ADOLESCENT CIGARETTE USE:
EFFECTS OF RACE/ETHNICITY AND SCHOOL
RACIAL/ETHNIC COMPOSITION

Robert A. Johnson
Abt Associates Inc.

Abstract

Research has shown there exists substantial variability in substance use among U.S. schools but has been unable to separate variability due to differences among schools from variability due to differences among the populations that are served by schools. To disentangle the effects of explanatory variables operating at the levels of individuals, families, and schools, we applied multilevel models (mixed models or hierarchical models) to self-report data on daily cigarette use collected in the eighth and tenth grade panels of the National Educational Longitudinal Study (NELS). The response variable is cigarette initiation (first use), coded "1" if the adolescent initiated daily cigarette use (at least one cigarette per day) during the interval between the baseline NELS interview and the reinterview conducted two years later; and coded "0" if the adolescent did not smoke on a daily basis at either wave. The final model used 18 explanatory variables- including 7 individual, 4 family, and 7 school variables- and incorporated "cross-level interaction effects"- interactions between school and family/individual variables- and variance components gauging differences among schools in the effects of individual and family variables.

This paper focuses on a subset of the results, those pertaining to the effects of racial/ethnic minority status and the interactions between minority status and the racial/ethnic composition of students in the school. The findings support research showing that Asian/Pacific Islander, black, and Hispanic adolescents are substantially less likely than white non-Hispanic adolescents to report cigarette use. In each of the NELS eighth and tenth grade panels, API and black adolescents are less than half as likely- and Hispanics about three-fourths as likely- to initiate daily cigarette use as other adolescents. The findings go beyond previous research by showing that the deterrent effects of minority status on cigarette use are much larger among minority adolescents who attend predominantly minority schools. Especially in the NELS tenth grade panel, where the effect is to reduce cigarette risk by one-half, minority schools appear to reinforce the effect of an individual's minority status, so minority students are at even lower risk of cigarette use if they attend schools with a high percentage of minority students. The results also suggest that minority schools help to mitigate the adverse effects on cigarette risk of two social conditions- fewer than two parents living at home and school dropout- that are prevalent within the largest racial/ethnic minority populations of the U.S.

1. Introduction. Research on adolescent cigarette use has been concerned mainly with the effects of individual and family variables. Many studies document associations between cigarette use and family attachment, school involvement, peer smoking, and other individual and family variables (e.g., Akers and Lee 1996; Ennett and Bauman, 1993). Recent research suggests that cigarette use also varies by type of school (Ennett et al. 1997; Skager and Fisher 1989).

A drawback is the failure to link school with individual and family explanatory variables. To reduce omitted-variables bias, factors operating at the different levels need to be included in the same model. School variables may also condition the effects of variables operating at the individual and family levels. For example, one might expect racial/ethnic differences in cigarette use to be intensified in schools with a high percentage of minority students. If minority adolescents are less likely than others to use cigarettes (Fendrich and Vaughan, 1994⁽¹⁾; Johnson and Larison, 1998), then predominantly minority schools may reinforce this effect by offering a normative climate in opposition to cigarette use, a "cross-level interaction effect" (Bryk and Raudenbush, 1992).

The multilevel modeling approach applied in this paper (Bryk and Raudenbush 1992; Goldstein 1995; Kreft and De Leeuw 1999) has a number of advantages, including variance estimates that take into account the data hierarchy, such as clustering of sample students within schools. For our purposes- and those of public policy research generally- the most important advantage may be that multilevel models can yield consistent estimates of cross-level interaction effects (see the next section). We applied multilevel models to the National Educational Longitudinal Study (NELS) to explore how individual, family, and school characteristics affect adolescent cigarette use. This paper reports our findings about the effects of race/ethnicity and school racial/ethnic composition.⁽²⁾

Adolescent cigarette use is only one of many social outcomes that are relevant to policies affecting the racial/ethnic composition of schools. Attempts to manipulate school composition as a matter of law and policy have a contentious history in the U.S. (Bok, 1996). Supreme Court rulings in the 1960s favored racial integration of the public schools, and the percentage of black students enrolled in predominantly (more than half) minority schools declined from about 77% in 1968 to 64% in 1973. However, school segregation has remained roughly constant since the early 1970s (Orfeld, 1993). Based on NELS, the percentages of API, black, Hispanic, and other eighth graders who were in predominantly minority schools in 1988 were 34%, 64%, 65%, and 40%, respectively. The corresponding percentages of tenth graders in 1990 were 41%, 66%, 68%, and 37%.

The following sections discuss the advantages of multilevel models for policy research, the data and methods of our application, and the results.

2. Multilevel models in policy research. The standard single-level regression model has long been the model of choice in policy-oriented research. Let y_ij be a continuous response measured for the i-th student in the j-th school; x_ij an individual-level explanatory variable measured for the same student; and z_j a school-level variable measured for the j-th school. The model can be written

______________ y_ij = a + b x_ij + c z_j + d (x_ij z_j ) + e_ij , _________________ (1)

where e_ij is student-level error with zero mean; and a, b, c, and d are regression coefficients. The error e_ij represents student-level variables that are not included in the model and that affect y_ij.

It is instructive to write the single-level model as a "pseudo two-level model" by defining a school-specific intercept a_j equal to (a + c z_j ):

______Level 1 (students):____ y_ij = a_j + b x_ij + d (x_ij z_j ) + e_ij

______Level 2 (schools):_____a_j = a + c z_j ___________________________(1')

But (1) is not a true multilevel model because there is no random error at Level 2. The single-level model allows unmeasured variables at the student level, but not at the school level.

The multilevel approach introduces the idea of separate regressions in each school or context:

______Level 1 (students):___y_ij = a_j + b_jx_ij + e_ij

______Level 2 (schools):____a_j = a + c z_j + u_1j

_{________________________}b_j = b + d z_j + u_2j _____________________ (2)

The key property of the level-1 equation is that the regression intercept and slope of y_ij on x_ij each have a subscript "j," which implies that these parameters can vary across schools. The level-1 regressions are linked by a level-2 model, where the regression coefficients of the level-1 model are themselves regressed on the school explanatory variable z_j . As in the single-level model, additional assumptions are needed to estimate the model, the main ones being that the level-1 error (e_ij) and level-2 errors (u_1j and u_2j) are uncorrelated with each other and with the explanatory variables.

For comparison with the single-level model, we can write the two-level model as a single equation by substituting the right-hand-sides of the level-2 equations for a_jand b_j in the level-1 equation:

_______________ y_ij = a + b x_ij + c z_j + d (x_ij z_j ) + (e_ij + u_1j + x_ij u_2j).______(2)

Comparing (1) with (2) shows the only difference is in the assumed error structure.

An important parameter for policy is d, the cross-level interaction. Individuals and families in the U.S. are afforded many legal protections, so schools are the principal lever of drug prevention policy. Cross-level interactions may be critical paths by which school policies can impact individual behavior. Yet, if d is estimated using (1) when the true error is that of (2), the estimate is inconsistent, because the error in (2) is correlated with (x_ij z_j ). Thus, if there exist unmeasured school variables- as there almost surely are- good estimates of cross-level interaction effects might not be possible using a single-level model.

3. Data and methods

a. Sample and data collection design. The longitudinal design of NELS (NCES, 1992) allows us to gauge changes in cigarette use between measurement waves and to control for whether or not respondents used cigarettes at the prior wave. Most research on cigarette use in the U.S. has used cross-sectional rather than longitudinal data, perhaps because the National Household Survey on Drug Abuse (NHSDA) and Monitoring the Future (MTF)-two major surveys designed to measure substance use-are cross-sectional in design. Yet retrospective reporting can bias responses about past drug use obtained from cross-sectional surveys (Fendrich and Vaughn, 1994; Johnson et al. 1998).

We split the NELS longitudinal file into two panels to investigate cigarette use separately among eighth graders in 1988 and tenth graders in 1990. Eighth grade panel members were first interviewed as eighth graders in Fall 1988 (Wave 1) and reinterviewed two years later (Wave 2). Tenth grade panel members were first interviewed as tenth graders in Fall 1990 (Wave 1) and reinterviewed two years later (Wave 2).⁽³⁾ Both panels followed up school drop-outs. About 6.8% of eighth grade panel members and 10.4% of tenth grade panel members dropped out before Wave 2. The adolescent interviews used traditional personal interviewing techniques. Personal interviews of parents and school administrators were also conducted at Wave 1 of each panel.

Both panels are based on a two-stage national probability sample of U.S. students: Stratified random sampling of schools was followed by random sampling of eligible students within schools. In our analysis, the eighth grade panel consists of 17,424 adolescents in 1,014 schools who responded to both interviews. The tenth grade panel consists of 16,542 adolescents in 1,464 schools who responded to both interviews. We used standard NELS weights (NCES, 1992) to adjust for unit nonresponse and unequal selection probabilities. We used techniques described in Pfeffermann et al. (1997)- as implemented in the program MLWIN (www.ioe.ac.uk/mlwin) to incorporate the NELS weights in the multilevel model estimation.

b. Measurement of daily cigarette use. Daily cigarette use at each wave of each panel was measured based on responses to the question "How many cigarettes do you usually smoke in a day?" We collapsed the response categories at each wave to form a binary variable: 1 = One or more cigarettes per day; 0 = Not a daily smoker. Item nonresponse was small, ranging from 2.3% at Wave 1 of the eighth grade panel to 7.1% at Wave 2 of the tenth grade panel. To impute the missing data, we used techniques for multilevel models described by Schafer (1996, 1997).⁽⁴⁾

c. Measurement of race/ethnicity and school racial/ethnic composition. The adolescent's race/ethnicity was measured at Wave 1 using response categories similar to the 1990 Census. This report distinguishes four categories- Asian and Pacific Islander (API), black, Hispanic, and other.⁽⁵⁾ School racial/ethnic composition is based on the Wave-1 school administrator interviews. The administrators were asked about the percentage of students who were non-white or Hispanic. To reduce measurement error, we recoded the responses to form a binary variable, equal to 0 if the minority percentage was less than 50% and equal to 1 otherwise. Missing values were imputed based on the distributions of races and ethnicities reported by sample students in the school.

d. Measurement of other explanatory variables. There are seven additional explanatory variables at the individual and family levels: gender (based on Wave-1 adolescent interviews); family income (Wave 1 parental interviews); two biological parents at home (Wave-1 adolescents)⁽⁶⁾;school dropout (Wave-2 follow up); negative peer relations scale (Wave-1 adolescents- number of affirmative answers to five questions about how school peers viewed the respondent, e.g., as a poor student); school participation scale (Wave-1 adolescents- number out of nine activities, e.g., athletics, music); and parental support scale (Wave-1 adolescents- number of affirmative answers to ten questions about the respondent's parent(s), e.g., whether a parent helped with home work).⁽⁷⁾There are six additional explanatory variables at the school level: region; type of place (central city v. other metro v. nonmetro); type of school (public v. Catholic v. other); school size; average beginning teacher salary; and student-teacher ratio. The first four are from the school sampling frame (NCES, 1992); the last two from administrator interviews.⁽⁸⁾ Descriptive statistics for all variables- means, variances, and intercorrelations- are in Johnson and Hoffmann (1999). Prior to analysis, continuous variables were "centered" by subtracting their means (Bryk and Raudenbush, 1992).

e. Statistical models. This paper presents results based on two multilevel models- called Model 1 and Model 2. Model 1 is a 2-level "variance-components model" with a logit-linked binary response and one fixed covariate. We use Model 1 to underscore the importance of cigarette initiation as a response variable, so it is convenient to present both the model and the results in this section. Let y_ij denote a binary (0-1) response variable indicating daily cigarette use at Wave 2 and let p_ij denote the corresponding probability of using cigarettes daily at Wave 2. Let x_ij denote a binary (0-1)

response variable indicating daily cigarette use at Wave 1. The model is written

____________Level 1 (adolescents):____y_ij = p_ij + e_ij

_________________________________logit(p_ij) = a_j + b x_ij

____________Level 2 (schools):_______a_j = a + u_j ,____________________(3)

where logit(p_ij) = log(p_ij /(1 - p_ij)); "log" denotes the natural logarithm; e_ij is a level-1 random error; and u_j is a level-2 random error. We assume that y_ij is distributed as an extra-Bernoulli variable with mean p_ij, so e_ij has mean 0 and variance SIGMA _e² = k p_ij (1 - p_ij ).⁽⁹⁾ We also assume that u_j is normal with mean 0 and variance _u² and that the level-1 and level-2 errors are independent.⁽¹⁰⁾

Table 1 shows the Model 1 parameter estimates.⁽¹¹⁾ The slope parameter b gauges the dependence of daily cigarette use on daily cigarette use two years earlier. For the eighth grade panel, the estimated b of 2.77 corresponds to an odds-ratio of current relative to past smoking of about exp(2.77) = 16. That is, an eighth grade panel member is about 16 times more likely to be a daily smoker if he (she) was a daily smoker two years ago than if he was not. For the tenth grade panel, the corresponding estimate equals about exp(2.99) = 20. The increase in the odds-ratio may reflect that addiction becomes more severe the longer an individual uses cigarettes. If so, it makes sense to try to prevent adolescents from ever using cigarettes for the first time.⁽¹²⁾ Past cigarette use is such a strong predictor of current use that it might be misleading to include past smokers and nonsmokers in the same model. We examined separate models for cigarette initiation and cessation and found that most explanatory variables interact with past use. NELS data on initiation are more plentiful than data on cessation, so this paper focuses on initiation.

Model 2 uses daily cigarette initiation between Waves 1 and 2 -rather than cigarette use at Wave 2- as the response variable. That is, y_ij equals 1 if the adolescent began daily cigarette use between Waves 1 and 2; and y_ij equals 0 if the adolescent was a daily nonsmoker at both waves. The analysis is restricted to daily nonsmokers at Wave 1, which reduces the sample size from 17,424 to 16,454 in the eighth grade panel and from 16,542 to 13,840 in the tenth grade panel. Model 2 also extends Model 1 by adding individual and family explanatory variables at level 1 and school explanatory variables at level 2. We assume P level-1 explanatory variables, denoted x_pij, p = 1, ..., P; and Q level-2 explanatory variables, denoted w_qj, q = 1...,Q. Model 2 is written:

______Level 1 (adolescents):____y_{ij ___}=_ _ij + e_ij

___________________________logit(p_ij) = a_j + SUM_p b_pj x_pij

______Level 2 (schools):____a_j = a + SUM_q c_0q w_qj + u_0j

________________________b_1j = b₁ + SUM_q c_1q w_qj + u_1j

_____________________________ . . .

________________________b_Pj = b_P + SUM_q c_Pq w_qj + u_{Pj ,} __________(4)

where the summations extend from p = 1 to p = P at Level 1 and from q = 1 to q = Q at Level 2.

The first level of (4) is similar to (3), except that p_ij - the probability of initiation- depends upon a school-specific intercept- a_j - and upon school-specific slopes- b_1j through b_Pj. In the (P + 1) level-2 equations, the level-1 regression intercept and slopes are themselves treated as response variables. Each is regressed on Q school-level explanatory variables. For example, in the equation for b_1j, b₁ is the average across schools of the slope of logit(p_ij) on x_1ij; c₁₁ is the effect on b_1j of a unit increase in w_1j; and u_1j is the level-2 random error associated with b_1j.⁽¹³⁾

Substituting the right-hand-side of each level-2 equation of (4) into Level 1 expresses logit(p_ij) as a function of the x_pij's, the w_qj's, and their products- the x_pijw_qj's: