Racial residential segregation (hereafter “segregation”) in the United States is a complex, multidimensional phenomenon1 best described as “an institutional manifestation of racism.”2 Segregation is commonly quantified using indices capturing a series of unique but related dimensions (evenness, exposure, clustering, concentration, and centralization1). Most racial/ethnic groups in the US are to some extent segregated, but black Americans have experienced uniquely sustained and extremely high levels of segregation over the past century2,3 due to a complex series of social, political, and historical forces. Research in economics suggests that segregation is a major contributor to black/white differences in socioeconomic position2 and a cause of adverse educational, occupational, and financial outcomes in young blacks.2,4,5 Given this impact on socioeconomic position and the relationship between socioeconomic position and health, epidemiologists and sociologists have examined segregation as a potential cause of racial differences in a variety of health outcomes. Findings have linked segregation to limited access to healthy foods,6–8 poorer-quality medical care,9 and deficiencies in the built and natural environments.7,10
Segregation has also been linked to racial differences in birth weight. The substantive importance of birth weight as an outcome is not without controversy,11 but a striking black/white birth weight difference among US-born infants has persisted over time: as recently as 2009, the rate of low birth weight (defined as <2,500 g) in US blacks was nearly four times that of whites.12 Existing research points to an association between segregation (specifically the dimensions of evenness [measured by the dissimilarity index] and exposure [measured by the isolation index]) and lower birth weight, or a higher probability of low birth weight in black infants.13–20 This association is evident after controlling for individual-level covariates.4,13 An important distinction is the level at which exposure is assessed: segregation is by definition an area-level phenomenon, marking the separation between groups over a broad geographical region (typically the metropolitan statistical area, or MSA). Studies at the city/neighborhood level exist,17,18 but may not model true patterns of segregation due to the exclusion of key surrounding areas economically integrated with the urban core.21 However, even among studies at the MSA level, it is difficult to provide an average estimate of the association between segregation and birth weight due to differences in the choice of segregation index, model covariates, model form, and conceptualization of the birth weight variable (binary vs. continuous).
A major barrier in identifying the causal effect of segregation on birth weight is unobserved confounding, particularly by neighborhood preference and self-sorting, as these factors may also be associated with birth outcomes. Furthermore, if segregation does affect birth weight, it likely does so through a number of mediating channels,22 especially given its causal link to adverse economic outcomes.5,23 Given these concerns, it is possible that existing estimates of the effect of segregation on birth weight are biased. Instrumental variable (IV) analysis24 can account for unobserved confounding by isolating the variance in the exposure attributable to its association with an exogenous variable (the instrument), essentially simulating exposure randomization under a specific set of assumptions. A key challenge of conducting an IV analysis is identifying a suitable instrument, which must (1) be associated with the exposure, (2) only affect the outcome through the exposure, and (3) not share any common causes with the outcome. The railroad division index, which measures the extent to which a metropolitan area was divided into subplots by railroad tracks (and thus the ease with which blacks were segregated into racially homogeneous enclaves), has been used as an instrument for dissimilarity in an economic context.23 Using the railroad division index as an instrument for segregation—specifically, the 1990 dissimilarity index—suggested a causal relationship between segregation and race-specific poverty rates in non-Southern locations.23 The present analysis aims to address recent calls for improved analytical methods in segregation research6 and clarify existing estimates of the effect of segregation on birth weight using the railroad division index as an instrument for segregation.
Institutional review was not required as this study relied on publicly available data. We obtained information on birth weight and maternal risk factors for the year 2000 from the National Vital Statistics System natality files, restricting to singleton births to non-Hispanic black and non-Hispanic white, US-born/resident women residing in non-Southern MSAs (we elaborate on this below). Viability bounds further restricted the sample to infants weighing 500 to 6,000 g, with gestational ages from 20 to 43 weeks. We assessed two outcome variables: individual-level birth weight (modeled continuously in grams), and the black/white difference in average birth weight at the MSA level. We modeled birth weight as a continuous variable to avoid imposing an arbitrary low birth weight cutoff on the data,13 and because our main interest was in the persistent difference in the black and white birth weight distributions. For individual-level analyses, we also (1) examined the effect of segregation at the upper and lower race-specific birth weight quantiles to assess potential effect heterogeneity across distributions, and (2) modeled birth weight restricting to term births (reported gestational age ≥ 37 weeks). Gestational age is a key determinant of birth weight but subject to considerable misclassification when based on last reported menstrual period,25 as is the case in the National Vital Statistics System data; as such, we present these findings only to demonstrate the robustness of our original estimates. We categorized infants as black or white based upon recorded maternal race; this method of classification is imperfect, but roughly 84% of infants in our sample for whom both maternal and paternal race data was available were born to same race parents.
The exposure of interest was MSA-level segregation, quantified via the 2000 dissimilarity index, which indicates the proportion of the population who would need to move to a different census tract in order for racial groups to be evenly distributed across an MSA.1 Although the dissimilarity index is one of many segregation indices, it was used in the original analysis of the railroad division index as an instrument for segregation23 and was selected here in the interest of consistency. We obtained dissimilarity calculations from the US Census Housing & Household Economic Statistics Division (non-Hispanic black file).26 We also retained two indices measuring other dimensions of segregation (isolation and spatial proximity) for comparative purposes. MSA-level demographic data were obtained from the Lewis Mumford Center at SUNY Albany.27 We used the railroad division index, conditional on total railroad track length, as an instrument for segregation in the IV analyses; both variables were obtained from Ananat’s publicly available dataset.23 Both the dissimilarity index and the railroad division index were multiplied by 100 to facilitate interpretation.
The railroad division index relies on pre-great migration railroad maps to capture track placement within a 4-kilometer radius of 19th century (non-Southern) urban centers, which would have had negligible black populations at the time of track placement. The original analysis of the railroad division index as an instrument for segregation23 used historical accounts and various falsification checks (including an assessment of the index’s relationship to population characteristics both before and after the great migration) to demonstrate that track placement in non-Southern areas was driven by geographical features—specifically ground slope—and was unrelated to a conscious desire to segregate an area’s inhabitants by race or other social factors (supporting the instrument’s theoretical exogeneity). Southern locations were excluded because track placement in the South, where the majority of the black US population resided before the great migration, may have been driven by a desire to segregate communities. As railroad division index calculations were dependent on map availability,23 some non-Southern MSAs were necessarily excluded from our analysis as they did not have a corresponding index (refer to eAppendix 1, http://links.lww.com/EDE/B57). Data sources were merged on Federal Information Processing Standard codes, yielding an analytical sample of 583,791 births occurring in 93 non-Southern MSAs in the year 2000.
We maintain, as have a number of researchers,10,13,14,28 that maternal- and MSA-level covariates likely mediate the relationship between segregation and birth weight. Controlling for effects of segregation would attenuate effect estimates, leading to erroneously downplayed notions of segregation’s overall impact. Moreover, introducing statistical control for many of these common covariates—for example, maternal education, age, MSA proportion black—is conceptually problematic, as it implies we can hold these factors constant without first intervening on segregation; the economic impact of segregation is so pervasive that this is likely unrealistic. Because these covariates could all be conceptualized as mediators, we did not control for them, opting instead to estimate the total causal effect of segregation on birth weight. We did collect covariate information for demographic purposes and instrument validation; the distribution of these variables is summarized in Table 1.
All analyses were conducted using Stata 13 (StataCorp LP, College Station, TX). We used ordinary least squares to obtain naïve estimates of the effect of segregation on both individual-level birth weight and the MSA-level birth weight gap between black and white infants. Two-stage least squares IV models were constructed using the railroad division index, conditional on total track length, as an instrument for dissimilarity. The first-stage regression, isolating the exogenous variation in segregation, can be expressed as
where Segj refers to the dissimilarity index in the jth MSA, RDI is the railroad division index, and length is the total historical track length. The second stage, which estimates the causal effect of segregation on birth weight (illustrated here at the individual level), is expressed as
where BWij is the birth weight for infant i in MSA j and
is the predicted level of segregation for infant i in MSA j. We ran these models in a single step using Stata’s ivregress command to ensure accurate estimation of standard errors. We conducted standard diagnostic analyses to assess the strength and performance of the railroad division index as an instrument.29 Ordinary least squares and two-stage least squares results were compared for both individual-level birth weight (n = 583,791 births) and the MSA-level birth weight gap (n = 93 MSAs), which was calculated as the population-weighted MSA-level black/white difference in average birth weight. Because the effect of segregation on infant and economic outcomes varies by race,3,15,23 and due to substantial demographic differences between black and white mothers in our sample, individual-level analyses were stratified by race. Cluster robust standard errors were used in all models to account for within-MSA correlation.
Sample characteristics are summarized in Table 1. Although this sample included only a portion of all US MSAs, levels of segregation were comparable with national segregation estimates for the year 2000.30 The sample dissimilarity index was 0.61 (national = 0.64), suggesting that an average of 61% of residents would need to move to different census tracts to achieve MSA-level racial balance in every census tract. Individual-level covariates differed substantially by race: compared with white women, black women in this sample were younger (mean age: 25 vs. 29), less likely to have completed high school (74% vs. 90%), less likely to be married at the time of delivery (26% vs. 77%), and more likely to have received inadequate prenatal care (11% vs. 3%). The prevalence of low birth weight, defined as less than 2,500 g but no less than 500, was considerably higher in black infants (11%) than white infants (5%), but gestational age was similar on average (38 vs. 39 weeks). Kernel smoothed race-stratified birth weight distributions are presented in Figure 1.
Instrument Strength and Performance
We assessed the first-stage regression (the effect of the instrument on exposure) and identified a positive linear relationship (Fig. 2): a one percentage-point increase in the railroad division index corresponded to a 0.39 percentage-point increase in dissimilarity. The first-stage (pooled) F statistic was 8.4 (Table 2), slightly below the established benchmark of 10 and potentially indicative of a weak instrument. We found similar directional trends but weaker first-stage relationships between two other segregation indices—isolation and spatial proximity—and the railroad division index (F statistics of 5.4 and 2.4, respectively). Given existing evidence of its strength as an instrument for dissimilarity (our analysis of the original railroad division index data23 yielded an F statistic of 13.5), as well as the proximity of the overall F statistic to the ideal cutoff, we considered the first-stage results reasonable, but not optimal. As levels of dissimilarity decreased slightly from 1990 to 2000, a corresponding decrease in the F statistic was anticipated.31 We also assessed the reduced form effect, or the total effect of the railroad division index on birth weight; this should be statistically different from zero if the instrument affects the exposure and the exposure affects the outcome. Table 2 indicates that this was essentially the case for blacks, but findings for whites and the combined sample were less definitive.
While it is not possible to empirically test whether a potential instrument only affects the outcome through the exposure, several analyses can inform whether such an assumption is reasonable. We assessed the change in birth weight-specific covariates across quantiles of the railroad division index after stratifying by race. If exposure assignment was as good as random, measured covariates should be similar across quantiles of the railroad division index; this would also suggest that the distribution of unmeasured confounders is potentially evenly balanced across levels of exposure. eTable 1 (http://links.lww.com/EDE/B57) illustrates the stability of most race-stratified covariates across railroad division index quantiles with the exception of MSA population size and the MSA percent of black residents, where the absolute change from quantile 1 to 4 was substantial but nonlinear, suggesting that these fluctuations may have arisen by chance.
Ordinary least squares and two-stage least squares results are compared in Table 3. All two-stage least squares analyses used the railroad division index as an instrument for the dissimilarity index. Differences between model estimates were notable for black infants: ordinary least squares estimated a 1.2 g decrease (95% confidence interval [CI]: −1.9, −0.50) in individual birth weight for every one-percentage point increase in segregation, whereas two-stage least squares estimated a 2.8 g decrease (95% CI: −6.0, 0.48). For white infants, ordinary least squares estimated a 0.53 g increase (95% CI: −0.23, 1.3) in birth weight with each one percentage-point increase in segregation; the two-stage least squares estimate was in the opposite direction (−0.68; 95% CI: −3.5, 2.1). Restricting to term births did not have a substantial impact on individual-level two-stage least squares estimates, although black and white estimates were slightly attenuated (shifting to −2.2 and −0.41 g, respectively). Table 3 also suggests that the effect of segregation was consistent across the black, but not the white, birth weight distribution.
At the MSA level, ordinary least squares estimated a 1.6 g widening of the black/white gap per each percentage-point increase in segregation (95% CI: 0.75, 2.4), indicative of an increased black/white birth weight difference with increasing levels of segregation. The two-stage least squares analysis estimated a slightly larger increase of 2.1 g (95% CI: −1.8, 6.0). Restricting to term births yielded a similar estimate (2.1, 95% CI: −1.8, 6.1). Two-stage least squares estimates of the MSA-level birth weight gap and black individual-level birth weight were generally robust to the inclusion of individual covariates (eTable 2; http://links.lww.com/EDE/B57), suggesting that the railroad division index likely operated only through segregation. In contrast, white individual-level estimates were generally not robust to the addition of other parameters. These point estimates changed considerably in models controlling for MSA population and percent black, respectively, but the corresponding CIs were wide.
Table 4 presents sensitivity analyses restricting the sample to MSAs with at least 5,000 black residents to account for the dissimilarity index’s sensitivity to population distribution. Nineteen MSAs in the original sample were omitted, resulting in a modified sample of 550,354 births (18% black) in 74 MSAs (refer to eTable 3 for demographic information; http://links.lww.com/EDE/B57). There were minor shifts in the ordinary least squares estimates but changes in two-stage least squares results, particularly for whites. This approach reduced the estimated effect of a one percentage-point increase in segregation on the birth weight gap from 2.3 to 1.8 g, and the estimate for whites shifted from −0.68 in the original analysis to −1.7 in the restricted analysis. The robustness of individual-level estimates for black infants lends support to the model specification for this group.
We found weak evidence of a causal effect of segregation on the MSA-level absolute black/white gap in birth weight in this sample, with two-stage least squares estimating a 2.1 g increase (95% CI: −1.8, 6.0) in the black/white birth weight gap for every percentage-point increase in segregation. Evidence at the individual level, however, was stronger: individual-level two-stage least squares predicted a 2.8 g decrease (95% CI: −6.0, 0.48) in black birth weight for every one percentage-point increase in segregation. This effect was consistent across the black birth weight distribution and robust to restricting to term births. Given the width and position of the CI, it is reasonable to conclude that the effect of segregation is likely negative for black birth weight, and ordinary least squares may underestimate this effect. Put into context, our estimates indicate that a one standard deviation (12 percentage point) increase in dissimilarity decreases black birth weight by roughly 33 g. Although this represents a small fraction of the variation in birth weight distribution, it is comparable with the estimated effect of environmental tobacco smoke.32
The direction of the first-stage IV estimates (0.32 to 0.39) was consistent with our expectations. Reduced form estimates of the relationship between the instrument and our outcome variable suggested that increased values of the railroad division index were associated with decreased birth weight at the individual level and increased black/white birth weight differences at the MSA level. Although imprecise, the upper bound of the black interval was close to zero, signifying that the association between the railroad division index and birth weight for blacks in this sample was likely negative. The overall individual-level F statistic was 8.4, and F statistics following stratification were stronger for whites (9.5) than for blacks (3.7). The railroad division index may have been a weaker predictor of dissimilarity in blacks because blacks consistently resided in more segregated MSAs overall (mean dissimilarity score: 68, SD: 12), whereas whites lived in MSAs with a wider range of segregation (mean dissimilarity score: 60, SD: 15). The available amount of exogenous variation for the railroad division index to isolate may have therefore been greater for whites. The correlation between dissimilarity and the railroad division index also fell at high values of dissimilarity; this is potentially problematic, as a substantial proportion of blacks in this sample resided in highly segregated MSAs.
This study had a number of limitations, most notably the exclusion of the South. Although essential in promoting the exogeneity of the railroad division index, excluding Southern MSAs affects the generalizability of our findings and prohibits us from describing the effects of segregation in the Southeastern US, which contains the country’s highest density of black residents. However, residential segregation in many Midwestern and Northeastern MSAs is currently (and has historically been) substantial33; although our sample did not capture all of these areas (index calculation was limited by map availability23 and was therefore not available for 77 non-Southern locations, some of which are highly populous), we are confident that our findings apply to a relevant and historically segregated portion of the US. The similarity between our sample dissimilarity index and the 2000 national index lends additional support to this assertion.
The cross-sectional design prevented us from assessing the length of exposure to segregated environments, but given the relative stability of segregation over time and the propensity of individuals to move within their existing MSA, this design may mimic a cumulative exposure model.22 Longitudinal designs should nevertheless be prioritized in future studies, particularly as it is unclear if segregation leads to poor outcomes, or if poor outcomes result in sustained patterns of segregation. Existing research on the causal effects of segregation suggests the former is the case,4,5 but it is reasonable to speculate that both dynamics play a role in the maintenance of segregation and black/white differences in health outcomes.
Instrument validity checks, specifically the race-stratified first-stage F statistics and the reduced form estimates, suggest that the railroad division index may have been a weak instrument.29,34 However, in the just-identified case (single instrument, single endogenous exposure), IV results are approximately unbiased29 unless the first stage is zero, in which case bias would push the IV estimate toward ordinary least squares. Because our approach is the just-identified case, and because our first stage was nonzero, the expectation is that our estimates are unbiased, even if relatively imprecise.
The sensitivity analysis restricting to MSAs with at least 5,000 black residents raises concerns that the original two-stage least squares estimates may have been biased by the dissimilarity index’s sensitivity to population distribution. There was a sizable increase in the magnitude of the white point estimate and a modest decrease in the MSA-level gap estimate, although both were fairly imprecise. Both the ordinary least squares and two-stage least squares point estimates for blacks, however, were robust. In our sample, restricting to MSAs with at least 5,000 black residents meant restricting to MSAs with larger populations and higher dissimilarity indices on average (eTable 3; http://links.lww.com/EDE/B57). There was also less overall variation in dissimilarity in MSAs with larger black populations. Our initial estimates therefore reflected a combination of MSAs in which segregation is an important feature of the social landscape, and areas in which segregation, particularly for whites, may have been less relevant. By eliminating subjects living in overwhelmingly white MSAs, the sensitivity analysis likely reflects a more accurate estimate of the effect of segregation on birth weight in areas where segregation is a social reality.
We selected birth weight as an outcome due to its association with both infant mortality and long-term health problems. This association may not be causal,11 but our priority was to isolate the causal effect of segregation on birth weight; examining the causal effect of birth weight on subsequent outcomes remains an active area of research and was beyond the scope of this study. Our decision not to control for covariates may have confounded ordinary least squares (but not IV) estimates if the excluded factors were in fact confounders instead of mediators. Future research should clarify the causal structure and carefully assess the role of relevant covariates in a race-specific context.
Finally, the multidimensional nature of segregation1 was inadequately addressed in this analysis. Our study lends some support to other research findings suggesting that dissimilarity may be harmful, but high levels of dissimilarity may be “balanced out” by high levels of clustering or other dimensions, if they are in fact protective.4 Two studies35,36 have reported a decreased risk of adverse birth outcomes when black women lived in high relative density black neighborhoods and had “positive income incongruity”—or resided in an area with a higher median household income than would be predicted, based on the woman’s level of education. The same research found that black women in higher income neighborhoods had better birth outcomes only when the neighborhood did not consist of a black minority,35 implying that neighborhood socioeconomic position may be protective for blacks only when the potential for increased racism and discrimination is not a factor.
Segregation is a long-standing issue with potentially deleterious consequences, particularly for black Americans: although recent census data suggests that US segregation is declining,37 it would take 20 years at the current rate of decline for black/white segregation to equal the present rate of Hispanic/white segregation,33 which is still substantial.38 Our results support previous findings of segregation’s detrimental effect on black birth weight13–20 and suggest that ordinary least squares analyses may underestimate this effect. Future studies should aim to conduct IV analyses with stronger instruments and a nationally representative sample to refine our understanding of the extent to which various dimensions of segregation influence birth weight. Finally, adopting a life-course approach22 would be valuable, as the effect of segregation is likely cumulative and perhaps intergenerational.
The authors thank Dr. Theresa Osypuk for her feedback on an earlier version of this study.
1. Massey DS, Denton NA. The dimensions of residential segregation. Soc Forces. 1988;67:281–315.
2. Williams DR, Collins C. Racial residential segregation: a fundamental cause of racial disparities in health. Public Health Rep. 2001;116:404–416.
3. Massey DS. 1993.Harvard University Press; Boston, MA: American Apartheid: Segregation and the Making of the Underclass
4. Ellen IG, Cutler DM, Dickens W. Is segregation bad for your health? the case of low birth weight [with comments]. Brookings-Wharton papers on urban affairs. 2000:203–38.
5. Cutler DM, Glaeser EL. Are ghettos good or bad? Q J Econ. 1997;112:827–72.
6. White K, Borrell LN. Racial/ethnic residential segregation: framing the context of health risk and health disparities. Health Place. 2011;17:438–448.
7. Williams DR, Mohammed SA, Leavell J, Collins C. Race, socioeconomic status, and health: complexities, ongoing challenges, and research opportunities. Ann N Y Acad Sci. 2010;1186:69–101.
8. Kwate NOA. Fried chicken and fresh apples: racial segregation as a fundamental cause of fast food density in black neighborhoods. Health Place. 2008;14:32–44.
9. Dimick J, Ruhter J, Sarrazin MV, Birkmeyer JD. Black patients more likely than whites to undergo surgery at low-quality hospitals in segregated regions. Health Aff (Millwood). 2013;32:1046–1053.
10. Kramer MR, Hogue CR. Is segregation bad for your health? Epidemiol Rev. 2009;31:178–194.
11. Wilcox AJ. On the importance–and the unimportance–of birthweight. Int J Epidemiol. 2001;30:1233–1241.
12. Mathews TJ, MacDorman MF. Infant mortality statistics from the 2009 period linked birth/infant death data set. Natl Vital Stat Rep. 2013;61:1–27.
13. Bell JF, Zimmerman FJ, Almgren GR, Mayer JD, Huebner CE. Birth outcomes among urban African-American women: a multilevel analysis of the role of racial residential segregation. Soc Sci Med. 2006;63:3030–3045.
14. Walton E. Residential segregation and birth weight among racial and ethnic minorities in the United States. J Health Soc Behav. 2009;50:427–442.
15. McFarland M, Smith CA. Segregation, race, and infant well-being. Popul Res Policy Rev. 2011;30:467–493.
16. Debbink MP, Bader MD. Racial residential segregation and low birth weight in Michigan’s metropolitan areas. Am J Public Health. 2011;101:1714–1720.
17. Anthopolos R, James SA, Gelfand AE, Miranda ML. A spatial measure of neighborhood level racial isolation applied to low birthweight, preterm birth, and birthweight in North Carolina. Spat Spatiotemporal Epidemiol. 2011;2:235–246.
18. Grady SC. Racial residential segregation impacts on low birth weight using improved neighborhood boundary definitions. Spat Spatiotemporal Epidemiol. 2010;1:239–249.
19. Grady SC, Ramírez IJ. Mediating medical risk factors in the residential segregation and low birthweight relationship by race in New York City. Health Place. 2008;14:661–677.
20. Nyarko KA, Wehby GL. Residential segregation and the health of African-American infants: does the effect vary by prevalence? Matern Child Health J. 2012;16:1491–1499.
21. Osypuk TL, Acevedo-Garcia D. Beyond individual neighborhoods: a geography of opportunity perspective for understanding racial/ethnic health disparities. Health Place. 2010;16:1113–1123.
22. Osypuk TL. Invited commentary: integrating a life-course perspective and social theory to advance research on residential segregation and health. Am J Epidemiol. 2013;177:310–315.
23. Ananat EO. The wrong side (s) of the tracks: The causal effects of racial segregation on urban poverty and inequality. Am Econ J Appl Econ. 2011;3:34–66.
24. Baiocchi M, Cheng J, Small DS. Instrumental variable methods for causal inference. Stat Med. 2014;33:2297–2340.
25. Savitz DA, Terry JW Jr, Dole N, Thorp JM Jr, Siega-Riz AM, Herring AH. Comparison of pregnancy dating by last menstrual period, ultrasound scanning, and their combination. Am J Obstet Gynecol. 2002;187:1660–1666.
26. US Census: Housing Patterns [Internet]. 2005. Washington DC: US Census; Updated 27 Jun 2005. Available at: https://www.census.gov/housing/housing_patterns/msa/tract2000nhbk.dat
. Accessed 1 September 2015.
27. Census 2000: State of the Cities file [Internet]. 2003. Albany, NY: Lewis Mumford Center. Updated 27 October 2003. Available at: http://mumford.albany.edu/census/CityProfiles/CPdownload.html
. Accessed 1 September 2015.
28. Acevedo-Garcia D, Osypuk TL. Invited commentary: residential segregation and health–the complexity of modeling separate social contexts. Am J Epidemiol. 2008;168:1255–1258.
29. Angrist JD, Pischke J. Mostly Harmless Econometrics: An Empiricist’s Companion. 2008.New Jersey: Princeton University Press.
30. Iceland J, Weinberg DH. Racial and ethnic residential segregation in the United States 1980–2000. Bureau of Census; 2002.
31. Iceland J, Sharp G, Timberlake JM. Sun belt rising: regional population change and the decline in black residential segregation, 1970-2009. Demography. 2013;50:97–123.
32. Leonardi-Bee J, Smyth A, Britton J, Coleman T. Environmental tobacco smoke and fetal health: systematic review and meta-analysis. Arch Dis Child Fetal Neonatal Ed. 2008;93:F351–F361.
33. Logan JR. The persistence of segregation in the 21(st) century metropolis. City Community. 2013;12: 160–168.
34. Angrist JD, Krueger AB. Instrumental variables and the search for identification: from supply and demand to natural experiments. J Econ Perspect. 2001;15:69–85.
35. Pickett KE, Collins JW Jr, Masi CM, Wilkinson RG. The effects of racial density and income incongruity on pregnancy outcomes. Soc Sci Med. 2005;60:2229–2238.
36. Vinikoor LC, Kaufman JS, MacLehose RF, Laraia BA. Effects of racial density and income incongruity on pregnancy outcomes in less segregated communities. Soc Sci Med. 2008;66:255–259.
37. Glaeser E, Vigdor J. The end of the segregated century: Racial separation in America’s neighborhoods, 1890–2010. 2012.New York, NY: Manhattan Institute for Policy Research.
38. Britton ML, Shin H. Metropolitan residential segregation and very preterm birth among African American and Mexican-origin women. Soc Sci Med. 2013;98:37–45.