We used data from the 2008 Boston Youth Survey (BYS), a cross-sectional survey of high school students in grades 9-12 in Boston Public Schools [16, 17]. Of the thirty-two public high schools in Boston, 69% (n=22) agreed to participate and were representative of all schools in the Boston area with respect to race/ethnicity of the students, school drop-out rates and other socio-demographic variables [18].
A self-administered questionnaire was developed using reliable and valid scales that measure behaviors and experiences in the neighborhood. Each school selected a list of classrooms stratified by grades. A random sample of classes were selected for participation until 100-120 students were identified per school. All students in selected classrooms were invited to complete a paper-and-pencil survey during the Spring of 2008 [18]. The sample size included 1,878 students; a response rate of 69%. We used multiple imputation to address missing socio-demographic and behavioral data. However, students who did not provide the location of their residence were excluded. The imputerd analytical sample, which included complete socio-demographic and individual-level social cohesion data, comprised of 1,506/1878 (80.2%) students. We created five multiply- imputed datasets. Multilevel regression analyses was then used to fit the pre-specified model to each of the imputed datasets. Next, we averaged the estimates to obtain estimated associations [19]. All analyses was completed using Stata 14.0. Those with missing data were more likely to be male, black, and older in age and to have immigrated to the USA within the last 4 years.
The Office of Human Research Administration of the Harvard School of Public Health approved all data collection procedures and research protocols for the BYS. All methods were carried out in accordance with relevant guidelines and regulations. Passive consent (i.e., upon informed consent, students' parents were required to return a form if they did not want their child to participate) was used and students were allowed to refuse to participate at any time before or during the survey administration [18].
Study variables
Outcome: Alcohol consumption behavior- Three alcohol behavior outcomes were assessed. Lifetime alcohol consumption (yes vs. no), alcohol consumption in the past 30 days (yes vs. no), and alcohol consumption more than 3 times in the past 30 days (yes vs. no). Given the age of the respondents (ages 13 to 19), each of these variables represents illegal behaviors since the minimum drinking age in Massachusetts is 21 years of age.
Area-level covariates- In order to geocode the each student's residence to U.S. Census tracts, investigators asked them for their nearest cross-street. Of the total sample, 85.9% (n=1,614) provided their locations. BYS investigators consulted with key informants from Boston neighborhoods to aggregate the 157 Boston Census tracts (each with a population of approximately 4,000) into 38 socially meaningful neighborhood clusters of tracts.[20] The details of this process are described elsewhere [18]. Neighborhood-level characteristics were then created for this investigation.
The main exposure of interest was income inequality within the Census Tract (CT), which was measured using the Gini coefficient. The Gini coefficient ranges from 0 (perfect equality, where every household in the CT has the exact same income) to 1.0 (perfect inequality, where households in the CT earn a wide range of incomes). In this investigation, the Gini coefficient was calculated for each census tract by the Boston Indicators Project (http://www.bostonindicators.org/), which was then linked to the BYS dataset. We categorized the Gini coefficient using the tertiles as threshold cutoffs. The Gini coefficient is based on the Lorenz curve, a cumulative frequency curve that compares the distribution of a specific variable with the uniform distribution that represents equality [21].
Using principal components analysis, a socioeconomic composite score, economic deprivation, was created for each of the 38 neighborhoods. Economic deprivation is comprised of U.S. Census indicators: proportion of residents living below the poverty level, proportion of households receiving public assistance, and proportion of families with a female head of household. The Cronbach a was .84, which is indicative of good internal consistency. A higher score was indicative of greater economic deprivation. Neighborhood economic deprivation was categorized into low, moderate, and high using tertiles as thresolds.
The Boston Neighborhood Survey (BNS) [18, 22] was a random-digit dial telephone survey, administered among adult residents (³18 years). Respondents were randomly selected from a list-assisted sampling frame, stratified proportional to population size of the 16 large neighborhoods defined by the Boston Redevelopment Authority, resulting in a sample size of 1,710 adults in 2008. The purpose of the BNS was to enrich the BYS data with contextual information about neighborhood-level conditions and social processes perceived by adult residents [18, 22]. For each of the 38 neighborhoods, neighborhood disorder, which is comprised of both social (i.e. presence or absence of drinking alcohol in public) and physical disorder (i.e., abandoned cars) scores were determined. A combined score was created using these two indicators, with higher scores indicating greater neighborhood disorder. Tertiles were used to categorize neighborhoods into low, moderate, and high neighborhood disorder.
We also used the BNS to measure neighborhood social cohesion by adapting a reliable and valid questionnaire [23]. Respondents were asked if they strongly agreed, agreed, disagreed, or strongly disagreed with five statements. For example: "People in my neighborhood can be trusted" and "People are willing to help their neighbors." A combined score was created and a greater score indicated higher social cohesion. Tertiles were used to categorize neighborhoods into low, moderate, and high neighborhood social cohesion.
Boston Police Department data were used to develop a global score for neighborhood danger in each of the 38 neighborhoods. Indicators included: counts of criminal homicide, robbery, aggravated assault, burglary, larceny theft, vehicle theft, and arson. The higher the score the greater the danger was within the neighborhood. These indicators were then matched to the U.S. Census tracts and used to characterize the 38 neighborhoods. Tertiles were used to categorize danger within the neighborhood into low, moderate, and high.
Students' age, nativity (U.S. born, foreign born, arrived £4 years, and foreign born arrived >4 years), and race or ethnicity (white, black, Asian, Hispanic, and other) were also measured at the student-level.
Student social cohesion was also measured by asking them for their perception of their neighborhood using five statements, which included: I live in a neighborhood where people know and like each other; People in my neighborhood generally get along with each other; People in my neighborhood generally share the same beliefs about what is right and wrong; People in my neighborhood can be trusted. Response options ranged from (1) strongly disagree to (4) strongly agree. The mean social cohesion score was 12.0 (standard deviation [SD]=2.9) and the range was 5-20. The items showed high internal consistency (Cronbach a=.80). Tertile cutoffs were used to categorize social cohesion into low, moderate, and high values.
Statistical Analysis
We used multilevel logistic modeling to investigate the relationship between neighborhood income inequality and alcohol consumption, while adjusting for both individual and area-level characteristics. Since students were nested within CTs, which were nested within neighborhoods, a three-level multi-level model was initially considered-i.e., models with a random intercept specified for each CT and each neighborhood [24]. However, because a small amount of variation in alcohol consumption behavior was explained at the CT-level (data not shown), we treated income inequality as an individual-level exposure resulting in a two-level model (with neighborhood as the level-two unit).
We fitted the following sequence of models to investigate the association between neighborhood income inequality and the three alcohol consumption outcomes. First, we conducted an intercept-only model, which allowed us to calculate the 95% plausible value range. This is an indication of the variability of likelihood of experiencing each outcome, similar to the ICC. Second, we determined the crude relationship between income inequality and each of the alcohol consumption outcomes. Third, we fitted models adding individual and neighborhood characteristics. Fourth, we added the sex*income inequality interaction term to determine if the association between income inequality and alcohol consumption behavior differed between boys and girls. Finally, we added students' perceptions of neighborhood social cohesion and individual depressive symptoms to determine if perceptions mediated the relationship between neighborhood income inequality and alcohol consumption [25]. These mechanisms were evaluated using the Baron and Kenny [25] method of testing and comparing results from three different models: (1) income inequality and each of the alcohol consumption outcomes, (2) income inequality and social cohesion and depressive symptoms, and (3) social cohesion and depressive symptoms and each of the alcohol consumption outcomes.