Overview Of Studies On Health
Studies of the relationship between abortion and health have progressed with advances in the field of applied microeconometrics. Borrowing from the medical sciences, random control trials (RCTs) have become the gold standard. RCTs remain rare in economics, but their acknowledged quality has pushed researchers to design studies with strong internal validity and transparent sources of identification. In this section, the author reviews the evolution of studies linking abortion and health through the improvement in research design. Early studies of abortion and health relied on cross-sectional variation to identify an association. The second phase of studies on abortion and health leveraged panel data and changes in policy to understand the determinants of abortion and its impact on fertility. A related group of studies used panel methods to estimate the cohort effect of abortion legalization on broad measures of well-being. The most recent set of studies has employed abortion legalization as an instrument for births in an effort to estimate changes not only in the health of birth cohorts exposed to legalized abortion in utero but also to estimate the potential health of children that were not born.
Early Studies Of Abortion And Health
As noted above, Grossman and Jacobowitz (1981) were the first to use the household production function framework to associate access to abortion with infant health. The empirical work involved regressions of county-level neonatal mortality rates averaged over 3 years from 1970 to 1972 on measures of the cost of fertility control and other inputs into the production of health. They used two measures of abortion availability: Dichotomous indicators of whether the county was in a state that had reformed or legalized abortion by 1970 and the 3 year average of the state abortion rate (abortions per 1000 live births) from 1970 to 1972. And they applied coefficients from the cross-sectional regression to estimate the reduction in neonatal mortality attributable to each input.
Overall the model could explain between 35% and 53% of the decline in neonatal mortality between 1971 and 1977. However the most striking result was that the abortion rate accounted for more than 50 of the explained decline for both White and non-Whites.
A series of papers followed the Grossman and Jacobowitz (1981) framework but with more recent data and greater attention to the endogeneity of abortion in the production of infant health. In one study economists estimated the reduced form production function of infant health. The outcome was again the county-level neonatal mortality rate averaged over 3 years (1976–78). They included proxies for the price of inputs such as the number of abortion providers in the county or the number of maternal and child health clinics. The results suggested that an increase in the number of abortion providers was strongly associated with decreases in neonatal mortality. Other economists used the county-level neonatal mortality rate in an effort to estimate the structural production function of infant health. They were interested in the pathways through which abortion affected survival. Thus, they also estimated structural models of low birth weight and preterm births. They included the abortion rate as well as the number of teenage users of family planning clinics as determinants of each outcome. They used two-stage least squares (TSLS) to account for the endogeneity of the abortion rate with number of abortion providers per county as an instrument (more on the instruments below). They found that state-level abortion rates were inversely correlated with neonatal mortality, low birth weight, and preterm birth. Moreover, they argued that abortion improved newborn survival by lowering the incidence of low birth weight births. Others followed this approach by estimating structural models of infant survival. However, their objective was to understand the relative contribution of government programs. These include participation in the Supplemental Nutrition Program for Women, Infants, and Children (WIC), inpatient days in neonatal intensive care units, use of family planning clinics, as well as maternal and child health clinics. As did other economists, these authors used TSLS with the availability of clinics, abortion providers, and neonatal beds as instruments. They reported that the abortion rate explained approximately half of the decline in neonatal mortality between 1964 and 1977 accounted for by the model.
The aforementioned studies used aggregate data to correlate the abortion rate with county-level measures of health. All reasoned that areas with higher abortion rates had a more optimal distribution of birth outcomes as less healthy or desired fetuses were aborted. An ecological approach appeared the only way to associate abortion to health. At the individual level, a pregnancy that is terminated is eliminated from the sample of births. There seemed to be no individual-level analog to the aggregate analysis. However, in two papers, economists applied the emerging econometrics on censored samples to analyze the effect of pregnancy resolution on birth outcomes (Grossman and Joyce, 1990). In both papers, the authors used individual-level data on births and abortions in New York City. The birth and abortion files contained information on age, race, marital status, parity, schooling, as well as measures of the availability of family planning and abortion services by neighborhood. The authors concatenated the files to create a sample of pregnancies that resulted in either an induced abortion or a live birth. They argued that the sample of births represented a nonrandom draw from the population of pregnancies. In one paper, the authors used the decision to give birth conditional on pregnancy as an expression of wantedness. Women who were selected in the birth sample were more likely to obtain timely prenatal care than those who aborted had they instead carried to term. They estimated the observed counterfactual by using the inverse Mill’s ratio to obtain the expected number of months a woman would have delayed prenatal care had she not aborted. The difference between the expected and actual months of delay for women with the same observables became an estimate of the impact of ‘wantedness’ on the demand for health-producing inputs. They found that women who had a greater probability of giving birth had less than expected delay in prenatal care.
Grossman and Joyce (1990) extended the model to include birth outcomes while treating prenatal care as an endogenous input into the production of health. They also provide a framework that signed the effect of changes in the cost of abortion, the cost of contraception, and underlying health endowment of the fetus. They treated contraception and abortion as substitutes. An increase in the cost of contraception or a decrease in the cost of abortion raises the probability of becoming pregnant. However, an increase in the cost of abortion holding the cost of contraception constant raises the probability of giving birth, conditional on becoming pregnant. For instance, assume that Black women face a higher cost of contraception due to less access and information. A decrease in the cost of abortion will lower the probability of giving birth conditional on pregnancy, increase the demand for healthy inputs, and increase birth weight. This is what the authors found for Black women but not for Whites.
These early papers were important because they tried to develop an empirical test of the association between abortion and health. They used the household production framework to incorporate the cost of fertility control in models of the quantity and quality of children. The statistical analyses became progressively more sophisticated as researchers applied recent advances in econometrics to account for the endogeneity of inputs. However, the identification strategies used then would never meet the standards of today. First, all data were cross-sectional. The lack of a panel precluded fixed effects, which would have limited the identifying variation to within-area changes in policy. Instead, authors compared the impact of abortion rates on birth outcomes in, for example, Utah relative to New York. Given the limited number of covariates, the likelihood of omitted variable bias was large. Even reduced-form analyses suffered from problems of endogeneity. The number of abortion providers in a state or county, for instance, represents the interplay of the supply and demand of abortion services instead of some exogenous measure of price. The sample selection models used by Joyce and Grossman (1990) were novel applications at the time but again lacked a credible identification strategy. More importantly, the robustness of these models depends on the availability of instruments that predict the probability of giving birth but which have no direct effect on the birth outcome. None of the instruments in the two papers could be credibly excluded from the birth outcome equation. Despite these serious drawbacks, this early work motivated subsequent studies that paid much greater attention to identification and for much of the 1990s focused on reduced-form policy questions.
A paper by economists in the mid-1990s provided a segue to the reduced-form policy-orientated papers that soon followed. The authors took the model of Grossman and Joyce (1990) as their starting point. They used individual data from the National Longitudinal Survey of Youth 1979 (NLSY79) to estimate the impact of the price of abortion on birth outcomes. State policies regarding the public financing of abortion through Medicaid served as proxies for the price of abortion in the reduced-form production function of infant health. They found no association between Medicaid financing restrictions and birth weight. In the second part of the paper, they estimated the birth probability equation and found a robust association between Medicaid financing of abortion and the decreased probability of giving birth. For Black women, the availability of Medicaid financing lowered the probability of birth by 0.10 over a mean of 0.88, which is a large effect.
Several features of the analysis are noteworthy. First, the authors used 10 years of data from the NLSY79 and were able to exploit changes in policy over time. Second, they focused on the reduced form instead of the structural production function of health. However, they used random effects instead of state-fixed effects to control for unobserved cross-state heterogeneity. A random effects specification assumes that unobserved state factors (the random effects) are uncorrelated with the policy under study; in this case Medicaid financing of abortions. This was unlikely because mostly liberal states continued to use public funds for abortion after the Hyde Amendment in 1976. In addition, there is very little within state variation in Medicaid financing of abortion. The big changes in Medicaid came in the late 1970s with the Hyde Amendment. In other words, despite the use of longitudinal data, their policy estimates are essentially obtained from cross-sectional variation in Medicaid financing of abortion. Nevertheless, the paper represented a bridge to subsequent papers in the 1990s that took advantage of panel data with state-fixed effects to eliminate confounding from hard-to-measure differences between states and counties.