Data
The data comes from the National Longitudinal Survey of Adolescent Health (AddHealth). The first wave of AddHealth was conducted in 1994-1995, and consisted of a nationally representative sample of school students in grades 7-12 in the United States. The sample was a school-based cluster design, which was selected to ensure a nationally representative population, by region, urbanicity, racial and ethnic composition, school type, and school size [20].
A total of 20,745 students completed the in-home interview. Parents of students who were selected for the in-home interview were also interviewed. The third wave of the study, which followed-up respondents from the Wave I interview, was completed in 2001-02, when respondents were 18-27 years old. A few respondents (72 students) with mental retardation were excluded from the sample, as their educational trajectories are likely to differ from those of their peers.
Estimation Models
The estimation models examine the relationship between adolescent SES and subsequent use of alcohol and illicit drugs, controlling for baseline use of alcohol and illicit drugs and mental health (depressive symptoms, delinquency and suicidality) and a set of other individual, family school and neighborhood characteristics.
The outcome variables of interest are binge drinking and illicit drug use in early adulthood. AddHealth asks respondents separately about their use of alcohol, marijuana, cocaine, crystal methamphetamine and any other drug use. As the use of these substances is measured separately in AddHealth, they are analyzed separately in these analyses. Substance use was measured as a binary indicator of use, as previous research has found that responders tend to report accurately whether they consume substances, but tend to under-report the amount consumed [6]. Sensitivity analyses found that results were qualitatively similar when examining a continuous measure of use (not shown).
The key independent variable was socioeconomic status at baseline, as measured by parental education and income. As these were assessed separately in AddHealth, they are analyzed separately in these analyses as well. The parents' socioeconomic status reflects the economic status of the family in which the adolescent resides at baseline. Previous research has shown that parental socioeconomic status reflects the child's socioeconomic status [21]. The highest educated parent was used rather than the education level of the mother and father to avoid losing respondents living in single parent families [22]. Household income was also included as a separate measure of SES, as household income captures smaller gradations in family well-being than parental income [23].
A number of the adolescent's baseline personal characteristics were also included in the model. The adolescent's baseline binge drinking and illicit drug use was included in the model in order that these models measure change in substance use from baseline to follow-up in early adulthood. In addition, other behavioral health factors at baseline were controlled for, including delinquency, suicidality and depressive symptoms, due to the high rates of comorbidity of mental health diagnoses among persons with substance abuse [24].
A number of demographic variables were also included in the model, including age, gender, and race/ethnicity. Additional individual characteristics that have been shown to be correlated with educational attainment were included, such the Peabody Picture Vocabulary (PPVT) score, which measures cognitive ability [25], and general health, which measures physical health and may account for endogeneity between mental health and labor market outcomes [26, 27].
A number of family background variables were also included in the model. An indicator of whether the respondent was the first child born in the family was included, as previous research has shown that children born first in the family tend to have better educational outcomes [28]. Family structure (single parent household, stepfamily and foster/other) is included, as children from non-traditional families often face barriers to schooling [29]. An indicator for whether the respondent's biological mother and father have alcoholism is included as alcoholism has a high genetic component [30].
The estimation model is of the form:
Where:
pi,t+1 = Probability of substance use at Wave III, as measured by binge drinking, marijuana use, cocaine use, crystal methamphetamine use and other drug use, estimated separately. Binge drinking was measured as a binary variable with a positive value indicating 5 or more drinks in one setting more than once a month in the past year. Marijuana, cocaine, crystal methamphetamine and other drug use was measured as a binary variable with a positive value indicating any use in the past 30 days.
SESi = Socioeconomic status at baseline, as measured by parental education and household income. Parental education was measured by the education level of the highest educated parent living with the adolescent, and was categorized by the highest educated parent having not completed high school, having completed high school only, having education beyond high school but less than a college degree, or having a college degree. Household income was measured continuously and was topcoded in AddHealth at $999,000, so reported income ranged from $0-$999,000.
BHi = A number of behavioral health (mental health and substance use) characteristics at baseline in adolescence (Wave I interview) were included in the model: binge drinking, marijuana, cocaine, inhalant and other drug use at baseline, as well as depressive symptoms, delinquency and suicidality. Each of these components was included separately in the model. As with the outcome variables, binge drinking in adolescence was measured as a binary variable with a positive value indicating 5 or more drinks in one setting more than once a month in the past year. Marijuana, cocaine, inhalant and other drug use was measured as a binary variable with a positive value indicating any use of these substances in the past 30 days. Depressive symptoms are measured by the Center for Epidemiologic Studies Depression scale (CES-D) [31, 32]. Delinquency is a scale of 0-12 of criminal behaviors in the past year [33]. Suicidality is measured by a binary variable with a positive value indicating a suicide attempt in the past year.
Xi = A number of individual and family characteristics included as control variables. These include demographic variables such as gender, age, age squared (to account for non-linearities in age), and race/ethnicity [34]. Race/ethnicity categories are defined as Hispanic or white non-Hispanic, black non-Hispanic, Asian non-Hispanic, or other race, non-Hispanic. The Peabody Picture vocabulary test score is measured on a continuous scale, and is standardized so that the mean is approximately 100. General health is categorized as excellent, very good, good, or fair/poor. Family structure is categorized as two biological parents, single parent household, stepfamily or foster/other family. A binary indicator indicates whether the adolescent is the first born in the family, and whether the adolescent has a biological mother or father with alcoholism.
These equations were analyzed using Stata 10.0. Logit models were analyzed because the outcome variables of interest were dichotomized, and odds ratios are reported in the tables. The research protocol was approved by the University of Chicago Social and Behavioral Sciences Institutional Review Board (#H06238).
Missing Observations
A total of 20,745 students completed the AddHealth in-home interview, and of these, 15,170 followed up at Wave III. There is thus attrition as well as item non-response. Item non-response is particularly problematic for the items asked of parents, parental income and alcoholism. The final sample, with no missing observations, was 9,872. Therefore, sample loss is a concern. This was addressed using re-weighting and imputation measures. Results were qualitatively similar under re-weighting and imputation, and so the results presented here use the AddHealth's weights and original sample to allow for greater replicability. The full detail is provided in Additional file 1.