- Email: [email protected]
Neighborhood deprivation and preterm birth: an application of propensity score matching
Annals of Epidemiology 25 (2015) 120e125
Contents lists available at ScienceDirect
Annals of Epidemiology journal homepage: www.annalsofepidemiology.org
Original article
Neighborhood deprivation and preterm birth: an application of propensity score matching Xiaoguang Ma MD, PhD a, b, *, Nancy L. Fleischer PhD b, c, Jihong Liu ScD b, James W. Hardin PhD b, Guang Zhao PhD b, Angela D. Liese PhD b, c a b c
Department of Nutrition and Food Hygiene, School of Public Health, Zhejiang University, Hangzhou, China Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia Center for Research in Nutrition and Health Disparities, Arnold School of Public Health, University of South Carolina, Columbia
a r t i c l e i n f o
a b s t r a c t
Article history: Received 11 July 2014 Accepted 24 November 2014 Available online 28 November 2014
Purpose: On the basis of a neighborhood deprivation index (NDI), this study aims to examine the association between neighborhood deprivation and preterm birth (PTB) by applying propensity score matching (PSM) methods. Methods: NDI was calculated for all census tracts in South Carolina based on the US Census data. Live births in South Carolina during 2008 to 2009 (n ¼ 98,456) were assigned to an NDI quartile group based on residential addresses. PSM was used to create matched pairs by NDI quartiles to avoid any potential inference on imbalanced data. The differences of prevalence of PTB were calculated for exposed and reference deprivation groups. Results: Neighborhood deprivation was higher among blacks than whites. The overall prevalence of PTB was 8.5% for whites and 12.6% for blacks. Living in neighborhoods with higher deprivation was associated with increased risk of PTB among blacks compared with living in neighborhoods with lower deprivation among blacks. However, random-effect regression models showed that the most deprived whites experienced 1.13 times the odds of having PTB than the least deprived whites. Conclusions: The racial disparities in adverse birth outcomes might be partially explained by neighborhood deprivation in South Carolina. PSM may be an appropriate approach to avoid imbalanced data inferences. Ó 2015 Elsevier Inc. All rights reserved.
Keywords: Neighborhood deprivation index Preterm birth Propensity score Matching Principal component analysis
Introduction Racial disparities in adverse birth outcomes are well documented but not well explained [1,2]. Racial disparities vary across geographic regions with different political, economic, and social contexts [3,4], which suggests that studies focusing on environmental factors, including of the neighborhood environment, are needed to explain racial disparities in birth outcomes [5]. The neighborhood socioeconomic environment may shape individual biological and behavior risk factors, which may cause adverse birth outcomes through a variety of biological mechanisms [6]. However, relationships between neighborhood socioeconomic factors and adverse birth outcomes are not consistent across
Conflict of interest: None declared. * Corresponding author. Department of Nutrition and Food Hygiene, School of Public Health, Zhejiang University, 866 Yuhangtang Road, Hangzhou, Zhejiang 310058, China. Tel.: þ86 571-8820-8591; fax: þ86 571-8820-8519. E-mail address: [email protected] (X. Ma). http://dx.doi.org/10.1016/j.annepidem.2014.11.021 1047-2797/Ó 2015 Elsevier Inc. All rights reserved.
studies. Some studies have identified a relationship between deprived neighborhood conditions and adverse birth outcomes [6e10], some have not [11], and some have only demonstrated the associations among certain racial groups [12e15]. A possible explanation for the inconsistency is that various indicators have been used to characterize the neighborhood socioeconomic context, making results difficult to interpret and compare across studies. The neighborhood deprivation index (NDI), which synthesizes multiple dimensions of the neighborhood socioeconomic context, allows comparisons across geographic areas [16]. This index has been linked to adverse birth outcomes, including low birth weight and preterm birth (PTB) [17e19]. However, the distribution of NDI quartiles can be extremely imbalanced across different racial groups; often, more white women live in less deprived areas and more minority women live in more deprived areas. With the addition of covariates in the adjusted models, certain covariate strata may contain imbalanced data or subjects who could never be exposed to a condition, leading to inference based on no actual data [20].
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
Propensity score matching (PSM) is a useful approach for dealing with these issues. A propensity score is defined as the conditional probability of being exposed to a condition [21,22]. The propensity score reduces the dimensionality of a large set of potential confounders to unity, making it conducive to simple pair matching [23]. After exposure groups are matched by propensity scores, they have been balanced on all available relevant covariates. In this way, we reduce the observable bias while maintaining the support of the data. This study aimed to examine the association between neighborhood deprivation and PTB based on all live births in 2008 to 2009 in South Carolina, stratified by racial groups (whites and blacks). PSM was used to avoid any sparse data among covariate categories caused by an imbalanced distribution of data across quartiles of neighborhood deprivation. Methods Study population Birth certificates of all live births in South Carolina from January 1, 2008, to December 31, 2009, were obtained from the South Carolina Department of Health and Environmental Control. Within the period, there were 123,759 live births. After excluding births without census tract information, multiple births, and births in Hispanic and other racial/ethnic groups (American Indian and Alaska Native, Asian Native Hawaiian, and other Pacific Islander), and extreme outliers of birth weight (3 SD) and gestational age (<20 weeks), 98,456 white and black births were included in the study. Measures PTB was defined as gestational age less than 37 weeks. The PTB data were from the obstetric estimate of gestation from the birth certificate. This measure was added from 2003 in birth certificate and it was based on birth attendant’s final estimate of gestation. The algorithm published by Messer et al. [16] was used to create the NDI for each census tract in South Carolina using principal component analysis. The eight census tract-level sociodemographic factors used to compute the NDI included percentage population with less than high school, percentage unemployed population, percentage males in management occupations, percentage crowded housing, percentage households in poverty, percentage female head households with children, percentage households earning less than $30,000 per year, and percentage households on public assistance. We used data from the 2000 census. The NDI was predicted based on the loadings of the eight factors in the first principal component. In this study, only the first principal component had an eigenvalue more than 1, accounting for 61.08% of the total variance. The NDI was standardized to have a mean of 0 and SD of 1 by dividing the index by the square of the eigenvalue. Quartiles of NDI were coded as Q1 (least deprived), Q2, Q3, and Q4 (most deprived). Q1 was considered the reference group. Principal component analysis analysis was conducted using the pca program in STATA (Version 10, StataCorp, College Station, TX). In the PSM analysis, to achieve the best model fit to predict propensity scores, we included all appropriate covariates which were predictive of the exposure of interest and occurred before the outcome of interest. We included all the sociodemographic variables available in the data set, including maternal age, maternal education, the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC) participation, and urbanicity. Other covariates included body mass index, maternal smoking, prenatal care, number of previous live births, number of previous PTBs, and
121
maternal risk factors such as infection, chronic and gestational hypertension, and diabetes. Statistical analysis Population characteristics were summarized for the pooled sample and for samples in each NDI quartile. Q2, Q3, and Q4 were compared with Q1 based on t tests for continuous variables and c2 tests for categorical variables. As shown in Figure 1, the distribution of NDI was imbalanced between white and black women, with approximately 50% of black women living in the most deprived areas. To avoid inference due to the imbalanced distribution of NDI, we used PSM to analyze the relationship between NDI and adverse birth outcomes stratified by race. We used logistic regression to estimate the predicted probability of a mother’s exposure to neighborhood deprivation to create matched pairs comparing NDI quartiles. All appropriate covariates discussed previously were included in the models to achieve the best fit. Propensity scores were estimated for each mother and computed separately for whites and blacks. We then matched the mothers living in deprived areas (Q2, Q3, and Q4, separately) with those living in the reference area (Q1) with the same propensity score. The matching procedure was conducted using the psmatch2 module in STATA. The mothers living in deprived areas were matched 1:1 with replacement (to yield the largest number of matched pairs, the matched observations were returned to the data pool and could be matched again if the propensity score could be matched) to mothers living in reference areas with the same predicted probability of exposure to neighborhood deprivation within a range of 0.01. We yielded 100% matching between the deprived group and the reference group because of the large sample size. Balance tests were performed to assure the balance of the data after matching. We compared the means and percentage bias before and after matching, and percentage bias reduction, with a goal of a percentage bias reduction of less than 10% indicating sufficient balance. The percentage bias is the percentage difference of the sample means in the deprived and reference group as a percentage of the square root of the average of the sample variances [24]. The graph of propensity score overlap was generated by level of neighborhood deprivation for each racial group. Generally, the graph showed the distribution of propensity score between exposed (deprived groups such as Q2, Q3, and Q4) and unexposed (Q1) groups. Adequate overlap between the distributions showed that we should be able to find a suitable match for an exposed subject. After the subjects were matched, the differences in prevalence of PTB were computed between the matched deprived and
Fig. 1. Distribution of NDI by race in South Carolina.
122
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
group; below the midline are for the reference group (Q1). Generally, there was adequate overlap between the two exposure groups. Most of the overlap was in the middle of the propensity score distribution for Q2 versus Q1 and Q3 versus Q1, whereas most of the overlap for Q4 versus Q1 was on the left side of the distribution among whites and the right side of the distribution among blacks. After matching, the percentage bias reduction for covariates ranged from 0% to 7.4%, which achieved the 10% goal discussed previously. For most covariates, the percentage bias was reduced after PSM (results not shown). Based on the matched pairs of deprived (Q2 or Q3 or Q4) and reference (Q1) mothers, prevalence differences were calculated for PTB among whites and blacks (Table 2). Among whites, the prevalence difference between the more deprived group and the reference group ranged from 0.38% (Q2 vs. Q1) to 1.42% (Q4 vs. Q1) for PTB. According to the 95% confidence intervals, no statistically significant difference was found. For blacks, compared with mothers living in the least deprived (Q1) areas, those living in the most deprived (Q4) areas experienced a 2.91% higher prevalence of PTB. No difference was found for other NDI quartiles for PTB. The results of multivariate random effects logistic regression models are shown in Table 3. In the models for whites, mothers living in Q4 areas had 1.13 times the odds of giving PTB births when compared with mothers living in Q1 areas. However in the analysis for blacks, no significant differences were found for PTB among different neighborhood deprivation areas.
reference groups. There are a variety of methods for the calculation of the standard error of propensity score matched effect estimates. In this study, we used the bootstrap method with 1000 repetitions to calculate bias-corrected 95% confidence intervals. To compare the PSM results to a typical regression analysis, we conducted a sensitivity analysis using random effects (women clustered in the Census tracts) multivariate logistic regression models to examine the association between NDI and adverse birth outcomes, stratified by race. The maternal age, maternal education, WIC participation, and urbanicity were adjusted as covariates in the models. The random effects regression models were fitted with xtlogit command for multilevel analysis in STATA. P values less than .05 were considered as significance level. Results Table 1 shows the characteristics of the sample. The overall prevalence of PTB was 8.5% for whites and 12.6% for blacks. Average NDI was higher among blacks than whites in South Carolina. Women residing in the Q2, Q3, and Q4 (most deprived) quartile of the NDI were more likely to be younger and have lower levels of education, a higher proportion of WIC participation and rural residence, and have babies with lower birth weight than women living in the least deprived quartile of NDI (Q1), regardless of race. The mean gestational age and the prevalence of PTB were different between deprived and reference areas among whites, but not among blacks. PSM yielded 100% matching between deprived quartiles (Q2eQ4) and the reference quartile (Q1) of the NDI. Figure 2 graphically depicts the propensity score overlap by NDI quartiles among whites (upper panel) and blacks (lower panel). The bars above the midline are propensity scores for the more deprived
Discussion In this study, we used PSM to examine the difference in the prevalence of PTB for mothers living in neighborhoods with different levels of deprivation. We found that neighborhood
Table 1 Characteristics of sample by quartiles of neighborhood deprivation index in South Carolina Variables
Mean (SD) or % Q1
P for trend Q2
Q3
Q4
Total
White
N ¼ 21,895
N ¼ 21,662
N ¼ 12,713
N ¼ 6153
N ¼ 62,423
Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural Birth weight, g Low birth weight Gestational age, w PTB NDI
29.1 (5.7)
26.7 (5.8)*
25.9 (5.7)*
25.5 (5.7)*
27.3 (5.9)
<.01
21.3 29.8 48.9 23.1 26.2 3388.2 (492.6) 4.98 38.6 (1.7) 7.79 1.19 (0.32)
41.4* 34.1 24.5 43.9* 52.7* 3348.0 (505.6)* 6.16* 38.5 (1.9)* 8.51* 0.37 (0.19)*
49.0* 32.7 18.4 52.8* 68.3* 3322.8 (506.0)* 6.43* 38.5 (1.8)* 8.81* 0.23 (0.18)*
55.5* 29.7 14.8 60.9* 62.3* 3290.4 (513.8)* 7.56* 38.4 (1.9)* 9.91* 1.00 (0.42)*
37.3 31.9 30.8 40.2 47.5 3351.3 (503.0) 5.94 38.5 (1.8) 8.46 0.40 (0.75)
<.01
<.01 <.01 <.01 <.01 <.01 <.01 <.01
Non-Hispanic black
N ¼ 5303
N ¼ 7482
N ¼ 9362
N ¼ 13,886
N ¼ 36,033
Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural Birth weight, g Low birth weight Gestational age, w PTB NDI
26.3 (6.2)
25.1 (5.8)*
24.5 (5.7)*
24.0 (5.5)*
24.7 (5.8)
<.01
38.8 37.8 23.3 61.1 17.5 3147.6 (499.5) 11.90 38.1 (2.4) 12.41 1.15 (0.28)*
49.5* 38.1 12.4 72.2* 45.5* 3129.0 (495.5)* 11.84 38.1 (2.4) 12.64 0.33 (0.19)*
58.1* 34.2 7.7 79.6* 61.8* 3099.9 (493.8)* 12.70 38.2 (2.4) 12.12 0.27 (0.17)*
67.4* 27.8 4.7 82.0* 53.7* 3088.1 (490.2)* 12.89 38.1 (2.4) 12.91 1.40 (0.75)*
57.1 33.1 9.8 76.3 48.8 3108.4 (494.5) 12.48 38.1 (2.4) 12.57 0.37 (1.04)
<.01
BMI ¼ body mass index; Q ¼ Neighborhood Deprivation Index quartiles (Q1, less deprived to Q4, more deprived). Q1 was used as the reference group, and all other three groups were compared with the reference. t test and c2 were used to compare for continuous and categorical variables, respectively. * P < .05.
<.01 <.01 <.01 <.05 .927 .390 <.01
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
123
Fig. 2. Propensity score overlap by level of neighborhood deprivation for whites (upper panel) and blacks (lower panel) (Y-axis is the number of births and X-axis is the propensity score. The area above middle axis is for deprived groups including Q2, Q3, and Q4, and the area below middle axis is for reference group Q1.).
deprivation was higher among blacks than whites. Living in neighborhoods with higher deprivation was associated with increased risk of PTB among blacks compared with living in neighborhoods with lower deprivation among blacks. A number of studies have demonstrated a positive association between neighborhood deprivation and adverse birth outcomes; however, results are inconsistent for different racial/ethnic groups [13,16e19]. Pearl et al. [14] found that in addition to individual socioeconomic characteristics, living in neighborhoods that are less socioeconomically advantaged may differentially influence birth weight, depending on women’s ethnicity and nativity. Messer et al. [25] and Kaufman et al. [8] found that living in less deprived or higher income neighborhoods was associated with an increased risk of PTB among African-American but not Caucasian women. Masi et al. [6] concluded that living in an economically disadvantaged neighborhood put African American women, but not Caucasian or Hispanic women, at an increased risk of having a PTB. However, the findings are difficult to compare because of different Table 2 Difference of prevalence of PTB between neighborhood deprivation index quartiles after propensity score matching by race in all births 2008 to 2009 in South Carolina Difference between NDI quartiles
Matched Prevalence Prevalence Prevalence Biasdifference, % corrected pairs in deprived in Q1, % 95% CI* group, %
Non-Hispanic white Q2 versus Q1 21,895 Q3 versus Q1 21,895 Q4 versus Q1 21,895 Non-Hispanic black Q2 versus Q1 7482 Q3 versus Q1 9362 Q4 versus Q1 13,886
8.51 8.81 9.90
8.13 8.57 8.48
0.38 0.24 1.42
0.77 to 1.61 1.90 to 1.30 0.46 to 2.84
12.65 12.11 12.91
11.22 12.13 10.00
1.43 0.02 2.91
1.22 to 2.87 2.91 to 1.64 1.48 to 4.92
CI ¼ confidence interval; Q ¼ Neighborhood Deprivation Index quartiles (Q1, less deprived to Q4, more deprived). Bolded means P < .05. * The bias-corrected 95% CIs were calculated by bootstrap method with 1000 replications.
neighborhood deprivation indicators. An index (NDI) was developed by Messer et al. [16] to synthesize multiple dimensions of the neighborhood context and allow comparisons across geographic areas. By using this index, Elo et al. [17] found that the association Table 3 The association between neighborhood deprivation index quartiles and PTB from random-effect logistic regressions by race in all births 2008 to 2009 in South Carolina Variables
OR (95% CI)
Non-Hispanic white
N¼ 57,608
NDI Q1 Q2 Q3 Q4 Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural
1.00 1.02 1.03 1.13 1.01
(0.94e1.12) (0.94e1.14) (1.01e1.27) (1.01e1.02)
1.00 0.85 0.65 0.95 1.08
(0.79e0.91) (0.60e0.72) (0.89e1.02) (1.01e1.16)
Non-Hispanic black
N¼34,356
NDI Q1 Q2 Q3 Q4 Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural
1.00 1.06 1.01 1.07 1.03
(0.94e1.20) (0.89e1.14) (0.96e1.21) (1.02e1.03)
1.00 0.87 0.61 0.72 1.00
(0.81e0.94) (0.54e0.70) (0.66e0.77) (0.92e1.07)
CI ¼ confidence interval; OR ¼ odds ratio; Q ¼ quartile. Adjusted variables are maternal age, maternal education, WIC participation, and urbanicity. Bolded means P < .05.
124
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
between neighborhood deprivation and small-for-gestational age did not vary significantly by race. In Janevic et al.’s study [18], associations were reported for PTB among Hispanic Caribbean women only. Several studies claimed larger neighborhood effects on PTB among whites than among blacks [16,19]. By using the same NDI measure, our results confirmed that the association between neighborhood deprivation and PTB varied by race, and that the results are different for whites and blacks. The reasons for inconsistent results may be because of the methodological innovation used in this study, PSM methods. Previous studies used logistic regression models to examine the relationship between NDI and birth outcomes. Compared with regression models, PSM weights the data differently and bases its inference on actual data only. To compare our results with previous studies, we analyzed our data using logistic regression models with a random effect. The logistic results showed different results for PTB. The effects of NDI on PTB were much larger among whites than among blacks, which are consistent with previous findings [16,19]. However, PSM methods showed different results compared with traditional logistic regression models. As discussed previously, PSM avoided relying on imbalanced data which were usually observed in research topic of adverse birth outcomes and racial disparities as well as neighborhood deprivation. In such topics, PSM may be an appropriate approach to solve the issue of inference on imbalanced data. Our findings are subject to several limitations. Although PSM was preferred for the data pattern in this study, there are some limitations for this method. PSM did not account for unobserved or unobservable characteristics. Rosenbaum has developed a method of sensitivity analysis to assess if one’s estimate based on matching is robust to the possible presence of an unobserved confounder [26]. Based on this sensitivity analysis, we yielded tight confidence bounds around the log odds of differential assignment due to unobserved factors. We also observed very small Hodgese Lehmann point estimates, which indicated that unmeasured confounding was inconsequential. Moreover, PSM did not incorporate the “clustering” of the neighborhood. However, small within-tract variance was found from multilevel logistic regression models in this study (data not shown). In addition to the limitations of PSM, we only had the census tract IDs in 2000 for the mothers in the database when the data were requested from South Carolina Department of Health and Environmental Control. To merge with birth data by census tract ID, we used 2000 census data to identify NDI. The birth data are closer to 2010; thus, using 2000 census data might cause bias. However, we used the quartiles of NDI, which reflected the relative rather than the actual difference among different tracts. Therefore, the 2000 census boundaries might not have a significant influence on the results of this study. Despite these limitations, this study had several strengths beyond previous studies. In general, 1:1 PSM often yields observations which cannot be matched. However in our study, the large sample size allowed for 100% matching between deprived and reference groups. PSM is not a new approach [21], but it has only started to be used in social epidemiology and reproductive health research in recent years [23,27,28]. However, to our knowledge, no studies have used PSM method to examine the association between NDI and adverse birth outcomes. In addition, this is the first study on neighborhood deprivation and birth outcomes in South Carolina, where racial disparities on adverse birth outcomes are a serious public health concern. Conclusions Neighborhood deprivation was associated with increased risk of PTB among blacks. The racial disparities on birth outcomes might
be partially explained by the neighborhood deprivation according to this study. PSM can be an appropriate approach to avoid inference on unbalanced data. Future research using PSM is encouraged to examine the effect of neighborhood characteristics on birth outcomes. Acknowledgments This work was not supported by any grants or funds. The authors thank James Hibbert (Center for Research in Nutrition and Health Disparities, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina) for geographic information system analysis and Dr. Daniela Nitcheva, Mr. Sung-Jun Kim, and Mr. John Allen (South Carolina Department of Health and Environmental Control, Columbia, South Carolina) for providing the data needed for this study. References [1] Goldenberg RL, Cliver SP, Mulvihill FX, Hickey CA, Hoffman HJ, Klerman LV, et al. Medical, psychosocial, and behavioral risk factors do not explain the increased risk for low birth weight among black women. Am J Obstet Gynecol 1996;175(5):1317e24. [2] Lu MC, Halfon N. Racial and ethnic disparities in birth outcomes: a life-course perspective. Matern Child Health J 2003;7(1):13e30. [3] Teitler JO, Reichman NE, Nepomnyaschy L, Martinson M. A cross-national comparison of racial and ethnic disparities in low birth weight in the United States and England. Pediatrics 2007;120(5):e1182e9. [4] Nepomnyaschy L. Race disparities in low birth weight in the US south and the rest of the nation. Soc Sci Med 2010;70(5):684e91. [5] Metcalfe A, Lail P, Ghali WA, Sauve RS. The association between neighbourhoods and adverse birth outcomes: a systematic review and meta-analysis of multi-level studies. Paediatr Perinat Epidemiol 2011;25(3):236e45. [6] Masi CM, Hawkley LC, Piotrowski ZH, Pickett KE. Neighborhood economic disadvantage, violent crime, group density, and pregnancy outcomes in a diverse, urban population. Soc Sci Med 2007;65(12):2440e57. [7] Agyemang C, Vrijkotte TG, Droomers M, van der Wal MF, Bonsel GJ, Stronks K. The effect of neighbourhood income and deprivation on pregnancy outcomes in Amsterdam, The Netherlands. J Epidemiol Community Health 2009;63(9): 755e60. [8] Kaufman JS, Dole N, Savitz DA, Herring AH. Modeling community-level effects on preterm birth. Ann Epidemiol 2003;13(5):377e84. [9] Messer LC, Kaufman JS, Dole N, Savitz DA, Laraia BA. Neighborhood crime, deprivation, and preterm birth. Ann Epidemiol 2006;16(6):455e62. [10] Schempf A, Strobino D, O’Campo P. Neighborhood effects on birth weight: an exploration of psychosocial and behavioral pathways in Baltimore, 1995e1996. Soc Sci Med 2009;68(1):100e10. [11] Cubbin C, Marchi K, Lin M, Bell T, Marshall H, Miller C, et al. Is neighborhood deprivation independently associated with maternal and infant health? Evidence from Florida and Washington. Matern child Health J 2008;12(1):61e74. [12] Buka SL, Brennan RT, Rich-Edwards JW, Raudenbush SW, Earls F. Neighborhood support and the birth weight of urban infants. Am J Epidemiol 2003;157(1):1e8. [13] Messer LC, Vinikoor LC, Laraia BA, Kaufman JS, Eyster J, Holzman C, et al. Socioeconomic domains and associations with preterm birth. Soc Sci Med 2008;67(8):1247e57. [14] Pearl M, Braveman P, Abrams B. The relationship of neighborhood socioeconomic characteristics to birth weight among 5 ethnic groups in California. Am J Public Health 2001;91(11):1808e14. [15] Pickett KE, Ahern JE, Selvin S, Abrams B. Neighborhood socioeconomic status, maternal race and preterm delivery: a case-control study. Ann Epidemiol 2002;12(6):410e8. [16] Messer LC, Laraia BA, Kaufman JS, Eyster J, Holzman C, Culhane J, et al. The development of a standardized neighborhood deprivation index. J Urban Health 2006;83(6):1041e62. [17] Elo IT, Culhane JF, Kohler IV, O’Campo P, Burke JG, Messer LC, et al. Neighbourhood deprivation and small-for-gestational-age term births in the United States. Paediatr Perinat Epidemiol 2009;23(1):87e96. [18] Janevic T, Stein CR, Savitz DA, Kaufman JS, Mason SM, Herring AH. Neighborhood deprivation and adverse birth outcomes among diverse ethnic groups. Ann Epidemiol 2010;20(6):445e51. [19] O’Campo P, Burke JG, Culhane J, Elo IT, Eyster J, Holzman C, et al. Neighborhood deprivation and preterm birth among non-Hispanic black and white women in eight geographic areas in the United States. Am J Epidemiol 2008;167(2):155e63. [20] Messer LC, Oakes JM, Mason S. Effects of socioeconomic and racial residential segregation on preterm birth: a cautionary tale of structural confounding. Am J Epidemiol 2010;171(6):664e73.
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125 [21] Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika 1983;70(1):41e55. [22] Rosenbaum PR, Rubin DB. Reducing bias in observational studies using subclassification on the propensity score. J Am Stat Assoc 1984;79(387):516e24. [23] Oakes JM, Kaufman JS. Methods in social epidemiology. San Francisco: JosseyBass; 2006. [24] Rosenbaum PR, Rubin DB. Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. Am Stat 1985;39(1):33e8.
125
[25] Messer LC, Kaufman JS, Dole N, Herring A, Laraia BA. Violent crime exposure classification and adverse birth outcomes: a geographically-defined cohort study. Int J Health Geogr 2006;5:22. [26] Rosenbaum PR. Heterogeneity and causality. Am Stat 2005;59(2):147e52. [27] Hearst MO, Oakes JM, Johnson PJ. The effect of racial residential segregation on black infant mortality. Am J Epidemiol 2008;168(11):1247e54. [28] Johnson PJ, Oakes JM, Anderton DL. Neighborhood poverty and American Indian infant death: are the effects identifiable? Ann Epidemiol 2008;18(7):552e9.
Contents lists available at ScienceDirect
Annals of Epidemiology journal homepage: www.annalsofepidemiology.org
Original article
Neighborhood deprivation and preterm birth: an application of propensity score matching Xiaoguang Ma MD, PhD a, b, *, Nancy L. Fleischer PhD b, c, Jihong Liu ScD b, James W. Hardin PhD b, Guang Zhao PhD b, Angela D. Liese PhD b, c a b c
Department of Nutrition and Food Hygiene, School of Public Health, Zhejiang University, Hangzhou, China Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia Center for Research in Nutrition and Health Disparities, Arnold School of Public Health, University of South Carolina, Columbia
a r t i c l e i n f o
a b s t r a c t
Article history: Received 11 July 2014 Accepted 24 November 2014 Available online 28 November 2014
Purpose: On the basis of a neighborhood deprivation index (NDI), this study aims to examine the association between neighborhood deprivation and preterm birth (PTB) by applying propensity score matching (PSM) methods. Methods: NDI was calculated for all census tracts in South Carolina based on the US Census data. Live births in South Carolina during 2008 to 2009 (n ¼ 98,456) were assigned to an NDI quartile group based on residential addresses. PSM was used to create matched pairs by NDI quartiles to avoid any potential inference on imbalanced data. The differences of prevalence of PTB were calculated for exposed and reference deprivation groups. Results: Neighborhood deprivation was higher among blacks than whites. The overall prevalence of PTB was 8.5% for whites and 12.6% for blacks. Living in neighborhoods with higher deprivation was associated with increased risk of PTB among blacks compared with living in neighborhoods with lower deprivation among blacks. However, random-effect regression models showed that the most deprived whites experienced 1.13 times the odds of having PTB than the least deprived whites. Conclusions: The racial disparities in adverse birth outcomes might be partially explained by neighborhood deprivation in South Carolina. PSM may be an appropriate approach to avoid imbalanced data inferences. Ó 2015 Elsevier Inc. All rights reserved.
Keywords: Neighborhood deprivation index Preterm birth Propensity score Matching Principal component analysis
Introduction Racial disparities in adverse birth outcomes are well documented but not well explained [1,2]. Racial disparities vary across geographic regions with different political, economic, and social contexts [3,4], which suggests that studies focusing on environmental factors, including of the neighborhood environment, are needed to explain racial disparities in birth outcomes [5]. The neighborhood socioeconomic environment may shape individual biological and behavior risk factors, which may cause adverse birth outcomes through a variety of biological mechanisms [6]. However, relationships between neighborhood socioeconomic factors and adverse birth outcomes are not consistent across
Conflict of interest: None declared. * Corresponding author. Department of Nutrition and Food Hygiene, School of Public Health, Zhejiang University, 866 Yuhangtang Road, Hangzhou, Zhejiang 310058, China. Tel.: þ86 571-8820-8591; fax: þ86 571-8820-8519. E-mail address: [email protected] (X. Ma). http://dx.doi.org/10.1016/j.annepidem.2014.11.021 1047-2797/Ó 2015 Elsevier Inc. All rights reserved.
studies. Some studies have identified a relationship between deprived neighborhood conditions and adverse birth outcomes [6e10], some have not [11], and some have only demonstrated the associations among certain racial groups [12e15]. A possible explanation for the inconsistency is that various indicators have been used to characterize the neighborhood socioeconomic context, making results difficult to interpret and compare across studies. The neighborhood deprivation index (NDI), which synthesizes multiple dimensions of the neighborhood socioeconomic context, allows comparisons across geographic areas [16]. This index has been linked to adverse birth outcomes, including low birth weight and preterm birth (PTB) [17e19]. However, the distribution of NDI quartiles can be extremely imbalanced across different racial groups; often, more white women live in less deprived areas and more minority women live in more deprived areas. With the addition of covariates in the adjusted models, certain covariate strata may contain imbalanced data or subjects who could never be exposed to a condition, leading to inference based on no actual data [20].
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
Propensity score matching (PSM) is a useful approach for dealing with these issues. A propensity score is defined as the conditional probability of being exposed to a condition [21,22]. The propensity score reduces the dimensionality of a large set of potential confounders to unity, making it conducive to simple pair matching [23]. After exposure groups are matched by propensity scores, they have been balanced on all available relevant covariates. In this way, we reduce the observable bias while maintaining the support of the data. This study aimed to examine the association between neighborhood deprivation and PTB based on all live births in 2008 to 2009 in South Carolina, stratified by racial groups (whites and blacks). PSM was used to avoid any sparse data among covariate categories caused by an imbalanced distribution of data across quartiles of neighborhood deprivation. Methods Study population Birth certificates of all live births in South Carolina from January 1, 2008, to December 31, 2009, were obtained from the South Carolina Department of Health and Environmental Control. Within the period, there were 123,759 live births. After excluding births without census tract information, multiple births, and births in Hispanic and other racial/ethnic groups (American Indian and Alaska Native, Asian Native Hawaiian, and other Pacific Islander), and extreme outliers of birth weight (3 SD) and gestational age (<20 weeks), 98,456 white and black births were included in the study. Measures PTB was defined as gestational age less than 37 weeks. The PTB data were from the obstetric estimate of gestation from the birth certificate. This measure was added from 2003 in birth certificate and it was based on birth attendant’s final estimate of gestation. The algorithm published by Messer et al. [16] was used to create the NDI for each census tract in South Carolina using principal component analysis. The eight census tract-level sociodemographic factors used to compute the NDI included percentage population with less than high school, percentage unemployed population, percentage males in management occupations, percentage crowded housing, percentage households in poverty, percentage female head households with children, percentage households earning less than $30,000 per year, and percentage households on public assistance. We used data from the 2000 census. The NDI was predicted based on the loadings of the eight factors in the first principal component. In this study, only the first principal component had an eigenvalue more than 1, accounting for 61.08% of the total variance. The NDI was standardized to have a mean of 0 and SD of 1 by dividing the index by the square of the eigenvalue. Quartiles of NDI were coded as Q1 (least deprived), Q2, Q3, and Q4 (most deprived). Q1 was considered the reference group. Principal component analysis analysis was conducted using the pca program in STATA (Version 10, StataCorp, College Station, TX). In the PSM analysis, to achieve the best model fit to predict propensity scores, we included all appropriate covariates which were predictive of the exposure of interest and occurred before the outcome of interest. We included all the sociodemographic variables available in the data set, including maternal age, maternal education, the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC) participation, and urbanicity. Other covariates included body mass index, maternal smoking, prenatal care, number of previous live births, number of previous PTBs, and
121
maternal risk factors such as infection, chronic and gestational hypertension, and diabetes. Statistical analysis Population characteristics were summarized for the pooled sample and for samples in each NDI quartile. Q2, Q3, and Q4 were compared with Q1 based on t tests for continuous variables and c2 tests for categorical variables. As shown in Figure 1, the distribution of NDI was imbalanced between white and black women, with approximately 50% of black women living in the most deprived areas. To avoid inference due to the imbalanced distribution of NDI, we used PSM to analyze the relationship between NDI and adverse birth outcomes stratified by race. We used logistic regression to estimate the predicted probability of a mother’s exposure to neighborhood deprivation to create matched pairs comparing NDI quartiles. All appropriate covariates discussed previously were included in the models to achieve the best fit. Propensity scores were estimated for each mother and computed separately for whites and blacks. We then matched the mothers living in deprived areas (Q2, Q3, and Q4, separately) with those living in the reference area (Q1) with the same propensity score. The matching procedure was conducted using the psmatch2 module in STATA. The mothers living in deprived areas were matched 1:1 with replacement (to yield the largest number of matched pairs, the matched observations were returned to the data pool and could be matched again if the propensity score could be matched) to mothers living in reference areas with the same predicted probability of exposure to neighborhood deprivation within a range of 0.01. We yielded 100% matching between the deprived group and the reference group because of the large sample size. Balance tests were performed to assure the balance of the data after matching. We compared the means and percentage bias before and after matching, and percentage bias reduction, with a goal of a percentage bias reduction of less than 10% indicating sufficient balance. The percentage bias is the percentage difference of the sample means in the deprived and reference group as a percentage of the square root of the average of the sample variances [24]. The graph of propensity score overlap was generated by level of neighborhood deprivation for each racial group. Generally, the graph showed the distribution of propensity score between exposed (deprived groups such as Q2, Q3, and Q4) and unexposed (Q1) groups. Adequate overlap between the distributions showed that we should be able to find a suitable match for an exposed subject. After the subjects were matched, the differences in prevalence of PTB were computed between the matched deprived and
Fig. 1. Distribution of NDI by race in South Carolina.
122
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
group; below the midline are for the reference group (Q1). Generally, there was adequate overlap between the two exposure groups. Most of the overlap was in the middle of the propensity score distribution for Q2 versus Q1 and Q3 versus Q1, whereas most of the overlap for Q4 versus Q1 was on the left side of the distribution among whites and the right side of the distribution among blacks. After matching, the percentage bias reduction for covariates ranged from 0% to 7.4%, which achieved the 10% goal discussed previously. For most covariates, the percentage bias was reduced after PSM (results not shown). Based on the matched pairs of deprived (Q2 or Q3 or Q4) and reference (Q1) mothers, prevalence differences were calculated for PTB among whites and blacks (Table 2). Among whites, the prevalence difference between the more deprived group and the reference group ranged from 0.38% (Q2 vs. Q1) to 1.42% (Q4 vs. Q1) for PTB. According to the 95% confidence intervals, no statistically significant difference was found. For blacks, compared with mothers living in the least deprived (Q1) areas, those living in the most deprived (Q4) areas experienced a 2.91% higher prevalence of PTB. No difference was found for other NDI quartiles for PTB. The results of multivariate random effects logistic regression models are shown in Table 3. In the models for whites, mothers living in Q4 areas had 1.13 times the odds of giving PTB births when compared with mothers living in Q1 areas. However in the analysis for blacks, no significant differences were found for PTB among different neighborhood deprivation areas.
reference groups. There are a variety of methods for the calculation of the standard error of propensity score matched effect estimates. In this study, we used the bootstrap method with 1000 repetitions to calculate bias-corrected 95% confidence intervals. To compare the PSM results to a typical regression analysis, we conducted a sensitivity analysis using random effects (women clustered in the Census tracts) multivariate logistic regression models to examine the association between NDI and adverse birth outcomes, stratified by race. The maternal age, maternal education, WIC participation, and urbanicity were adjusted as covariates in the models. The random effects regression models were fitted with xtlogit command for multilevel analysis in STATA. P values less than .05 were considered as significance level. Results Table 1 shows the characteristics of the sample. The overall prevalence of PTB was 8.5% for whites and 12.6% for blacks. Average NDI was higher among blacks than whites in South Carolina. Women residing in the Q2, Q3, and Q4 (most deprived) quartile of the NDI were more likely to be younger and have lower levels of education, a higher proportion of WIC participation and rural residence, and have babies with lower birth weight than women living in the least deprived quartile of NDI (Q1), regardless of race. The mean gestational age and the prevalence of PTB were different between deprived and reference areas among whites, but not among blacks. PSM yielded 100% matching between deprived quartiles (Q2eQ4) and the reference quartile (Q1) of the NDI. Figure 2 graphically depicts the propensity score overlap by NDI quartiles among whites (upper panel) and blacks (lower panel). The bars above the midline are propensity scores for the more deprived
Discussion In this study, we used PSM to examine the difference in the prevalence of PTB for mothers living in neighborhoods with different levels of deprivation. We found that neighborhood
Table 1 Characteristics of sample by quartiles of neighborhood deprivation index in South Carolina Variables
Mean (SD) or % Q1
P for trend Q2
Q3
Q4
Total
White
N ¼ 21,895
N ¼ 21,662
N ¼ 12,713
N ¼ 6153
N ¼ 62,423
Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural Birth weight, g Low birth weight Gestational age, w PTB NDI
29.1 (5.7)
26.7 (5.8)*
25.9 (5.7)*
25.5 (5.7)*
27.3 (5.9)
<.01
21.3 29.8 48.9 23.1 26.2 3388.2 (492.6) 4.98 38.6 (1.7) 7.79 1.19 (0.32)
41.4* 34.1 24.5 43.9* 52.7* 3348.0 (505.6)* 6.16* 38.5 (1.9)* 8.51* 0.37 (0.19)*
49.0* 32.7 18.4 52.8* 68.3* 3322.8 (506.0)* 6.43* 38.5 (1.8)* 8.81* 0.23 (0.18)*
55.5* 29.7 14.8 60.9* 62.3* 3290.4 (513.8)* 7.56* 38.4 (1.9)* 9.91* 1.00 (0.42)*
37.3 31.9 30.8 40.2 47.5 3351.3 (503.0) 5.94 38.5 (1.8) 8.46 0.40 (0.75)
<.01
<.01 <.01 <.01 <.01 <.01 <.01 <.01
Non-Hispanic black
N ¼ 5303
N ¼ 7482
N ¼ 9362
N ¼ 13,886
N ¼ 36,033
Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural Birth weight, g Low birth weight Gestational age, w PTB NDI
26.3 (6.2)
25.1 (5.8)*
24.5 (5.7)*
24.0 (5.5)*
24.7 (5.8)
<.01
38.8 37.8 23.3 61.1 17.5 3147.6 (499.5) 11.90 38.1 (2.4) 12.41 1.15 (0.28)*
49.5* 38.1 12.4 72.2* 45.5* 3129.0 (495.5)* 11.84 38.1 (2.4) 12.64 0.33 (0.19)*
58.1* 34.2 7.7 79.6* 61.8* 3099.9 (493.8)* 12.70 38.2 (2.4) 12.12 0.27 (0.17)*
67.4* 27.8 4.7 82.0* 53.7* 3088.1 (490.2)* 12.89 38.1 (2.4) 12.91 1.40 (0.75)*
57.1 33.1 9.8 76.3 48.8 3108.4 (494.5) 12.48 38.1 (2.4) 12.57 0.37 (1.04)
<.01
BMI ¼ body mass index; Q ¼ Neighborhood Deprivation Index quartiles (Q1, less deprived to Q4, more deprived). Q1 was used as the reference group, and all other three groups were compared with the reference. t test and c2 were used to compare for continuous and categorical variables, respectively. * P < .05.
<.01 <.01 <.01 <.05 .927 .390 <.01
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
123
Fig. 2. Propensity score overlap by level of neighborhood deprivation for whites (upper panel) and blacks (lower panel) (Y-axis is the number of births and X-axis is the propensity score. The area above middle axis is for deprived groups including Q2, Q3, and Q4, and the area below middle axis is for reference group Q1.).
deprivation was higher among blacks than whites. Living in neighborhoods with higher deprivation was associated with increased risk of PTB among blacks compared with living in neighborhoods with lower deprivation among blacks. A number of studies have demonstrated a positive association between neighborhood deprivation and adverse birth outcomes; however, results are inconsistent for different racial/ethnic groups [13,16e19]. Pearl et al. [14] found that in addition to individual socioeconomic characteristics, living in neighborhoods that are less socioeconomically advantaged may differentially influence birth weight, depending on women’s ethnicity and nativity. Messer et al. [25] and Kaufman et al. [8] found that living in less deprived or higher income neighborhoods was associated with an increased risk of PTB among African-American but not Caucasian women. Masi et al. [6] concluded that living in an economically disadvantaged neighborhood put African American women, but not Caucasian or Hispanic women, at an increased risk of having a PTB. However, the findings are difficult to compare because of different Table 2 Difference of prevalence of PTB between neighborhood deprivation index quartiles after propensity score matching by race in all births 2008 to 2009 in South Carolina Difference between NDI quartiles
Matched Prevalence Prevalence Prevalence Biasdifference, % corrected pairs in deprived in Q1, % 95% CI* group, %
Non-Hispanic white Q2 versus Q1 21,895 Q3 versus Q1 21,895 Q4 versus Q1 21,895 Non-Hispanic black Q2 versus Q1 7482 Q3 versus Q1 9362 Q4 versus Q1 13,886
8.51 8.81 9.90
8.13 8.57 8.48
0.38 0.24 1.42
0.77 to 1.61 1.90 to 1.30 0.46 to 2.84
12.65 12.11 12.91
11.22 12.13 10.00
1.43 0.02 2.91
1.22 to 2.87 2.91 to 1.64 1.48 to 4.92
CI ¼ confidence interval; Q ¼ Neighborhood Deprivation Index quartiles (Q1, less deprived to Q4, more deprived). Bolded means P < .05. * The bias-corrected 95% CIs were calculated by bootstrap method with 1000 replications.
neighborhood deprivation indicators. An index (NDI) was developed by Messer et al. [16] to synthesize multiple dimensions of the neighborhood context and allow comparisons across geographic areas. By using this index, Elo et al. [17] found that the association Table 3 The association between neighborhood deprivation index quartiles and PTB from random-effect logistic regressions by race in all births 2008 to 2009 in South Carolina Variables
OR (95% CI)
Non-Hispanic white
N¼ 57,608
NDI Q1 Q2 Q3 Q4 Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural
1.00 1.02 1.03 1.13 1.01
(0.94e1.12) (0.94e1.14) (1.01e1.27) (1.01e1.02)
1.00 0.85 0.65 0.95 1.08
(0.79e0.91) (0.60e0.72) (0.89e1.02) (1.01e1.16)
Non-Hispanic black
N¼34,356
NDI Q1 Q2 Q3 Q4 Mother’s age, y Mother’s education High school or less Some college Bachelor or above WIC participation Living in rural
1.00 1.06 1.01 1.07 1.03
(0.94e1.20) (0.89e1.14) (0.96e1.21) (1.02e1.03)
1.00 0.87 0.61 0.72 1.00
(0.81e0.94) (0.54e0.70) (0.66e0.77) (0.92e1.07)
CI ¼ confidence interval; OR ¼ odds ratio; Q ¼ quartile. Adjusted variables are maternal age, maternal education, WIC participation, and urbanicity. Bolded means P < .05.
124
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125
between neighborhood deprivation and small-for-gestational age did not vary significantly by race. In Janevic et al.’s study [18], associations were reported for PTB among Hispanic Caribbean women only. Several studies claimed larger neighborhood effects on PTB among whites than among blacks [16,19]. By using the same NDI measure, our results confirmed that the association between neighborhood deprivation and PTB varied by race, and that the results are different for whites and blacks. The reasons for inconsistent results may be because of the methodological innovation used in this study, PSM methods. Previous studies used logistic regression models to examine the relationship between NDI and birth outcomes. Compared with regression models, PSM weights the data differently and bases its inference on actual data only. To compare our results with previous studies, we analyzed our data using logistic regression models with a random effect. The logistic results showed different results for PTB. The effects of NDI on PTB were much larger among whites than among blacks, which are consistent with previous findings [16,19]. However, PSM methods showed different results compared with traditional logistic regression models. As discussed previously, PSM avoided relying on imbalanced data which were usually observed in research topic of adverse birth outcomes and racial disparities as well as neighborhood deprivation. In such topics, PSM may be an appropriate approach to solve the issue of inference on imbalanced data. Our findings are subject to several limitations. Although PSM was preferred for the data pattern in this study, there are some limitations for this method. PSM did not account for unobserved or unobservable characteristics. Rosenbaum has developed a method of sensitivity analysis to assess if one’s estimate based on matching is robust to the possible presence of an unobserved confounder [26]. Based on this sensitivity analysis, we yielded tight confidence bounds around the log odds of differential assignment due to unobserved factors. We also observed very small Hodgese Lehmann point estimates, which indicated that unmeasured confounding was inconsequential. Moreover, PSM did not incorporate the “clustering” of the neighborhood. However, small within-tract variance was found from multilevel logistic regression models in this study (data not shown). In addition to the limitations of PSM, we only had the census tract IDs in 2000 for the mothers in the database when the data were requested from South Carolina Department of Health and Environmental Control. To merge with birth data by census tract ID, we used 2000 census data to identify NDI. The birth data are closer to 2010; thus, using 2000 census data might cause bias. However, we used the quartiles of NDI, which reflected the relative rather than the actual difference among different tracts. Therefore, the 2000 census boundaries might not have a significant influence on the results of this study. Despite these limitations, this study had several strengths beyond previous studies. In general, 1:1 PSM often yields observations which cannot be matched. However in our study, the large sample size allowed for 100% matching between deprived and reference groups. PSM is not a new approach [21], but it has only started to be used in social epidemiology and reproductive health research in recent years [23,27,28]. However, to our knowledge, no studies have used PSM method to examine the association between NDI and adverse birth outcomes. In addition, this is the first study on neighborhood deprivation and birth outcomes in South Carolina, where racial disparities on adverse birth outcomes are a serious public health concern. Conclusions Neighborhood deprivation was associated with increased risk of PTB among blacks. The racial disparities on birth outcomes might
be partially explained by the neighborhood deprivation according to this study. PSM can be an appropriate approach to avoid inference on unbalanced data. Future research using PSM is encouraged to examine the effect of neighborhood characteristics on birth outcomes. Acknowledgments This work was not supported by any grants or funds. The authors thank James Hibbert (Center for Research in Nutrition and Health Disparities, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina) for geographic information system analysis and Dr. Daniela Nitcheva, Mr. Sung-Jun Kim, and Mr. John Allen (South Carolina Department of Health and Environmental Control, Columbia, South Carolina) for providing the data needed for this study. References [1] Goldenberg RL, Cliver SP, Mulvihill FX, Hickey CA, Hoffman HJ, Klerman LV, et al. Medical, psychosocial, and behavioral risk factors do not explain the increased risk for low birth weight among black women. Am J Obstet Gynecol 1996;175(5):1317e24. [2] Lu MC, Halfon N. Racial and ethnic disparities in birth outcomes: a life-course perspective. Matern Child Health J 2003;7(1):13e30. [3] Teitler JO, Reichman NE, Nepomnyaschy L, Martinson M. A cross-national comparison of racial and ethnic disparities in low birth weight in the United States and England. Pediatrics 2007;120(5):e1182e9. [4] Nepomnyaschy L. Race disparities in low birth weight in the US south and the rest of the nation. Soc Sci Med 2010;70(5):684e91. [5] Metcalfe A, Lail P, Ghali WA, Sauve RS. The association between neighbourhoods and adverse birth outcomes: a systematic review and meta-analysis of multi-level studies. Paediatr Perinat Epidemiol 2011;25(3):236e45. [6] Masi CM, Hawkley LC, Piotrowski ZH, Pickett KE. Neighborhood economic disadvantage, violent crime, group density, and pregnancy outcomes in a diverse, urban population. Soc Sci Med 2007;65(12):2440e57. [7] Agyemang C, Vrijkotte TG, Droomers M, van der Wal MF, Bonsel GJ, Stronks K. The effect of neighbourhood income and deprivation on pregnancy outcomes in Amsterdam, The Netherlands. J Epidemiol Community Health 2009;63(9): 755e60. [8] Kaufman JS, Dole N, Savitz DA, Herring AH. Modeling community-level effects on preterm birth. Ann Epidemiol 2003;13(5):377e84. [9] Messer LC, Kaufman JS, Dole N, Savitz DA, Laraia BA. Neighborhood crime, deprivation, and preterm birth. Ann Epidemiol 2006;16(6):455e62. [10] Schempf A, Strobino D, O’Campo P. Neighborhood effects on birth weight: an exploration of psychosocial and behavioral pathways in Baltimore, 1995e1996. Soc Sci Med 2009;68(1):100e10. [11] Cubbin C, Marchi K, Lin M, Bell T, Marshall H, Miller C, et al. Is neighborhood deprivation independently associated with maternal and infant health? Evidence from Florida and Washington. Matern child Health J 2008;12(1):61e74. [12] Buka SL, Brennan RT, Rich-Edwards JW, Raudenbush SW, Earls F. Neighborhood support and the birth weight of urban infants. Am J Epidemiol 2003;157(1):1e8. [13] Messer LC, Vinikoor LC, Laraia BA, Kaufman JS, Eyster J, Holzman C, et al. Socioeconomic domains and associations with preterm birth. Soc Sci Med 2008;67(8):1247e57. [14] Pearl M, Braveman P, Abrams B. The relationship of neighborhood socioeconomic characteristics to birth weight among 5 ethnic groups in California. Am J Public Health 2001;91(11):1808e14. [15] Pickett KE, Ahern JE, Selvin S, Abrams B. Neighborhood socioeconomic status, maternal race and preterm delivery: a case-control study. Ann Epidemiol 2002;12(6):410e8. [16] Messer LC, Laraia BA, Kaufman JS, Eyster J, Holzman C, Culhane J, et al. The development of a standardized neighborhood deprivation index. J Urban Health 2006;83(6):1041e62. [17] Elo IT, Culhane JF, Kohler IV, O’Campo P, Burke JG, Messer LC, et al. Neighbourhood deprivation and small-for-gestational-age term births in the United States. Paediatr Perinat Epidemiol 2009;23(1):87e96. [18] Janevic T, Stein CR, Savitz DA, Kaufman JS, Mason SM, Herring AH. Neighborhood deprivation and adverse birth outcomes among diverse ethnic groups. Ann Epidemiol 2010;20(6):445e51. [19] O’Campo P, Burke JG, Culhane J, Elo IT, Eyster J, Holzman C, et al. Neighborhood deprivation and preterm birth among non-Hispanic black and white women in eight geographic areas in the United States. Am J Epidemiol 2008;167(2):155e63. [20] Messer LC, Oakes JM, Mason S. Effects of socioeconomic and racial residential segregation on preterm birth: a cautionary tale of structural confounding. Am J Epidemiol 2010;171(6):664e73.
X. Ma et al. / Annals of Epidemiology 25 (2015) 120e125 [21] Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika 1983;70(1):41e55. [22] Rosenbaum PR, Rubin DB. Reducing bias in observational studies using subclassification on the propensity score. J Am Stat Assoc 1984;79(387):516e24. [23] Oakes JM, Kaufman JS. Methods in social epidemiology. San Francisco: JosseyBass; 2006. [24] Rosenbaum PR, Rubin DB. Constructing a control group using multivariate matched sampling methods that incorporate the propensity score. Am Stat 1985;39(1):33e8.
125
[25] Messer LC, Kaufman JS, Dole N, Herring A, Laraia BA. Violent crime exposure classification and adverse birth outcomes: a geographically-defined cohort study. Int J Health Geogr 2006;5:22. [26] Rosenbaum PR. Heterogeneity and causality. Am Stat 2005;59(2):147e52. [27] Hearst MO, Oakes JM, Johnson PJ. The effect of racial residential segregation on black infant mortality. Am J Epidemiol 2008;168(11):1247e54. [28] Johnson PJ, Oakes JM, Anderton DL. Neighborhood poverty and American Indian infant death: are the effects identifiable? Ann Epidemiol 2008;18(7):552e9.