Study design
This study was designed as a retrospective cross-sectional study linking data collected as part of a prospective cohort study to geocoded administrative data on violent crime.
Study Setting
This study was conducted in the municipality of São Paulo, Brazil. In São Paulo State, the proportion of preterm births increased from 7.7% in 2004 (7.7%) to 12.4% in 2013 [39]. At the same time, preterm birth increased from 8.6% in 2004 to 11.6% in 2013 in the São Paulo municipality. In 2016, preterm rates were 10.9% and 10.7% in São Paulo State and São Paulo municipality, respectively [39]. Regarding low birth weight, both State and municipality had low variation across time, considering the available data, with the highest proportion observed in 2012, 9.5% and 9.7%, respectively [40]. Data on children born small for gestational age is not publicly available for either area.
Like other LMICs, Brazil has experienced a rapid and mostly unmanaged urbanization process in the past decades, accompanied by large increases in social and economic inequality [41, 42]. The rates of homicides in Brazil rose steadily over the past years, reaching a record level of 31.6 homicides per 100,000 inhabitants in 2017 [43]. With 10.3 homicides per 100,000 inhabitants, the São Paulo municipality had one of the lowest homicide rates in Brazil in 2017 overall, but still experienced very high crime rates in areas of low socioeconomic development [43]. In the Western Region of São Paulo, where this study was conducted, homicide rates ranged between 5 and 20 homicides per 100,000 inhabitants across neighborhoods [44].
Study Population
The São Paulo Western Region Birth Cohort (Região Oeste Coorte – ROC-Cohort) enrolled locally resident infants born at the University Hospital of the School of Medicine of the University of São Paulo between April 1, 2012, and March 31, 2014. The study’s main objective was to evaluate the impact of environmental and social determinants of health on child development. A total of 7,066 births occurred in the period, and 6,207 mother-child pairs who resided in the studied setting were enrolled in the cohort. The cohort is still active, with children currently under the 72 months follow-up. A full description of the cohort can be found elsewhere [45].
Data
Hospital electronic birth records were available for all children in the cohort. The electronic medical registry includes birth characteristics such as type of delivery, gestational length, weight at birth, and others. During the postpartum hospital stay, all mothers were invited to complete a short-structured questionnaire administered by a trained interviewer to collect information on socio-demographic characteristics and health during pregnancy. The questionnaire was applied after maternal consent or a parent if the mother was too young to consent. The questionnaire can be found in Supplementary file 1, and the rates of interview and process can be found in the cohort profile [45].
Outcomes
The analysis focused on three adverse birth outcomes: low birth weight (LBW), preterm delivery (PT), and small-for-gestational-age (SGA). Birth weight was measured by the Hospital’s neonatology team immediately after birth using standard hospital equipment. Gestational length in weeks was estimated using the New Ballard Score [46]. LBW was defined as birth weight < 2,500 grams, and SGA was defined as weight-for-gestational age < 10th percentile based on the Intergrowth-21st growth reference tables [47].
Exposure – violence in the neighborhood
Data on violence is routinely collected and made publicly available by the Secretariat of Public Safety of the State of São Paulo (www.ssp.sp.gov.br). The system collects detailed information on all willful murders, femicides (gender-based murder), robberies followed by death, bodily injuries followed by death, deaths resulting from police intervention, suspicious deaths, fatal vehicle accidents, and cellphone and car theft occurring in São Paulo State. Data on other crimes such as intimate partner abuse (physical, emotional, financial), rape, and child sexual assault are not collected in this system.
Each reported incident record contains the date, time, and address of the crime. Following most of the external violence literature, we focused on violent crime in our analysis, which includes willful murder, robbery followed by death, bodily injury followed by death and death resulting from police intervention but excludes other crimes such as robberies without injuries. Feminicide was not included in the analysis because this data was only available after 2015.
We extracted data on violent crimes between 2011 and 2014 to cover all ROC-Cohort children’s pregnancy period. The address of each reported crime was geocoded with latitude and longitude coordinates using the ggmap package [48], and a point-layer for each time point (month-year) was generated using R map tools [49]. The maternal residential address was collected at birth and geocoded as an additional point-layer. Each residential address was treated as a centroid point. We then computed the number of crimes within a 1-kilometer spatial buffer by day, month, and year. Exposure to violence during pregnancy was estimated as the sum of violent crimes in the first two trimesters of pregnancy. Violence in the third trimester was not considered because total exposure to violence after week 24 of gestation directly depends on gestational length in weeks (the outcome variable). The sum of violent crimes within 1-kilometer of each mother’s residence during the first two trimesters was categorized into five equally sized quintiles for analysis. This categorization was chosen to allow us to compare the 20% of births with the lowest external violence exposure to families in higher exposure quintiles. We provide a detailed discussion of the exposure differences between these quintiles below.
Statistical analyses
We first computed the average number of crimes for each of the exposure quintiles. Next, we tested crude associations of exposure to violence quintiles with each outcome (LBW, SGA, and PT birth). In the fully adjusted multivariable models, we included all covariates highlighted in the existent literature as risk factors for adverse birth outcomes, including maternal age (<20, >=35), educational level (incomplete primary, complete secondary and tertiary), alcohol consumption (yes/no), smoke (yes/no), diabetes (yes/no), hypertension (yes/no) and exposure to physical violence (yes/no) during pregnancy – reported by the mother in the postpartum questionnaire, and socioeconomic (SES) index. The construction of a socioeconomic (SES) index was based on the approach proposed by Vyas and Kumaranayake [50], which relies on the use of principal component analysis (PCA) of the selected variables. The SES index was based on questions regarding asset holdings such as cars, family income, and household characteristics. The PCA’s first principal component was then divided into five SES quintiles - low, medium-low, medium, medium-high, and high SES status. Maternal age was categorized into adolescence (<20), core reproductive age (22-34) and ages > 35.
Given that the postpartum questionnaire was not available for the whole sample, we initially screened the dataset for missing data and patterns of non-responsiveness in the postpartum questionnaire using t-test and chi-square to test if the assumptions of missing at random (MAR) were met. The reasons for not completing the postpartum questionnaire vary from being too young to consent to being discharged before the interview could be completed. The details can be found elsewhere [45]. As the assumptions of MAR in our sample were met, we implemented the MICE to complete the missing data in the postpartum questionnaires to all cohort participants. We implemented multiple imputations by chained equations (MICE) with a fully conditional specification of prediction equations using variables that potentially predicted nonresponse or the outcome. The MICE method accounts for statistical uncertainty in the imputations and ensures more reliable inference in settings with missing data. All imputed models had a relative efficiency higher than 99%.
We performed the MICE using Stata’s mi package [51]. The variables used in the MICE models included exposures (e.g., violence during pregnancy), birth outcomes (PT, LBW and SGA), maternal characteristics (e.g., age, socioeconomic, education), pregnancy variables (e.g., diabetes, hypertension), and neonatal variables (e.g., birth weight, length, and sex). We generated 50 independent datasets (M = 50) and followed Rubin’s rules to aggregate results across imputed datasets [52]. All analyses were performed using Stata version 14 [53]. Further details on these imputations are provided in Table S1 (Supplementary file 2).