Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

The Impact of Legalized Abortion on Crime over the Last Two Decades

The Impact of Legalized Abortion on Crime over the Last Two Decades Abstract Donohue and Levitt (2001) presented evidence that the legalization of abortion in the early 1970s played an important role in the crime drop of the 1990s. That paper concluded with a strong out-of-sample prediction regarding the next two decades: “When a steady state is reached roughly twenty years from now, the impact of abortion will be roughly twice as great as the impact felt so far. Our results suggest that all else equal, legalized abortion will account for persistent declines of 1% a year in crime over the next two decades.” Estimating parallel specifications to the original paper, but using the seventeen years of data generated after that paper was written, we find strong support for the prediction and the broad hypothesis, while illuminating some previously unrecognized patterns of crime and arrests. We estimate that overall crime fell 17.5% from 1998 to 2014 due to legalized abortion—a decline of 1% per year. From 1991 to 2014, the violent and property crime rates each fell by 50%. Legalized abortion is estimated to have reduced violent crime by 47% and property crime by 33% over this period, and thus can explain most of the observed crime decline. 1. Introduction Donohue and Levitt (2001) proposed a link between the legalization of abortion and future crime. The theory motivating that analysis is simple: decades of social scientific research have demonstrated that unwanted children are at an elevated risk for less favorable life outcomes on multiple dimensions including criminal involvement,1 and the legalization of abortion appears to have dramatically reduced the number of unwanted births.2 consequence, cohorts exposed to legalized abortion would be expected to exhibit less criminal behavior than would have been the case absent the legalization of abortion. Using a range of empirical identification strategies, Donohue and Levitt (2001) argued that legalized abortion might well be the single most important factor in reducing crime in the 1990s—perhaps accounting for as much as half of the drop in crime observed in the United States between 1991 and 1997, the endpoint of their data. The claims in Donohue and Levitt (2001) proved to be highly controversial. The research triggered numerous critical academic comments3 and subsequent replies (Dills and Miron 2006; Joyce 2006; Lott and Whitley 2007; Chamlin et al. 2008; Foote and Goetz 2008), as well as numerous extensions consistent with the original findings (Hay and Evans 2006; Sen 2007; François et al. 2014; Shoesmith 2015).4 To this day, there remains a great diversity of views on the merits of the hypothesis among academics. There are a number of reasons why Donohue and Levitt (2001) provoked such a strong academic response.5 First, the magnitude of the results was both large and surprising. At the time, a voluminous academic literature had developed to address the question of understanding fluctuations in crime, including the reasons for the dramatic crime reduction observed during the 1990s. Prior to Donohue and Levitt (2001), there was no mention in this literature of a link between abortion and crime. For a previously unrecognized mechanism to account for possibly half of the largest crime reduction in American history posed a fundamental challenge to the existing scholarship on crime. Second, the evidence presented in Donohue and Levitt (2001) was suggestive, but not definitive. The identification of the estimates was derived neither from a randomized experiment nor even from a credibly exogenous natural experiment (with the possible exception of the 1973 Supreme Court decision in Roe v. Wade, 410 U.S. 113 (1973)). Instead, Donohue and Levitt (2001) presented evidence from a collage of different sources of variation, each of which had its weaknesses.6 Third, the timing of the crack epidemic—which coincided with the peak-crime ages of the first cohort exposed to legalized abortion—increased the difficulty of teasing out the causal impact of legalized abortion. Fourth, at the time, it was rare for economists to posit theories with such long lags between a stimulus (in this case, abortion) and an outcome (in this case, crime roughly two or more decades later).7 All of the existing explanations for fluctuations in crime focused on more proximate causes, e.g., the number of police, expected punishment, or the state of the labor market. Finally, the results of Donohue and Levitt (2001) were based on a short time window of abortion exposure. The original paper could only rely on crime data through 1997 and arrest data through 1996. At that time, the first nationwide cohort of individuals exposed to legalized abortion was only in their early twenties. In the face of these inherent challenges, reasonable people might disagree as to the persuasiveness of the evidence presented in Donohue and Levitt (2001). But the Donohue–Levitt theory makes a strong out-of-sample prediction, which was advanced almost two decades before the full of impact of abortion on crime would be felt. In the conclusion to their paper, Donohue and Levitt wrote: “Roughly half of the crimes committed in the United States are done by individuals born prior to the legalization of abortion. As these older cohorts age out of criminality and are replaced by younger offenders born after abortion became legal, we would predict that crime rates will continue to fall. When a steady state is reached roughly twenty years from now, the impact of abortion will be roughly twice as great as the impact felt so far. Our results suggest that all else equal, legalized abortion will account for persistent declines of 1% a year in crime over the next two decades.” In this article, we analyze the extent to which the nearly 20 years of crime data generated after our analysis was completed support or refute the hypothesized link between abortion and crime. Our methodology is straightforward: we reproduce the primary tables presented in Donohue and Levitt (2001), while extending the data set to cover the period from 1998 to 2014. The choice of specification in the original paper provides a strong degree of discipline on the exercise we carry out. In contrast to the typical empirical economics paper, where the researchers run many specifications and only report a few of those, we constrain ourselves by the choices made in the original paper.8 In addition, we report updated results using specifications suggested by the subsequent exchange between Foote and Goetz (2008) and Donohue and Levitt (2008). The results obtained provide strong support for the hypothesized link between abortion and crime. For most of the specifications reported in the original paper, the point estimates are larger in the out-of-sample 1998–2014 period than in the original publication. This finding is particularly striking because the tables use very different sources of identification (e.g. the natural experiment associated with early legalization, cross-state differences in abortion rates after legalization, within-state differences in crime rates for those born just before or after legalized abortion, etc.). Consequently, it appears that the predictions made in Donohue and Levitt (2001) for the next two decades were borne out. The remainder of the article is as follows. Section 2 provides background on the key legal and institutional factors relating to abortion in the United States and also describes the data used in this analysis. Sections 3 and 4 replicate and extend the results in Donohue and Levitt (2001) to cover the additional years of data from 1998 to 2014. Section 3 uses crime data at the state-year level to estimate the impact of legalized abortion on crime and illustrates the different paths in crime and abortion over our data period for the high- and low-abortion rate states. Section 4 then analyzes age-specific arrest data to link abortion rates of birth cohorts with their arrest rates when aged 15–24. Section 5 shows how the relatively weaker effects of abortion on property crime in our second period is explained by the divergence that emerged over the last 15 years of our data as property crime fell by almost 30% (and perhaps more according to the National Crime Victimization Survey), while property arrests remained at roughly the same level in 2000 and 2014. The section ends with a discussion of, the selective underreporting of crime in low-abortion states identified by Boylan (2019). Section 6 discusses the effect of lead reductions on crime, and Section 7 concludes, placing these findings into the broader context of abortion-related research that has been conducted following our 2001 article. 2. Background and Data For centuries, English common law (and thus early American law) permitted abortion prior to “quickening” (usually in the fourth month of pregnancy). During the 19th century American states began to outlaw the practice, but what ultimately became a uniform nation-wide prohibition started to break down in the United States when then-Governor Ronald Reagan signed the first major liberalizing abortion law in 1967. In late 1969 and 1970, five states (California, New York, Alaska, Hawaii, and Washington) fully legalized abortion. In 1973, the Supreme Court decision in Roe v. Wade made abortion legal nationwide, with the court emphasizing that denying the ability to seek an abortion could impose “a distressful life and future” for the mother and “the distress, for all concerned, associated with [bringing an] unwanted child ... into a family already unable, psychologically and otherwise, to care for it.”9 Roughly 18 years after Roe v. Wade, crime unexpectedly began to fall—even as some of the most prominent criminologists of the day were predicting crime was about to explode.10 From 1991 to 1997, the last year of data used in Donohue and Levitt (2001), violent crime fell 20%, property crime fell 16%, and homicide fell 30%.11 It took almost a decade after legalization for the number of abortions performed to reach a steady state, due to a lack of available providers as well as evolving norms. As the rate of abortion rose and cohorts born subsequent to legalization moved through their high-crime years, crime continued to decline. Between 1997 and 2014 (the last year of data included in our analysis), violent and property crime per capita fell by 40% and homicide declined by 35%, according to the Uniform Crime Reports. The path to legalization did not create particularly compelling quasi-randomized variation in abortion exposure. Although the early legalization of abortion in five states might appear to serve as a natural experiment, these five states are clear outliers. Even after steady state abortion rates are reached in the 1980s, the effective abortion rate for violent crime in the early-legalizing states are almost double those in the rest of the nation. Indeed, one striking feature of the data is the enormous heterogeneity in abortion usage across states. Dividing states into thirds according to the number of abortions per live birth, the 17 states with the lowest abortion rates have a steady-state effective abortion rate for violent crime that is roughly one-half that of the middle 17 states and one-third that of the 17 states with the highest abortion rates. As a consequence, Donohue and Levitt (2001) relied on a collage of individually imperfect sources of variation in an effort to discern the causal impact of abortion on crime. These consisted of a comparison of early-legalizing states to the rest of the country, a comparison of states with high- and low-abortion rates after abortion became legal everywhere, differences in crime patterns within states for cohorts born before and after legalization, and differences in arrest rates within states by single year of age. Because the impact of abortion on crime is not expected to be immediate, but rather is only felt when cohorts exposed to a legalized abortion regime in utero reach an age at which crimes are committed, the specification of the abortion measure is more complex than that of other variables in a typical panel data model of crime.12 At least initially, the expected impact of abortion on crime increases gradually as more and more of the crime-age cohorts have been exposed to legalized abortion and as these cohorts transition from the relatively low-crime ages of the early teenage years to the peak-crime ages in the late teens and early twenties. Thus, the hypothesized impact of abortion on crime emerges only incrementally; the full impact is not felt for many decades. To capture the extent to which legalized abortion would be expected to influence crime in a given state and year, Donohue and Levitt (2001) developed a metric they named the “effective abortion rate” per 1,000 live births. The “effective abortion rate” is the weighted average of the abortion rates of the birth cohorts in a state, with the weights determined by the 1985 share of total arrests nationally for a particular crime category of individuals of that age. More formally, $$ \begin{equation} \label{eq:1} { Effective \; abortion}_{t} = \sum_{a}{{ Abortion}_{t-a}} \cdot \bigg( \frac{{ Arrests}_{a}}{{ Arrests}_{ total}} \bigg), \end{equation} $$(1) where |$t$| indexes years and |$a$| indexes the age of a cohort. Abortion is the number of abortions per 1,000 live births, and the 1985 ratio of arrests inside the parentheses is the fraction of arrests for a given crime involving individuals with age |$a$|⁠.13 In a steady state with all cohorts subjected to the same abortion rate, the effective abortion rate is equal to the actual abortion rate. For many years following the introduction of legalized abortion, the effective abortion rate will be below the actual abortion rate since many active criminal cohorts are too old to have been affected by legalized abortion. For instance, following Roe v. Wade, the actual abortion rate (per 1,000 live births) rose to a steady state of about 368 by the early 1980s. Yet, we estimate that the effective abortion rate in 1991 was only about 24 for homicide, 51 for violent crime, and 110 for property crime. Because property crime is disproportionately done by the young, the effect of abortion legalization is felt earlier for that crime category. The effective rates grew steadily, rising to 132, 170, and 247, respectively, by 1997. In 2014, the effective abortion rates for these three crime categories had risen to 330, 341, and 338, respectively. If legalized abortion reduces crime, then crime should continue to fall (all else equal) as long as the effective abortion rate is rising. Throughout this article, we attempt to mirror the specifications of Donohue and Levitt (2001) as closely as possible, in order to tie our hands with respect to ex post facto model selection. We follow one major upgrade in our core abortion data that we initiated in 2004. In our original 2001 paper, we used abortion data that reflected the state in which an abortion was performed. This was less than ideal for our purposes because a substantial number of women travel across state lines to have an abortion. A much more natural metric for constructing an abortion rate would use the mother’s state of residence.14 This latter measure only became available from the Alan Guttmacher Institute after our initial research was published. We have consistently used this abortion by state of residence measure since it became available (see Donohue and Levitt 2004; Donohue and Levitt 2008; Donohue et al. 2009) and continue to do so in this article.15 3. Results Revealing the Abortion-Crime Link 3.1. Crime Fell Earlier and Further for the Five Early-Legalizing States We begin by looking at the patterns of crime in the five states (Alaska, California, Hawaii, New York, and Washington) that legalized or quasi-legalized abortion around 1970 relative to crime patterns in the rest of the nation where abortion did not become legal until the Supreme Court decision in Roe v. Wade of January 1973. Table 1 provides an updated version of Table I in Donohue and Levitt (2001). For each crime category (violent, property, and two measures of murder), we present percent changes in crime between 1976 and 1982, between 1982 and 1997, and between 1997 and 2014 for early-legalizers and the rest of the country. We then show the difference in these percent changes in crime between early-legalizing states and the rest of the nation. The first two columns correspond to data available in Donohue and Levitt (2001). The third and fourth columns, which report how crime changed in the early-legalizers versus the rest of the country in the period 1997–2014 and cumulatively between 1982 and 2014, are new. Table 1. Crime Trends for States Legalizing Abortion Early vs. the Rest of the US Natural Log of Differences in Crime Rates over Various Periods . . . 1997–2014 . Cumulative . Legalization Group . 1976–82 . 1982–97 . (New) . (1982–2014) . Violent crime   Early legalizers 15.8 -12.9 -61.7 -74.7   Rest of U.S. 20.9 14.5 -41.7 -27.1   Difference -5.1 -27.5 -20.1 -47.5   SE 5.1 7.3 8.6 11.8   P-value 0.3 0.0 0.0 0.0 Property crime   Early legalizers 0.8 -44.3 -54.4 -98.6   Rest of U.S. 5.2 -9.5 -52.3 -61.8   Difference -4.3 -34.7 -2.1 -36.8   SE 2.7 5.7 4.8 8.8   P-value 0.1 0.0 0.7 0.0 Murder (UCR)   Early legalizers 5.4 -40.8 -62.4 -103.3   Rest of U.S. 0.2 -24.7 -33.1 -57.7   Difference 5.3 -16.2 -29.3 -45.5   SE 7.3 10.7 6.9 11.4   P-value 0.5 0.1 0.0 0.0 Murder (VS)   Early legalizers 8.4 -38.3 -58.3 -96.7   Rest of U.S. 4.2 -24.6 -27.3 -51.9   Difference 4.2 -13.7 -31.1 -44.8   SE 6.1 9.9 6.1 10.4   P-value 0.5 0.2 0.0 0.0 Effective abortion rate at end of period   Early legalizers 1.6 281.0 514.4 514.4   Rest of U.S. 0.1 139.4 294.6 294.6   Difference 1.5 141.6 219.8 219.8 . . . 1997–2014 . Cumulative . Legalization Group . 1976–82 . 1982–97 . (New) . (1982–2014) . Violent crime   Early legalizers 15.8 -12.9 -61.7 -74.7   Rest of U.S. 20.9 14.5 -41.7 -27.1   Difference -5.1 -27.5 -20.1 -47.5   SE 5.1 7.3 8.6 11.8   P-value 0.3 0.0 0.0 0.0 Property crime   Early legalizers 0.8 -44.3 -54.4 -98.6   Rest of U.S. 5.2 -9.5 -52.3 -61.8   Difference -4.3 -34.7 -2.1 -36.8   SE 2.7 5.7 4.8 8.8   P-value 0.1 0.0 0.7 0.0 Murder (UCR)   Early legalizers 5.4 -40.8 -62.4 -103.3   Rest of U.S. 0.2 -24.7 -33.1 -57.7   Difference 5.3 -16.2 -29.3 -45.5   SE 7.3 10.7 6.9 11.4   P-value 0.5 0.1 0.0 0.0 Murder (VS)   Early legalizers 8.4 -38.3 -58.3 -96.7   Rest of U.S. 4.2 -24.6 -27.3 -51.9   Difference 4.2 -13.7 -31.1 -44.8   SE 6.1 9.9 6.1 10.4   P-value 0.5 0.2 0.0 0.0 Effective abortion rate at end of period   Early legalizers 1.6 281.0 514.4 514.4   Rest of U.S. 0.1 139.4 294.6 294.6   Difference 1.5 141.6 219.8 219.8 Early legalizing states are Alaska, California, Hawaii, New York, and Washington. These five states legalized abortion in late 1969 or 1970. In the remaining states, abortion became legal in 1973 after Roe v. Wade. Percent change in crime rate is calculated by subtracting the fixed 1985 population-weighted average of the natural log of crime rate at the beginning of the period from the fixed 1985 population-weighted average of the natural log of crime rate at the end of the period. The rows labeled “Difference” are the difference between early-legalizers and the rest of the United States. The bottom panel of the table presents the effective abortion rate for violent crime, as calculated using equation (1), based on the observed age distribution of national arrests for violent crime in 1985. Entries in the table are fixed 1985 population-weighted averages of the states. Abortion data are from the Alan Guttmacher Institute (by mother’s state of residence); crime data are from the Uniform Crime Reports or the National Vital Statistics System. Open in new tab Table 1. Crime Trends for States Legalizing Abortion Early vs. the Rest of the US Natural Log of Differences in Crime Rates over Various Periods . . . 1997–2014 . Cumulative . Legalization Group . 1976–82 . 1982–97 . (New) . (1982–2014) . Violent crime   Early legalizers 15.8 -12.9 -61.7 -74.7   Rest of U.S. 20.9 14.5 -41.7 -27.1   Difference -5.1 -27.5 -20.1 -47.5   SE 5.1 7.3 8.6 11.8   P-value 0.3 0.0 0.0 0.0 Property crime   Early legalizers 0.8 -44.3 -54.4 -98.6   Rest of U.S. 5.2 -9.5 -52.3 -61.8   Difference -4.3 -34.7 -2.1 -36.8   SE 2.7 5.7 4.8 8.8   P-value 0.1 0.0 0.7 0.0 Murder (UCR)   Early legalizers 5.4 -40.8 -62.4 -103.3   Rest of U.S. 0.2 -24.7 -33.1 -57.7   Difference 5.3 -16.2 -29.3 -45.5   SE 7.3 10.7 6.9 11.4   P-value 0.5 0.1 0.0 0.0 Murder (VS)   Early legalizers 8.4 -38.3 -58.3 -96.7   Rest of U.S. 4.2 -24.6 -27.3 -51.9   Difference 4.2 -13.7 -31.1 -44.8   SE 6.1 9.9 6.1 10.4   P-value 0.5 0.2 0.0 0.0 Effective abortion rate at end of period   Early legalizers 1.6 281.0 514.4 514.4   Rest of U.S. 0.1 139.4 294.6 294.6   Difference 1.5 141.6 219.8 219.8 . . . 1997–2014 . Cumulative . Legalization Group . 1976–82 . 1982–97 . (New) . (1982–2014) . Violent crime   Early legalizers 15.8 -12.9 -61.7 -74.7   Rest of U.S. 20.9 14.5 -41.7 -27.1   Difference -5.1 -27.5 -20.1 -47.5   SE 5.1 7.3 8.6 11.8   P-value 0.3 0.0 0.0 0.0 Property crime   Early legalizers 0.8 -44.3 -54.4 -98.6   Rest of U.S. 5.2 -9.5 -52.3 -61.8   Difference -4.3 -34.7 -2.1 -36.8   SE 2.7 5.7 4.8 8.8   P-value 0.1 0.0 0.7 0.0 Murder (UCR)   Early legalizers 5.4 -40.8 -62.4 -103.3   Rest of U.S. 0.2 -24.7 -33.1 -57.7   Difference 5.3 -16.2 -29.3 -45.5   SE 7.3 10.7 6.9 11.4   P-value 0.5 0.1 0.0 0.0 Murder (VS)   Early legalizers 8.4 -38.3 -58.3 -96.7   Rest of U.S. 4.2 -24.6 -27.3 -51.9   Difference 4.2 -13.7 -31.1 -44.8   SE 6.1 9.9 6.1 10.4   P-value 0.5 0.2 0.0 0.0 Effective abortion rate at end of period   Early legalizers 1.6 281.0 514.4 514.4   Rest of U.S. 0.1 139.4 294.6 294.6   Difference 1.5 141.6 219.8 219.8 Early legalizing states are Alaska, California, Hawaii, New York, and Washington. These five states legalized abortion in late 1969 or 1970. In the remaining states, abortion became legal in 1973 after Roe v. Wade. Percent change in crime rate is calculated by subtracting the fixed 1985 population-weighted average of the natural log of crime rate at the beginning of the period from the fixed 1985 population-weighted average of the natural log of crime rate at the end of the period. The rows labeled “Difference” are the difference between early-legalizers and the rest of the United States. The bottom panel of the table presents the effective abortion rate for violent crime, as calculated using equation (1), based on the observed age distribution of national arrests for violent crime in 1985. Entries in the table are fixed 1985 population-weighted averages of the states. Abortion data are from the Alan Guttmacher Institute (by mother’s state of residence); crime data are from the Uniform Crime Reports or the National Vital Statistics System. Open in new tab As noted above, these five early-legalizing states not only legalized abortion early, but continued to have higher abortion rates throughout the period. The bottom panel of the table presents the effective abortion rate for violent crime for the two sets of states at the end of each time period, calculated using equation (1). The gap in the effective abortion rate between the early-legalizers and the rest of the country has continued to grow over the entire time period, albeit more slowly in the later period. In 1997, the difference in the effective abortion rate between these two sets of states was 141.6; by 2014 the difference had increased to 219.8. Our theory predicts no difference in crime patterns across early-legalizers and the rest of the country prior to 1982 (just before the first abortion-exposed cohort in the early-legalizing states first reaches a crime-committing age), but greater decreases in crime for all periods since then. The results in Table 1 confirm that prediction. Prior to 1982, there are no statistically different crime trends across early-legalizing and all other states. Property and violent crime were increasing at a slower rate in early-legalizing states between 1976 and 1982, whereas murder was rising faster in early-legalizing states, whether measured by UCR or Vital Statistics data.16 Between 1982 and 1997, violent crime fell by 27.5% (⁠|${\rm se} = 7.3$|⁠) in early-legalizing states relative to the rest of the country. The parallel numbers for property crime and homicide are |$-$|34.7% (⁠|${\rm se} = 5.7$|⁠) and |$-$|16.2% (⁠|${\rm se} = 10.7$|⁠) with UCR data and |$-$|13.7% (⁠|${\rm se} = 9.9$|⁠) using Vital Statistics data. Of greatest interest, however, are the new results presented in column 3. Violent crime fell by an additional 20.1% (⁠|${\rm se} = 8.6$|⁠) in early-legalizing states relative to the rest of the nation between 1997 and 2014. While the difference in property crime was not statistically significant over the recent time period (⁠|$-$|2.1%; |${\rm se} = 4.8$|⁠); the gap in homicide was large (roughly |$-$|30% with either measure) and highly significant. The cumulative differences across the entire time period are enormous and highly statistically significant for all four crime measures, |$-$|47.5% (⁠|${\rm se} = 11.8$|⁠) and |$-$|36.8% (⁠|${\rm se} = 8.8$|⁠) for violent crime and property crime, respectively, and |$-$|45.5% (⁠|${\rm se} = 11.4$|⁠) and |$-$|44.8% (⁠|${\rm se} = 10.4$|⁠) for our two measures of homicide. 3.2. Crime Fell More in High-Abortion States Than in Low-Abortion States A second source of variation for identifying a link between abortion and crime is a comparison of crime patterns across states with differing levels of abortion usage post-legalization. Following Donohue and Levitt (2001), we rank order states by their effective abortion rates for violent crime in 1997 and partition the states into three categories with equal numbers of states in each category: low, medium, and high.17 The three top panels of Table 2 report the percent changes in high, medium, and low-abortion states for violent crime, property crime, and homicide respectively, for the periods 1973|$-$|85, 1985–97, and 1997–2014. The bottom panel of the table reports the mean effective abortion rate at the end of the relevant period for the three groups of states. Table 2. Crime Changes 1985–2014 in Low-, Medium-, and High-Abortion Rate States . . . 1997–2014 . Cumulative . Abortion Frequency (1997) . 1973–85 . 1985–97 . (New) . (1985–2014) . Violent crime   Lowest 32.9 26.3 -23.3 3.1   Medium 28.5 20.6 -36.2 -15.6   Highest 28.6 -1.5 -60.7 -62.2 Property crime   Lowest 33.6 10.8 -42.1 -31.2   Medium 27.4 2.9 -45.7 -42.8   Highest 13.2 -22.2 -61.5 -83.7 Murder (UCR)   Lowest -23.5 7.4 -32.3 -24.9   Medium -20.8 -12.7 -32.4 -45.2   Highest -11.9 -25.3 -46.7 -71.9 Murder (VS)   Lowest -17.0 13.7 -27.0 -13.3   Medium -22.8 -12.3 -25.5 -37.7   Highest -7.4 -23.6 -42.3 -65.9 Effective abortion per 1,000 at end of period   Lowest 0.8 77.0 179.8 179.8   Medium 1.4 125.6 265.6 265.6   Highest 5.4 232.6 450.8 450.8 . . . 1997–2014 . Cumulative . Abortion Frequency (1997) . 1973–85 . 1985–97 . (New) . (1985–2014) . Violent crime   Lowest 32.9 26.3 -23.3 3.1   Medium 28.5 20.6 -36.2 -15.6   Highest 28.6 -1.5 -60.7 -62.2 Property crime   Lowest 33.6 10.8 -42.1 -31.2   Medium 27.4 2.9 -45.7 -42.8   Highest 13.2 -22.2 -61.5 -83.7 Murder (UCR)   Lowest -23.5 7.4 -32.3 -24.9   Medium -20.8 -12.7 -32.4 -45.2   Highest -11.9 -25.3 -46.7 -71.9 Murder (VS)   Lowest -17.0 13.7 -27.0 -13.3   Medium -22.8 -12.3 -25.5 -37.7   Highest -7.4 -23.6 -42.3 -65.9 Effective abortion per 1,000 at end of period   Lowest 0.8 77.0 179.8 179.8   Medium 1.4 125.6 265.6 265.6   Highest 5.4 232.6 450.8 450.8 States are ranked by effective abortion rates for violent crime in 1997, with the 17 states with lowest abortion rates classified as “lowest,” the next 17 states classified as “medium,” and the highest 17 states (including District of Columbia) classified as “highest.” The effective abortion rate is the weighted average abortion rate per 1,000 live births (number of abortions per state according to mother’s state of residence), as calculated using equation (1), weighted using the observed age distribution of national arrests for violent crime in 1985. All values in the table are weighted averages using 1985 state populations as weights. Percent change in crime rate is calculated by subtracting the fixed 1985 population-weighted average of the natural log of crime rate at the beginning of the period from the fixed 1985 population-weighted average of the natural log of crime rate at the end of the period. Because crime rates are extremely low until the mid-teenage years, legalized abortion is not predicted to have had a substantial impact on crime in the period 1973–85, but would be predicted to affect crime in the period 1985–2014. Abortion data are from the Alan Guttmacher Institute; crime data are from Uniform Crime Reports or the National Vital Statistics System. Open in new tab Table 2. Crime Changes 1985–2014 in Low-, Medium-, and High-Abortion Rate States . . . 1997–2014 . Cumulative . Abortion Frequency (1997) . 1973–85 . 1985–97 . (New) . (1985–2014) . Violent crime   Lowest 32.9 26.3 -23.3 3.1   Medium 28.5 20.6 -36.2 -15.6   Highest 28.6 -1.5 -60.7 -62.2 Property crime   Lowest 33.6 10.8 -42.1 -31.2   Medium 27.4 2.9 -45.7 -42.8   Highest 13.2 -22.2 -61.5 -83.7 Murder (UCR)   Lowest -23.5 7.4 -32.3 -24.9   Medium -20.8 -12.7 -32.4 -45.2   Highest -11.9 -25.3 -46.7 -71.9 Murder (VS)   Lowest -17.0 13.7 -27.0 -13.3   Medium -22.8 -12.3 -25.5 -37.7   Highest -7.4 -23.6 -42.3 -65.9 Effective abortion per 1,000 at end of period   Lowest 0.8 77.0 179.8 179.8   Medium 1.4 125.6 265.6 265.6   Highest 5.4 232.6 450.8 450.8 . . . 1997–2014 . Cumulative . Abortion Frequency (1997) . 1973–85 . 1985–97 . (New) . (1985–2014) . Violent crime   Lowest 32.9 26.3 -23.3 3.1   Medium 28.5 20.6 -36.2 -15.6   Highest 28.6 -1.5 -60.7 -62.2 Property crime   Lowest 33.6 10.8 -42.1 -31.2   Medium 27.4 2.9 -45.7 -42.8   Highest 13.2 -22.2 -61.5 -83.7 Murder (UCR)   Lowest -23.5 7.4 -32.3 -24.9   Medium -20.8 -12.7 -32.4 -45.2   Highest -11.9 -25.3 -46.7 -71.9 Murder (VS)   Lowest -17.0 13.7 -27.0 -13.3   Medium -22.8 -12.3 -25.5 -37.7   Highest -7.4 -23.6 -42.3 -65.9 Effective abortion per 1,000 at end of period   Lowest 0.8 77.0 179.8 179.8   Medium 1.4 125.6 265.6 265.6   Highest 5.4 232.6 450.8 450.8 States are ranked by effective abortion rates for violent crime in 1997, with the 17 states with lowest abortion rates classified as “lowest,” the next 17 states classified as “medium,” and the highest 17 states (including District of Columbia) classified as “highest.” The effective abortion rate is the weighted average abortion rate per 1,000 live births (number of abortions per state according to mother’s state of residence), as calculated using equation (1), weighted using the observed age distribution of national arrests for violent crime in 1985. All values in the table are weighted averages using 1985 state populations as weights. Percent change in crime rate is calculated by subtracting the fixed 1985 population-weighted average of the natural log of crime rate at the beginning of the period from the fixed 1985 population-weighted average of the natural log of crime rate at the end of the period. Because crime rates are extremely low until the mid-teenage years, legalized abortion is not predicted to have had a substantial impact on crime in the period 1973–85, but would be predicted to affect crime in the period 1985–2014. Abortion data are from the Alan Guttmacher Institute; crime data are from Uniform Crime Reports or the National Vital Statistics System. Open in new tab There should be little or no impact of abortion on crime prior to 1985, because effective abortion rates are extremely low in 1985, even in high-abortion states. The results in column 1 for 1973–85 are consistent with that conjecture. Violent crime rate patterns are very similar across low, medium, and high-abortion states. Property crime rises less in high-abortion states than low-abortion states, but the opposite pattern is true for homicide, where crime declines are smallest in the high-abortion states. The crime changes between 1985 and 1997 reveal a very different pattern. For each crime category, high-abortion states experience more favorable crime trends than medium abortion states, with low-abortion states faring the worst. For all three crime categories, the difference between high-abortion states and low-abortion states is |$\sim$|30 percentage points, and somewhat larger for VS murder. Column 3 of Table 2 presents results for the time period that post-dates the publication of the original paper. Across all three crime categories, the decline in crime is considerably greater for the high-abortion states than for the low-abortion states, as theory would predict. The magnitude of the differences are substantial: violent crime has fallen an additional 35 percentage points since 1997 in high-abortion states relative to low-abortion states. For property crime that difference is over 19 percentage points, and for homicide it is |$\sim$|15 percentage points. Aggregating over the entire time period 1985 to 2014 (Column 4), high-abortion states have experienced a reduction in crime relative to low-abortion states of |$-$|65.3, |$-$|52.5, and |$-$|47.0 percentage points for UCR violent crime, property crime, and homicide, respectively. With Vital Statistics data, the homicide differential is |$-$|52.6 percentage points. 3.3. Abortion is Highly Significant in Explaining Crime Reductions in Panel Data A third source of variation comes from panel data analysis that allows us to control for other factors, in addition to abortion rates, that influence crime. The specification estimated takes the form: $$ \begin{equation}\label{eq:2} ln({ CRIME}_{st}) = \beta_{1}{ ABORT}_{st} + X_{st}\Theta + \gamma_{s} + \lambda_{t} + \epsilon_{st}. \end{equation} $$(2) The dependent variable is the respective logged per capita crime rate in state |$s$| at time |$t$|⁠. Our main independent variable of interest is the effective abortion rate for a given state, year, and crime category.18|$X$| is a vector of state-level controls, including prisoners and police per capita, a set of variables capturing state economic conditions, lagged state welfare generosity, an indicator for the presence of concealed handgun laws, and per capita beer consumption. Both state and year fixed effects are included, represented by |$\gamma_{s}$| and |$\lambda_{t}$|⁠, respectively. All regressions are weighted by state population and adjusted for serial correlation using the method outlined by Bhargava et al. (1982). Summary statistics for the full estimating sample are provided in Table 3.19 We present both overall standard deviations and within-state standard deviations, which is the more relevant measure since state-fixed effects are included in all specifications. The effective abortion rates differ across crime categories because the age distribution of arrests differs across crimes. Table 3. Summary Statistics, 1985–2014 . . Standard Deviation . Standard Deviation . Variable . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 540.93 238.43 156.97 Property crime per 100,000 residents 3,882.96 1,215.86 968.50 Murder per 100,000 residents (UCR) 6.59 3.60 2.33 Murder per 100,000 residents (VS) 7.03 3.53 2.25 EAR: Violent crime 203.64 152.38 128.47 EAR: Property crime 240.93 151.27 117.39 EAR: Murder 179.33 148.15 128.97 Prisoners per 1,000 residents (⁠|$t-1$|⁠) 3.83 1.61 1.05 Police per 1,000 residents (⁠|$t-1$|⁠) 3.08 0.71 0.37 Real state personal income per capita 17,045.93 2,914.40 1,942.68 Real AFDC generosity per recipient family/1,000 (⁠|$t-15$|⁠) 3.76 1.73 1.05 State unemployment rate (%) 6.20 1.92 1.73 Beer consumption per capita (Gallons of ethanol) 1.22 0.19 0.09 Poverty rate 13.47 3.24 1.77 . . Standard Deviation . Standard Deviation . Variable . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 540.93 238.43 156.97 Property crime per 100,000 residents 3,882.96 1,215.86 968.50 Murder per 100,000 residents (UCR) 6.59 3.60 2.33 Murder per 100,000 residents (VS) 7.03 3.53 2.25 EAR: Violent crime 203.64 152.38 128.47 EAR: Property crime 240.93 151.27 117.39 EAR: Murder 179.33 148.15 128.97 Prisoners per 1,000 residents (⁠|$t-1$|⁠) 3.83 1.61 1.05 Police per 1,000 residents (⁠|$t-1$|⁠) 3.08 0.71 0.37 Real state personal income per capita 17,045.93 2,914.40 1,942.68 Real AFDC generosity per recipient family/1,000 (⁠|$t-15$|⁠) 3.76 1.73 1.05 State unemployment rate (%) 6.20 1.92 1.73 Beer consumption per capita (Gallons of ethanol) 1.22 0.19 0.09 Poverty rate 13.47 3.24 1.77 All values reported are means of annual, state-level observations for the period 1985–2014 with the following exceptions. In 1996, there was a transition from the annual program AFDC to TANF. From 1998 onwards, the AFDC variable reflects TANF assistance. It is lagged by 15 years, measured in thousands of dollars and indexed at 1982–84 values. The police and prisoners data are both logged and once-lagged, so correspond to the years 1984–2013. The values reported in the table are population-weighted averages. The effective abortion rate is a weighted average of the abortion rate for each cohort born in a state, with weights determined by the percentage of arrests by age for a given crime category in the United States in 1985 as shown by equation (1). Open in new tab Table 3. Summary Statistics, 1985–2014 . . Standard Deviation . Standard Deviation . Variable . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 540.93 238.43 156.97 Property crime per 100,000 residents 3,882.96 1,215.86 968.50 Murder per 100,000 residents (UCR) 6.59 3.60 2.33 Murder per 100,000 residents (VS) 7.03 3.53 2.25 EAR: Violent crime 203.64 152.38 128.47 EAR: Property crime 240.93 151.27 117.39 EAR: Murder 179.33 148.15 128.97 Prisoners per 1,000 residents (⁠|$t-1$|⁠) 3.83 1.61 1.05 Police per 1,000 residents (⁠|$t-1$|⁠) 3.08 0.71 0.37 Real state personal income per capita 17,045.93 2,914.40 1,942.68 Real AFDC generosity per recipient family/1,000 (⁠|$t-15$|⁠) 3.76 1.73 1.05 State unemployment rate (%) 6.20 1.92 1.73 Beer consumption per capita (Gallons of ethanol) 1.22 0.19 0.09 Poverty rate 13.47 3.24 1.77 . . Standard Deviation . Standard Deviation . Variable . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 540.93 238.43 156.97 Property crime per 100,000 residents 3,882.96 1,215.86 968.50 Murder per 100,000 residents (UCR) 6.59 3.60 2.33 Murder per 100,000 residents (VS) 7.03 3.53 2.25 EAR: Violent crime 203.64 152.38 128.47 EAR: Property crime 240.93 151.27 117.39 EAR: Murder 179.33 148.15 128.97 Prisoners per 1,000 residents (⁠|$t-1$|⁠) 3.83 1.61 1.05 Police per 1,000 residents (⁠|$t-1$|⁠) 3.08 0.71 0.37 Real state personal income per capita 17,045.93 2,914.40 1,942.68 Real AFDC generosity per recipient family/1,000 (⁠|$t-15$|⁠) 3.76 1.73 1.05 State unemployment rate (%) 6.20 1.92 1.73 Beer consumption per capita (Gallons of ethanol) 1.22 0.19 0.09 Poverty rate 13.47 3.24 1.77 All values reported are means of annual, state-level observations for the period 1985–2014 with the following exceptions. In 1996, there was a transition from the annual program AFDC to TANF. From 1998 onwards, the AFDC variable reflects TANF assistance. It is lagged by 15 years, measured in thousands of dollars and indexed at 1982–84 values. The police and prisoners data are both logged and once-lagged, so correspond to the years 1984–2013. The values reported in the table are population-weighted averages. The effective abortion rate is a weighted average of the abortion rate for each cohort born in a state, with weights determined by the percentage of arrests by age for a given crime category in the United States in 1985 as shown by equation (1). Open in new tab Regression results are shown in Table 4. The dependent variable in columns 1 and 2 is (logged) violent crime. Columns 3 and 4 present results for (logged) property crime, while columns 5 and 6 reflect (logged) UCR homicide and columns 7 and 8 reflect (logged) VS homicide. For each of the four crime measures, two different specifications are reported. The odd-numbered columns present results without control variables (other than the state- and year-fixed effects); the even columns add the full set of controls. The top two rows present the two effective abortion rate measures, one corresponding to the time period included in our earlier study and the other capturing the period since that time. In Appendix B, we also report a modified version of Table 4 that estimates a single abortion variable for the entire 1985–2014 time period, instead of estimates for the two separate time periods. Table 4. Panel-data Estimates of the Relationship between Abortion Rates and Crime, 1985–2014 . Dependent variable: Log per capita value of... . . Violent crime . Property crime . UCR murder . VS murder . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Effective abortion rate ’85–’97 -0.184** -0.178** -0.138** -0.152** -0.087* -0.100* -0.098** -0.116** (0.022) (0.022) (0.017) (0.016) (0.038) (0.040) (0.034) (0.036) Effective abortion rate ’98–’14 -0.192** -0.189** -0.149** -0.168** -0.131** -0.152** -0.141** -0.164** (0.019) (0.019) (0.016) (0.015) (0.017) (0.021) (0.016) (0.019) ln(lagged prisoners per capita) 0.007 -0.111** -0.121* -0.133** (0.037) (0.034) (0.056) (0.051) ln(lagged police per capita) -0.015 -0.027 -0.137* -0.186** (0.015) (0.014) (0.053) (0.049) Unemployment rate -0.027 0.536 1.212 1.084 (0.356) (0.314) (0.716) (0.657) Ln(Real per capita income) 0.003 -0.076 0.329 0.172   1982–84 $\$$ (0.129) (0.114) (0.224) (0.203) Poverty rate -0.002 -0.001 -0.003 -0.001 (0.001) (0.001) (0.003) (0.003) Real AFDC generosity 0.004 -0.001 -0.008 -0.003   1982-1984 $\$$ (0.003) (0.003) (0.008) (0.007) Shall-issue concealed weapons 0.014 0.019 -0.042 -0.023   law (0.015) (0.011) (0.023) (0.022) Beer consumption per capita 0.077 0.026 0.286** 0.286** (0.050) (0.044) (0.106) (0.096) Year FE? Yes Yes Yes Yes Yes Yes Yes Yes State FE? Yes Yes Yes Yes Yes Yes Yes Yes Observations 1,530 1,517 1,530 1,517 1,530 1,517 1,512 1,499 R|$^{2}$| 0.815 0.844 0.974 0.980 0.853 0.865 0.877 0.887 Adjusted R|$^{2}$| 0.805 0.834 0.973 0.978 0.845 0.857 0.870 0.880 . Dependent variable: Log per capita value of... . . Violent crime . Property crime . UCR murder . VS murder . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Effective abortion rate ’85–’97 -0.184** -0.178** -0.138** -0.152** -0.087* -0.100* -0.098** -0.116** (0.022) (0.022) (0.017) (0.016) (0.038) (0.040) (0.034) (0.036) Effective abortion rate ’98–’14 -0.192** -0.189** -0.149** -0.168** -0.131** -0.152** -0.141** -0.164** (0.019) (0.019) (0.016) (0.015) (0.017) (0.021) (0.016) (0.019) ln(lagged prisoners per capita) 0.007 -0.111** -0.121* -0.133** (0.037) (0.034) (0.056) (0.051) ln(lagged police per capita) -0.015 -0.027 -0.137* -0.186** (0.015) (0.014) (0.053) (0.049) Unemployment rate -0.027 0.536 1.212 1.084 (0.356) (0.314) (0.716) (0.657) Ln(Real per capita income) 0.003 -0.076 0.329 0.172   1982–84 $\$$ (0.129) (0.114) (0.224) (0.203) Poverty rate -0.002 -0.001 -0.003 -0.001 (0.001) (0.001) (0.003) (0.003) Real AFDC generosity 0.004 -0.001 -0.008 -0.003   1982-1984 $\$$ (0.003) (0.003) (0.008) (0.007) Shall-issue concealed weapons 0.014 0.019 -0.042 -0.023   law (0.015) (0.011) (0.023) (0.022) Beer consumption per capita 0.077 0.026 0.286** 0.286** (0.050) (0.044) (0.106) (0.096) Year FE? Yes Yes Yes Yes Yes Yes Yes Yes State FE? Yes Yes Yes Yes Yes Yes Yes Yes Observations 1,530 1,517 1,530 1,517 1,530 1,517 1,512 1,499 R|$^{2}$| 0.815 0.844 0.974 0.980 0.853 0.865 0.877 0.887 Adjusted R|$^{2}$| 0.805 0.834 0.973 0.978 0.845 0.857 0.870 0.880 Note: |$^{*}\textit{P} < 0.05; ^{**}\textit{P} < 0.01$|⁠. The dependent variable is the log in the per capita crime rate named at the top of each pair of columns. The first column in each pair presents results from the specifications in which the only additional covariates are state- and year-fixed effects. The second column presents results using the full specification. The data set is comprised of annual state-level observations (including the District of Columbia) for the period 1985–2014. State- and year-fixed effects are included in all specifications. The prison and police variables are once-lagged to minimize endogeneity. Real AFDC generosity per recipient family is lagged by 15 years, measured in thousands of dollars and indexed at 1982–84 values. Estimation is performed using a two-step procedure. In the first step, weighted least squares estimates are obtained, with weights determined by state population. In the second step, a panel data generalization of the Prais–Winsten correction for serial correlation developed by Bhargava et al. (1982) is implemented. Standard errors are in parentheses. Open in new tab Table 4. Panel-data Estimates of the Relationship between Abortion Rates and Crime, 1985–2014 . Dependent variable: Log per capita value of... . . Violent crime . Property crime . UCR murder . VS murder . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Effective abortion rate ’85–’97 -0.184** -0.178** -0.138** -0.152** -0.087* -0.100* -0.098** -0.116** (0.022) (0.022) (0.017) (0.016) (0.038) (0.040) (0.034) (0.036) Effective abortion rate ’98–’14 -0.192** -0.189** -0.149** -0.168** -0.131** -0.152** -0.141** -0.164** (0.019) (0.019) (0.016) (0.015) (0.017) (0.021) (0.016) (0.019) ln(lagged prisoners per capita) 0.007 -0.111** -0.121* -0.133** (0.037) (0.034) (0.056) (0.051) ln(lagged police per capita) -0.015 -0.027 -0.137* -0.186** (0.015) (0.014) (0.053) (0.049) Unemployment rate -0.027 0.536 1.212 1.084 (0.356) (0.314) (0.716) (0.657) Ln(Real per capita income) 0.003 -0.076 0.329 0.172   1982–84 $\$$ (0.129) (0.114) (0.224) (0.203) Poverty rate -0.002 -0.001 -0.003 -0.001 (0.001) (0.001) (0.003) (0.003) Real AFDC generosity 0.004 -0.001 -0.008 -0.003   1982-1984 $\$$ (0.003) (0.003) (0.008) (0.007) Shall-issue concealed weapons 0.014 0.019 -0.042 -0.023   law (0.015) (0.011) (0.023) (0.022) Beer consumption per capita 0.077 0.026 0.286** 0.286** (0.050) (0.044) (0.106) (0.096) Year FE? Yes Yes Yes Yes Yes Yes Yes Yes State FE? Yes Yes Yes Yes Yes Yes Yes Yes Observations 1,530 1,517 1,530 1,517 1,530 1,517 1,512 1,499 R|$^{2}$| 0.815 0.844 0.974 0.980 0.853 0.865 0.877 0.887 Adjusted R|$^{2}$| 0.805 0.834 0.973 0.978 0.845 0.857 0.870 0.880 . Dependent variable: Log per capita value of... . . Violent crime . Property crime . UCR murder . VS murder . . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . Effective abortion rate ’85–’97 -0.184** -0.178** -0.138** -0.152** -0.087* -0.100* -0.098** -0.116** (0.022) (0.022) (0.017) (0.016) (0.038) (0.040) (0.034) (0.036) Effective abortion rate ’98–’14 -0.192** -0.189** -0.149** -0.168** -0.131** -0.152** -0.141** -0.164** (0.019) (0.019) (0.016) (0.015) (0.017) (0.021) (0.016) (0.019) ln(lagged prisoners per capita) 0.007 -0.111** -0.121* -0.133** (0.037) (0.034) (0.056) (0.051) ln(lagged police per capita) -0.015 -0.027 -0.137* -0.186** (0.015) (0.014) (0.053) (0.049) Unemployment rate -0.027 0.536 1.212 1.084 (0.356) (0.314) (0.716) (0.657) Ln(Real per capita income) 0.003 -0.076 0.329 0.172   1982–84 $\$$ (0.129) (0.114) (0.224) (0.203) Poverty rate -0.002 -0.001 -0.003 -0.001 (0.001) (0.001) (0.003) (0.003) Real AFDC generosity 0.004 -0.001 -0.008 -0.003   1982-1984 $\$$ (0.003) (0.003) (0.008) (0.007) Shall-issue concealed weapons 0.014 0.019 -0.042 -0.023   law (0.015) (0.011) (0.023) (0.022) Beer consumption per capita 0.077 0.026 0.286** 0.286** (0.050) (0.044) (0.106) (0.096) Year FE? Yes Yes Yes Yes Yes Yes Yes Yes State FE? Yes Yes Yes Yes Yes Yes Yes Yes Observations 1,530 1,517 1,530 1,517 1,530 1,517 1,512 1,499 R|$^{2}$| 0.815 0.844 0.974 0.980 0.853 0.865 0.877 0.887 Adjusted R|$^{2}$| 0.805 0.834 0.973 0.978 0.845 0.857 0.870 0.880 Note: |$^{*}\textit{P} < 0.05; ^{**}\textit{P} < 0.01$|⁠. The dependent variable is the log in the per capita crime rate named at the top of each pair of columns. The first column in each pair presents results from the specifications in which the only additional covariates are state- and year-fixed effects. The second column presents results using the full specification. The data set is comprised of annual state-level observations (including the District of Columbia) for the period 1985–2014. State- and year-fixed effects are included in all specifications. The prison and police variables are once-lagged to minimize endogeneity. Real AFDC generosity per recipient family is lagged by 15 years, measured in thousands of dollars and indexed at 1982–84 values. Estimation is performed using a two-step procedure. In the first step, weighted least squares estimates are obtained, with weights determined by state population. In the second step, a panel data generalization of the Prais–Winsten correction for serial correlation developed by Bhargava et al. (1982) is implemented. Standard errors are in parentheses. Open in new tab All of the coefficients on abortion in both Table 4 and Appendix B are negative, implying that higher abortion rates are associated with lower crime. These estimated effects of abortion are in all cases highly statistically significant—more so than any other variable included in the analysis. Eleven of the twelve estimates based on specifications with a full set of controls in these two tables are significant at the 0.01 level (with the first-period effect on UCR murder significant at the 0.05 level). Notably, the coefficients on abortion in the time period that post-dates our initial study are larger in magnitude in all eight Table 4 specifications, implying that the out of sample results are even stronger than those in the original paper.20 The real-world magnitude implied by the coefficients on abortion is substantial. An increase in the effective abortion rate of 100 per 1,000 live births (the mean effective abortion rate in 2014 for violent crime is 340.95 with a standard deviation of 131.37 across states) is associated with a reduction of roughly 10–20% in crime. Looking at the estimates with controls for the entire time period (Appendix B), we see that the estimated drop in crime from an increase of 100 in the effective abortion rates ranges from a low of 15.8% for UCR murder to 18.9% for violent crime. The consistency of the strong negative relationship between abortion and crime shown in Table 4 is striking. Note that adding controls to the panel data models (the even columns) has little effect on the estimated abortion effect for violent crime, and it increases the effect for property crime and the two measures of murder. We also experimented with different specifications and additional controls and the Table 4 findings remained remarkably robust. For example, when we used the controls from Donohue et al. (2019) instead of those from our original (and current) Table 4, the results are essentially the same (see Appendix C).21 Specifically, the negative effect of abortion on crime becomes modestly stronger for violent and property crime, and overall for five of the eight estimates, while modestly weaker in three. Every estimated abortion effect using the Table 4 or the modified Appendix C set of controls is statistically significant at at least the 0.01 level for violent crime, property crime and Vital Statistics murder.22 3.4. Illustrating the Greater Drops in Crime in High-Abortion States To illustrate how tightly the relative increases in abortions in high-abortions states correspond to the relative drops in crime in these states relative to low-abortion rate states, we divided the states into two groups of roughly equal population in 1985 according to the number of abortions per 1,000 live births based on state of residence.23 Both groups had roughly 119 million residents in 1985, although their growth rates differed over time. The low-abortion states had a somewhat higher population in 1977 but had almost 18 million fewer residents than the high-population states by 2014, as shown in the table below: . 1977 . 1985 . 2014 . Low-abortion states 112.1 118.9 150.5 High-abortion states 107.6 119.0 168.4 . 1977 . 1985 . 2014 . Low-abortion states 112.1 118.9 150.5 High-abortion states 107.6 119.0 168.4 Notes: Population in millions. Open in new tab . 1977 . 1985 . 2014 . Low-abortion states 112.1 118.9 150.5 High-abortion states 107.6 119.0 168.4 . 1977 . 1985 . 2014 . Low-abortion states 112.1 118.9 150.5 High-abortion states 107.6 119.0 168.4 Notes: Population in millions. Open in new tab Note that our crime rates for the following figures are crimes per 100,000 population. Figure 1 shows the difference in effective abortion rates between our high- and low-abortion states, weighted to reflect when legalized abortion would be expected to influence violent crime. The impact of legalized abortion starts slowly but rises substantially in both sets of states after 1990, although obviously much more rapidly in the high-abortion states. Our thesis predicts that while both sets of states would experience downward pressure on crime by virtue of the growing effective abortion rates, the impact on crime would be substantially greater in the high-abortion states. Figure 1 Open in new tabDownload slide Effective Abortion Rates in High- and Low-Abortion States (Weighted by Violent Crime). Figure 1 Open in new tabDownload slide Effective Abortion Rates in High- and Low-Abortion States (Weighted by Violent Crime). Figure 2 illustrates the pattern of violent crime over the period from 1977 to 2014 in these two sets of states. The first aspect to note in this figure is the striking parallel trends in the pattern of violent crime in these two sets of states in the roughly 15 years prior to the impact of legalized abortion began to take effect. While both sets were initially battered by sharply rising violent crime, the pattern reversed in the early 1990s, and the high-abortion states experienced much more dramatic drops in crime for the remainder of our data period. While in the late 1970s and up to the early 1990s, violent crime was substantially higher in high-abortion states, the persistent faster decline over the next quarter-century largely eliminated that gap. Figure 2 Open in new tabDownload slide Violent Crime Rates in High- and Low-Abortion States, 1977–2014. Figure 2 Open in new tabDownload slide Violent Crime Rates in High- and Low-Abortion States, 1977–2014. The interesting connection between the greater increase in abortion and the greater drops in violent crime experienced in the high-abortion states can be seen by plotting the difference in violent crime rates on the same graph as the difference in the abortion rates in these two sets of states, as shown in Figure 3. The violent crime rate vacillated between 250 and 300 more crimes per 100,000 population in the high-abortion states in the decade prior to 1990. But just as the greater increases in abortion in the high-abortion states took hold, continuing for the remainder of our data period, the more sharply declining violent crime rate in the high-abortion states virtually eliminated the crime differential by 2014. Figure 3 Open in new tabDownload slide The Growing Abortion Disparity Corresponds to a Relative Decline in the Violent Crime Rate, 1977–2014. Figure 3 Open in new tabDownload slide The Growing Abortion Disparity Corresponds to a Relative Decline in the Violent Crime Rate, 1977–2014. The same underlying patterns in violent crime shown in Figures 1–3 for high- and low-abortion states also exist for murder and property crimes. Figures 4 and 5 replicate the Figure 3 juxtaposition showing the pronounced opposing movements of the rising gap in effective abortion rates between the two groups of states with the simultaneously declining gap in murder and property crime, respectively. Indeed, as Figure 4 reveals, the relatively greater drop in murder in the high-abortion states was so substantial that by the end of our data period the higher murder rates of the high-abortion states had not only been eliminated but had actually been reversed. By 2014, the high-abortion states had a VS murder rate that was 0.709 per 100,000 lower than the murder rate in low-abortion states.24 Figure 4 Open in new tabDownload slide The Growing Abortion Disparity Corresponds to a Relative Decline in the VS Murder Rate, 1977–2014. Figure 4 Open in new tabDownload slide The Growing Abortion Disparity Corresponds to a Relative Decline in the VS Murder Rate, 1977–2014. Figure 5 Open in new tabDownload slide The Growing Abortion Disparity Corresponds to a Relative Decline in the Property Crime Rate, 1977–2014. Figure 5 Open in new tabDownload slide The Growing Abortion Disparity Corresponds to a Relative Decline in the Property Crime Rate, 1977–2014. Visually, the same pattern of a higher initial crime rate in the high-abortion states that declines and was ultimately reversed at the same time that the abortion differential grew is seen in Figure 5 for property crime. Note that the murder rate gap was eliminated in roughly 2008 and turned in favor of high-abortion states thereafter, while in the case of property crime this contemporaneous decline in the higher initial crime was eliminated almost a decade earlier. Interestingly, the higher property crime and murder rate differentials of the high-abortion rate states were both eliminated at about the time that their respective effective abortion rate differentials (over the low-abortion rate states) reached 175 per 1,000 births. Since property crime tends to be committed by younger criminals, the property crime effective abortion rate differential reached 175 roughly 10 years earlier than the murder effective abortion rate. Of course, these graphs are only juxtaposing the contemporaneous growth in abortion rates with the greater crime drops of the high-abortion rate states, but we know that the rates of police staffing and incarceration were growing very substantially over this period. One could imagine that the high-abortion rate states simply grew their police forces and incarceration rates faster than low-abortion rate states starting around 1990. In this event, these other policies might explain all or most of the relative crime drop that we have paired with the rising relative abortion rates. To explore this possibility, we plot the relative changes in our two sets of states for rates of incarceration and police staffing Figures 6 and 7, respectively. The figures clearly document the very substantial expansions in these two crime-fighting technologies but two points underscore why these factors do not undermine the hypothesized link between legalized abortion and crime. First, the steady increases in both incarceration and police cannot explain the sudden and unanticipated decline in the crime rate starting in roughly 1990. Second, we have just seen that the crime drops were substantially greater in high-abortion states, so if the abortion-crime thesis is to be undermined by increasing incarceration or growing police forces we would need to see greater increases in these factors in the high-abortion states. But the figures refute this proposition. Figure 6 shows that incarceration rates rose more sharply and are now substantially higher in the low-abortion states. Thus, the greater crime-reducing increases in incarceration in the low-abortion states would suggest that the relative crime improvements in the high-abortion states depicted in Figures 3–5 that we attribute to increased abortion are, if anything, understated. Figure 6 Open in new tabDownload slide Incarceration Rate Trends in High- and Low-Abortion States. Note: The incarceration rate is state and federal prisoners per 100,000 population. Figure 6 Open in new tabDownload slide Incarceration Rate Trends in High- and Low-Abortion States. Note: The incarceration rate is state and federal prisoners per 100,000 population. Figure 7. Open in new tabDownload slide Police Staffing Rate Trends in High- and Low-Abortion States. Note: The policing rate is law enforcement employees per 100,000 population. Figure 7. Open in new tabDownload slide Police Staffing Rate Trends in High- and Low-Abortion States. Note: The policing rate is law enforcement employees per 100,000 population. Figure 7 reaffirms that both sets of states have substantially increased police staffing over our data period, at least until the financial crisis led to budget cutbacks and both groups experienced dips in the ensuing years. While the high-abortion states have always had higher police staffing rates, at least over the last 15 years the police staffing gap between high- and low-abortion states has narrowed. Again the relatively greater crime-reducing expenditures on police employment in low-abortion states would be the opposite of the pattern needed to explain away the link between higher abortion and lower crime. 4. Linking Abortion Rates to Arrests By Age 4.1. The Benefits of Shifting From Crime Data to Arrest Rate Data The analysis up to this point has established the relationship between rising effective abortion rates and declining rates of violent crime, property crime, and murder. Crime rates are the obvious and conceptually appropriate outcome to focus on, but they have the unfortunate limitation that we do not know the crime rate by age of offender because the perpetrator is frequently unknown. Consequently, any analysis using crime rates is restricted to having state-year as the unit of analysis. This level of analysis does not allow us to take advantage of the unusual richness in the predictions of the abortion-crime hypothesis, which argues that crime patterns should differ by cohort, even in a given state and year, depending on the abortion rate when that cohort was in utero. Since it is impossible to develop accurate crime data by age, we turn to arrest data, which enable us to test the hypothesis with a level of specificity not possible with aggregate crime data. For the subset of crimes in which an arrest is made, the age of the individuals arrested is reported. Thus, we can analyze arrest data at the level of state |$\times$| year |$\times$| single year of age. This allows us to include in our panel data analysis of arrests by age dummy variables for state |$\times$| age, age |$\times$| year, and state |$\times$| year. The precise specification estimated is: $$ \begin{equation}\label{eq:3} ln({ ARRESTS}_{sta}) = \beta_{1}ABORT_{sta} + \gamma_{sa} + \lambda_{at} + \Theta_{st} + \epsilon_{sta}, \end{equation} $$(3) where |$s$|⁠, |$t$|⁠, and |$a$| index state, year, and age, respectively. The variable |$ARRESTS$| is the raw number of arrests for a given crime. As our measure of the abortion rate for a particular cohort, we use the abortion rate in the state where the arrest was made, in the calendar year most likely to have preceded the arrestee’s birth. State |$\times$| age, age |$\times$| year, and state |$\times$| year dummies absorb variation along those different dimensions. All of the variation in the covariates used in the panel crime regressions estimated above is at the state |$\times$| year level, so no variation remains in those covariates in these specifications. Data by single year of age are available only for ages 15–24 (for older and younger ages the data are grouped, typically into 5-year age windows) so we limit our sample to that age range. Table 5 presents the results. The dependent variable is the natural log of the number of arrests for the crime category listed at the top of the column. Following the original paper, we present results for violent crime (columns 1–3) and property crime (columns 4–6), but not for homicide, because homicide is rare enough (with arrest often rarer) that many state-year-age cells are empty. The set of covariates included grows moving from left to right for a given crime category, as noted in the bottom portion of the table. The top two rows present the coefficient on the abortion rate in the period covered by the original data (top row) and in later years (second row). Only the coefficient on the abortion rate is shown in the table. As we did for our Table 4 crime regression, we also estimated our arrest regression using the single abortion variable specification, which we report in the third row of Table 5. Table 5. The Relationship between Abortion Rates and Arrests, by Single Year of Age, 1985–2014 . Dependent variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . Abortion rate ’85–’97 (⁠|$\times$|100) -0.033 -0.056 -0.031 -0.048 -0.032 -0.029 (0.006)** (0.008)** (0.006)** (0.007)** (0.006)** (0.004)** [0.012]** [0.024]* [0.014]* [0.017]** [0.012]** [0.009]** Abortion rate ’98–’14 (⁠|$\times$|100) -0.042 -0.049 -0.057 -0.086 -0.080 -0.044 (0.006)** (0.006)** (0.007)** (0.005)** (0.005)** (0.007)** [0.024] [0.026] [0.017]** [0.015]** [0.015]** [0.019]* Abortion rate ’85–’14 (⁠|$\times$|100) -0.039 -0.051 -0.038 -0.074 -0.067 -0.033 (0.005)** (0.006)** (0.005)** (0.005)** (0.005)** (0.003)** [0.019]* [0.025]* [0.013]** [0.012]** [0.012]** [0.010]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 . Dependent variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . Abortion rate ’85–’97 (⁠|$\times$|100) -0.033 -0.056 -0.031 -0.048 -0.032 -0.029 (0.006)** (0.008)** (0.006)** (0.007)** (0.006)** (0.004)** [0.012]** [0.024]* [0.014]* [0.017]** [0.012]** [0.009]** Abortion rate ’98–’14 (⁠|$\times$|100) -0.042 -0.049 -0.057 -0.086 -0.080 -0.044 (0.006)** (0.006)** (0.007)** (0.005)** (0.005)** (0.007)** [0.024] [0.026] [0.017]** [0.015]** [0.015]** [0.019]* Abortion rate ’85–’14 (⁠|$\times$|100) -0.039 -0.051 -0.038 -0.074 -0.067 -0.033 (0.005)** (0.006)** (0.005)** (0.005)** (0.005)** (0.003)** [0.019]* [0.025]* [0.013]** [0.012]** [0.012]** [0.010]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 Note: |${}^{*} P < 0.05; ^{**} P < 0.01$|⁠. Results in the table are coefficients from estimation of equation (3). The unit of observation in the regression is annual arrests by state by single year of age. The sample covers the period of 1985–2014 for ages 15–24, and the top panel of the table estimates the effect of abortion both for our initial period (1985–97) and for the remainder of our full data period (1998–2014). The bottom panel (the third row) estimates a single abortion variable model over the entire 1985–2014 time period. The abortion rate for a cohort of age a in state s in year y is the number of abortions per 1000 live births in state s in year y – a |$-$| 1. Note that this is the actual abortion rate, rather than the “effective” abortion rate used in preceding tables. Therefore, the coefficients in this table are not directly comparable to those of earlier tables. If data were available for all states, years, and ages, the total number of observations would be 15,300. Due to missing arrest data and occasional zero values for arrests, the actual number of observations is somewhat smaller. A complete set of year-birth cohort interactions are included in all specifications to capture national changes in the shape of the age-crime profile over time. The table indicates the various fixed effects included in each column. Estimation is weighted least squares, with weights determined by total population by state-year-age. Standard errors clustered by cohort year of birth and state are included in parentheses; this accounts for correlation over time within a given birth cohort in a particular state. Such a correction is necessary because the abortion rate for any given cohort is fixed over time, but multiple observations corresponding to different years of age are included in the regression. Standard errors clustered by state are included in square brackets below the first set of clustered standard errors. Open in new tab Table 5. The Relationship between Abortion Rates and Arrests, by Single Year of Age, 1985–2014 . Dependent variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . Abortion rate ’85–’97 (⁠|$\times$|100) -0.033 -0.056 -0.031 -0.048 -0.032 -0.029 (0.006)** (0.008)** (0.006)** (0.007)** (0.006)** (0.004)** [0.012]** [0.024]* [0.014]* [0.017]** [0.012]** [0.009]** Abortion rate ’98–’14 (⁠|$\times$|100) -0.042 -0.049 -0.057 -0.086 -0.080 -0.044 (0.006)** (0.006)** (0.007)** (0.005)** (0.005)** (0.007)** [0.024] [0.026] [0.017]** [0.015]** [0.015]** [0.019]* Abortion rate ’85–’14 (⁠|$\times$|100) -0.039 -0.051 -0.038 -0.074 -0.067 -0.033 (0.005)** (0.006)** (0.005)** (0.005)** (0.005)** (0.003)** [0.019]* [0.025]* [0.013]** [0.012]** [0.012]** [0.010]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 . Dependent variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . Abortion rate ’85–’97 (⁠|$\times$|100) -0.033 -0.056 -0.031 -0.048 -0.032 -0.029 (0.006)** (0.008)** (0.006)** (0.007)** (0.006)** (0.004)** [0.012]** [0.024]* [0.014]* [0.017]** [0.012]** [0.009]** Abortion rate ’98–’14 (⁠|$\times$|100) -0.042 -0.049 -0.057 -0.086 -0.080 -0.044 (0.006)** (0.006)** (0.007)** (0.005)** (0.005)** (0.007)** [0.024] [0.026] [0.017]** [0.015]** [0.015]** [0.019]* Abortion rate ’85–’14 (⁠|$\times$|100) -0.039 -0.051 -0.038 -0.074 -0.067 -0.033 (0.005)** (0.006)** (0.005)** (0.005)** (0.005)** (0.003)** [0.019]* [0.025]* [0.013]** [0.012]** [0.012]** [0.010]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 Note: |${}^{*} P < 0.05; ^{**} P < 0.01$|⁠. Results in the table are coefficients from estimation of equation (3). The unit of observation in the regression is annual arrests by state by single year of age. The sample covers the period of 1985–2014 for ages 15–24, and the top panel of the table estimates the effect of abortion both for our initial period (1985–97) and for the remainder of our full data period (1998–2014). The bottom panel (the third row) estimates a single abortion variable model over the entire 1985–2014 time period. The abortion rate for a cohort of age a in state s in year y is the number of abortions per 1000 live births in state s in year y – a |$-$| 1. Note that this is the actual abortion rate, rather than the “effective” abortion rate used in preceding tables. Therefore, the coefficients in this table are not directly comparable to those of earlier tables. If data were available for all states, years, and ages, the total number of observations would be 15,300. Due to missing arrest data and occasional zero values for arrests, the actual number of observations is somewhat smaller. A complete set of year-birth cohort interactions are included in all specifications to capture national changes in the shape of the age-crime profile over time. The table indicates the various fixed effects included in each column. Estimation is weighted least squares, with weights determined by total population by state-year-age. Standard errors clustered by cohort year of birth and state are included in parentheses; this accounts for correlation over time within a given birth cohort in a particular state. Such a correction is necessary because the abortion rate for any given cohort is fixed over time, but multiple observations corresponding to different years of age are included in the regression. Standard errors clustered by state are included in square brackets below the first set of clustered standard errors. Open in new tab Across both crime categories and all specifications, Table 5 reports abortion coefficients that are negative and highly statistically significant. The inclusion of additional covariates does not have an obvious impact on the magnitude of the abortion coefficients. Consistent with the Table 4 regression results, the point estimates on the abortion coefficient are larger in magnitude in the later period than in the initial sample period in five of the six columns in Table 5. In some of the specifications, those differences are statistically significant.25 The third row of Table 5 shows effects of abortion on crime estimated over the entire 1985–2014 time span. Note that columns 3 and 6, which estimate the effect of abortions on arrests with the full set of fixed effects, show an overall effect of |$-$|0.038 for violent crime and |$-$|0.033 for property crime. Both are highly statistically significant using either of the two presented standard error estimates: clustered by birth cohort and state (in parentheses) and by state (in square brackets).26 4.2. Improving the Precision of Our Abortion Measures The results presented thus far have directly mimicked the specifications and data definitions of Donohue and Levitt (2001) in order to make the comparison of the new results to the original results as clear as possible. As noted, the primary exception to this was to use the superior AGI abortion by state of residence data that was furnished to us after our initial publication and which we have exclusively relied on as our main abortion measure since that time. In the years since that first paper was published, we have also tried to address the problem of measurement error in the abortion variable in three ways with two improvements to our variable construction to more closely link these variables to what the theory suggests are the appropriate proxies and by using an instrumental variable for our residence abortion measure. First, we constructed an abortion measure that better corresponds to the actual month and year of birth of the individual. Second, we have adjusted our abortion measure to take into account cross-state mobility between birth and adolescence. Third, recognizing the noise in our abortion proxy (based on Alan Guttmacher Institute data), we have used another independently generated estimate of the abortion rate (from the Centers for Disease Control) as an instrumental variable.27 Table 6 illustrates the impact of running the same Table 5 regressions while using these two adjustments to our abortion measure and instrumenting to address the impact of measurement error in our abortion proxy. We report our instrumental variables (IV) estimates in two ways (as we did in Table 5): first separating our abortion effect into 1985–1997 and 1998–2014 in the first two rows, and then estimating a single abortion variable for 1985–2014, which is reported in the third row. Comparing Tables 5 and 6, one sees that the 18 estimated effects of abortion on crime are larger using the better abortion measure and the instrumenting—often doubling in magnitude—in every case except the second-period effect on property arrests with the state |$\times$| year interaction (column 6), which is essentially zero.28 All the overall period estimates in row 3 are statistically significant for both sets of standard error estimates, and considerably larger than the corresponding Table 5 values. Table 6. Estimated Effects of Abortion on Crime with Measurement Error Adjustments, 1985–2014 . Dependent Variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.114 -0.115 -0.100 -0.074 -0.080 (0.013)** (0.018)** (0.026)** (0.013)** (0.014)** (0.017)** [0.017]** [0.025]** [0.029]** [0.026]** [0.025]** [0.018]** IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.067 -0.068 -0.074 -0.166 -0.154 0.003 (0.010)** (0.012)** (0.021)** (0.014)** (0.016)** (0.022) [0.032]* [0.034]* [0.034]* [0.056]** [0.062]* [0.028] IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.066 -0.087 -0.108 -0.135 -0.121 -0.065 (0.010)** (0.013)** (0.022)** (0.013)** (0.015)** (0.014)** [0.021]** [0.029]** [0.030]** [0.041]** [0.052]* [0.017]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 . Dependent Variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.114 -0.115 -0.100 -0.074 -0.080 (0.013)** (0.018)** (0.026)** (0.013)** (0.014)** (0.017)** [0.017]** [0.025]** [0.029]** [0.026]** [0.025]** [0.018]** IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.067 -0.068 -0.074 -0.166 -0.154 0.003 (0.010)** (0.012)** (0.021)** (0.014)** (0.016)** (0.022) [0.032]* [0.034]* [0.034]* [0.056]** [0.062]* [0.028] IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.066 -0.087 -0.108 -0.135 -0.121 -0.065 (0.010)** (0.013)** (0.022)** (0.013)** (0.015)** (0.014)** [0.021]** [0.029]** [0.030]** [0.041]** [0.052]* [0.017]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 Note: |$^{* }$|P |$<$| 0.05; |$^{** }$|P |$ < $| 0.01. Table 6 is exactly comparable to Table 5, except that it adjusts the abortion measure to better link the timing of abortion with each relevant age cohort and to reflect the inter-state movement from birth state to where individuals live when we are measuring state arrest rates. Our instrumental variables (IV) estimate uses a CDC abortion measure as an instrument for our AGI measure. The top panel of the table estimates the effect of abortion both for our initial period (1985–97) and for the remainder of our full data period (1998–2014). The bottom panel (the third row) estimates a single abortion variable model over the entire 1985–2014 time period. Open in new tab Table 6. Estimated Effects of Abortion on Crime with Measurement Error Adjustments, 1985–2014 . Dependent Variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.114 -0.115 -0.100 -0.074 -0.080 (0.013)** (0.018)** (0.026)** (0.013)** (0.014)** (0.017)** [0.017]** [0.025]** [0.029]** [0.026]** [0.025]** [0.018]** IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.067 -0.068 -0.074 -0.166 -0.154 0.003 (0.010)** (0.012)** (0.021)** (0.014)** (0.016)** (0.022) [0.032]* [0.034]* [0.034]* [0.056]** [0.062]* [0.028] IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.066 -0.087 -0.108 -0.135 -0.121 -0.065 (0.010)** (0.013)** (0.022)** (0.013)** (0.015)** (0.014)** [0.021]** [0.029]** [0.030]** [0.041]** [0.052]* [0.017]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 . Dependent Variable: . . ln(Violent arrests) . ln(Property arrests) . . (1) . (2) . (3) . (4) . (5) . (6) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.114 -0.115 -0.100 -0.074 -0.080 (0.013)** (0.018)** (0.026)** (0.013)** (0.014)** (0.017)** [0.017]** [0.025]** [0.029]** [0.026]** [0.025]** [0.018]** IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.067 -0.068 -0.074 -0.166 -0.154 0.003 (0.010)** (0.012)** (0.021)** (0.014)** (0.016)** (0.022) [0.032]* [0.034]* [0.034]* [0.056]** [0.062]* [0.028] IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.066 -0.087 -0.108 -0.135 -0.121 -0.065 (0.010)** (0.013)** (0.022)** (0.013)** (0.015)** (0.014)** [0.021]** [0.029]** [0.030]** [0.041]** [0.052]* [0.017]** Year * Age? Yes Yes Yes Yes Yes Yes State fixed effects? Yes Implied Implied Yes Implied Implied State * Age? No Yes Yes No Yes Yes State * Year? No No Yes No No Yes Observations 13,765 13,765 13,765 13,770 13,770 13,770 Note: |$^{* }$|P |$<$| 0.05; |$^{** }$|P |$ < $| 0.01. Table 6 is exactly comparable to Table 5, except that it adjusts the abortion measure to better link the timing of abortion with each relevant age cohort and to reflect the inter-state movement from birth state to where individuals live when we are measuring state arrest rates. Our instrumental variables (IV) estimate uses a CDC abortion measure as an instrument for our AGI measure. The top panel of the table estimates the effect of abortion both for our initial period (1985–97) and for the remainder of our full data period (1998–2014). The bottom panel (the third row) estimates a single abortion variable model over the entire 1985–2014 time period. Open in new tab The move from Table 5 to Table 6 illuminates some aspects of our data and an important social phenomenon: the declining interstate mobility since the early 1970s. All six first-period estimates at least doubled by introducing the Table 6 corrections, while none of the second-period estimates did. Focusing on the violent crime estimates, the three first-period estimates grew by at least 100%, while the second-period increases ranged from 30% to 60%. This is not surprising since our abortion data are clearly less well-measured in the early days of legalization, and therefore we would expect that instrumenting to improve the accuracy of the abortion data would have a bigger effect on our first-period estimates. Similarly, our Table 6 migration adjustments are more consequential in the first period when the amount of inter-state migration was much greater.29 As a result, our effort to link the abortion rate for a given cohort to the state in which the cohort members ultimately reside improves the quality of our measures, thereby raising the magnitude of the estimated effect of abortion on crime.30 Foote and Goetz (2008) argued that our state panel data analysis of crime rates (Table 4 in this paper) and our regressions explaining arrests by age (Tables 5 and 6 in this paper) might not fully establish the abortion-crime link that we posited. The heart of their critique was that the abortion rate may be proxying for some state-specific omitted variable, and that therefore the “crucial” test needed to eliminate the “potential omitted variable bias on the state-year level” would focus on a per capita arrest rate regression that included state-year fixed effects. This regression would control for whatever factors were influencing crime in a given state and a given year and determine whether the abortion rate at the time of any birth cohort would correlate with the arrest rate for that cohort years later as it moved through ages 15–24 (for which we have age-specific arrest rates). Foote and Goetz (2008) also suggested that concerns about residual-independence assumptions that emerged after “[Donohue and Levitt] published their 2001 paper” made it advisable to provide “a second set of standard errors [that would] cluster the standard errors by state.” We address these issues in our current Table 7, which continues our practice (in Tables 5 and 6) of showing two sets of standard errors in all our arrest rate regressions, including their preferred clustering by state. Every regression in Table 7 also includes the “crucial” state-year fixed effects. Accordingly, Table 7 presents in columns 2 and 4 exactly what Foote and Goetz stated would establish or refute the abortion-crime link. The initial, short answer is that, using the precise per capita arrest rate variable and cluster adjustment that Foote and Goetz advocate, the abortion and crime effect that we identified in 2001 remains strong and statistically significant at the 0.05 level for violent crime (see the value of |$-$|0.05 in row 4, column 2) and is negative but not statistically significant for property crime (a value of |$-$|0.007 in row 4, column 4) when estimated for the 1985–2014 period. While 17 years of additional data have strengthened the evidence in support of the abortion-crime hypothesis, the current Table 7 finding is essentially the same as what we showed in our 2008 response to Foote and Goetz.31 Table 7. Distinguishing Between the Channels through Which Abortion Affects Crime, 1985–2014 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.041 -0.044 -0.006 (0.025)* (0.021)* (0.017)** (0.013) [0.028]* [0.029] [0.019]* [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.084 -0.089 -0.004 -0.011 (0.021)** (0.021)** (0.021) (0.022) [0.029]** [0.029]** [0.031] [0.037] ln(SEER population) 0.680 0.486 (0.066)** (0.053)** [0.125]** [0.134]** IV Abortion effect ’85–14 (x100) -0.069 -0.050 -0.036 -0.007 (0.021)** (0.018)** (0.014)* (0.012) [0.023]** [0.023]* [0.019] [0.020] ln(SEER population) 0.672 0.503 (0.061)** (0.051)** [0.124]** [0.137]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 13,765 13,765 13,770 13,770 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.041 -0.044 -0.006 (0.025)* (0.021)* (0.017)** (0.013) [0.028]* [0.029] [0.019]* [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.084 -0.089 -0.004 -0.011 (0.021)** (0.021)** (0.021) (0.022) [0.029]** [0.029]** [0.031] [0.037] ln(SEER population) 0.680 0.486 (0.066)** (0.053)** [0.125]** [0.134]** IV Abortion effect ’85–14 (x100) -0.069 -0.050 -0.036 -0.007 (0.021)** (0.018)** (0.014)* (0.012) [0.023]** [0.023]* [0.019] [0.020] ln(SEER population) 0.672 0.503 (0.061)** (0.051)** [0.124]** [0.137]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 13,765 13,765 13,770 13,770 Note: |$^{* }$|P |$<$| 0.05; |$^{** }$|P |$<$| 0.01 Table 7 modifies the column 3 and 6 specifications from Table 6 in two ways to remove the cohort-size effect of abortion on arrests by single year of age for ages 15–24. Columns 1 and 3 of Table 7 simply add a control for the population of each state by single year of age. The fact that the estimated values for this population control are substantially below 1 (see rows 3 and 5) illustrates the presence of measurement error in our population variable. Columns 2 and 4 control for population by changing the dependent variable to ln(per capita arrest rate) by single year of age, and these estimates will suffer from ratio bias because of the observed measurement error in the population variable that appears in the denominators of both the dependent variable and the abortion independent variable. Standard errors clustered by cohort year of birth and state are included in parentheses, with standard errors clustered by state in square brackets directly below. Open in new tab Table 7. Distinguishing Between the Channels through Which Abortion Affects Crime, 1985–2014 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.041 -0.044 -0.006 (0.025)* (0.021)* (0.017)** (0.013) [0.028]* [0.029] [0.019]* [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.084 -0.089 -0.004 -0.011 (0.021)** (0.021)** (0.021) (0.022) [0.029]** [0.029]** [0.031] [0.037] ln(SEER population) 0.680 0.486 (0.066)** (0.053)** [0.125]** [0.134]** IV Abortion effect ’85–14 (x100) -0.069 -0.050 -0.036 -0.007 (0.021)** (0.018)** (0.014)* (0.012) [0.023]** [0.023]* [0.019] [0.020] ln(SEER population) 0.672 0.503 (0.061)** (0.051)** [0.124]** [0.137]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 13,765 13,765 13,770 13,770 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.065 -0.041 -0.044 -0.006 (0.025)* (0.021)* (0.017)** (0.013) [0.028]* [0.029] [0.019]* [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.084 -0.089 -0.004 -0.011 (0.021)** (0.021)** (0.021) (0.022) [0.029]** [0.029]** [0.031] [0.037] ln(SEER population) 0.680 0.486 (0.066)** (0.053)** [0.125]** [0.134]** IV Abortion effect ’85–14 (x100) -0.069 -0.050 -0.036 -0.007 (0.021)** (0.018)** (0.014)* (0.012) [0.023]** [0.023]* [0.019] [0.020] ln(SEER population) 0.672 0.503 (0.061)** (0.051)** [0.124]** [0.137]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 13,765 13,765 13,770 13,770 Note: |$^{* }$|P |$<$| 0.05; |$^{** }$|P |$<$| 0.01 Table 7 modifies the column 3 and 6 specifications from Table 6 in two ways to remove the cohort-size effect of abortion on arrests by single year of age for ages 15–24. Columns 1 and 3 of Table 7 simply add a control for the population of each state by single year of age. The fact that the estimated values for this population control are substantially below 1 (see rows 3 and 5) illustrates the presence of measurement error in our population variable. Columns 2 and 4 control for population by changing the dependent variable to ln(per capita arrest rate) by single year of age, and these estimates will suffer from ratio bias because of the observed measurement error in the population variable that appears in the denominators of both the dependent variable and the abortion independent variable. Standard errors clustered by cohort year of birth and state are included in parentheses, with standard errors clustered by state in square brackets directly below. Open in new tab But while the data through 2014 clearly meets the “crucial” test that Foote and Goetz articulated for the link between abortion and declining violent crime, it should also be noted that their recommended per capita arrest rate regression understates the impact of legalized abortion on crime in multiple ways. First, legalized abortion impacted crime in the 1990s and therefore contributed to the momentous crime drop in a way that is universally acknowledged: it reduced the size of cohorts moving into their high-crime teenage years beginning in the early 1990s.32Table 7 regression will not capture that effect because it only captures the per capita crime rate of each cohort, thereby neglecting any cohort-size effect. Second, and more critically, the per capita regression that Foote and Goetz advocate is biased against a finding that legalized abortion has a selection effect that leads to a per capita reduction in crime. The reason is straightforward: the denominator of the dependent variable in the per capita regression is the size of the cohort born in year t, which is also the identical denominator of the abortion rate independent variable for that same cohort. In other words, we have the same population variable in the denominator of both the dependent variable and the key independent variable of interest. Since Foote and Goetz acknowledge that this population variable “is measured with error,” the estimated effect of abortion on crime in columns 2 and 4 will be upward biased, thereby understating or obscuring any dampening effect that legalized abortion had on the per capita arrest rate.33 To illustrate the presence of this “ratio bias,” we include in Table 7 a regression with the count of arrests by cohort (the dependent variable) regressed on the abortion rate and a control for the size of that cohort. Since the coefficients on the population variable (shown in the third and fifth rows of Table 7) are substantially below 1, being roughly two-thirds for violent crime (column 1) and about one-half for property crime (column 3), we know—as Foote and Goetz acknowledged from the same evidence—that these results reflect measurement error in the population by state and age data. But this measurement error confirms the presence of ratio bias that attenuates the estimated abortion coefficient in the per capita regressions of Table 7 for which Foote and Goetz advocate, with that bias presumably greater for property crime by virtue of the considerably smaller column 3 coefficient on population.34 But while the columns 1 and 3 regressions reveal the attenuation in the estimated influence of population on arrests, these regressions are superior to the columns 2 and 4 per capita regressions in that they do not suffer from the ratio bias that understates the true selection effect of abortion on crime in the per capita regressions. First, note that while both of the overall (row 4) estimates of the impact of abortion on violent crime arrests are highly significant using either standard error measure, the column 1 estimate is almost 40% larger in absolute value than the column 2 estimate. The column 1 violent crime arrest estimates are also statistically significant in each of our two time periods, again using either standard error measure. Second, the column 3 regression generates a sizeable negative estimated effect of abortion on property crime arrests for the entire period with a P-value of 0.083 when clustering by state (significant at the 0.05 level when clustering by birth cohort per state). This column 3 property crime arrest estimate for the entire period is more than five times larger than the column 4 estimate that is marred by ratio bias. Note also that the column 3 first-period estimate of abortion on property crime arrests is negative and statistically significant using either standard error measure. 4.3. Robustness Check of the Abortion-Arrest Rate Link One might be concerned that our arrest by age results could be driven by an outlier large state like New York that had a particularly dramatic drop in crime or a jurisdiction with an unusually high number of abortions like D.C. To test this possibility, we ran our exact Table 7 analysis on the 26 states for which complete arrest data are available. The results are shown in Table 8, which drops a very eclectic, and non-selected (by us) sample of 25 states, including New York, Florida, D.C., Pennsylvania, and Colorado. Comparing the four abortion estimates for the entire 1985–2014 period depicted in Tables 7 and 8, we see that all are more negative for the states with complete arrest data, and the negative abortion coefficient in the property per capita arrest regression (row 4, column 4) triples in absolute value in Table 8 and is significant at the 0.05 level with clustering by birth-cohort by state. Moreover, the 1985–2014 abortion estimates for both violent and property crime that are not marred by attenuation bias (row 4 for columns 1 and 3) are both statistically significant at the 0.05 level even with the more stringent clustering. In other words, our arrest regressions findings are not only robust, but are strengthened even with the diminished sample size when we limit the sample to the 26 states with the most complete arrest data.35 We also report our Tables 5 and 6 results with our set of 26 states with complete arrest data in Appendices E and F, respectively. Table 8. Distinguishing Between the Channels through Which Abortion Affects Crime 26 States with Complete Arrest Data, 1985–2014 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.060 -0.046 -0.038 -0.011 (0.020)** (0.018)** (0.010)** (0.009) [0.038] [0.040] [0.020] [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.100 -0.107 -0.042 -0.056 (0.022)** (0.023)** (0.021) (0.023)* [0.036]** [0.034]** [0.023] [0.027]* ln(SEER population) 0.798 0.602 (0.060)** (0.052)** [0.125]** [0.173]** IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.071 -0.060 -0.039 -0.021 (0.017)** (0.016)** (0.010)** (0.010)* [0.031]* [0.032] [0.016]* [0.016] ln(SEER population) 0.767 0.600 (0.056)** (0.053)** [0.129]** [0.170]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 7,800 7,800 7,800 7,800 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.060 -0.046 -0.038 -0.011 (0.020)** (0.018)** (0.010)** (0.009) [0.038] [0.040] [0.020] [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.100 -0.107 -0.042 -0.056 (0.022)** (0.023)** (0.021) (0.023)* [0.036]** [0.034]** [0.023] [0.027]* ln(SEER population) 0.798 0.602 (0.060)** (0.052)** [0.125]** [0.173]** IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.071 -0.060 -0.039 -0.021 (0.017)** (0.016)** (0.010)** (0.010)* [0.031]* [0.032] [0.016]* [0.016] ln(SEER population) 0.767 0.600 (0.056)** (0.053)** [0.129]** [0.170]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 7,800 7,800 7,800 7,800 Note: |$^{* }$|P |$<$| 0.05; |$^{** }$|P |$<$| 0.01 Table 8 replicates the Table 7 analysis but only using the 26 states for which complete arrest data is available. The 25 excluded states are listed in the notes to Appendix E. Standard errors clustered by cohort year of birth and state are included in parentheses, with standard errors clustered by state in square brackets directly below. Open in new tab Table 8. Distinguishing Between the Channels through Which Abortion Affects Crime 26 States with Complete Arrest Data, 1985–2014 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.060 -0.046 -0.038 -0.011 (0.020)** (0.018)** (0.010)** (0.009) [0.038] [0.040] [0.020] [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.100 -0.107 -0.042 -0.056 (0.022)** (0.023)** (0.021) (0.023)* [0.036]** [0.034]** [0.023] [0.027]* ln(SEER population) 0.798 0.602 (0.060)** (0.052)** [0.125]** [0.173]** IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.071 -0.060 -0.039 -0.021 (0.017)** (0.016)** (0.010)** (0.010)* [0.031]* [0.032] [0.016]* [0.016] ln(SEER population) 0.767 0.600 (0.056)** (0.053)** [0.129]** [0.170]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 7,800 7,800 7,800 7,800 . Dependent Variable: . . ln(Viol. . ln(Viol. . ln(Prop. . ln(Prop. . . arrests) . arrests pc) . arrests) . arrests pc) . . (1) . (2) . (3) . (4) . IV Abortion effect ’85–’97 (⁠|$\times$|100) -0.060 -0.046 -0.038 -0.011 (0.020)** (0.018)** (0.010)** (0.009) [0.038] [0.040] [0.020] [0.020] IV Abortion effect ’98–’14 (⁠|$\times$|100) -0.100 -0.107 -0.042 -0.056 (0.022)** (0.023)** (0.021) (0.023)* [0.036]** [0.034]** [0.023] [0.027]* ln(SEER population) 0.798 0.602 (0.060)** (0.052)** [0.125]** [0.173]** IV Abortion effect ’85–’14 (⁠|$\times$|100) -0.071 -0.060 -0.039 -0.021 (0.017)** (0.016)** (0.010)** (0.010)* [0.031]* [0.032] [0.016]* [0.016] ln(SEER population) 0.767 0.600 (0.056)** (0.053)** [0.129]** [0.170]** Year * Age? Yes Yes Yes Yes State fixed effects? Implied Implied Implied Implied State * Age? Yes Yes Yes Yes State * Year? Yes Yes Yes Yes Ln Population? Yes No Yes No Observations 7,800 7,800 7,800 7,800 Note: |$^{* }$|P |$<$| 0.05; |$^{** }$|P |$<$| 0.01 Table 8 replicates the Table 7 analysis but only using the 26 states for which complete arrest data is available. The 25 excluded states are listed in the notes to Appendix E. Standard errors clustered by cohort year of birth and state are included in parentheses, with standard errors clustered by state in square brackets directly below. Open in new tab 5. Evaluating Our Overall Regression Results 5.1. Comparing Our Crime and Arrest Rate Results Taking stock of our two strands of regression analysis, we first showed that there is a continuing drop in crime that is proportional to our abortion measure in each state as cohorts born following the legalization of abortion entered adolescence and then their high-crime years (Table 4). No other explanatory variable in our panel data model of crime comes close to the statistical power of abortion, and this result is robust with an alternative set of controls (Appendix C) or dropping 5 states that had consistent problems reporting crime data under the Uniform Crime Reports for multiple years. We then switched to an analysis of arrest data so that we could directly link arrests of 15 year olds in a given state with the relevant abortion rate for their birth cohort, and so on by single year of age through age 24 (the end of the UCR’s arrest data by single year of age). Table 5 established the inverse relationship between the natural log of arrests for a single year of age and the abortion rate for the birth year of that arrest cohort. We showed that this relationship held and was highly statistically significant (using two different standard error measures), even when controlling for state-year fixed effects. Our findings based on arrests dramatically support the earlier findings of our panel data analysis of state-year crime data, but we should pause to emphasize how striking the arrest results are and how they rely on a very different source of variation, with a substantially different abortion measure that has a sharply different time trend. As one can see in Figure 1, the basic pattern for the effective abortion rate that we use as the key explanatory variable in our state-year crime regressions is predominantly rising throughout our data period, although of course differently across different states (and with a few states experiencing a downturn in the effective abortion rate for property crime late in our data period).36 But the pattern for the abortions per 1,000 live births that we use for our arrest rate regressions by single age of arrest peaks far earlier at 432 in 1981 and then declines over the relevant portion of our data (the 1999 abortion rate would be the last year relevant to the arrest of 15 year olds in 2014). Of course, these patterns vary substantially by state, but for some years the youngest cohorts will have the highest abortion rates that we link with their arrests and sometimes it will be the oldest cohorts that have the highest abortion rates. Given all of the imperfections in abortion and arrest data, the lapse of time between birth and arrests with inter-state mobility, and the fact that we are controlling for the level of crime in the particular state and year, the fact that the negative relationship between abortion and arrests emerges so powerfully is really quite stunning. We continued this arrest rate analysis by trying (1) to improve the measurement of the appropriate abortion rate for any particular arrest cohort to better reflect the timing of the relevant abortions and the level of arrests years later for the corresponding cohort and (2) to capture the abortion rate in the state of birth for those who had subsequently moved to a different state. In addition, we used an instrumental variable approach to address imperfections in our abortion measure. Again, the results were highly statistically significant (even using the most stringent standard error measures), and the magnitudes of the estimated abortion effects grew substantially with these better measures (Table 6). While the impact of abortion on property crime was large and highly statistically significant when estimated for the entire period from 1985 to 2014 and for the initial period from 1985 to 1997, the one deviation from the uniformly strong abortion effects was a near-zero estimate for property crime arrests from 1998 to 2014 when state-year controls were added—an anomaly that we discuss below. Finally, we followed Foote and Goetz in estimating per capita arrest rates and showed that what they viewed to be the critical test of the abortion-crime link was met for violent crime arrests—even with the most stringent standard error measures. Although the property crime estimates were negative but not statistically significant with this per capita specification, we showed that this model had the unattractive quality of using the relevant population for each particular age group in the denominators of both the arrest rate (the dependent variable) and the abortion rate (the key independent variable). Since the resulting ratio bias would tend to obscure the abortion-crime link given the documented errors in our population measures, we addressed this concern by estimating an alternative arrest rate regression based on arrest counts that directly controlled for the relevant population. The resulting estimates in columns 1 and 3 of Table 7 further strengthened the already strong violent crime results and generated a highly statistically significant estimated effect of abortions on property crime in the first period (with either standard error measure). Again, we see a weak property arrest rate effect in the second period, as in Table 6.37 5.2. Our Arrest Rates Results for Violent and Property Crime What then are we to make of the strong findings that legalized abortion substantially reduced all crime, reduced violent crime arrests for both periods, and reduced property crime arrests—but only for the first period? One possibility is that the poorer quality of arrest data by age versus state crime data may explain the weaker property crime arrest estimates in the second period. Support for this view comes from re-estimation of Table 7 using only the 26 states that had complete arrest rate data. This approach increases all of the violent and property overall period estimates and generates a statistically significant impact of abortion on property arrests, even using the most stringent standard error measure (Table 8, Column 3). But in light of our Table 4 estimates showing the dampening effect of abortion is roughly the same for both violent and property crime, it remains a bit of a puzzle why our estimated abortion effects would be greater for arrests for violent crime than for property crime (Tables 7 and 8). Two factors may be relevant to this issue, which we discuss in turn. 5.2.1. The Divergence Between Property Arrests and Property Crime in Our Second Period. Recall that the abortion-crime hypothesis posits that the legalization of abortion will reduce crime many years later as post-legalization birth cohorts move through their years of highest criminal activity. In the first portion of the article, we addressed this issue directly looking at crime data by state and year. Our subsequent arrest analysis gave us the advantage of being able to link abortion rates of birth cohorts to specific ages in which the members of that birth cohort were arrested. While this aided our effort to test the abortion-crime hypothesis in a more precise way, it was premised on the assumption that the relationship between arrests and crime would be stable over time. While this assumption is true for the relationship between violent crime and violent arrests, and was true for the first-period of our data analysis for property crime, the arrest/crime ratio for property crime shifted sharply during our second data period. Figure 8 below has two lines that document the substantial decline in property crime rates in the United States from 2000 to 2014: one reflects the police-reported property crime rate as published by the UCR and one reflects the decline documented by the National Crime Victimization Survey (NCVS). As the figure shows, property crime itself fell by roughly 30% or more. The top line in the graph, which measures property crime arrests per capita, shows a very different path. Despite the very substantial drop in property crime over this period, there was virtually no drop in property arrests over this period. Indeed, property crime arrests rose sharply starting in 2006, even as property crime continued to fall substantially.38 Figure 8 Open in new tabDownload slide Property Crime and Arrests Rates, 2000–14 (Indexed to 1 in 2000). Note: Values in 2014: Arrests = 0.986; Crime = 0.711 (UCR), and = 0.620 (NCVS). Figure 8 Open in new tabDownload slide Property Crime and Arrests Rates, 2000–14 (Indexed to 1 in 2000). Note: Values in 2014: Arrests = 0.986; Crime = 0.711 (UCR), and = 0.620 (NCVS). Since the abortion and crime hypothesis posits that the increased abortion rate will lead to reductions in crime, but our arrest rate regressions link abortion rates with arrests, Figure 8 illustrates that our arrest rate regressions should be considerably weakened in our second period. In fact, this is the exact pattern we observe in Table 7: our first-period effect is negative and statistically significant with even the most stringent standard errors, but the second-period effect is dramatically smaller and not statistically significant. Figure 9 shows the comparable graph for violent crime and arrests from 2000 to 2014. Even though there is some divergence starting in 2006, the trends in UCR violent crime and arrests are virtually identical thereafter. Moreover, the overall disparity of a 19% decline in violent crime arrests versus a 29% drop in violent crime is only about one-third the size of the arrest-crime disparity for property crime over this period. Figure 9 Open in new tabDownload slide Violent Crime and Arrests Rates, 2000–14 (Indexed to 1 in 2000). Note: Values in 2014: Arrests = 0.809; Crime = 0.714 (UCR), and = 0.536 (NCVS). Figure 9 Open in new tabDownload slide Violent Crime and Arrests Rates, 2000–14 (Indexed to 1 in 2000). Note: Values in 2014: Arrests = 0.809; Crime = 0.714 (UCR), and = 0.536 (NCVS). 5.2.2. Selective Under-Reporting of Crime. A new paper by Richard Boylan provides additional illumination on the abortion-crime relationship (Boylan 2019). Boylan begins by using two tests to show that police departments are less likely to submit statistics when crime is high. Since our hypothesis is that higher abortion rates reduce crime, it is conceivable that there might be a selection effect operating that could influence our estimates of the impact of abortion on crime. Boylan finds this to be the case, and then goes on to show that studies such as ours that rely on UCR crime data will tend to underestimate crime and thereby tend to understate the impact of policies on crime. We alluded to this effect in footnote 16 where we noted that the low-abortion states were more likely to under count UCR homicides, which led to a stronger estimated impact of abortion on murder in Table 4 when we used Vital Statistics data, which is not collected by the police and is subject to mandatory reporting requirements. In other words, the police departments that disproportionately under-report to the UCR tend to be in lower-abortion rate states. Boylan describes his findings as follows: “Depending on the manner that I account for missing statistics, an increase in the effective abortion rate of 100 per 1000 live births is associated with a reduction in property crimes by 10–11%, while accounting for sample selection leads to a 15% reduction in property crime. Thus, the results show that missing reports are biasing downwards the relation between abortion and crime.” In summary, Boylan provides evidence suggesting that the true crime-reducing impact of increased abortion is greater than our estimates using UCR crime and arrest data (on which we exclusively rely in all of our violent and property crime and arrest models). Indeed, the bias towards zero that Boylan identifies in the relation between abortion and crime is particularly noteworthy for property crime. Accordingly, it will require further exploration to determine whether the weaker second-period effects of abortion on property crime observed in Table 7 reflect a shift in policing tactics or clearance rates that alters the relationship between arrests and property crime or defects in the quality of arrest data (as suggested in Table 8) or crime data (as suggested by Boylan). 6. Considering the Impact of Lead on Crime We have already noted that the conventional explanations for both the enormous drop in crime after 1992 and the wide differences among states in the degree of this decline do not have anywhere near the explanatory power of the abortion effect that we documented in Table 4’s state panel data model for the period 1985–2014. Another novel theory posits that the efforts to reduce lead exposure have also contributed importantly to this crime decline. This raises the obvious question: might the abortion effect that we document simply be proxying for the lead effect, which is in fact the true causal force behind the post-1992 crime decline? Reyes (2007) is one of the most important papers illustrating the impact of childhood lead exposure on crime. Reyes is the first researcher to present a state-level panel data analysis that links lead levels in early childhood to subsequent changes in crime, and she specifically explores whether controlling for the lead effect undermines the abortion and crime link. The short answer is that it does not. Figure 10 reproduces Table 5 from Reyes (2007), which essentially conforms to the specification of our Table 4 over the period 1980–2002. The first row of the Reyes table estimates the elasticity of crime with respect to lead and the second row introduces the elasticity of crime with respect to abortion to test whether the abortion effect is explained by the lead effect. The highlighted estimates make it clear that this is not the case. The abortion effect on crime is extremely strong and highly statistically significant for violent crime, property crime, and murder. Figure 10 Open in new tabDownload slide Reyes (2007) Explores Lead and Abortion Effect, 1985–2002. Figure 10 Open in new tabDownload slide Reyes (2007) Explores Lead and Abortion Effect, 1985–2002. When one considers the large increase in the number of abortions that occurred after legalization, the Reyes (2007) results showing that a doubling of the abortion rate would lead to almost a 25% drop in both violent crime and murder and a nearly 15% drop in property crime is impressive. Moreover, the property estimate is significant at the 0.01 level and the violent crime and murder estimates are significant at well below the 0.0001 level! Moreover, introducing the abortion variable into the panel data model leaves no estimated lead effect on crime to be statistically significant at the 0.05 level, so it is clear that the lead hypothesis does not undermine the link between abortion and crime. 7. Conclusion It is rare for an economic theory to make predictions for twenty years into the future that are both bold and precise. The abortion-crime hypothesis of Donohue and Levitt (2001), however, did just that. Based on an extrapolation that assumed the same point estimates in the ensuing two decades as were estimated in the original sample, Donohue and Levitt (2001) predicted that crime would fall an additional 20% in the United States. The results in this paper provide strong support for that prediction. Using the same specifications as Donohue and Levitt (2001), but extended to a sample that includes an additional 17 years of data, in almost all cases the point estimates are at least as large as in the original analysis, and in many cases the point estimates are bigger. From 1997 to 2014, the effective abortion rate for violent crime rose from roughly 170 to 341 and the effective abortion rate for property crime rose from 247 to 348. Using the preferred specifications in Table 4—the same specifications upon which the original predictions were based—the implied overall crime decline due to legalized abortion over the ensuing 17 years was 17.5%.39 From the 1991 peak of crime in the United States, the rates of violent and property crime each fell by 50% and VS murder fell by 52% by 2014. Over that same period, we estimate that legalized abortion reduced violent crime by 47%, property crime by 33%, and VS murder by 41%. Thus, while many other factors were operating to stimulate or suppress crime, legalized abortion can explain most of the observed crime decline.40 The strong evidence of the impact of legalized abortion on crime in the United States would of course be strengthened by similar evidence from a different continent where the timing of abortion legalization and frequency of abortions varies greatly from ours. In fact, François et al. (2014) provide such evidence with a panel data analysis with country and year fixed effects from 1990–2007 for 16 Countries in Western Europe. The paper “confirm[s] the negative impact of abortion on crime for both homicides and thefts....” While the authors do not compute the impact of their regression coefficients and even speculate that their estimates are smaller than ours, their model showing the impact on crime 15 years after abortion legalization implies that over the ensuing decade, abortion legalization reduced homicides by 12–40% and reduced theft by 23–43%. These estimates are roughly comparable to and therefore provide significant support for our own estimates on data from the United States. An enormous literature has developed showing that optimizing the circumstances of pregnancy and early childhood can improve life prospects on everything from cognitive development and physical and mental health to educational success, earnings, and avoidance of crime (Almond, Currie, and Duque 2018). Since legalization of abortion provides a vehicle to delay childbirth until a time when these critical environmental and family circumstances would be relatively more favorable or prevent it if they are particularly dire, this growing literature on improving adult outcomes is supportive of the underlying mechanism of the abortion-crime hypothesis.41 As we have noted previously, our study has tried to elucidate one previously unidentified factor that can provide insight into the otherwise unexplained drop in crime over the last two decades. All of the estimated crime-reducing effects from legalized abortion could be generated by reducing unwanted pregnancies and births.42 But as Darroch et al. (2001) noted, “U.S. teenagers [had] the highest rates of pregnancy, childbearing, and abortion” from 1970 to 2000 compared to England, Canada, Sweden, and France primarily because of less contraceptive use. Overall, 18.8% of pregnancies in the United States ended in abortions in 2014 (Jones et al. 2007).43 Restraining access to abortion without reducing unwanted pregnancies is both personally and socially costly. Appendix A Summary Statistics for High- versus Low-Abortion Rate States, 1985–2014 . High Abortion Rate States . Low Abortion Rate States . . . Std Dev . Std Dev . . Std Dev . Std Dev . Variable . Mean . (Overall) . (Within State) . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 601.3 259.4 193.2 475.4 193.1 104.2 Property crime per 100,000 residents 3,964.5 1,357.0 1,140.7 3,794.4 1,033.9 737.4 Murder per 100,000 residents (UCR) 6.9 4.0 2.7 6.3 3.1 1.8 Murder per 100,000 residents (VS) 7.3 3.7 2.6 6.8 3.3 1.8 EAR: Violent crime 264.3 169.4 156.5 137.8 94.6 88.5 EAR: Property crime 311.5 160.8 142.0 164.3 91.6 82.8 EAR: Murder 233.2 167.8 157.5 120.8 92.9 88.1 Prisoners per 1000 residents (t-1) 3.8 1.3 0.8 3.9 1.9 1.2 Police per 1000 residents (t-1) 3.3 0.7 0.3 2.9 0.7 0.4 Real state personal income per capita 18,276.5 2,832.4 2,005.0 15,709.8 2,361.3 1,872.7 Real AFDC generosity per recipient family/1,000 (⁠|$t-15$|⁠) 4.4 1.7 1.1 3.1 1.6 1.0 State unemployment rate (percent) 6.3 2.1 1.9 6.1 1.7 1.5 Beer consumption per capita (Gallons of ethanol) 1.2 0.2 0.1 1.3 0.2 0.1 Poverty rate 13.1 2.9 1.7 13.9 3.5 1.8 . High Abortion Rate States . Low Abortion Rate States . . . Std Dev . Std Dev . . Std Dev . Std Dev . Variable . Mean . (Overall) . (Within State) . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 601.3 259.4 193.2 475.4 193.1 104.2 Property crime per 100,000 residents 3,964.5 1,357.0 1,140.7 3,794.4 1,033.9 737.4 Murder per 100,000 residents (UCR) 6.9 4.0 2.7 6.3 3.1 1.8 Murder per 100,000 residents (VS) 7.3 3.7 2.6 6.8 3.3 1.8 EAR: Violent crime 264.3 169.4 156.5 137.8 94.6 88.5 EAR: Property crime 311.5 160.8 142.0 164.3 91.6 82.8 EAR: Murder 233.2 167.8 157.5 120.8 92.9 88.1 Prisoners per 1000 residents (t-1) 3.8 1.3 0.8 3.9 1.9 1.2 Police per 1000 residents (t-1) 3.3 0.7 0.3 2.9 0.7 0.4 Real state personal income per capita 18,276.5 2,832.4 2,005.0 15,709.8 2,361.3 1,872.7 Real AFDC generosity per recipient family/1,000 (⁠|$t-15$|⁠) 4.4 1.7 1.1 3.1 1.6 1.0 State unemployment rate (percent) 6.3 2.1 1.9 6.1 1.7 1.5 Beer consumption per capita (Gallons of ethanol) 1.2 0.2 0.1 1.3 0.2 0.1 Poverty rate 13.1 2.9 1.7 13.9 3.5 1.8 Real AFDC generosity per recipient family is measured in thousands of dollars and indexed at 1982-1984 values. The 32 states in the low-abortion rate group are: Alabama, Alaska, Arkansas, Delaware, Idaho, Illinois, Indiana, Iowa, Kansas, Kentucky, Louisiana, Maine, Minnesota, Mississippi, Missouri, Montana, Nebraska, New Mexico, North Dakota, Ohio, Oklahoma, Oregon, Pennsylvania, South Carolina, South Dakota, Tennessee, Texas, Utah, Vermont, West Virginia, Wisconsin, and Wyoming. The 19 high-abortion rate states are: Arizona, California, Colorado, Connecticut, District of Columbia, Florida, Georgia, Hawaii, Maryland, Massachusetts, Michigan, Nevada, New Hampshire, New Jersey, New York, North Carolina, Rhode Island, Virginia, and Washington. Open in new tab Appendix A Summary Statistics for High- versus Low-Abortion Rate States, 1985–2014 . High Abortion Rate States . Low Abortion Rate States . . . Std Dev . Std Dev . . Std Dev . Std Dev . Variable . Mean . (Overall) . (Within State) . Mean . (Overall) . (Within State) . Violent crime per 100,000 residents 601.3 259.4 193.2 475.4 193.1 104.2 Property crime per 100,000 residents 3,964.5 1,357.0 1,140.7 3,794.4 1,033.9 737.4 Murder per 100,000 residents (UCR) 6.9 4.0 2.7 6.3 3.1 1.8 Murder per 100,000 residents (VS) 7.3 3.7 2.6 6.8 3.3 1.8 EAR: Violent crime 264.3 169.4 156.5 137.8 94.6 88.5 EAR: Property crime 311.5 160.8 142.0 164.3 91.6 82.8 EAR: Murder 233.2 167.8 157.5 120.8 92.9 88.1 Prisoners per 1000 residents (t-1) 3.8 1.3 0.8 3.9 1.9 1.2 Police per 1000 residents (t-1) 3.3 0.7 0.3 2.9 0.7 0.4 Real s