric difficulties, the data challenges, and the deficiencies, just [take] the basic idea or the hypothesis that legalized abortion some ten to fifteen to twenty years later leads to a reduction in crime – do you think that’s a reasonable theory or hypothesis or do you think it’s nonsense?  Raise your hand if you think it’s reasonable.  That it could be true - about two-thirds of the room.
 
Second question, which is not unrelated to the first question.  When you first heard of this idea that legalized abortion led to crime reduction some time later, did you have a gut response?  How many of you were repulsed by the idea?  About a quarter of the room.
Well, when I read about this paper for the first time, which was shortly after it appeared in the QJE, I have to say that my responses were I thought it could be possible that there’s a linkage between the two – I was open-minded about that – and I didn’t really have an emotional response or a gut response, other than thinking that other people would have an emotional response.  I’m certain I’m right about that.  Not quite as emotional perhaps as the right to carry handguns, and John could talk more about that.  So I admit, that’s what was the impetus for thinking about trying to test this model, this idea, in a different environment, different data set.

So the three of us – David, Rob, and I – embarked on this study of abortion and crime in the UK.  A couple of things will make this comparison between the U.S. and UK difficult, the data difficulties in how crime data in particular is collected.  A couple of things about the history – well, a brief synopsis of what legalization in the UK is all about.  There was a legalization act in 1967 which came into force in April of 1968.  I’m going to refer to the UK, and what I mean is England and Wales.  Scotland and Northern Ireland are not part of this data set.  The first full year of legalized abortion was 1969.  We can see from the figures up here that it went from about 50,000 in that year to a little over 100,000 by 1972, up to 170,000 by 1989, where it starts to level out.  So we have significant evidence that there was this rather significant ramping upward on the abortion rate, which is something required if we’re going to find any kind of effect on crime sometime later.

One of the things that makes the UK data a little more desirable than the U.S. data is that regardless of how abortions are performed – and they’re performed basically in two ways, either on the National Health Service (NHS) for zero price or performed privately – every abortion, NHS or private, has to be reported.  So we have a fairly good data set, fairly accurate data set, on the number of abortions performed by region as well as a couple other measures of the woman who’s seeking an abortion – marital status, age, residence and so on.  So it gives us a slightly more reliable data set than I think the U.S. data in that respect.

We had two goals in mind when we started this research project.  The first was to do just a straightforward test, as close as we could match, to the original Donohue and Levitt model.  Then the second goal was to think about shortcomings in their methodology and problems that might exist in their data – Chris has already talked about this a little bit, the issue of endogeneity.  We look at it in a slightly different way, I guess, or try to deal with it a little differently.  But we compute effective abortion rates as they do, using this kind of arcane equation here, which I won’t spend time on.  But it’s kind of a weighted average of past abortions and how they may affect crime today.

One of the things that we can try to do when it comes this issue of endogeneity is by the fact that we have this better information about the woman who’s having an abortion, is to try to identify abortion access a little better than simply abortion.  That’s what’s really key, right? We have a measure that we call ratio home, which we composed using the data.  I should let you know that the data in the UK, unlike the U.S. where you have the 50 states or 51 if you count D.C., the UK doesn’t have an equivalent set of data like that.  Our crime data is by police force area, which is essentially a collection of counties.  The abortion data, as I recall, is by county.  So it has to be aggregated up to this police force area.  Ratio home shows the proportion of abortions performed on women within a [region] that actually had their abortion in that [region].  So the higher the ratio, the more women that are having abortions in that region or having them near their homes.  Not having to go elsewhere.  Implying, we believe, that there’s greater access within that region.  That’s the focus in our minds, access.

Got a little graph here which shows the effective abortion rate as we computed it using the Donohue and Levitt method.  We’ve also got the abortion ratio.  Let’s move on to their model.  Another model of crime which bears a close resemblance to the Donohue and Levitt model in 2001.  We model crime as a function of the effective abortion rate.  Then this X, which has other control measures that might affect crime rates across time and space, with controls for region-specific effects and time effects as well.  This has been discussed by others on the panel.

We have some co-variants.  These are the measures that make up this thing X in the previous slide.  A lot of these things are really not that important to talk about.  They’re standard controls.  The work is motivated by Becker and Ehrlich in the 1960s and 1970s.  Wages within regions at a particular point in time, unemployment rates.  These essentially show motivations for seeking income in a criminal way rather than a legal way.  We have a measure of wealth that we use as a proxy for that, cars per thousand of people.  Unlike the state data where you could find better measures of wealth, we don’t have that luxury with police force area data.  We’ve got presence of police force.

One measure that we have that none of the other papers has, which makes ours a little unique, I guess, and also a measure that turns out to be kind of a champ in our results, is this percentage of children in local authority care.  We include this as a kind of measure or representation of social deprivation within a region and time.  It’s computed as the percent of children that are in local care.  Being in local care means that their parents were either unfit or in jail or for some other reason [gave] up their rights or abilities to care for their children to the state, or the local authority I should say.

This kind of describes the data set.  We’ve got, again, 42 police force areas spanning the time period 1983-1997.  Part of the problem we have is that, and you’ll see this toward the end, we don’t have the data of crimes by single age of year.  So we can’t replicate the latter part of the results that appear in the original Donohue and Levitt paper, which have been the focus today.  Although we’ve been promised by the Home Office in the UK – the office that controls the data – that they will make it available to us sometime in the near future, hopefully.  So that will become part of our paper at a later time.

We have total crimes, which we look at, and then we break it down as six sub-categories, which are listed here.  We’ve got movements of crime, percentage changes across various chunks of time, and then the entire period, 1983-97, in the last column.  A couple of things that stand out dramatically are violence and robbery crimes, which have gigantic increases over this 15-year span.  That’s something that’s not the same for the U.S. data and adds a different dimension to our analysis.

This is not very interesting, but this is some of the statistics that make up our data set.  Some descriptive statistics of the measures that we use.  On total recorded offenses, these are crimes that are reported to the police but with little information about the perpetrator.  Then we have a second data set which are called cautions and guilty rates, which we were just discussing earlier what that means.  This is a similar set of data for the U.S. of the arrests data in the U.S.  Cautions are scenarios where someone has been caught doing something wrong and they’re given the option of either receiving a caution, which is a kind of penalty that goes into their file and they have a criminal record but they don’t serve jail time, or if they don’t receive a caution they can actually have to go to court and perhaps be sentenced for a crime. 

We combine these together.  This is where, if we had [data] by age [and] year, we could more closely match the latter results in the Donohue and Levitt paper, but we don’t.  We have age bands: 14-17, 17-21, 21 and over.  In fact, we have a 10-13 band, which is not shown here.  That’s our control group.  And then some measures on the co-variants that are included.
A couple of important statistics with regard to crime, effective abortion rate, and our ratio home measure, which we’re going to work with in a few minutes.  The point of this additional summary table is to show that for each one of these important measures that’s going to appear in our model, there’s significant variation between, within and across time and space for the effective abortion rate and particularly for the ratio home measure, which we’re going to use to work with the possible endogeneity problem which was mentioned earlier for the effective abortion rate.  So we’ve got significant variation across time and space to help us identify this effective abortion rate measure.

We put together a couple of slides, time-casts of crime, to show the similarities and dissimilarities between the UK and the U.S.  Probably the easiest [dissimilarity] to spot, is the case of violent crime, which sums together violence and robbery, which unlike the U.S. doesn’t have this turndown anywhere, it’s just continuing to rise throughout the time period we’re looking at.  The property crime and burglary shown on the other two profiles have that similar shape, that kind of hump shape, where there’s a dramatic and visually clear turndown in the mid-1990s here.

Part of the problem, however, is when we go to the cautions and guilty data, they don’t seem to have that same kind of behavior.  In particular, this is total cautions and guilty by age group.  We’ve got a couple of conundrums.  One is that the age group 10-15, which is the solid line here, has this dramatic drop, starting around 1985, in crime.  But the timing for that group, if abortion is to explain that, is off.  It should have occurred a couple of years earlier.  So you’ve got an inconsistency in the timing.  For the other two age groups, there doesn’t seem to be any downward trend to match an abortion effect when it should have had it as an effect, if it does at all.

This next slide was produced to reemphasize this inconsistency in timing.  The most important plots up here are for property crime in the U.S. and property crime in the UK.  We don’t have murder rates – not quite the same quality of crime data and availability of crime data for the UK as there are for the U.S.  So we don’t have murder rates to show here.  But property crimes in the U.S. and UK are both collected and we’re able to compare them. What I’ve got on the horizontal axis is years since legalization.  So it was 1969 for the UK and 1973 for the first full year of legalization for all 50 states in the U.S.  If there’s an abortion effect, you’d expect the timing of the downturn to coincide or be close, but we’ve got a significant gap of about five years between what happens in the U.S., the downturn time in the U.S., versus what’s happening in the UK.  Gives us some pause about this linkage for the UK between abortion and crime rates.

Our approach [has] two ways of estimating this relationship.  First we tried to estimate it with the methodology and approach that’s quite similar to the Donohue and Levitt, with the exception we used something called a panel corrected standard error, which is designed to deal with the problem of what we call spatial correlation or spillover effects in a particular time period between, say, neighboring counties or neighboring police force areas.  This is likely to be more important for the UK given the smallness of the country in comparison to the U.S., [which is] more spread out.  So this is supposed to control for that.

Then we take another estimation approach called an instrumental variable, where we try to deal with the potential endogeneity of the effective abortion rate by using ratio home as an instrument for the ERA measure.  We had a full set of controls for year and area fixed effects.  A lot of numbers.  Don’t need to worry about all the numbers, especially given the time that I have.

If you look at the top row, that’s the effective abortion rate effect.  We got a consistent result with the Donohue and Levitt result, a negative significant coefficient for total crime.  Kind of a bizarre effect with the sex crimes, but it too has this uncharacteristic upward trend to it.  So there’s something else going on there.  But for burglary, robbery, theft, you’ve got the negative insignificant coefficients.  The implied elasticities are a little large in comparison to Donohue and Levitt, maybe casting some doubt here.  But there seems to be, using their method, or more or less their method, and a model that matches theirs somewhat well, that we find a similar effect.

I’ll point out that unemployment, children in care and to a lesser extent police all have the expected impacts on crime.  In fact, that’s almost throughout all regressions that we see in our analysis.

The next, even more complicated table hopes to deal with the potential endogeneity.  The story of endogeneity, I think, is best understood with an example of how you can get the spurious relationship between crime and abortion rates.  If you think of, for example, things that contemporaneously affect perhaps both.  Studies that looked at demand for abortion services by state across time often find that income is a determining factor of abortion rates. 
Education is a determining factor of abortion rates.  But those factors would also, we argue, likely determine what kind of home environment a child who’s not aborted would experience in their home, which may affect their proclivity towards criminal behavior some 15-20 years later.  So if you grow up in a high-income home, that would affect your criminal intent or likelihood of becoming a criminal some 15-20 years later.  But it also affects whether or not a woman actually has that child when they find themselves pregnant.  So that’s the mechanism by which we think abortion seems to be endogenous to crime.

You can kind of ignore the top panel. The more important one is the lower panel, where we try to strip out this endogeneity effect.  What we find is that at least for total crimes, there’s not a significant – not only is the coefficient positive, but it’s not statistically important in explaining crime in the UK.  We’ve got an unusual effect of violence but again, that’s one of the categories of crime in the UK that behaves very differently from the other categories.  The other co-variants, as I mentioned before, seems to be behaving predictably.

This one is a whole series of robustness checks.  Taking out London, because London is peculiar in many ways, kind of obliterates the abortion-crime relationship.  It’s no longer significant. 

Estimating it using abortion rates for single women, who would arguably be more at risk of producing children who would become criminals perhaps later in life, strengthens the result a bit.  However, we can’t rule out -- some of the tests here show us that we can’t rule out that that’s a spurious relationship, because there’s evidence that there’s endogeneity.  So the next column over shows that there’s not a significant effect.  This kind of pattern repeats for the other robustness checks, which are looking at teenage abortion rates, something called area-specific trends, and then estimating this with model inference differences.

Lastly, with our cautions and guilty data set – and this is over the four age bands we have, which we hope to someday soon have the separate years, age years – the results are rather bleak when it comes to this linkage between abortion and crime.  There doesn’t seem to be any evidence that there’s a causal effect between legalized abortion and crime some 15-20 years later.

Summary.  These three co-variants that I mentioned before come through as predicted in virtually every regression that we run.  So [we] feel good about that result.  [As to] the effects of abortion, when we try to match the Donohue and Levitt model as best we can, we find similar results.  But then when we start to consider robustness checks and the issue of potential endogeneity, that relationship seems to go away.  Bottom line: we don’t know.

MR. KLICK:  Thank you very much, Leo.  Now we’re going to have some discussion of these papers and these talks, from two discussants.  The first is going to be Phil Levine, who is the William R. Keenan Junior Professor of Economics at Wellesley College.  He’s also, if you take all the economics papers on abortions that weren’t written by Ted Joyce, they’re pretty much written by Phil Levine.  In fact, he’s the guy who’s written the book on the economic effects of abortion.  Phil?

PHILLIP LEVINE:  Thank you.  I wanted to thank the conference organizers for inviting me here today.  I’m going to be glad to share some of my insights and hopefully offer something to the topic.

I do have to admit though that this is a rather challenging assignment that you gave me.  I have twelve minutes to discuss three different papers.  The three papers combined levy roughly 5,000 criticisms about the Donohue and Levitt work.  So it’s about a second and a half per criticism, so I think that’s probably not a good strategy for going forward.  What you don’t know, however, is that I’ve also been doing some work on the topic of abortion and selection on my own earlier, and in a recent paper I finished last week, too late for this conference.  Finished a new one which also touches on or is about abortion and selection more broadly, but also touches on the abortion and crime debate.  So not only do I have to comment on all three of these other papers, but I have to put in a plug for my own work as well.  So I’ve got a lot to do and about ten minutes left to do it, so I think I should probably get started.

Basically, I’m going to try to simplify this into what I consider to be two fundamental things to think about in terms of this work.  The first one is, what belongs on the right-hand side of this regression?  Is the abortion rate the thing that we’re interested in or is fertility the thing that we’re interested in?  Of course the intuition that we’re looking for is about fertility.  So I took a couple quotes from the papers I read.  The Kahane paper: “A necessary condition for abortion to affect crime is that abortion should affect fertility.”  In Ted’s paper: “It is difficult to demonstrate an association between abortion rates and fertility rates beyond the immediate period of legalization.”

I think clearly there’s an emphasis in this literature that there’s got to be something going on with fertility, because if there’s nothing going on with fertility, there can’t be any selection.  So I see this as really a fundamental issue.

The thing you have to be a little bit careful of, though, the reason why fertility matters so much, is we need to have an effect on unwanted births for there to be the selection effect.  The problem is that we don’t really ever get to observe unwanted births in the data.  All we ever get to observe are births. 

The problem there is that you can be having two things going on at the same time that are canceling each other out.  In particular, one thing that we might worry about in terms of changes in abortion policy and increasing its access is that pregnancies may go up.  So if pregnancies go up, that would [mean] more people get pregnant; one thing that that might lead to is additional births. 

I want to comment on something in your introductory remarks – those would be wanted births, not unwanted births, the additional births that took place.  An example of this is a woman gets pregnant.  Not knowing her situation in her relationship, she tells her boyfriend that she’s pregnant.  They decide to get married.  The baby comes along.  That pregnancy wouldn’t have, or may not have, occurred in the first place.  So it’s possible that in some instances you may get more wanted births that will cancel out the reduction in unwanted births that are sort of the primary effect people think of in the selection literature.  So it’s not obvious what we should expect to see in terms of births that will be connected to the reduction in unwanted fertility.

So I think this matters a lot, because in some sense the thing that really does belong on the right-hand side of these regressions is the abortion rate as Donohue and Levitt have incorporated.  Because it’s that abortion rate that maybe will get at this notion that changes in fertility don’t pick up all the changes in unwanted births.

That said, there are significant problems to putting the abortion rate on the right-hand side of one of these regressions.  We’ve heard comments about that already.  The obvious problem here from an econometric point of view is that the abortion rate’s endogenous.  So in the Kahane paper, he writes, “A key issue not addressed in the existing literature is the possibility that abortion rates may be endogenous to crime.”  Joyce writes, “The other advantage of a quasi-experimental design” – this is in part of the paper he distributed that you didn’t really talk about so much today – “is the change in the abortion rate during this period, the early 1970s, is large and more plausibly endogenous than changes in the later years.”  So he’s worried about the endogeneity of changes in abortion in the late 1970s.

I think Leo did a very nice job of describing how this endogeneity might come into play.  If there are other things going on in the late 1970s – for example, changes in culture and other social outcomes – occurring at exactly the same time, you may be capturing some of the effects of those things and not necessarily the abortion rate.

So there are a couple solutions proposed in these papers that we just heard about.  In the Kahane paper, there’s an instrumental variables approach employed, where the ratio home variable is what the instrument is.  It’s the number of abortions performed in the health authority of residence relative to the number outside the area.  While I’m sympathetic to the idea of trying to instrument for the abortion rate, which is something I’ll talk more about in a few minutes, it’s not obvious to me that this is solving the problem.  I think the thing I’d be concerned about here is that basically any measure of abortion use is going to be capturing both supply features and demand features.  As soon as you put demand of abortion as an instrument, it loses its ability to serve as an instrument because it’s also capturing some of these other events.

In the later part of his paper, Ted is very concerned about this issue of endogeneity.  In his paper, what he does is he takes a very narrow focus and looks right around the period that abortion was originally legalized in the early 1970s and basically looks at slices of data right around that period of time and argues for exogeneity there.  Which I’m much more sympathetic towards in terms of I think it probably does do a very good job of getting at exogenous variation, which I have used in the past. 

On the other hand, to look at just that really very narrow focus is also throwing out a lot of data, so one of the things I’d be worried about there is that maybe you’re throwing out the baby with the bath water in the sense that you’re losing a lot of statistical power, to be able to take that very narrow lens in the data period that you’re looking at.

So what I want to spend a couple of minutes talking about next is in our paper, what we did to solve the endogeneity problem and what the results were of that analysis.  This is a paper, a joint work with Liz Ananat, Jon Gruber and Doug Staiger.  The title of the paper is “Abortion and Selection,” because it’s about selection more broadly, not just focused specifically on crime.  We look at other outcomes like educational attainment, welfare receipt, single parenthood, that sort of thing.

We’re worried about the same sort of econometric issues.  We also adopt an IV strategy in our analysis.  So to take a step back, what constitutes a good instrument?  You need something that’s going to be correlated with the abortion rate, because that’s what you’re trying to replace in some sense on the right-hand side of this regression.  It also has to be uncorrelated with children’s outcomes, with potentially criminal activity being one of them, because that’s what’s causing the bias.

So what is it that we used?  Like Ted, we also adopted the approach of using state variation in the timing of abortion legalization, because that really does do a very nice job of providing exogenous variation in the early 1970s.  The other thing we do which I don’t think Ted does in this paper, but I think [he has] done in others, is to also incorporate for the states where abortion was illegal during the early 1970s how far they were from a legal abortion state.  If you lived in New Jersey, abortion may not have been legal in your state, but it would have been with a subway train ride away, whereas if you were in Texas, it was a much more significant obstacle.  So we used that and that seems like exogenous variation as well.

The problem in all of these things is trying to figure out what to do about the late 1970s in a way to instrument for the variation in abortion during that period.  That was sort of the difficulty we encountered.

What we did to solve that was to sort of take a step back and think about what life was like in the late 1960s, when abortion was illegal nationwide.  If you think about the cross-state variation, the laws were the same in all of those states but that doesn’t mean that the environment was the same in all of those states, in the sense that if you could magically eliminate those abortion bans, women in some states would use abortion more than others because the constraint was less in terms of public attitudes toward abortion and culture and that sort of thing.  So the constraints were more serious in some states than in other states. 
So one thing that you can think about is in some sense what’s happening in the late 1970s is you just lifted that constraint and all the states’ abortion rates are going to what they would have been in the 1960s had you been able to eliminate those bans.  So a good instrument might be one that actually captures this variation in public sentiment toward abortion in the 1960s.  So this is not correlated with any of the social outcomes in the 1970s, because this is happening ten years before.

So that’s what we used as our instrument to capture changes in abortion use in the late 1970s.  We incorporate all three of those sets of instruments in our analysis.  What we find when we re-estimate the Donohue and Levitt models is that for the most part the Donohue and Levitt results are robust to our IV strategy.  So the fact that the abortion rate is on the right-hand side and endogenous really does not seem to be driving a lot of the results.  If we use an IV strategy that we think is a legitimate one, it turns out that we get virtually the same results anyway.  So that doesn’t seem to be a significant problem.

On the other hand, we do find that the results are still sensitive to the Foote and Goetz critique that when we estimate these models with the numbers of crimes on the left-hand side, we get the Donohue and Levitt results, and when we put the crime rates on the left-hand side, it turns out that everything really sort of does again become statistically insignificant.  So I’m sure that’s going to be a significant focus of the discussion this afternoon, so I’ll leave that comment at that.

I should also say that our analysis focuses just on that one specific issue of the endogeneity and really does not touch on the many others that were raised here this morning and that potentially we’ll hear more of this afternoon.  But you can only do so many things in one paper and so we did our one.  We’ll leave the rest to others.  Thank you.

MR. KLICK:  Thanks a lot, Phil.  Our last discussant for this panel will be Steve Sailer.  Steve is a journalist, he writes a column for the American Conservative.  He also has the distinction, I think, of being the first person to engage Donohue and Levitt on their results, entering into an Internet debate with Levitt back in 1999.  So we’ll hear now from Steve.

STEVE SAILER:  Hi, I’m Steve Sailer.  I’d like to thank everybody for inviting me here, especially because I’m not an economist, I’m a journalist, to discuss the papers here and offer comments on them.  The statistical analysis is way over my head.  So what I’m going to direct this toward is journalists in general.  I’m very impressed by the work of the economists who put in just tremendous amounts of effort, but I have to say about my fellow journalists who have helped make the “abortion cut crime” theory into conventional wisdom that it’s really not been an impressive effort by us journalists.

There’s a big issue here, which is, what is the burden of proof?  Where does it lie?  Did legalizing abortion drive down the crime rate or did it drive it up, as Dr. Lott is going to suggest?  Or did it have some other effect?  For example, perhaps – if you look at the data, there’s good evidence that it may have driven the homicide rate up in the early years and possibly down in the later years.  All that could be investigated.

But the main point is that large assertions require large evidence.  With about 15 minutes of work on Google, people could figure out that this is still a very live controversy, that we haven’t reached any kind of consensus among social scientists yet.  But if you look at things like all the book reviews for Freakonomics a year ago and so forth, the vast majority of them just accepted it as fact.  Any kind of criticism of the empirical validity of this theory was basically [treated as if it were] driven by right-to-life moralism, and [people make it seem] that there’s really nothing out there factually to argue about anymore.  So I think we can say, as this conference is showing, that for seven years nobody yet has met the burden of proof.

So what I’m going to talk about is how journalists can reality check social science theories.  There’s a couple predilections that social scientists have.  First is a tendency – they like complex statistical models.  For example, we heard a lot here about state-level analysis of data.  Very powerful, but there’s also some problems with it. For example, just because it’s so complicated, mistakes can be made.  There’s more moving parts.  Murphy’s Law comes into effect, as we saw, as Dr. Foote and Dr. Goetz showed last fall, the original Donohue and Levitt analysis had two technical flaws in it which pretty much wiped it out.  Dr. Donohue will come back this afternoon and talk about a new database they’ve got that they believe upholds their original idea.  But it was kind of unfortunate that after all these years in which this theory became just about conventional wisdom, it turned out the original analysis had two major flaws in it.

Economists particularly like simple human behavioral models, such as this theory of wantedness, which is, making abortion legalized should decrease the number of unwanted births, meaning the average child is more wanted.  Wanted children should be raised better and therefore they should have less crime.  This makes sense to social scientists, makes sense to journalists.  It’s kind of like, this is how we think, this is how everybody we know thinks.   However, we’re not typically the parents of the next generation of potential criminals.  People at different levels of society often think very differently from us.

So what journalists could do is simple statistical reality checks.  For example, let’s look at the crime rate at the national level.  Did it follow the “abortion cut crime” theory?  Second one is, we can try to use some more sophisticated empathy, try to put ourselves in the shoes of the people who are really the ones we’re trying to understand here, rather than just assume that the rest of the world behaves like people with graduate degrees.

Okay, let’s go through three reality checks.  Donohue and Levitt said crime started falling about eighteen years after Roe v. Wade.  That’s pretty interesting.  But one obvious question is, did the crime decline happen among those born in the mid to late 1970s?  As we know, abortion was legalized in three big media capitals of New York, Los Angeles, and de facto in Washington, D.C., in 1970.  Then across the country on January 22, 1973.  So big question is, let’s look at a specific age group and see, did their crime rate go down?

Let’s take a look at 14 to 17-year-olds.  This data comes from the federal Bureau of Justice Statistics.  What I’ve done here is I’ve showed – this is when the homicides took place by 14 to 17-year-olds and then subtracting 16 years from there to get the average birth date.  So here’s 1970, when abortion starts to be legalized.  A very crude version of the theory is, well, the crime rate should be going down for 14 to 17-year-olds as abortion works its beneficial effects on who gets born.  As it turns out, boom, it went straight upwards, it basically tripled.  You had your peak years for juvenile murders in 1993 and 1994, which is two decades after Roe v. Wade.

Let’s take a look at a different set of data.  This comes from the FBI’s annual study of 50,000 Americans, asking if they were victims of crimes.  The blue bars are total number of incidents of serious violent crime.  This is not including homicide because it’s kind of hard to interview homicide victims.  But it’s basically other felonies.  This is the total number of incidents in which a 12 to 17-year-old was involved as a criminal.  As you can see, the peak year once again was 1993.  Number two was 1994.  The red line is [the] percentage of total crime and you can see that it peaked out with juvenile serious violent crime being 27 percent of all total violent crime in 1993.  Once again, it’s going the absolute opposite direction in which way it should be going for the “abortion cut crime” theory.

Let’s look at a sub-group: black youth, a very high abortion sub-group.  As you recall, back in September, former Education Secretary William Bennett got into all sorts of trouble for making a philosophical point, saying that abortion wasn’t justified even if just aborting all blacks would reduce the crime rate.  He got into all sorts of trouble like that.  But this is a basic theory here, and we have to admit this has been one of the reasons why the “abortion cut crime” theory has been very popular.  I first heard this argument from my in-laws in Chicago back in the 1980s after they had a couple drinks.

One of the things we can note is crimes took off among non-whites, which back then was mostly blacks, faster than among whites.  The abortion rate among non-whites peaked in 1977, and a few years later for whites.  So the black-to-white ratio should have fallen in the early 1990s among youth.  Dr. Donohue and Levitt logically theorized back in their 2001 paper that legalizing abortion should have driven down the black youth murder rate.  They wrote, “Fertility declines for black women” – this is following legalization – “are three times greater than for whites, 12 percent compared to 4 percent.  Given that homicide rates for black youths are roughly nine times higher than those of white youths, racial differences in the fertility effects of abortion are likely to translate into greater homicide reductions.”  We may not like that kind of argument, but it does make a lot of sense on the surface.

But did it happen?  Here what I’ve done is I’ve taken the black to white homicide ratio among 14 to 17-year-old males, which was typically running around 6 or so among youths born before abortion was legalized, and then after it was legalized the black rate just shot up.  This is of course due to the great crack wars that started largely in the three big cities where abortion was legalized early – LA, New York and Washington, D.C., which are also the media capitals. 

I think future research would be wise to take a look at the effect of things like gangster rap, the West Coast and East Coast gangster rap which started with the “Straight Out of Compton” album in 1988, that really spread the crack dealer glamour across the country.  I think that’s going to have a bigger explanatory value for a lot of these statistics than abortion.

Let’s do reality check number three, the wantedness theory.  The book Freakonomics dropped the racial explanation and just relied solely on wantedness, citing a lot of European studies suggesting abortion increased wantedness of babies born and in general increased the quality of the upbringing.  American studies weren’t cited because the ones that have been done typically came to the opposite conclusion.  They said that young women who got abortions typically tended to be more organized, more future-oriented and so forth, than the ones who just went ahead and had a baby, once you adjust for all else being equal, such as socioeconomic stats.  The European argument made perfect sense to book reviewers of Freakonomics because basically people who review books for a living tend to think like Europeans rather than inner-city Americans.

One obvious check is, let’s talk about wantedness by fathers.  Did legalizing abortion drive down the illegitimacy rate?  That’s typically what abortion is supposed to do, for girls in trouble.  No, as far as we can tell, [it had] no obvious effect whatsoever.  Illegitimacy had been going up since about 1964, when oral contraceptives were introduced, and really shot upwards all the way into the early 1990s.  And still has continued to go up since then.  So [there is] no evidence there whatsoever for the wantedness theory.

Why did Roe have so little impact on this kind of data?  Why isn’t this kind of – whether you want to call it eugenic or culturalist or selectionist logic, which really when we get down to it does make a lot of sense – why didn’t it really work in any way that we can really see in the big picture?  Here’s an important quote.  “Conceptions rose by nearly 30 percent after the legalization of abortion but births actually fell by 6 percent.”  That comes from Dr. Levitt in Freakonomics.  Essentially, legalized abortion was to Americans somewhat like alcohol is to Homer Simpson.  Homer famously said, “To alcohol – the cause of and the cure for all of life’s problems.”  Legal abortion was supposed to be the cure for but then it also turned out to be the cause of tens of millions of unwanted pregnancies.  Basically, as Dr. [Klick’s] study pointed out, as we saw gonorrhea and things like that went up, so people just had more unsafe sex.  So it had this huge impact not on who actually got born but how many people got conceived.

Future research:  I think it’s important to take a look at class differences.  My guess, and this is simplifying extremely, is that the introduction of oral contraceptives in 1964 lowered the fertility of the middle class.  Legalization of abortion in the early 1970s really cut down on fertility of the working class.  And neither had much impact on the fertility of the underclass until maybe 15 or 20 years later.

Then another crucial thing to talk about is the crack wars.  We had a lot of talk about prenatal executions here, that abortion is supposed to eliminate potential murderers and criminals before they’re ever born.  But during the crack wars, when the level of violence in the inner cities just went up enormously, what you’re seeing is a lot of postnatal abortions.  They were extremely selective because it was typically guys with guns shooting other guys with guns.  So as well as a huge increase in the imprisonment rate, you had an enormous increase in the killing of people who might have turned out to be killers.  I think you may well have seen a real hollowing out of the black underclass in 1986 to 1995.  So what you started out with was the underclass became a larger proportion of the black population due to more abortion among black working class and middle class people, and then due to imprisonment, murder and so forth, the underclass started to decline during the crack wars.

One of the real controversial questions is what happened to the crime rate in New York City, which has really fallen.  Jonathan Tilove, a reporter for Newhouse News Service, has pointed out one really important fact, which is that the number of black