The declining contribution of socioeconomic disparities to the racial gap in infant mortality rates, 1920-1970.

1. Introduction

Great improvements in physical health are among the 20th century's most impressive social achievements. In the United States, life expectancy at birth increased by more than 25 years (Atack and Passell 1994), average height increased by about 7 centimeters for native-born white males (Costa and Steckel 1997), and a number of deadly diseases and debilitating illnesses were all but vanquished. As the population's overall level of health improved, racial gaps in health outcomes persisted. Such gaps remain a serious concern for U.S. policymakers (Department of Health and Human Services 2000; Levine et al. 2001; Byrd and Clayton 2002). In large part, this is because physical health is a direct determinant of well being. Additionally, because one's health may influence educational attainment, labor force participation and productivity, and one's children's health, the consequences of intergroup health disparities may be quite far reaching (Grossman 1975; Edwards and Grossman 1979; Currie and Hyson 1999).

In discussions of intergroup health differences, infant mortality is a frequently referenced barometer. As shown in Figures 1 and 2, the overall infant mortality rate (deaths under 1 year of age per 1000 live births) and the absolute (nonwhite-white) racial gap in infant mortality rates fell over time, but nonwhite infants were always more likely to die than white infants. In fact, on average, nonwhite infants were about 75% more likely to die than white infants during the 20th century. When translated Into numbers of "excess" deaths, it is clear that the racial gap remained nontrivial even as it narrowed in absolute terms: In 1940, approximately 8900 more nonwhite infants died than would have if they had had the same mortality rate as whites; in 1970, the excess was approximately 8400 infants. (1)

The literature on infant mortality is voluminous and multifaceted, but relatively few studies have systematically examined how racial gaps in infant mortality changed in the early and mid-20th century, choosing instead to focus on more recent data (see inter alia, Rochester 1923; Woodbury 1925; Shapiro, Schlesinger, and Nesbitt 1968; Chase 1972; Shin 1975; Grossman and Jacobowitz 1981; Ewbank 1987; David and Collins 1997; Department of Health and Human Services 2000). Understanding the racial gap and its movement in the earlier period is important for at least three reasons. First, by far the largest improvements in infant mortality, and the largest declines in the black-white gap, occurred before 1970. Historical experience therefore provides perspective on the magnitude of contemporary infant mortality levels, changes, and disparities. Second, the most striking, and puzzling, aspect of the post-1970 experience is that racial disparities in socioeconomic characteristics account for less than half of the racial gap in infant mortality rates (Hecht and Cutright 1979; Miller 2001; Cutler and Meara 2003). We find that this was not always the case; rather, the large "unexplained" portion of the gap emerged gradually in the postwar period. Consequently, researchers and policymakers interested in understanding the origins of the large unexplained residual may benefit from closer empirical scrutiny of the pre-1970 period. Third, for most of the 20th century, prevailing patterns of racial discrimination embedded themselves in the health care system, perhaps compounding the disadvantages that African Americans already faced due to their geographic distribution and relative lack of financial and educational resources. The health implications of racial segregation (and desegregation) have not been adequately documented, especially for the period when segregation was most intense.

In this paper, we explore state-level infant mortality data in a framework that sheds light on both the declining overall level of infant mortality and the racial gap. After discussing the interaction of race, history, and infant mortality in the context of a simple model of demand for and supply of health care, we construct and analyze a panel of state-level data for whites and nonwhites. Our analysis focuses on the 1920 to 1970 period because, as noted already, it is underexplored despite experiencing levels and changes in mortality that dwarf those of the last 30 years. We make efforts to link our findings to the existing literature on the post-1970 period, but we do not attempt to reexamine data for the later period. In part, this is because major shifts in social policy, especially the legalization of abortion, add a layer of complexity that cannot be adequately treated in the space of this paper. Additionally, large infant-level datasets are available to researchers for analysis of the later period, and it is unlikely that extending a state-level analysis to later years could yield more insight than the micro-level studies already have.

We find that differences in income, women's education, urban residence, and the supply of physicians can account for a large portion of the racial gap in infant mortality rates from 1920 to 1945, but that they account for a declining proportion of the gap thereafter. We also find that although there was a strong secular decline in infant mortality for both races, the racial gap did not narrow continuously over time. We discuss the post-1940 period in light of changes in birth weight, maternal characteristics, smoking and breast-feeding behavior, air pollution, and institutional changes that may have influenced access to professional medical care. Some of these trends tended to widen the racial infant mortality gap even as movements in fundamental socioeconomic characteristics tended to narrow it.

2. Conceptual Framework and Historical Context

Our interpretation is guided by a simple model of infant mortality in which the likelihood of an infant's death is influenced the family's consumption of quality-adjusted units of nutrition, housing, health-related information, and health services. For convenience, we refer to this bundle of goods as health care (broadly speaking), and we think of it as an intermediate input in an infant health production function. (2) Infant health may also vary because of changes in the disease environment and because of variation in families' (especially mothers') initial endowments of human capital (a combination of health, knowledge, and responsiveness to health-related information). Of course, it may also be influenced by changes in technology, policy, and health-related institutions. This section begins by discussing the technology of infant medical care. Next, it identifies several factors that influenced racial disparities in health care (and therefore in health outcomes) throughout the 1920 to 1970 period. Finally, it highlights the major governmental interventions that targeted maternal and infant health.

The Technology of Infant Medical Care

Medical technology can reduce infant mortality in either the neonatal period (birth to 28 days) or the postneonatal period (28 days to 1 year). Neonatal infant mortality typically results from congenital defects, birth injury, and low birth weight (under 2500 grams), while postneonatal infant mortality is more likely to result from environmental factors such as communicable disease. Early in the 20th century, medical technology held few benefits for either neonates or postneonates. In 1900, only about 5% of all births occurred in hospitals (Wertz and Wertz 1989). At the time, the benefits of hospital and physician-attended birth were slim, and hospitals and physicians could do little to improve the health of unhealthy newborns. The better hospitals strove to be aseptic, but general practitioners were poorly trained and probably no better at delivering infants than midwives (Wertz and Wertz 1989).

After 1900, the accumulation of clinical experience at hospitals, reforms in medical education following the Flexner Report in 1910, and the rise of specially trained obstetricians increased the advantages of hospital births, particularly for complicated deliveries. Increases in urbanization and insurance coverage, along with refinements in pain-relieving procedures, meant that birth became a predominantly hospital-based activity by midcentury. In 1955, 94% of deliveries took place in hospitals where doctors could assist mothers with anesthetics, drugs to speed up delivery, X-ray machines to detect pelvic abnormalities, fetal heart monitoring equipment, and cesarean sections (Wertz and Wertz 1989). These improvements helped reduce neonatal infant mortality from birth injuries from 4.8 to 2.4 deaths per 1000 births from 1930 to 1960 (Linder and Grove 1943; Grove and Hetzel 1968).

The slow accumulation of clinical experience (e.g., in feeding) and technological advances (e.g., in incubators and drug therapies) contributed to a steady decline in neonatal infant mortality due to noninjury reasons (predominantly, low birth weight) up to 1950 (Costa 1998). But great strides in caring for unhealthy neonates, particularly those with low birth weights, were made relatively recently. While special nurseries equipped with incubators existed in many cities in the 1940s (Corwin 1952; Cone 1985), the mortality rate for low birth weight infants remained high: In both 1950 and 1960, about 170 per 1000 low birth weight infants died as neonates (Chase 1972). It was not until the introduction of modern neonatal intensive care units (NICUs) in the late 1960s--and significant refinements in ventilation techniques--that the mortality rate of low birth weight infants declined significantly. By 1983, the mortality rate of low birth weight infants had fallen to 96 deaths per 1000 births (National Center for Health Statistics 2001).

Improved medical technologies contributed to dramatic reductions in posmeonatal infant mortality before midcentury. In particular, two leading causes of postneonatal infant death, pneumonia and gastroenteritis, could be effectively treated in hospitals by the mid-1940s (Almond, Chay, and Greenstone 2001). Infants with pneumonia benefited from intravenous fluid replacement, breathing assistance, and (after 1945) antibiotics. Intravenous fluid replacement was especially helpful to infants with gastroenteritis since they dehydrated rapidly otherwise. Overall, the postneonatal infant mortality rate fell from about 29 to 9 per 1000 births between 1930 and 1950, and the posneonatal proportion of all infant deaths fell from 45% to 30%.

Technological innovation and improved medical care had clear implications for the overall level of infant mortality in the 20th century. The implications for the racial gap in mortality rates are less obvious. In part, they hinge upon racial disparities in access to high-quality care, and the next section suggests several factors that limited blacks' ability to attain such care.

Income, Education, Discrimination, and Health Care

Throughout the period under study, the average income of nonwhites lagged behind that of whites (Smith 1984), implying tighter budget constraints and less demand for all normal goods, including those we have bundled together as health care. Additionally, for several reasons discussed below, nonwhites faced relatively high costs for any given quantity of health care. Thus, both demand and supply conditions tended to dampen nonwhites' consumption of health-related goods relative to whites'.

These conditions were inextricably linked to the population's geographic distribution and to the historical phenomena, such as slavery, that determined that distribution. In 1920, 85% of African Americans resided in the South (compared to 25% of whites), and in 1960, 60% still resided there (compared to 27% of whites). Throughout this period, southern per capita income was relatively low. In 1940, for example, it was about 64% of the national average (Mitchener and McLean 1999). In the absence of large federal transfers (discussed below), this implied tightly constrained public supplies of health-related services, including hospitals, subsidized care for the poor, and investments in sanitation and education (Beardsley 1987). (3) Within the South, blacks' economic resources were further constrained relative to whites' by slavery's legacy of poverty, reinforced by discriminatory practices in labor, land, and credit markets, as well as in the educational system. (4)

The shortcomings of the southern educational system, including its discriminatory treatment of blacks, have been well documented (Margo 1990). According to the 1920 census, approximately one in four blacks over the age of 10 was illiterate. Educational disparities had an indirect effect on health care consumption by perpetuating income and wealth disparities. More directly, blacks may have had less exposure to health-related information that was transmitted through schools, and once out of school, poorly educated blacks might have found it difficult to acquire and evaluate health-related information. For example, Ewbank and Preston (1990) discuss the early 20th-century diffusion of information regarding healthy feeding practices and hygienic conditions through pamphlets, magazines, newspapers, books, and schools, and they argue that this information helped lower the infant mortality rate, at least for families who received and responded to the information. Educational deficiencies might also have facilitated the continuing influence of superstition, foil remedies, and lay referral even when modern medical care and information was available (Beardsley 1987). Along these lines, in recent studies, better educated people appear to be more behaviorally responsive to health-related information than others (Berger and Leigh 1989; Kenkel 1991; Meara 2001). In nearly all studies that examine the determinants of infant mortality, including those for developing countries (Hobcraft, McDonald, and Rutstein 1984, 1985; Subbarao and Raney 1995; Preston 1996), the mother's education level is found to be an important factor.

Not only were southern blacks concentrated at the bottom end of the socioeconomic ladder, they also tended to reside in more rural areas, especially early in the period under study. Therefore, they often lived far from hospitals and doctors and faced considerable costs (in terms of time, effort, and money) when seeking professional medical services. The likelihood of physician attendance at delivery in southwest Tennessee in 1940 illustrates the importance of proximity to care for nonwhites. Physicians attended 97% of nonwhite births to residents of Memphis, which had several large hospitals, attended 43% of nonwhite births to residents of Shelby County (excluding residents of Memphis), and attended only 15% of nonwhite births to residents of Fayette County, the next county to the east of Shelby, which had no general, maternal, or children's hospitals (State of Tennessee 1941). (5)

Discriminatory practices within the medical system (which were not limited to the South) also may have effectively raised the cost of health care for black households. Such discrimination took many forms, including restricting blacks' access to health care facilities, severely limiting the number and quality of black doctors (both directly through the medical school admissions process and indirectly through the provision of inferior elementary and secondary schools), and using black patients for clinical studies with dubious ethical standards (Johnson 1949; Seham 1964; Beardsley 1987; Smith 1999). (6)

Under the circumstances, it is not surprising that the available evidence indicates that southerners, particularly black southerners, trailed the rest of the country in the consumption of health care. In 1945, only about 25% of nonwhite southern births and 68% of white southern births occurred in hospitals, compared to about 81% of nonwhite and 91% of white births outside the South. (7) In 1940, the southern level of physicians per capita was only about 60% of that in the rest of the country, almost surely implying less consumption of physician services. Also in 1940, only 35% of southern homes had flush toilets (an indicator of housing quality and sanitation), about half the level for the rest of the country, (8) Nationally, as late as the 1960 birth cohort, adult black males remained more than a centimeter shorter than whites on average (an indicator of childhood nutrition and health) (Costa and Steckel 1997).

Between 1920 and 1970, millions of rural southerners moved to other parts of the country, and black migrants went primarily to central-city neighborhoods. These migrants might have benefited from proximity to medical facilities and from higher incomes, but urban life may have been associated with health behaviors (like smoking) and environmental conditions (like air pollution) that are now known to be detrimental to maternal health and fetal development. Moreover, discriminatory practices underpinning patterns of urban residential segregation artificially constrained the supply of housing for blacks, thereby raising its price and exacerbating crowded, unhealthy living conditions in the emerging ghettos (Massey and Denton 1993). The movement might also have required adjusting to a new disease environment, forfeiting support from extended family networks, and taking on physically debilitating industrial jobs (see Maloney and Whatley 1995).

Institutional Interventions

Against this backdrop, several charitable organizations made efforts to improve the health and education of poor children. The Rockefeller Sanitary Commission for the Eradication of Hookworm (1909-1915), the Rosenwald Fund, and the Duke Endowment all made efforts to improve poor children's health and educational status (Beardsley 1987). Between 1926 and 1942, the Rosenwald Fund contributed to black hospitals in both the North and the South, funded public health training and research, contributed to thousands of black schools in the South, and was an early participant in the battles against tuberculosis and syphilis. The Duke Endowment, established in 1924, was instrumental in funding hospitals and health care for the poor in the Carolinas.

Perhaps more importantly, the federal government became increasingly engaged in the study, funding, and provision of health care for infants and mothers (see Meckel 1998). Congress established the Children's Bureau in 1912 to "investigate and report ... upon all matters pertaining to the welfare of children," with special attention to the issue of infant mortality (Bradbury 1956). Under the auspices of the Children's Bureau, the Sheppard-Towner Act of 1921 provided matching grants to states for educational materials and instruction on maternal and infant nutrition, care, and hygiene (Lemons 1969). Although it was discontinued in 1929, the Sheppard-Towner program laid the political and administrative foundations for subsequent and more extensive federal health initiatives (Sinai and Anderson 1948; Bradbury 1974).

The Social Security Act of 1935 featured a federal-state program for maternal and child health that, in terms of both spending and scope of activity, eclipsed Sheppard-Towner. In the late 1930s, proposals to expand the government's role in funding and providing universal maternal and infant health care met with staunch resistance from the American Medical Association (Meckel 1998). Nevertheless, from 1943 to 1949, the Children's Bureau administered an extensive Emergency Maternal and Infant Care (EMIC) program that funded medical care for the pregnant wives and infant children of low-ranking servicemen. EMIC handled approximately 1.5 million cases and covered nearly one out of every seven U.S. births at its peak (Sinai and Anderson 1948; Bradbury 1956). Just as importantly, the program directly elevated the standard of prenatal and obstetric care through its process of review and approval of medical facilities and services, especially in the South (Sinai and Anderson 1948; Beardsley 1987). (9)

The EMIC program also provided an indirect, long-term benefit by casting light on the general inadequacy of the nation's hospital system. In response, Congress launched the colossal Hill-Burton program (1946-1971) to inventory and modernize existing hospitals and to fund new construction in areas (like much of the South) with relatively few hospital beds per capita (Lave and Lave 1974). The establishment of Medicaid in 1965, an ambitious component of the Great Society legislative agenda, further extended government spending on health care for relatively poor families within a federal-state matching program (see Gruber 2000). A disproportionate number of poor families resided in the South, but because several southern states waited until 1969 or later to initiate the program, many poor southerners had to wait to enjoy the program's benefits.

Other Factors

In addition to income, education, place of residence, and institutional changes, several other factors may help explain the racial gap in infant mortality rates. For example, differences in the age distribution of mothers may be important, since infants born to teenagers or to women over 30 are more likely to die than infants born to mothers in their 20s (Pampel and Pillai 1986; Loudon 1991; David and Collins 1997). Some studies suggest that the availability of family planning services may also affect infant mortality rates (e.g., Subbarao and Raney 1995) and that maternal health both before and during pregnancy is a key factor in infant health (Kempe et al. 1992). Finally, maternal behaviors such as smoking and breastfeeding undoubtedly impact infant health, and changes in these behaviors over time and across races could have affected the racial gap in infant mortality rates. In general, these factors cannot be captured in our econometric analysis, but we will return to these points in discussing the growing unexplained portion of the racial infant mortality gap.

3. State-Level Data

In the absence of family- or infant-level data, we exploit variation across states and over time at 5-year intervals to measure the empirical strength of the links between socioeconomic variables and infant mortality. (10) At each point in time, our dataset includes (by race and state) the infant mortality rate, estimates of real per capita income, the proportion of the population residing in urban areas, average years of education for women between 20 and 40 years of age, and the number of physicians per capita (not race specific). These variables reflect the basic financial, environmental, and medical resources available to families having children. (11) The intermediating variables between these resources and infant mortality (such as birth weight) will be discussed when and where evidence is available (see section 5), but they do not enter the econometric exploration directly. (12)

Calculating the infant mortality rate requires both death- and birth-registration data. In the United States, a national death-registration area, covering just 10 states and some cities in other states, was established in 1900. A national birth-registration area was established in 1915, also with geographically limited coverage. Over time, states were added to the official registration areas as they passed laws requiring the registration of deaths and births and as they demonstrated to the Census Bureau that they met minimum requirements for completeness of coverage (90%).

By 1933, all states were covered, but there is reason to believe that the extent of underreporting was sometimes significant (Linder and Grove 1943; Demeny and Gingrich 1967). Using 1940 census data to verify the birth-registration data, the Census Bureau found that about 94% of white births were indeed registered, and 82% of nonwhite births were. By 1950, 99% of white births were registered and 94% of nonwhite births were (U.S. Bureau of the Census 1950). The underregistration of births is not a problem (for the purposes of this study) if births and infant deaths were underreported by the same proportion within states. Otherwise, the infant mortality rate would be mismeasured, and if this mismeasurement is not random, it could confound the econometric analysis. For example, if relatively poor states had relatively weak registration systems, and if infant mortality rates were systematically underestimated for those states, we would tend to observe a spurious positive correlation between income and infant mortality. Unfortunately, no checks exist on the completeness of infant death registration. So, for the most part, like Shin (1975), who discusses the issue at some length, we must take the infant mortality data at face value. However, we check some econometric specifications that take account of information from the Census Bureau's 1940 accuracy study. We also estimate relationships after excluding the early (and potentially inaccurate) years from the sample.

We discuss the construction of each of the regressions' explanatory variables and data sources in the Appendix. A household's ability to pay for food, housing, and medical services depends in large part on its income, and so we include measures of per capita income by state, race, and year. (13) Education may affect infant mortality rates both indirectly, through its effect on earnings, and directly, through its effects on health-related knowledge and responsiveness to that knowledge (Berger and Leigh 1989; Kenkel 1991; Meara 2001). Therefore, even though education and income tend to be strongly correlated, we include measures of both in the regressions below, focusing on the educational attainment of women between 20 and 40 years of age.

Prior to 1920, it is clear that urban residence was relatively hazardous for infants' health. Haines (2001, p. 3) notes that "A variety of circumstances contributed to the excess mortality of cities: greater density and crowding, leading to the more rapid spread of infection; a higher degree of contaminated water and food; garbage and carrion in streets and elsewhere not properly disposed of; larger inflows of foreign migrants, both new foci of infection and new victims; and also migrants from the countryside who had not been exposed to the harsher urban disease environment." The unconditional gap between urban and rural mortality rates eroded during the early decades of the 20th century as public health initiatives took root, medical science progressed, water and food supplies improved, and urban hospitals increased the volume and quality of available health care. We take a broad interpretation of the estimated urban effect because the price level adjustments to nominal income do not distinguish between more and less urban states. Thus, the urban variable may capture not only health disamenities associated with urban life, but also the relatively high cost of living in urban areas.

Finally, the econometric exploration includes the number of physicians per capita to reflect the availability of medical care. Ceteris paribus, we expect a larger number of physicians per capita to lower the effective cost of medical care, and therefore to lower the rate of infant mortality. A higher proportion of doctors in the population could also increase general awareness of public health issues (and remedies). Also, to the extent that doctors are attracted to areas with plentiful and state-of-the-art medical facilities, the physicians-per-capita variable will reflect the supply of medical facilities (Lave and Lave 1974). Of course, there is no quality-of-training or vintage-of-knowledge adjustment built into this measure of health service supply, and so we must rely on time-period dummy variables to capture such trends.

Table 1 reports summary statistics for our basic data set for each race category. Some states had very few nonwhite residents, and so the nonwhite sample includes fewer state-year cells than the white sample. Each state-race-year observation is weighted by the number of births that it represents. Not surprisingly, on average, nonwhites in this sample had lower incomes, less education, and higher infant mortality rates than whites. Although initially less urban than whites, nonwhites urbanized quickly during this period, and by 1970 the proportion of nonwhites in urban areas exceeded the proportion of whites. The physicians-per-capita variable is identical within state-year ceils for nonwhites and whites, and so the relatively low average value for nonwhites is driven primarily by nonwhites' concentration in states with relatively few physicians per capita.

4. Results: Accounting for the Racial Gap in Infant Mortality

Our basic regression equation is expressed in log form, implicitly assuming constant elasticities between the dependent and independent variables:

(1) ln IM[R.sub.it] = (ln[X.sub.it])[beta] + [[gamma].sub.t] + [e.sub.it],

where i indexes states, t indexes time periods, X is a set of race-state-year characteristics, and IMR stands for infant mortality rate. We run the regressions separately for whites and nonwhites, thereby allowing the coefficients to differ between race categories. (14) We add the time-period dummy variables ([gamma]) to absorb unobserved period-specific factors influencing infant mortality (such as advancing medical technology), and in some specifications we add region or state dummy variables to absorb area-specific effects. The form allows for nonlinearities that clearly exist in the relationships, and it does so in a parsimonious way. It also produces coefficients that are straightforward to interpret and tractable for decomposition analysis. (15) Pritchett and Summers (1997) chose a similar functional form in their analysis of an international panel dataset, as did Flegg (1982). Later in this section, we split the sample at 1950 and allow the coefficients to vary between the early and later periods. The main results are unaffected by the split.

Table 2 reports regression results from three different log specifications. Columns 1 and 2 correspond to the basic specification described by Equation 1. Columns 3 and 4 add state dummy variables to the specification, and columns 5 and 6 include a dummy variable for the southern region (rather than a full set of state dummies). (16) In general, the coefficients reported in Table 2 have the expected signs: Ceteris paribus, higher levels of income, women's education, and physicians per capita tend to lower infant mortality rates, whereas higher levels of urbanization are correlated with higher infant mortality rates. When we forfeit a great deal of the cross-state variation by including state-fixed effects (columns 3 and 4), the standard errors of the coefficient estimates increase, and some of the coefficients change noticeably in magnitude, particularly for whites.

Comparing the coefficients across racial groups (columns 1 and 2), it appears that urban residence was more detrimental for nonwhites than for whites, and that education was more beneficial for whites than for nonwhites. These gaps persist, though their magnitudes change, in columns 3, 4, 5, and 6, when state or region dummies are included in the regressions. The point estimates imply that a 10% increase in urbanization was associated with a 1.05% increase in white infant mortality and a 2.86% increase in nonwhite infant mortality. This might reflect the relatively poor living conditions common to many nonwhite urban neighborhoods in the post-World War I period (Massey and Denton 1993). A 10% increase in education was associated with a 10.2% decrease in the white infant mortality rate, but only a 4.5% decrease in nonwhite infant mortality. The comparatively low returns (in terms of mortality) to education for nonwhites might reflect the relatively low quality of education received by nonwhites (see Margo 1986, 1990). Conditional on income, education, and urbanization, a 10% rise in the number of physicians per capita was associated with a 13% decline in infant mortality for nonwhites and an 18% decline for whites.

Columns 5 and 6 include a southern region dummy variable to provide a sense of how different, on average, infant mortality was in the South compared to elsewhere when accounting for state-level differences in income, education, urbanization, and physicians per capita. For whites (column 5), there is no significant difference in southern infant mortality after accounting for each state's socioeconomic characteristics. For nonwhites, however, the infant mortality rate is nearly 15% higher in the South, ceteris paribus.

There is evidence that southern states had less complete birth-registration statistics than others, at least in the prewar period (Linder and Grove 1943). If this translated into a systematic underreporting of infant mortality, then the South coefficients may understate the true degree of southern mortality disadvantage. We tested this hypothesis by adding the estimated proportion of all births registered in 1940 (from Linder and Grove) to the regressions specified in columns 5 and 6.17 In the white sample, after conditioning on the quality of the birth-registration system, the South coefficient remained insignificant. In the nonwhite sample, the South coefficient increased slightly to 16.5%.

Even with controls for income, education, urban residence, and supply of physicians, a strong secular trend in infant mortality is manifested in the downward march of the coefficients on the time-period dummies. Two aspects of the time-period coefficients are especially interesting. First, even as the colossal Hill-Burton program pumped federal funds into the health care system (Lave and Lave 1974), the secular decline in infant mortality stagnated for nonwhites from 1950 to 1965 and for whites between 1955 and 1965. Thus, the regressions suggest that the midcentury plateau evident in Figure 1 is not due to adverse movements in the independent variables offsetting an underlying downward trend. Second, during some intervals, the nonwhite time-period coefficient fell by more than the white coefficient, and during others the white time-period coefficient fell by more than the nonwhite one. We discuss these uneven movements at length in the paper's next section.

To what extent can racial disparities in the independent variables account for the gap in infant mortality? In general, the answer to this kind of question depends somewhat on whether coefficients from the white or the nonwhite regression are used to weight racial differences in characteristics (Oaxaca 1973). Rather than arbitrarily choosing one set of coefficients or the other, we use an average of the two sets of coefficients. In Figure 3a we plot the proportion of the gap that is accounted for by racial differences in observed characteristics at each point in time. From this perspective, around 80% of the gap can be accounted for by differences in characteristics up to 1945, but by 1970, only about a third of the gap can be explained.

Table 3 decomposes the infant mortality gap, variable by variable, for each year. Through most of the period under study, the largest contributor to the racial gap in infant mortality is the gap in education levels of women aged 20 to 40. Up to 1935, the difference in years of education accounts for more than 70% of the observed infant mortality gap (for example, in 1930, 0.343/0.493 = 0.70). But the racial gap in years of education narrowed quickly over time, and by 1970 the educational gap can account for only about 12% of the infant mortality gap. Up to 1945, the income gap accounts for about 25% of the overall infant mortality gap, but by 1970 income accounts for only 15% of the gap. All else equal, whites' greater likelihood of urban residence tended to narrow the infant mortality gap early in the period under study, but the nonwhite population urbanized quickly and surpassed the white urban proportion by 1960. Finally, the geographic distribution of physicians accounts for 7% of the infant mortality gap early in the period under study, but with the redistribution of nonwhites (and to some extent, of physicians), the contribution diminished over time.

These findings fit reasonably well with work on more recent data that suggests that differences in basic socioeconomic characteristics can explain only part of the racial infant mortality gap (e.g., Hecht and Cutright 1979; Miller 2001; Cutler and Meara 2003). (18) The apparent decline in the significance of socioeconomic differences in explaining the infant mortality gap could be interpreted in different ways. It may be that unobserved factors influencing infant health diverged even as the observed socioeconomic factors converged. Alternatively, it may be that the relationship between the observed characteristics and infant mortality changed in the postwar period, and that, consequently, the coefficients estimated over the full 1920-1970 period are inaccurate descriptors of the 1950-1970 experience.

To test the latter hypothesis, we reestimated regression equations separately for the 1920 to 1945 and the 1950 to 1970 periods, thereby allowing the coefficients to change from the early to the later period. Results are reported in Table 4. Although the coefficients do change somewhat from the early to the late period, the changes are not statistically significant (at the 5% level). (19) The differences in the point estimates are nontrivial, however, and therefore for Figure 3b we constructed a two-part series that is similar in nature to that presented in Figure 3a, but that is based on the separate sets of coefficients for the early and late periods. The pattern in Figure 3b is strikingly similar to that of Figure 3a: From 1950 to 1970, racial disparities in observable socioeconomic characteristics explained less and less of the racial infant mortality gap. (20)

A related hypothesis is that the observable characteristics might still be strong predictors of postneonatal mortality from 1950 to 1970 but weak predictors of neonatal mortality (Stockwell 1962; Jiobu 1972). Coupled with the declining proportion of all infant deaths in the postneonatal category, perhaps this is why the socioeconomic characteristics appear to lose their explanatory power for overall infant mortality. After running separate regressions and undertaking separate decompositions for neonatal and postneonatal mortality rates from 1950 to 1970, we find that socioeconomic characteristics do explain more of the postneonatal gap than the neonatal gap. But for both rates, the proportion of the gap accounted for by socioeconomic characteristics declines over time (from about 63% to 27% for neonates, and from about 90% to 40% for postneonates). Again, this suggests that something other than the relationship between the observed socioeconomic characteristics and infant mortality must have changed over time. (21)

5. Discussion: Exploring the Unexplained Changes in the Infant Mortality Gap, 1940-1970

In this section, we discuss several forces that may have influenced the racial gap in infant mortality but that were not observed in the econometric investigation above. After 1940, Figure 2 features two periods of abrupt decline (1941-1946 and 1966-1971) and one period of prolonged stagnation (1948-1965). The two episodes of rapid decline depicted in Figure 2 coincide almost exactly with the only two periods of significant decline in the racial wage gap (for men) (Donohue and Heckman 1991), and, in fact, Table 3 shows that income and educational convergence helped drive much of the convergence in infant mortality from 1940 to 1945. (22) In this period, as shown in Table 5, the largest declines in nonwhite infant mortality occurred in the South, and it is possible that the federal government's expanded military and administrative presence in the South disproportionately benefited nonwhites. During the war, government efforts virtually eliminated malaria, aggressively targeted sexually transmitted diseases and tuberculosis, and supported the extensive Emergency Maternal and Infant Care program for the wives and children of low-ranking servicemen, a program with (potentially) positive spillovers to the general quality of infant care (Sinai and Anderson 1948; Beardsley 1987).

Table 3 implies that given the convergence in socioeconomic characteristics, the racial gap should have declined by 0.13 log points instead of rising by 0.13 log points from 1950 to 1965. Alternatively, the nonwhite/white infant mortality ratio should have been about 1.24 in 1965 rather than 1.88. While white infant mortality continued to fall for both neonatal and postneonatal infants through the 1950s, albeit slowly compared to the 1940s, the nonwhite rates were nearly constant. This trend was especially evident in the South (Table 5), where the nonwhite infant mortality rate increased slightly between 1950 and 1960 while the white rate declined by 6.1 per 1000 births.

After 1950, we can observe several variables that had more proximate connections to infant mortality than the basic socioeconomic characteristics had, including birth weight distributions, maternal characteristics, maternal behaviors, environmental factors, and insurance coverage. Trends in these variables might provide some clues as to why racial convergence in basic socioeconomic characteristics did not result in more compression of the infant mortality gap. In general, the data are unsuited for the regression analysis of the previous section (they are often not available by race, state, and year), but the basic trends and empirical links to infant health are clear enough to indicate whether or not they could have significantly influenced the racial gap in infant mortality rates.

Birth Weight Distributions

The widening infant mortality gap after 1948, at least among neonates, can be accounted for in large part by widening racial differences in birth weight distributions. (23) Comprehensive birth weight statistics became available during the 1950s, and Chase and Byrnes (1972) noted an increase in the proportion of low birth weight neonates (under 2500 grams) among nonwhites from 10.2% of births in 1950 to 12.8% in 1960 and 13.8% in 1966. Among whites, the low birth weight proportion fell from 7.1% in 1950 to 6.8% in 1960 and then increased slightly to 7.2% in 1966. Although these changes in the weight distributions may appear small, they may have had important consequences for the racial gap in infant mortality because the neonatal mortality rate of low birth weight infants was very high: around 160 per thousand births during the 1950s (for nonwhites) compared to about 10 per thousand for neonates weighing more than 2500 grams (Chase 1972). In fact, Table 6 shows that given the 1950 distribution of nonwhite neonatal mortality rates across weight categories, the change in the distribution of nonwhite birth weights from 1950 to 1960 by itself would have raised the nonwhite neonatal infant mortality rate by 5.7 per thousand live births. This upward force completely offset the benefits from declining mortality rates within each weight category.

An immediate concern is that the apparent shift in the nonwhite birth weight distribution is a statistical artifact associated with the increasing proportion of nonwhite births occurring in hospitals. However, the proportion of low birth weight nonwhite infants increased in all regions, including those where birth-registration rates were already very high in 1950 and where there were relatively small changes in the proportion of infants delivered in hospitals. (24) Chase and Byrnes (1972) evaluate the measurement error hypothesis from a variety of perspectives and conclude that the birth weight shift is unlikely to be solely an artifact of the data. Cutler and Meara (2003) also note that the shift in the birth weight distribution is continuous over the entire 1960 to 2000 period for very low birth weight infants, reinforcing the notion that the shift was not a one-time occurrence resulting from better reporting of births.

Even if the rising proportion of low birth weight infants among nonwhites is the proximate cause of the widening neonatal infant mortality gap in the 1950s, it leads immediately to a more difficult question: Why did the birth weight distribution change? Furthermore, as shown in Table 5, the postneonatal mortality gap, which (in this period) is generally regarded as being fairly insensitive to birth weight, also failed to narrow during the 1950s. Thus, it seems that even in the proximate sense, there must be more to the story than adverse changes in the distribution of nonwhites' birth weight.

Maternal Characteristics: Fertility, Age, Marital Status, and Birth Order

Correlations between certain maternal characteristics and infant health have been documented and appreciated for decades (e.g., see Rochester 1923). Could racial divergence in maternal characteristics have driven the divergence in birth weight distributions, or, ultimately, the divergence in infant mortality rates in the 1950s as the baby boom reached its peak? On the basis of observable trends in age, marital status, and birth order, the answer appears to be no. The data suggest that the trends made relatively small contributions to changes in the birth weight distributions and infant mortality rates, at least up to 1960.

White and nonwhite fertility rates (births per 1000 women ages 15-44) were highly correlated during the period under study (0.92 in annual time-series for 1920-1970), but there was considerable divergence during the 1950s as the fertility rate gap increased from 27.3 in 1948 to 47.6 in 1959, before declining to 28.9 in 1970. (25) In theory, this divergence in relative fertility rates could have affected the racial gap in infant mortality rates through at least two channels. First, if less healthy women were selected into motherhood as fertility rates increased, then the average health of black mothers relative to white mothers could have fallen. Trends in detailed measures of maternal health are difficult to document in this period, but age and marital status can be observed and are correlated with maternal health and infant health outcomes. Second, within the group of women who were mothers, higher fertility rates may have been associated with higher average birth order, and high-order births (5 and above) were associated with higher risk of infant mortality. (26) We examine the available age, marital status, and birth-order data in turn.

Differential racial shifts in maternal age distributions could have widened the infant mortality gap because younger mothers (under 20 years old) and older mothers (over 40) were more likely to deliver low birth weight infants than other women. Column 4 of Table 7 reflects the impact of changes in maternal age distributions (1950 to 1967) on the proportion of children with low birth weight, given the probability of low birth weight in each age category in 1950. For both nonwhites and whites, the impact is small (see sum rows). Column 6 captures the impact of changes in the probability of low birth weight, given the age distribution of mothers in 1950. For whites, the impact is again small, but for nonwhites the impact was relatively large, reflecting increases in the likelihood of low birth weight in every age category for nonwhite mothers.

The link between neonatal mortality and maternal age is examined directly in Table 8. Given the 1960 distribution of neonatal mortality rates across maternal age categories for white and nonwhite infants, it appears that the shift in the age distribution of nonwhite mothers had an insignificant effect on the infant mortality rate. The proportion of infants born to older mothers increased only slightly, and the proportion born to young mothers fell. The net impact on infant mortality was close to zero.

The mortality rate of illegitimate infants was higher than for other infants, reflecting the socioeconomic characteristics of the mothers, as well as their age and perhaps their attitude toward the (often unintended) pregnancy (Shapiro, Schlesinger, and Nesbitt 1968; Department of Health and Human Services [DHHS] 2000). The Vital Statistics coverage for this variable is limited both geographically and temporally, but the major post-1940 trends are readily discernable. In 1940, the nonwhite proportion of illegitimate births was much higher than the proportion among whites in the states reporting the variable (15.6% vs. 1.55%). Both proportions increased slowly up to 1960 (18.5% vs. 1.69%), and then increased sharply until 1970 (35.2% vs. 5.45%). (27) Rough estimates of the elevation in neonatal mortality risk associated with illegitimacy imply that the 1940 to 1960 trends had a small positive effect on the neonatal infant mortality gap (widening it by about 0.44), but that the post-1960 influence was considerably stronger (widening the gap by about 2.1). (28) Given that upward force, the strong racial convergence that did occur during the late 1960s is all the more remarkable.

Table 9 (columns 1 and 2) shows that high-order infants (fifth and above) had elevated risks of death. Therefore, different racial distributions over the birth-order range might have influenced the racial gap in infant mortality rates. Columns 5 and 6, however, demonstrate that in 1960 differences in mortality rates within birth-order categories were far more important than differences in the distribution of births across categories in explaining the gap in neonatal mortality rates. Moreover, in columns 9 and 10, it appears that changes in the distribution of birth order during the 1950s had a small impact on neonatal mortality rates. (29)

Health-Related Behavior: Smoking and Breast Feeding 

A separate set of hypotheses suggests that differential racial trends in health-related behavior could have influenced the infant mortality figures. For example, nationally, per adult consumption of factory-produced cigarettes increased dramatically between 1920 (665 per year) and 1945 (3449 per year) and peaked during the 1960s (at about 4300 per year) (Giovino et al. 1994). Using 1988 data, Meara (2001) documents a strong correlation between smoking during pregnancy and the probability of having a low birth weight infant. If nonwhites' cigarette consumption increased at a faster rate than whites' consumption, particularly among young women, then the nonwhite birth weight distribution may have tended to shift leftward relative to the white distribution.

The earliest statistical characterization of smoking by race (to our knowledge) suggests that by 1955 the proportion of young nonwhite women who smoked was as high as that of whites (35.6% vs. 34.2% among 25-34 year old women), although white women were more likely to smoke heavily (Haenszel, Shimkin, and Miller 1956). The cross-sectional evidence also suggests that the farm population was less likely to smoke than the urban population, suggesting that the rapid urbanization of blacks from 1920 to 1970 might have been accompanied by a relatively rapid increase in cigarette consumption. However, it appears that from 1950 to 1965 the smoking proportions of white and black young women were nearly equal (Bums et al. 1997), and at the same time, the low birth weight proportion of nonwhites continued to increase relative to whites (Chase and Bymes 1972). Data from the post-1970 period suggest that black women smoke less than whites during pregnancy (Cutler and Meara 2003), and while we do not have direct evidence on how much pregnant women (or their spouses) smoked before 1970, the evidence that does exist does not offer much support for the hypothesis that differential trends in cigarette smoking drove the divergence in birth weight distributions during the 1950s and early 1960s.

This period was also marked by a dramatic decline in the proportion of mothers who breast fed their infants, and the magnitude of the change in behavior appears to have been larger for nonwhites than for whites. According to a large survey of ever-married women aged 15 to 44 in 1973 (Cycle I of the National Survey of Family Growth) regarding the feeding of their first-born infant, approximately 73% of (first born) nonwhite infants born before 1950 were breast fed (for any duration), whereas only 14% of those born between 1966 and 1970 were (Hirschman and Hendershot 1979). For whites, the drop over the same period was from 56% to 29%. The decline is evident (though not equal in magnitude) among all occupational categories (including never worked), all family income groups, and all education groups, except for college educated women. Early studies of the Children's Bureau found a strong empirical relationship between feeding practices and infant mortality (Rochester 1923; Woodbury 1925). More recently, Forste, Weiss, and Lippincott (2001) have also documented a strong correlation between breast feeding and postneonatal infant mortality. If this correlation is truly causal, then the precipitous decline in breast feeding among nonwhites would have propped up the nonwhite infant mortality rate through 1970. A convincing effort to establish such causality is beyond the scope of this paper, but more detailed research into variation in breast-feeding practices and infant health trends from 1940 to the present could illuminate a potentially important aspect of health history.

Environmental Quality

Chay and Greenstone (1999, 2001) and Wolpaw Reyes (2001) have argued that environmental conditions, in particular air pollution, are causally linked to infant mortality. If so, then the massive redistribution of blacks out of the rural South might have had adverse health consequences that tended to offset health improvements associated with proximity to hospital care and/or higher incomes. The urban variable used in the econometric exploration should have captured some of this effect, but it is a coarse characterization of environmental quality.

Are the links between African Americans' geographic redistribution, pollution exposure, and infant mortality strong enough to matter? The earliest available Environmental Protection Agency (EPA) measures of ambient air quality are for 1960, before the Clean Air Act of 1970 induced large declines in air pollution (U.S. EPA 1973). Measures of total suspended particulates (TSP) are reported separately for urban (central city) and nonurban areas, and the gap was quite large, approximately 110 versus 25 micrograms per cubic meter (U.S. EPA 1973). Among blacks, the proportion of births to metropolitan area residents increased from about 32% to about 75% between 1940 and 1970. (30) Chay and Greenstone (2001) estimate that circa 1970 an increase of 1 microgram of TSP per cubic meter increased infant mortality by 0.05 to 0.08 per 1000 live births. (31) Thus, choosing the midpoint of Chay and Greenstone's range, the redistribution of nonwhites to places with worse air quality may have increased the nonwhite infant mortality rate by about 2.4 per thousand live births, a nontrivial upward force.

The Supply-Side of Health Care: Access, Insurance, and Technology

An alternative set of hypotheses concentrates on the rapidly changing health care system including the extension of the hospital system through the Hill-Burton program, the ascendance of private health insurance, and the implications for access to high-quality medical care. Did white infants benefit disproportionately from Hill-Burton funding of the hospital system's postwar expansion? Unlike the services available to southern blacks prior to World War II, Beardsley argues that "In new federally sponsored hospitals black patients, if still segregated, at least had benefit of modern facilities and enjoyed roughly equal treatment" (Beardsley 1987, p. 256). The available evidence supporting this claim is mixed. In the South, where Hill-Burton made its largest impact, the proportion of nonwhite hospital births increased from 24% to 74% between 1945 and 1960, whereas the white proportion increased from 68% to 97%. Of course, these figures say nothing about the quality of hospital care, nor about the availability of prenatal and postpartum care, but they do suggest a rapid increase in hospital and physician services for nonwhite infants. On the other hand, improvements in postneonatal death rates from influenza, pneumonia, other respiratory diseases, and gastrointestinal illnesses, all of which were treatable in hospitals, were larger for whites than for nonwhites, suggesting that whites may have benefited more from improved hospital services than nonwhites.

A second major change in the health care system in this period was the rise of private health insurance, often provided as a benefit through one's employer (Thomasson 2002). Prior to 1940, relatively few people were covered by health insurance (about 9% in 1940). By 1962, however, about 74% of whites had hospital insurance, whereas only 46% of nonwhites did (Hoffmann 1964). Did the racial insurance gap have implications for maternal and infant care, and ultimately for infant mortality? (32) In the absence of data on treatment intensity by race and insurance status, it is not possible to answer this question directly. However, within weight classes (that is, roughly conditional on health at birth), the decline in neonatal mortality in the 1950s was larger for nonwhite infants than for white infants. (33) On the basis of such evidence, it would be difficult to argue that white newborn infants benefited more from improvements in medical care than nonwhite newborns, even with the insurance gap. However, as noted above, white postneonatal infant mortality rates improved more rapidly than those of nonwhites during the 1950s (Grove and Hetzel 1968). (34)

Convergence: 1965 to 1970

After nearly 20 years of stagnation, the nonwhite-white infant mortality gap plunged from 1966 to 1971. This decline has three important features: The overall decline was somewhat stronger for postneonates (from 9.0 to 4.8) than for neonates (from 9.2 to 6.6); the decline in nonwhite mortality outside the South (by 8 per thousand births) was nearly as large as in the South (by 10 per thousand births); and the decline in nonwhite postneonatal mortality was the dominant feature of the southern data, whereas the decline in nonwhite neonatal mortality was the dominant feature of the nonsouthern data.

Importantly, between 1965 and 1970, the relative movements in the proportion of low birth weight infants were small. There were slight declines for both whites and nonwhites. Therefore, the overall decline and observed convergence of infant mortality rates in the late 1960s appears to have had little to do with changes in infants' health at birth. Rather, it appears that, in addition to the renewed convergence in income and the continued convergence in women's educational attainment, institutional and technological changes may have contributed to the relative improvement of nonwhite infant mortality rates in the late 1960s.

First, Almond, Chay, and Greenstone (2001) argue that the desegregation of southern hospitals played an important role in the racial mortality gap's decline in the South, particularly in rural areas. Postneonatal infant mortality due to diseases such as influenza, pneumonia, and gastroenteritis fell for nonwhites in Mississippi following the passage of the Civil Rights Act in 1964 (and the implementation of Medicare in 1965, which enabled the government to enforce desegregation by withholding Medicare funds from hospitals that discriminated against nonwhites). Second, under the Medicaid program implemented in 1966, the federal government began to give money to states to assist them in funding medical care for low-income families, particularly low-income pregnant women and children. While some states did not choose to implement the program until after 1970, Medicaid may have been responsible for some convergence in the racial gap in infant mortality rates given the relative concentration of nonwhites in the low income group. Third, according to Cone (1985), relatively rapid technological progress in the treatment of low birth weight infants began in the mid-1960s with the institution of modern NICUs. Given the relatively high proportion of low birth weight nonwhite infants, such advances could have differentially benefited nonwhite infants relative to white infants. Thus a number of forces--economic, institutional, political, and technological--appear to have contributed to the rapid narrowing of the racial infant mortality gap in the late 1960s.

6. Conclusion

Health and economic status are interconnected not only contemporaneously, but also intertemporally through families, communities, and institutions. History, therefore, contains valuable information for understanding intergroup health disparities. In this case, we focus on the racial gap in infant mortality rates between 1920 and 1970, a period that witnessed the rapid expansion of the health care system, the Great Migration of African Americans, the emergence of the Civil Rights Movement, and a significant and persistent gap in the mortality rates of nonwhite and white infants.

Clearly, the rapid decline of infant mortality rates during this period benefited both whites and nonwhites. At each point in time, however, nonwhites were disadvantaged in terms of income, education, and location relative to whites. Using a panel of state-level race-specific data, we found that a large portion of the racial gap in log infant mortality rates can be accounted for by differences in those characteristics, especially between 1920 and 1945. However, after 1945, group differences in observable characteristics lost much of their explanatory power, suggesting that either unobserved factors worked to offset the tendency toward convergence in infant mortality rates, or the relationship between the observed characteristics and infant mortality changed significantly after World War II. Comparing Figures 3a and 3b, we concluded that the former possibility is more likely to have been the case.

We then went on to investigate several hypotheses regarding additional factors that could have influenced the post-1940 gap. Proximately, the neonatal gap may have been strongly influenced by a leftward shift in the nonwhite birth weight distribution during the 1950s. By itself, however, this shift cannot explain the persistence in the racial mortality gap. The postneonatal gap, which is rather insensitive to weight at birth, also failed to narrow. Moreover, birth weight itself is not an ultimate cause of infant mortality; rather, it is an intermediating variable between the ultimate causes and the mortality outcome.

In theory, aside from the basic socioeconomic characteristics, the ultimate causes could have included independent changes in maternal characteristics, maternal and familial behaviors, environmental quality, and access to new medical technologies. Drawing on a wide range of evidence on these trends and on the strength of their connections with infant mortality, the paper provides empirical perspective on their potential importance in sustaining the racial infant mortality gap despite the convergence in socioeconomic characteristics.

Changes in observable maternal characteristics appear to have had relatively small impacts on the racial gap in infant mortality rates, at least up to 1960, from which time the proportion of nonwhite children born to unwed mothers increased rapidly. Changes in maternal behaviors are more difficult to document, but we found little evidence of racial divergence in the prevalence of cigarette smoking by young women after 1950. It does appear, however, that the decline in breast feeding among nonwhites was larger than among whites, and this may have tended to widen the mortality gap. Also, the geographic redistribution of blacks to large urban centers may have increased blacks' relative exposure to air pollution, though it also may have increased blacks' relative incomes and access to hospital care. Finally, on the basis of changes in neonatal mortality rates within birth weight categories after 1950, it would be difficult to argue that nonwhite neonates benefited less than white neonates did from the hospital system's expansion and new medical technologies. (35)

Additional research along these lines, perhaps with a local or regional focus, could further enhance our understanding of the course of racial health disparities in the postwar period. For example, the local impacts of Hill-Burton expansions and, later, Medicaid funding await more detailed study. Similarly, given the increasing concentration of black families in central cities during this period, efforts to link long-run urban economic growth (and in many cases, decline) with the health outcomes of residents could prove illuminating.

Appendix: Data Construction

The infant mortality figures are taken from standard sources of vital statistics for the United States, as described in the text. The appendix details the construction of independent variables used in the econometric analysis. These data are state-race-time specific measures of several of the prime determinants of infant mortality. The income measures are the most complex in practice, though the idea is very simple: We attempt to split standard measures of state income per capita into a portion received by whites and portion received by blacks on the basis of racial differences in occupational status in each state.

Income

Because the census first inquired about income in 1940, compiling any such measure by race and state for the full 1920 to 1970 period is difficult. We have constructed measures of per capita income by race, state, and year ([Y.sub.W.it] and [Y.sub.NW,it]) that ultimately depend on estimates of a few variables in the following two equations. The first is simply a weighted average of white and nonwhite income per capita.

(1) [Y.sub.it] = [[theta].sub.it][Y.sub.W,it] + (1 - [[theta].sub.it])[Y.sub.NW,it],

(2) [Z.sub.it] = [Y.sub.W,it]/[Y.sub.NW,it].

[Y.sub.it] is nominal income per capita in state i at time t, as reported by the Bureau of Labor Statistics (after 1929) and by Mitchener and McLean (1999) in 1920. [theta] is the proportion of the population that is white. In Equation 2, Z is the ratio of white income per capita ([Y.sub.w,it]) to nonwhite income per capita ([Y.sub.NW,it]).

We estimate Z for each state and census yeas using the detailed occupational information for individuals reported in the Integrated Public Use Microdata Series (IPUMS, Ruggles and Sobok 1997). In this regard, our methodology is similar in spirit to that of Smith (1984). We assign every person in the microdata samples an income proxy based on his or her observed characteristics. Specifically, this income proxy is the median value of total annual income earned by workers of that sex, of that race, in that region of the country, in that three-digit occupation category in 1960. In other words, the 1960 workers were sorted into a number of sex-byrace-by-region-by-occupation cells, and the median income in each cell was recorded. Then, in other years, observations were sorted into the same matrix of ceils and assigned the 1960 median income value. (36) Since we are interested in a per capita ratio (rather than a per worker ratio), those who were not working were assigned zero income. After averaging the income proxies with each race/state/ year category, one can form an estimate of [Z.sub.it], the ratio of nonwhite/white income per capita.

Given measures of [Y.sub.it] and [Z.sub.it], and using the number of IPUMS observations for each race category to measure [theta], we have enough information to calculate [Y.sub.W,it] and [Y.sub.NW,it]. Essentially, we are simply using Z and [theta] to figure out how to split the state's income between whites and nonwhites. Although estimates of [Y.sub.it] are available in every year, Z and [theta] can only be estimated directly for census years, and we interpolate between those years to get middecade figures. These parameters (Z and [theta]) appear to change relatively slowly over time, so their interpolation seems reasonable. This procedure yields estimates of nominal income per capita by race, state, and year. To get real income estimates, we adjust for price level changes over time using national consumer price indices (U.S. Department of Commerce 1975).

Education

Again, estimating race- and state-specific education levels is a challenge because the census did not inquire about years of education prior to 1940. Instead, the pre-1940 census reports literacy rates, but there is no simple way to move from literacy rates to average years of education. We estimated years of education for the years prior to 1940 by using the individual-level information contained in the 1940 IPUMS sample. For example, to estimate years of education for women who were 20 to 40 years old in 1930 for each state, we calculated the average level of education for women who were 30 to 50 in 1940 for each state. This approach has defects: First, it does not account for migration (or death); second, it does not account for upward biases in self-reported education as age rises. But it does provide a much richer measure of educational attainment than the literacy measure, and on the basis of previous studies of education and infant health, it can certainly be argued that the focus on young women is warranted.

In some state-year cells, the number of nonwhite women in the relevant age category was small. In cases where education information was available for less than 50 women, we did not form estimates of average education and therefore did not use those cells in the econometric estimates. (37) Later, we checked the results using all observations that had education estimates based on 10 women or more; they were quite similar to those with the higher cutoff.

Urban Residence and Physicians per Capita

We use census measures of the urban proportion of the population by race and state, with an adjustment to the data starting in 1960 to reflect a nonnegligible change in the census definition of urban. The change was introduced in 1950, and in that year figures for both new and old definitions of urban are reported. The 1960 urban figures are scaled according to the ratio between the new and old measures for each state and race in 1950. The figures from 1930 to 1950 come from the 1950 state-level census volumes. The figures for 1920 come from the 1940 state-level census volumes. For the sake of consistency, we made an effort to avoid using the 1930 volumes because in them, Mexicans are counted as nonwhite, but in other years they are counted as white.

We use the census occupation tables to obtain the number of physicians per capita in each state. The physicians-per-capita measure is not calculated on a race-specific basis, but since all the regressions are run separately for whites and nonwhites, we do not assume that the effect of more physicians per capita is the same for both race categories. The census also reports the number of nurses and medical technicians, but we found those occupational categories difficult to compile in a consistent manner over time.

Table 1. Summary Statistics, 1920-1970
                                              White          Nonwhite

Summary statistics of log state-year values entered in regressions)

  In infant mortality                         3.450           3.900
                                             (0.481)         (0.404)
  In income                                   7.557           6.352
                                             (0.434)         (0.482)
  In education                                2.346           2.114
                                             (0.119)         (0.259)
  In urban                                    4.049           3.991
                                             (0.327)         (0.521)
  In physicians                              -6.689          -6.831
                                             (0.257)         (0.314)
  N                                            478             344

Summary statistics of state-year values

  Infant mortality                            35.71           54.03
                                             (19.59)         (25.87)
  Income                                      2085            1124
                                            (806.5)         (655.1)
  Education                                   10.52            8.549
                                              (1.194)         (2.030)
  Urban                                       60.11           60.85
                                             (16.72)         (26.21)
  Physicians                                   0.001286        0.001136
                                              (0.000329)      (0.00037)
  N                                              478             344

Each observation represents a particular race/state/year cell.
Observations are weighted by the number of births. Standard
deviations are in parentheses. The regressions in the next table are
run using log values of variables. Figures in the bottom half of
the table are reported for ease of interpretation. Note that the log
figures in the bottom half of the table will not equal the figures in
the top half of the table (because the log of an average value is not
equal to the average of log values).

Sources: See the text for description of the dataset.

Table 2. Infant Mortality Regressions, 1920-1970

                     Equation 1

               White (1)    Nonwhite (2)

Income          -0.1265       -0.1363
                (0.0875)      (0.1314)
Education       -1.0199       -0.4508
                (0.2050)      (0.2133)
Urban            0.1046        0.2857
                (0.0639)      (0.1368)
Physicians      -0.1332       -0.1842
                (0.0519)      (0.0994)
South              -             -

1925            -0.1280       -0.1612
                (0.0196)      (0.0598)
1930            -0.2305       -0.3095
                (0.0306)      (0.0780)
1935            -0.3431       -0.5032
                (0.0463)      (0.0926)
1940            -0.4542       -0.5761
                (0.0634)      (0.0784)
1945            -0.5637       -0.7590
                (0.0869)      (0.0924)
1950            -0.8134       -0.9857
                (0.0887)      (0.0857)
1955            -0.8917       -0.9690
                (0.0899)      (0.0859)
1960            -0.8903       -0.9232
                (0.0948)      (0.0915)
1965            -0.8934       -0.9243
                (0.1091)      (0.1052)
1970            -1.0312       -1.1347
                (0.1306)      (0.1200
Constant         6.1471        4.2068
                (0.8065)      (0.9744)
State dummies      No            No
[R.sup.2]        0.96          0.89
N                 478           344
Mean dep. var.   3.450         3.900

                 Equation 1 and State
                   Dummy Variables

              White (3)     Nonwhite (4)

Income          -0.0131       -0.1574
                (0.0657)      (0.1068)
Education       -0.6918       -0.3418
                (0.2085)      (0.1312)
Urban           -0.0648        0.5729
                (0.0695)      (0.1297)
Physicians      -0.3601       -0.1636
                (0.0693)      (0.1071)
South              -             -

1925            -0.1468       -0.1634
                (0.0325)      (0.0808)
1930            -0.2725       -0.2886
                (0.0318)      (0.0794)
1935            -0.4023       -0.5034
                (0.0408)      (0.0850)
1940            -0.5529       -0.5962
                (0.0487)      (0.0835)
1945            -0.7153       -0.8028
                (0.0608)      (0.0781)
1950            -0.9691       -1.0612
                (0.0636)      (0.0767)
1955            -1.0620       -1.0704
                (0.0706)      (0.0798)
1960            -1.0702        -1.045
                (0.0752)      (0.0851)
1965            -1.0898       -1.0620
                (0.0858)      (0.0920)
1970            -1.2476       -1.2806
                (0.0966)      (0.1002)
Constant         3.8247        3.1985
                (0.7677)      (0.7669)
State dummies     Yes           Yes
[R.sup.2]        0.98          0.95
N                 478           344
Mean dep. var.   3.450         3.900

                   Equation 1 and
                 Dummy Variables for
                     the South

               White (5)    Nonwhite (6)

Income          -0.1249        0.0680
                (0.0859)      (0.1721)
Education       -0.9659       -0.4577
                (0.2558)      (0.2150)
Urban            0.1106        0.2330
                (0.0653)      (0.1366)
Physicians      -0.1314       -0.2210
                (0.0517)      (0.1007)
South            0.0129        0.1489
                (0.0397)      (0.0608)
1925            -0.1307       -0.1557
                (0.0206)      (0.0594)
1930            -0.2361       -0.2790
                (0.0352)      (0.0795)
1935            -0.3515       -0.4640
                (0.0480)      (0.0936)
1940            -0.4653       -0.5712
                (0.0654)      (0.0755)
1945            -0.5778       -0.8484
                (0.0918)      (0.0985)
1950            -0.8294       -1.0571
                (0.0943)      (0.0840)
1955            -0.9096       -1.0606
                (0.0994)      (0.0879)
1960            -0.9100       -1.0210
                (0.1074)      (0.0935)
1965            -0.9153       -1.0566
                (0.1235)      (0.1169)
1970            -1.0553       -1.3033
                (0.1451)      (0.1407)
Constant         6.0077        2.7701
                (0.8827)      (1.2647)
State dummies      No            No
[R.sup.2]        0.96          0.90
N                 478           344
Mean dep. var.   3.450         3.900

State-race-year cells are weighted by the number of births. Robust
standard errors with adjustment for clustering in states
(except for specifications with state dummies) are in parentheses.

Sources: See the text for description of the dataset.

Table 3. Accounting for the Racial Infant Mortality Gap, 1920-1970

                        1920   1925    1930    1935    1940    1945

Total log IMR gap      0.455   0.462   0.493   0.447   0.521   0.454
Gap explained by
  Income               0.110   0.106   0.134   0.131   0.132   0.116
  Education            0.337   0.293   0.343   0.329   0.307   0.265
  Urban               -0.106  -0.079  -0.082  -0.069  -0.056  -0.037
  Physicians           0.034   0.027   0.034   0.038   0.042   0.037
Total explained gap    0.375   0.347   0.429   0.429   0.425   0.381
Total unexplained gap  0.080   0.115   0.064   0.018   0.096   0.073

                       1950    1955    1960    1965    1970

Total log IMR gap      0.496   0.583   0.626   0.630   0.559
Gap explained by
  Income               0.111   0.102   0.097   0.091   0.082
  Education            0.223   0.166   0.121   0.093   0.069
  Urban               -0.014   0.000   0.011   0.019   0.029
  Physicians           0.029   0.022   0.016   0.012   0.007
Total explained gap    0.349   0.290   0.245   0.215   0.187
Total unexplained gap  0.147   0.293   0.381   0.415   0.372

The total infant mortality rate gap is the difference between the
average (weighted by population) log white and nonwhite infant
mortality rates in each year. Each component of the gap explained
by section is the product of the relevant coefficients from Table 2
and the difference in the variable's average value for whites and
nonwhites (in that year).

Sources: See the text for description of the dataset.

Table 4. Infant Mortality Regressions, 1920-1945,
1950-1970, and 1920-1970

                   White       White       White
                 1920-1970   1920-1945   1950-1970
                    (1)         (2)         (3)

Income            -0.1265     -0.1486     -0.1106
                  (0.0875)    (0.1078)    (0.1033)
Education         -1.0199     -1.0014     -1.0238
                  (0.2050)    (0.2300)    (0.2967)
Urban              0.1046      0.1550      0.0730
                  (0.0639)    (0.0962)    (0.0614)
Physicians        -0.1332     -0.2094     -0.1000
                  (0.0519)    (0.1157)    (0.0453)
Year dummies        Yes         Yes         Yes
State dummies       No          No          No
[R.sup.2]          0.96        0.85        0.80
N                   478         238         240
Mean dep. var.     3.450       3.950       3.102

                 Nonwhite    Nonwhite    Nonwhite
                 1920-1970   1920-1945   1950-1970
                    (4)         (5)         (6)

Income            -0.1363     -0.0666     -0.1849
                  (0.1314)    (0.1811)    (0.1162)
Education         -0.4508     -0.4283     -0.5182
                  (0.2133)    (0.2183)    (0.3081)
Urban              0.2857      0.3444      0.1627
                  (0.1368)    (0.1482)    (0.1423)
Physicians        -0.1842     -0.2732     -0.0494
                  (0.0994)    (0.1776)     -0.0812
Year dummies        Yes         Yes         Yes
State dummies       No          No          No
[R.sup.2]          0.89        0.78        0.66
N                   344         156         188
Mean dep. var.     3.900       4.374       3.674

State-race-year cells are weighted by the number of
births. Robust standard errors with adjustment for
clustering in states are in parentheses.

Sources: See the text for description of the dataset.

Table 5. Neonatal and Postneonatal Mortality, 1940-1970

                      1940    1945    1950    1955
White, South
  Neonatal            30.1    24.7    20.8    18.3
  Postneonatal        21.7    16.1     9.8     7.0
  Sum                 51.8    40.8    30.6    25.3
  Weight               0.31    0.30    0.29    0.28

White, non-South
  Neonatal            25.9    22.6    18.8    17.5
  Postneonatal        14.1    10.8     6.5     5.5
  Sum                 40.0    33.4    25.3    23.0
  Weight               0.69    0.70    0.71    0.72
White, national       43.6    35.6    26.8    23.6

Nonwhite, South
  Neonatal            40.6    30.8    27.3    26.4
  Postneonatal        35.2    25.1    18.3    17.7
  Sum                 75.8    55.9    45.6    44.1
  Weight               0.78    0.74    0.67    0.61

Nonwhite, non-South
  Neonatal            36.4    34.3    28.2    28.3
  Postneonatal        31.1    24.9    14.6    12.7
  Sum                 67.5    59.2    42.8    41.0
  Weight               0.22    0.26    0.33    0.39
Nonwhite, national    73.9    56.8    44.7    42.9

                      1960    1965    1970
White, South
  Neonatal            18.1    16.8    14.5
  Postneonatal         6.4     5.9     4.2
  Sum                 24.5    22.7    18.7
  Weight               0.28    0.28    0.29

White, non-South
  Neonatal            16.9    15.8    13.4
  Postneonatal         5.4     5.3     3.9
  Sum                 22.3    21.1    17.3
  Weight               0.72    0.72    0.71
White, national       22.9    21.5    17.7

Nonwhite, South
  Neonatal            27.0    25.0    22.0
  Postneonatal        20.0    18.1    11.0
  Sum                 47.0    43.1    33.0
  Weight               0.56    0.51    0.49

Nonwhite, non-South
  Neonatal            26.9    26.3    21.7
  Postneonatal        11.9    11.6     8.3
  Sum                 38.8    37.9    30.0
  Weight               0.44    0.49    0.51
Nonwhite, national    43.4    40.6    31.5

The weight is the proportion of the relevant
race-category's births in that region.

Table 6. Nonwhite Birth Weight and
Neonatal Infant Mortality, 1950-1960

                       1950
                     Nonwhite
                     Neonatal             1950
                    Mortality           Nonwhite
                     Rate (1)       Distribution (2)

<1000 g                821               0.006
1000-1500              507               0.009
1501-2000              196               0.020
2001-2500               50               0.068
2501-3000               15               0.214
3001-3500               10               0.354
3501-4000               11               0.228
4001-4500               13               0.069
>4500 g                 20               0.033
Sum                                      1.00

                       1960             Col. 1 X
                     Nonwhite           (Col.3 -
                 Distribution (3)      Col.2) (4)

< 1000 g              0.010               3.28
1000-1500             0.011               1.01
1501-2000             0.025               0.98
2001-2500             0.083               0.75
2501-3000             0.253               0.59
3001-3500             0.371               0.20
3501-4000             0.189              -0.43
4001-4500             0.046              -0.30
>4500 g               0.013              -0.40
Sum                   1.00                5.68

Column 4 reflects the impact of the change in
birth weight distribution on the neonatal infant
mortality rate. Sources: Neonatal IMR is from
Chase (1972). Birth weight distributions are
from Chase and Byrnes (1972).

Table 7. Age of Mother and Incidence of
Low Birth Weight, 1950-1967

                Percent LBW    Age Dist.     Age Dist.
                 1950 (1)      1950 (2)      1967 (3)

Nonwhite

  Under 15         14.7          0.007         0.010
  15-19            12.0          0.206         0.269
  20-24             9.6          0.326         0.324
  25-29             8.4          0.232         0.196
  30-34             8.8          0.132         0.116
  35-39             9.0          0.075         0.063
  40-44             8.9          0.020         0.020
  45 and over       7.4          0.002         0.001
  Sum

White

  Under 15         15.9          0.001         0.001
  15-19             8.0          0.104         0.149
  20-24             6.9          0.318         0.382
  25-29             6.5          0.297         0.257
  30-34             7.0          0.174         0.127
  35-39             7.5          0.084         0.065
  40-44             7.5          0.021         0.019
  45 and over       5.7          0.001         0.001
  Sum

                  Col.1 x                     Col.2 x
                 (Col.3 -     Percent LBW    (Col.5 -
                Col.2) (4)     1967 (5)      Col.1) (6)

Nonwhite

  Under 15         0.044         19.5          0.034
  15-19            0.756         15.7          0.762
  20-24           -0.019         13.2          1.174
  25-29           -0.302         11.8          0.788
  30-34           -0.141         12.6          0.502
  35-39           -0.108         13.3          0.323
  40-44            0.000         12.2          0.066
  45 and over     -0.007         10.8          0.007
  Sum              0.223                       3.656

White

  Under 15         0.000         12.5         -0.003
  15-19            0.360          8.5          0.052
  20-24            0.442          6.7         -0.064
  25-29           -0.260          6.5          0.000
  30-34           -0.329          7.0          0.000
  35-39           -0.143          8.3          0.067
  40-44           -0.015          9.1          0.034
  45 and over      0.000          8.1          0.002
  Sum              0.055                       0.088

Column 4 reflects the impact of changes in the age
distribution on the proportion of low birth weight
infants (under 2500 grams). Column 6 reflects the
impact of changes in the likelihood of low birth
weight within age categories on age categories on
the overall proportion of low birth weight infants.
Summing column 4 and column 6 gives the total change
in the percentage of infants born with low both weight.

Source: Data are from Chase and Byrnes (1972).

Table 8. Age of Mother and Neonatal Mortality, 1950-1960

                Neonatal
                Mortality                            Col.1 x
                  Rate,     Age Dist.   Age Dist.    (Col.3 -
                1960 (1)    1950 (2)    1960 (3)    Col.2) (4)

Nonwhite

  Under 20        31.4        0.219       0.203       -0.502
  20-24           25.3        0.325       0.315       -0.253
  25-29           24.2        0.229       0.229        0.000
  30-34           26.4        0.131       0.151        0.528
  35-39           27.3        0.074       0.080        0.164
  40-44           29.4        0.019       0.021        0.059
  45 and over     28.9        0.002       0.001       -0.029
  Sum                                                 -0.034

White

  Under 20        20.4        0.107       0.128        0.428
  20-24           15.9        0.316       0.339        0.366
  25-29           15.3        0.295       0.262       -0.505
  30-34           17.0        0.173       0.163       -0.170
  35-39           18.4        0.085       0.085        0
  40-44           22.0        0.022       0.022        0
  45 and over     31.8        0.001       0.001        0
  Sum                                                  0.119

The age distribution figures for 1950 differ slightly from those
in the previous table. This table's age figures are based on the
first 3 months of 1950; the previous table's age figures are for
the full year.

Sources: Age distribution figures are from Chase (1972). Mortality
by age category are from Chase (1972).

Table 9. Birth Order and Neonatal Mortality, 1950-1960

              White     Nonwhite     White     Nonwhite
            Neonatal    Neonatal     Birth      Birth
            Mortality   Mortality    Order      Order
              Rate,       Rate,      Dist.,     Dist.,
            1960 (1)    1960 (2)    1960 (3)   1960 (4)

First         15.0        25.7       0.262      0.210
Second        16.1        27.2       0.248      0.183
Third         15.9        25.8       0.196      0.153
Fourth        17.2        25.4       0.127      0.121
Fifth +       21.8        28.0       0.167      0.334
Sum
                                     White
            (Col.4 -    (Col.2 -     Birth
             Col.3)      Col.1)      Order
            x Col. 2    x Col. 3     Dist.,
               (5)         (6)      1950 (7)

First         -1.34       2.80       0.323
Second        -1.77       2.75       0.313
Third         -1.11       1.94       0.173
Fourth        -0.15       1.04       0.083
Fifth +        4.68       1.04       0.108
Sum            0.31       9.57

            Nonwhite
              Birth     (Col.3 -    (Col.4 -
              Order      Col.7)      Col.8)
             Dist.,     x Col.1     x Col.2
            1950 (8)      (9)        (10)

First         0.246      -0.915      -0.925
Second        0.219      -1.047      -0.979
Third         0.162       0.366      -0.232
Fourth        0.108       0.757       0.330
Fifth +       0.264       1.286       1.960
Sum                       0.447       0.154

Columns 5 and 6 decompose the racial gap in neonatal mortality
rates in 1960. Column 5 reflects the importance of differences
in the birth-order distributions conditional on mortality rates
within categories. Column 6 reflects the importance of differences
in mortality rates conditional on birth order. Using alternative
weights for the decomposition (multiplying by column 1 rather than
column 2 in column 5, and by column 4 rather than column 3 in
column 6) would yield a sum of 0.95 in column 5 and 8.85 in
column 6. Columns 9 and 10 reflect the importance of changes in
the white and nonwhite birth-order distributions between 1950 and
1960, given the race-specific mortality rates in each category
for 1960.

Source: Underlying data are from Chase (1972).

We appreciate suggestions from Kenneth Chay, Dora Costa, Janet Currie, Leemore Dafny, Joseph Ferric, Robert Fogel, Claudia Goldin, Susan Hautaniemi, Ellen Meara, William Sundstrom, Werner Troesken, two referees, and participants in the National Bureau of Economic Research Summer Institute (NBER DAE Program 2001), the Social Science History Association meetings (2001), the Allied Social Sciences Association meetings (2002), and seminars at Yale University, Harvard University, Northwestern University, and the University of Chicago. We are also grateful for data supplied by David Mintz at the Environmental Protection Agency, Ian McLean and Kris Mitchener, and Damien de Walque. The authors are faculty research fellows at the NBER and are grateful for support as national fellows. Both authors acknowledge support from the National Science Foundation. Collins also acknowledges research support from the Oak Ridge Associated Universities. The views expressed herein are those of the authors, not necessarily those of the NBER, ORAU, or NSF. Kathleen Albers, Nicole Bologna-Emrick, and Timothy Watts provided excellent research assistance.

(1) Although our primary interest is in black relative to white health, infant mortality data are reported by white and nonwhite, and so our statistical analysis proceeds accordingly. Integrated public use microdata series (IPUMS, Ruggles and Sobek 1997) samples of children aged 0 or 1 (excluding Hawaii and Alaska) indicate that in 1920, 95% of nonwhite infants were black; in 1970, 92% of nonwhite infants were black.

(2) See Grossman (1972), Rosanzweig and Schultz (1983), Berger and Leigh (1989), Kenkel (1991), and Goldman and Lakdawalla (2001) on household production of health.

(3) On average, lower income states had fewer hospital facilities per capita (Lave and Lave 1974). But even if southern income per capita had been as high as elsewhere, political economy considerations suggest that public spending on health initiatives might have remained low in the South in this period (see Alston and Ferrie 1999).

(4) In this context we attach a fairly narrow meaning to "legacy of poverty," that is, that blacks were emancipated with essentially zero net wealth. Even by 1995, the black/white ratio of median household wealth was only 0.12 (Wolff 1998).

(5) In contrast, nearly all white births in Memphis were physician attended, along with 98% of white births in Shelby County (exclusive of Memphis), and 83% of white births in Fayette County.

(6) In the South, some hospitals, typically private, refused black patients (except, perhaps, for emergencies). Most public hospitals treated black patients in inferior segregated wards. Almost all major facilities refused to allow black physicians to practice there (until the 1960s). There were a number of relatively small black hospitals (administered by black doctors) operating in the South, but, on average, they were poor institutions. Beardsley (1987) reports that only 9 of about 100 black hospitals in the South received passing marks from the American College of Surgeons in 1928.

(7) Washington, DC, is not included in the averages. 1945 is the first year for which we have such data by race and state, but Tandy reports that in 1935, only 17% of all black births occurred in hospitals compared to 40% of all white births (Tandy 1937).

(8) Physician data are from the occupational tables of the 1940 U.S. Census of Population. Plumbing data are from the 1940 U.S. Census of Housing, Volume II, part 1.

(9) For example, drawing from interviews in a southern state, Sinai and Anderson (1948) report "Not a few physicians practicing in the area have asked what the prenatal visits are for and what they are supposed to do at such visits. They have practiced for years without giving any thought to prenatal care" (p. 165). The value of prenatal exams in detecting eclampsia, venereal disease, and malformed pelvises was apparent to specialists by 1910 (Wertz and Wertz 1989). See Speert (1980) on the history and value of prenatal care.

(10) Ideally, we would prefer to use data at a lower level of aggregation. Unfortunately, family- and infant-level data sets do not exist during the period under study, and county-level vital statistics data are not reported until the 1940s (and even then, often do not report vital statistics by race for counties outside the South).

(11) Variables discussed in section 2 that we do not include explicitly in the model are controls for the availability of family planning services and maternal age, health, and behavior (e.g., smoking and breastfeeding). Time-series data at the state level simply do not exist for many of these variables, such as maternal health and behavior and family planning services. We did collect state-level data on the age distribution of mothers. After excluding some suspicious outliers from the 1920s, the impact of the age distribution on infant mortality in the regressions was negligible and statistically insignificant; the other explanatory variables' coefficients were unaffected. We gather more evidence on these factors for the postwar period, as discussed in section 5.

(12) For example, low income might lead to poor nutrition, low birth weight infants, and higher risk of infant mortality. We can observe income and mortality, but until the 1950s we do not observe birth weight. Thus, our initial strategy is to focus on the ultimate causes of infant mortality rather than the proximate causes, which we take up later. Because unobserved behaviors may be correlated with observed socioeconomic characteristics, we do not claim to measure narrowly defined treatment effects in the regressions.

(13) The left tail of the income distribution may be especially relevant to the study of infant mortality. For 1950, 1960, and 1970, we used the IPUMS samples to calculate the proportion of men (ages 20 to 60, not in school) who were in the lower one-fifth of U.S. annual income distribution by race and state. We interpolated for 1955 and 1965 and entered the measures in the 1950-1970 base regressions for infant mortality. The coefficients are negative for whites (-0.12, t-stat = 0.6) and positive for nonwhites (0.35, t-stat = 1.5).

(14) Two common econometric issues also deserve mention. First, there may be measurement error in the independent variables. Even if measurement error in a particular variable is random, it implies some degree of attenuation bias (toward zero) to that variable's coefficient and an unknown direction of bias to the other coefficients (Greene 1993). Second, because health and economic conditions are interrelated in complex ways. one could argue that the regressions' explanatory variables are, to some extent, endogenous. In theory, an instrumental variable approach could help circumvent these concerns, but plausible instruments are scarce in this paper's context.

(15) [P.sub.E] tests of the linear versus log form favor the log specification (Greene 1993). Additionally, link tests, essentially regressions of the dependent variable on fitted values and fitted values squared, do not suggest misspecification. The relative stability of the coefficient estimates when the sample is split at 1950, particularly for whites, also suggests that the specification is reasonable.

(16) The nonwhite results were sensitive to outlying observations from Arizona, New Mexico, South Dakota, and Montana, all of which had few black residents and low black/nonwhite ratios. The nonwhite results reported in Table 2 omit these states, representing approximately 1% of total nonwhite births in the sample. Results are unaffected by dropping other states with few blacks and low black/nonwhite ratios (e.g., Idaho, North Dakota, Utah). The white results were not sensitive to the inclusion of these states, and they are included in Table 2's regressions.

(17) The coefficient on the percentage of births recorded in 1940 is positive for nonwhites (0.004, t-stat = 0.72) and negative for whites (-0.006, t-stat = 1.33). The other coefficients are relatively stable.

(18) Using a cross-section of county groups in 1969, Hecht and Cutright (1979) account for about two-thirds of the racial infant mortality gap on the basis of differences in health and socioeconomic characteristics. However. birth weight is one of the control variables, and subtracting its contribution drops the proportion of the gap explained to less than 40%, close to what we report for 1970.

(19) To facilitate testing the significance of the coefficient changes, we ran the regressions with data for the entire period and interactions for the variables in the post-1945 period.

(20) The robustness of the results is also reassuring given the fact that there are many missing data points in the 1920s and early 1930s (because several states did not enter the birth-registration area until after 1920). The robustness of the results suggests that the missing observations do not unduly influence the results.

(21) One might imagine that as absolute levels of infant mortality fell that random variation became more important relative to systematic variation associated with X variables. But if so, we would expect a decline in the accounted for portion throughout the 1920 to 1970 period, and especially in the 1920 to 1950 period when the absolute declines were largest. In fact, we do not see that. Moreover, the absolute racial gap did not fall at all between 1950 and 1965, and so there is no reason to think that there is just less to explain over that period, throughout which the explained portion of the gap fell.

(22) The development of penicillin and other antibiotics in the 1940s had a strong impact on the level of infant mortality, especially for postneonatal infants (CDC 1999). However, the timing and magnitude of the racial gap's decline implies that penicillin cannot be the whole story behind the 1941 to 1946 convergence, or perhaps even a large part of it. In particular, the racial gap fell mostly before the mid-1940s, that is, before antibiotics became widely available for civilian use.

(23) For a recent discussion of the medical literature on genetic versus behavioral, environmental, and socioeconomic origins of the racial disparity in birth weights, see David and Collins (1997). They find that women born in Africa (but residing in the United States) have birth outcomes that are more similar to those of white women than to those of African-American women, suggesting that the gap is not driven largely by underlying genetic differences.

(24) In the mid-Atlantic states (Pennsylvania, New York, and New Jersey), for example, in 1950 about 95% of nonwhite births occurred in hospitals, and 1965 about 98% did. The proportion of low birth weight nonwhite infants still increased from 13.3% to 15.6%.

(25) The nonwhite fertility rate was 131.6 in 1948 and 162.2 in 1959. The white rate was 104.3 in 1948 and 114.6 in 1959.

(26) A third avenue could be through more compressed timing of births. But from the late 1940s (1945-1949) to the late 1950s (1955-1959), whites' birth intervals shortened by more than nonwhites' based on information from the 1970 census on spacing between first and second births (U.S. Department of Commerce 1975). In levels, blacks' spacing remained shorter than whites' (23 vs. 28 months for 1955-1959).

(27) These figures are taken at face value from the Vital Statistics volumes, without adjustments for states entering and leaving the sample. A similar trend emerges when we use fixed samples of states.

(28) Shapiro, Schlesinger, and Nesbitt (1968) cite figures from a study on New York City in the early 1960s. Calculations in the text set the neonatal mortality rate for illegitimate nonwhite infants at 41 and for other infants at 25.

(29) We explored the potential impact of differences in high-order births and short spacing by identifying all children of ages 0 and 1 in the 1950, 1960, and 1970 IPUMS samples, counting the number of children under the age of 10 living in the households of 0 and 1 year olds, and then averaging those counts for race/state/year cells for the sake of incorporation in regression equations and decompositions similar to those in Table 4 and Figure 3b (interpolating for 1955 and 1965). In the state-level regressions, higher counts of young children were associated with higher infant mortality rates (t-stat = 1.79 for whites and 1.51 for nonwhites), and nonwhite families tended to have more young children. Therefore, in a decomposition of the racial gap for 1950 to 1970, disparities in this measure helped explain part of the gap in infant mortality rates. However, the explained portion of the total gap still fell greatly after 1950, from over 80% of the total to just over 40%.

(30) In 1970, Vital Statistics reports births by metropolitan status for blacks (as opposed to nonwhites). In 1940, the metropolitan area definitions were not in place, but they have been retroactively imposed on the 1940 IPUMS data (Ruggles and Sobek 1997). We estimate the proportion of black metropolitan area births in 1940 as the proportion of very young (under 3 years of age) black children resident in metro areas.

(31) This approach builds in a bias that might work against finding a large pollution effect. In particular, Chay and Greenstone's studies focus on the early 1970s and early 1980s, by which time unhealthy infants were benefiting from significant technological advances in neonatal intensive care.

(32) In a recent study of changes in Medicaid eligibility, Currie and Gruber (1997, p. 32) found evidence that "suggests that insured and uninsured populations have differential access to NICUs [neonatal intensive care units] and related interventions, and that this difference has real implications for health outcomes." Whether insurance had an effect on infant health prior to the establishment of modern NICUs (in the 1960s) is an open question.

(33) Between 1950 and 1960, the neonatal mortality rate for nonwhites less than 2500 grams at birth fell from 164.7 per thousand to 154.8; for whites it increased from 175.8 per thousand to 177.4. For nonwhites over 2500 grams, it fell from 11.9 per thousand to 7.7; for whites it fell from 7.1 per thousand to 5.1 (Chase 1972).

(34) The cause-of-death data must be interpreted with care because the residual category for nonwhites fell over time, implying that increases in certain disease categories may be due to more complete classification.

(35) This does not imply that nonwhite infants received the same level or quality of medical attention as whites; rather, the interpretation pertains to changes in the levels of attention.

(36) For this exercise, we cannot use the 1940 census because only wage and salary income is reported in that year. We cannot use the 1950 census because only sample line observations have total income reported, and there are not enough observations for meaningful sex/race/region/occupation estimates.

(37) We allowed some exceptions in 1950, when the education variable is available only for sample line individuals, to avoid losing several states from the sample.

References

Almond, Douglas V., Kenneth Y. Chay, and Michael Greenstone. 200l, Civil rights, the war on poverty, and black-white convergence in infant mortality rates in Mississippi. Working paper, August 2001, University of California, Berkeley.

Alston, Lee J., and Joseph P. Ferrie. 1999. Southern paternalism and the American welfare state. New York: Cambridge University Press.

Atack, Jeremy, and Peter Passell. 1994. A new economic view of American history: From colonial times to 1940, 2nd edition. New York: Norton.

Beardsley, Edward H. 1987. A history of neglect: Health care for blacks and mill workers in the twentieth-century South. Knoxville, TN: University of Tennessee Press.

Berger, Mark C., and J. Paul Leigh. 1989. Schooling, self-selection, and health. Journal of Human Resources 24:433-55.

Bradbury, Dorothy E. 1956. Four decades of action for children: A short history of the Children's Bureau. Washington, DC: Children's Bureau Publication No. 358.

Bradbury, Dorothy E. 1974. Five decades of action for children: A history of the Children's Bureau. In Children and youth: Social problems and social policy, edited by Robert H. Bremner. New York: Arno Press, pp. 1-126.

Burns, David M., Lora Lee, Larry Z. Shen, Elizabeth Gilpin, H. Dennis Tolley, Jerry Vaughn, and Thomas G. Shanks. 1997. Cigarette smoking behavior in the United States. In Monograph 8: Changes in cigarette-related disease risks and their implications for prevention and control, edited by David M. Burns, Lawrence Garfinkel, and Jonathan M. Samet. Bethesda, MD: National Cancer Institute, chapter 2.

Byrd, W. Michael, and Linda A. Clayton. 2002. An American health dilemma, volume II: Race, medicine, and health care in the United States 1900-2000. New York: Routledge.

Center for Disease Control. 1999. Achievements in public health, 1900-1999: Healthier mothers and babies. Morbidity and Mortality Weekly Report 48:849-58.

Chase, Helen C. 1972. A study of infant mortality from linked records: Comparison of neonatal mortality from two cohort studies, United States, January-March 1950 and 1960. DHEW Publication No. (HSM) 72-1056. Department of Health, Education, and Welfare. Public Health Service. Series 20, No. 13. June 1972.

Chase, Helen C., and Mary E. Byrnes. 1972. Trends in 'Prematurity', United States: 1950-67. DHEW Publication No. (HSM) 72-1030. Department of Health, Education, and Welfare. Public Health Service. Series 3, No. 15. January 1972.

Chay, Kenneth Y., and Michael Greenstone. 1999. The impact of air pollution on infant mortality: Evidence from geographic variation in pollution shocks induced by a recession. NBER Working Paper No. 7442.

Chay, Kenneth Y., and Michael Greenstone. 2001. Air quality, infant mortality, and the Clean Air Act of 1970. Working paper, July 2001, University of California, Berkeley.

Cone, Thomas E. 1985. History of the care and feeding of the premature infant. Boston: Little, Brown and Co.

Corwin, E. H. L. 1952. Infant and maternal care in New York City. Committee on Public Health Relations, New York Academy of Medicine. New York: Columbia University Press.

Costa, Dora L. 1998. Unequal at birth: A long-term comparison of income and birth weight. Journal of Economic History 38:987-1009.

Costa, Dora L., and Richard H. Steckel. 1997. Long-term trends in health, welfare, and economic growth in the United States. In Health and welfare during industrialization, edited by R. Steckel and R. Floud. Chicago: University of Chicago Press, pp. 47-89.

Currie, Janet, and Jonathan Gruber. 1997. The technology of birth: Health insurance, medical interventions, and infant health. Cambridge, MA: NBER Working Paper No. 5985.

Currie, Janet, and Rosemary Hyson. 1999. Is the impact of health shocks cushioned by socioeconomic status? The case of low birthweight. American Economic Review 89:245-50.

Cutler, David M., and Ellen Meara. 2003. The rise in the ratio of black to white infant mortality since 1960. Presentation at the Allied Social Science Association Meetings, Washington, DC.

David, Richard J., and James W. Collins, Jr. 1997. Differing birth weight among infants of U.S.-born blacks, African-born blacks, and U.S.-born whites. New England Journal of Medicine 337:1209-14.

Demeny, Paul, and Paul Gingrich. 1967. A reconsideration of Negro-White mortality differentials in the United States. Demography 4:820-37.

Department of Health and Human Services (DHHS). 2000. Racial and ethnic disparities in infant mortality. Available: http:// www.raceandhealth.hhs.gov/3rdpgblue/infant/red.htm.

Donohue, John H., and James Heckman. 1991. Continuous versus episodic change: The impact of Civil Rights policy on the economic status of blacks. Journal of Economic Literature 29:1603-43.

Edwards, Linda N., and Michael Grossman. 1979. The relationship between children's health and intellectual development. In Health: What is it worth? edited by Selma Mushkin and David Dunlop. Elmsford, NY: Pergamon Press, pp. 273-314.

Ewbank, Douglas C. 1987. History of black mortality and health before 1940. Milbank Quarterly 65:100-28.

Ewbank, Douglas C., and Samuel H. Preston. 1990. Personal health behavior and the decline in infant and child mortality: The United States, 1900-1930. In What we know about health transition: The cultural, social, and behavioral determinants of health, edited by John Caldwell, et al. Canberra: Health Transition Centre, Australian National University, pp. 116-49.

Flegg, A. T. 1982. Inequality of income, illiteracy, and medical care as determinants of infant mortality in underdeveloped countries. Population Studies 36:441-58.

Forste, Renata, Jessica Weiss, and Emily Lippincott. 2001. The decision to breastfeed in the United States: Does race matter? Pediatrics 108:291-6.

Giovino, Gary A., M. W. Schooley, S. P. Zhu, J. H. Chrismon, S. L. Tomar, J. P. Peddicord, R. K. Merritt, C. G. Husten, and M. P. Ericksen. 1994. Surveillance for selected tobacco use behaviors: United States, 1900-1994. Morbidity and Mortality Weekly Report 43:1-43.

Goldman, Dana, and Darius Lakdawalla. 2001. Understanding health disparities across education groups. NBER Working Paper No. 8328.

Greene, William H. 1993. Econometric analysis. New York: Macmillan Publishing Company.

Grossman, Michael. 1972. On the concept of health capital and the demand for health. Journal of Political Economy 80:223-55.

Grossman, Michael. 1975. The correlation between health and schooling. In Household production and consumption, edited by Nestor E. Terleckyj. New York: Columbia University Press, pp. 147-211.

Grossman, Michael, and Steven Jacobowitz. 1981. Variations in infant mortality rates among counties of the United States: The roles of public policies and programs. Demography 18:695-713.

Grove, Robert D., and Alice M. Hetzel. 1968. Vital statistics rates in the United States, 1940-1960. Washington, DC: U.S. Department of Health, Education, and Welfare; Public Health Service.

Gruber, Jonathan. 2000. Medicaid. NBER Working Paper No. 7829.

Haenszel, William, Michael B. Shimkin, and Herman P. Miller. 1956. Tobacco smoking patterns in the United States. Washington, DC: Public Health Service, Munograph 45.

Haines, Michael R. 2001. The urban mortality transition in the United States, 1800-1940. NBER Working Paper No. H0134

Hecht, Pamela K., and Phillips Cutright. 1979. Racial differences in infant mortality rates: United States, 1969. Social Forces 57:1180-93.

Hirschman, Charles, and Gerry E. Hendershot. 1979. Trends in breast feeding among American mothers. Washington, DC: Public Health Service, series 23, number 3.

Hobcraft, J. N., J. W. McDonald, and S. O. Rutstein. 1984. Socio-economic factors in infant mortality: A cross-national comparison. Population Studies 38:193-223.

Hobcraft, J. N., J. W. McDonald, and S. O. Rutstein. 1985. Demographic determinants of infant and early childhood mortality: A comparative analysis. Population Studies 39:363-85.

Hoffmann, Carolane H. 1964. Health insurance coverage. Washington, DC: National Center for Health Statistics, series 10, number 11.

Jiobu, Robert M. 1972. Urban determinants of racial differentiation in infant mortality. Demography 9:603-15.

Johnson, Joseph L. 1949. Supply of Negro health personnel Physicians. Journal of Negro Education 18:346-56.

Kempe, Allison, Paul H. Wise, and Susan E. Barkan, et al. 1992. Clinical determinants of the racial disparity in very low birth weight. New England Journal of Medicine 327:969-73.

Kenkel, Donald S. 1991. Health behavior, health knowledge, and schooling. Journal of Political Economy 99:287-305.

Lave, Judith R., and Lester B. Lave. 1974. The hospital construction act: An evaluation of the Hill-Burton program, 1948-1973. Washington, DC: American Enterprise Institute for Public Policy Research.

Lemons, J. Stanley. 1969. The Sheppard-Towner Act: Progressivism in the 1920s. Journal of American History 55:776-86.

Levine, Robert S., James E. Foster, Robert E. Fullilove, Mindy T. Fullilove, Nathaniel C. Briggs, Pamela C. Hull, A. Husiani, and Charles H. Hennekens. 2001. Black-White inequalities in mortality and life expectancy, 1933-1999: Implications for healthy people 2010. Public Health Reports 116:474-83.

Linder, Forrest E., and Robert D. Grove. 1943. Vital statistics rates in the United States, 1900-1940. Washington, DC: Government Printing Office (GPO).

Loudon, Irvine. 1991. On maternal and infant mortality 1900-1960. Social History of Medicine 4:29-73.

Maloney, Thomas N., and Warren C. Whatley. 1995. Making the effort: The contours of racial discrimination in Detroit's labor markets, 1920-1940. Journal of Economic History, 55:465-93.

Margo, Robert A. 1986. Race, educational attainment, and the 1940 census. Journal of Economic History 46:189-98.

Margo, Robert A. 1990. Race and schooling in the South, 1880-1950. Chicago: University of Chicago Press.

Massey, Douglas S., and Nancy A. Denton. 1993. American apartheid: Segregation and the making of the underclass. Cambridge, MA: Harvard University Press.

Meara, Ellen. 2001. Why is health related to socioeconomic status? The case of pregnancy and low birth weight. NBER Working Paper No. 8231.

Meckel, Richard A. 1998. Save the babies: American public health reform and the prevention of infant mortality, 1850-1929. Ann Arbor, MI: University of Michigan Press.

Miller, Doug. 2001. What underlies the Black-White infant mortality gap? The importance of birthweight, behavior, environment, and health care. Working paper, University of California at Berkeley.

Mitchener, Kris James, and Ian W. McLean. 1999. U.S. regional growth and convergence, 1880-1980. Journal of Economic History 59:1016-42.

National Center for Health Statistics. 2001. Health, United States, 2001, with urban and rural health chartbook. Hyattsville, MD: Centers for Disease Control.

Oaxaca, Ronald. 1973. Male-Female wage differentials in urban labor markets. International Economic Review 14:693-709.

Pampel, Jr., Fred C., and Vijayan K. Pillai. 1986. Patterns and determinants of infant mortality in developed nations, 1950-1975. Demography 23:525-42.

Preston, Samuel H. 1996. Population studies of mortality. Population Studies 50:525-36.

Pritchett, Lent, and Lawrence H. Summers. 1997. Wealthier is healthier. Journal of Human Resources 31:841-68.

Rochester, Anna. 1923. Infant mortality: Results of afield study in Baltimore, MD, based on births in one year. U.S. Department of Labor, Children's Bureau. Bureau Publication 119. Washington, DC: GPO.

Rosenzweig, Mark R., and T. Paul Schultz. 1983. Estimating a household production function: Heterogeneity, the demand for health inputs, and their effects on birth weight. Journal of Political Economy 91:723-46.

Ruggles, Steven, and Matthew Sobek. 1997. Integrated public use microdata series. Minneapolis, MN: Historical Census Projects, University of Minnesota.

Seham, Max. 1964. Discrimination against Negroes in hospitals. New England Journal of Medicine 271:940-3.

Shapiro, Sam, Edward R. Schlesinger, and Robert E. L. Nesbitt, Jr. 1968. Infant, perinatal, maternal, and childhood mortality in the United States. Cambridge, MA: Harvard University Press.

Shin, Eui Hang. 1975. Black-White differentials in infant mortality in the South, 1940-1970. Demography 12:1-19.

Sinai, Nathan, and Odin W. Anderson. 1948. EMIC: A study of administrative experience. Ann Arbor, MI: University of Michigan School of Public Health.

Smith, David Barton. 1999. Health care divided: Race and healing a nation. Ann Arbor, MI: University of Michigan Press.

Smith, James P. 1984. Race and human capital. American Economic Review 74:685-98.

Speert, Harold. 1980. Obstetrics and gynecology in America: A history. Chicago: American College of Obstetricians and Gynecologists.

State of Tennessee, Department of Public Health. 1941. Annual bulletin of vital statistics. Nashville, TN: State of Tennessee.

Stockwell, Edward G. 1962. Infant mortality and socio-economic status: A changing relationship. Milbank Memorial Fund Quarterly 40:101-111.

Subbarao, K., and Laura Raney. 1995. Social gains from female education: A cross-national study. Economic Development and Cultural Change 44:105-28.

Tandy, Elizabeth C. 1937. Infant and maternal mortality among Negroes. Journal of Negro Education July:322-49.

Thomasson, Melissa A. 2002. From sickness to health: The twentieth century development of U.S. health insurance. Explorations in Economic History 39:233-53.

United States Bureau of the Census. Various years. Vital statistics of the United States. Washington, DC: GPO.

United States Department of Commerce, Bureau of the Census. 1975. Historical statistics of the United States, colonial times to 1970. Washington, DC: GPO.

United States Environmental Protection Agency (U.S. EPA). 1973. The national air monitoring program: Air quality and emissions trends: Annual report, volume I. Springfield, VA: U.S. Department of Commerce, National Technical Information Service.

Wertz, Richard W., and Dorothy C. Wertz. 1989. Lying in: A history of childbirth in America. New Haven: Yale University Press.

Wolff, Edward J. 1998. Recent trends in the size distribution of household wealth. Journal of Economic Perspectives 12: 131-150.

Wolpaw Reyes, Jessica. 2001. The impact of prenatal lead exposure on infant health. Working paper, 2001, Harvard University.

Woodbury, Robert M. 1925. Causal factors in infant mortality. Children's Bureau Publication No. 142. Washington, DC: GPO.

William J. Collins * and Melissa A. Thomasson ([dagger])

* Department of Economics, Box 351819-B, Vanderbilt University, Nashville, TN 37235, USA; E-mail: william. [e-mail].

([dagger]) Department of Economics, Miami University, Oxford, OH 45056, USA; E-mail: [e-mail]; corresponding author.

Received January 2003; accepted July 2003.