Advertisement

Socioeconomic Position and Adolescent Trajectories in Smoking, Drinking, and Psychiatric Distress

Open AccessPublished:May 03, 2013DOI:https://doi.org/10.1016/j.jadohealth.2013.02.023

      Abstract

      Purpose

      Smoking, drinking, and psychiatric distress are inter-related and may also be associated with socioeconomic position (SEP). This paper investigates the role of SEP in adolescent development across all three of these outcomes.

      Methods

      Data were self-reported by adolescents in the Twenty-07 Study (N = 1,515) at ages 15, 17, and 18 years. Latent class analysis was used to identify homogeneous subgroups of adolescents with distinct developmental patterns. Associations between developmental patterns and a range of socioeconomic indicators were then tested.

      Results

      Five classes were identified. A Low Risk class had low levels for all outcomes. A High Distress class had persistently high levels of distress, but was otherwise similar to the Low Risk group. A High Drinking class drank alcohol earlier and more heavily but also had higher levels of distress than the Low Risk group. Smokers were grouped in two classes, Early Smokers and Late Smokers, and both also had raised levels of drinking and distress. Early Smokers tended to begin earlier and smoke more heavily than Late Smokers. Relative to the Low Risk class, adolescents in a disadvantaged SEP were more likely to be Early Smokers and somewhat less likely to be in the High Drinking class. SEP was not consistently associated with membership in the High Distress or Late Smokers classes.

      Conclusions

      Associations with SEP are evident in opposing directions or absent depending on the combination and timing of outcomes, suggesting that a disadvantaged SEP is not a simple common cause for all three outcomes.

      Keywords

      Implications and Contribution
      A disadvantaged socioeconomic position is specifically associated with a developmental pattern where smoking begins early and higher levels of drinking and distress follow. Outside of this pattern, drinking and distress appear somewhat more common among more affluent adolescents. Such opposing processes are only apparent when examining these outcomes in combination.
      Smoking and excessive alcohol consumption (hereafter referred to as drinking) are related to psychiatric distress (or symptoms of anxiety and depression) in both adolescent and adult populations. These behaviors and symptoms usually begin in adolescence and continue into adulthood [
      • Kessler R.C.
      • Berglund P.
      • Demler O.
      • et al.
      Lifetime prevalence and age-of-onset distributions of DSM-IV disorders in the national comorbidity survey replication.
      ,
      • Kandel D.B.
      • Logan J.A.
      Patterns of drug use from adolescence to young adulthood: I. Periods of risk for initiation, continued use, and discontinuation.
      ]. Prospective data from adolescents suggest reciprocal relationships with distress leading to smoking and drinking and vice versa [
      • Chaiton M.
      • Cohen J.
      • O'Loughlin J.
      • et al.
      A systematic review of longitudinal studies on the association between depression and smoking in adolescents.
      ,
      • Armstrong T.D.
      • Costello E.J.
      Community studies on adolescent substance use, abuse, or dependence and psychiatric comorbidity.
      ,
      • Mathers M.
      • Toumbourou J.W.
      • Catalano R.F.
      • et al.
      Consequences of youth tobacco use: A review of prospective behavioural studies.
      ]. Alcohol and tobacco may be used as forms of “self-medication” to manage psychiatric distress, and/or the use of these substances may pre-dispose a person to developing psychiatric symptoms, either through the physiological effects of substance use, or via the disruption of social relationships [
      • Mason A.W.
      • Kosterman R.
      • Haggerty K.P.
      • et al.
      Dimensions of adolescent alcohol involvement as predictors of young-adult major depression.
      ,
      • Cerda M.
      • Sagdeo A.
      • Galea S.
      Comorbid forms of psychopathology: Key patterns and future research directions.
      ,
      • Kuntsche E.
      • Knibbe R.
      • Gmel G.
      • et al.
      Who drinks and why? A review of socio-demographic, personality, and contextual issues behind the drinking motives in young people.
      ]. All three outcomes represent important public health problems: all are associated with mortality [
      • Hart C.L.
      • Smith G.D.
      • Hole D.J.
      • et al.
      Alcohol consumption and mortality from all causes, coronary heart disease, and stroke: Results from a prospective cohort study of Scottish men with 21 years of follow up.
      ,
      • Robinson K.L.
      • McBeth J.
      • Macfarlane G.J.
      Psychological distress and premature mortality in the general population: A prospective study.
      ,
      • Jarvis M.J.
      • Wardle J.
      Social patterning of individual health behaviours: The case of cigarette smoking.
      ], smoking and drinking carry risks for chronic disease [
      • Becker U.
      • Deis A.
      • Sorensen T.I.
      • et al.
      Prediction of risk of liver disease by alcohol intake, sex, and age: A prospective population study.
      ,
      • Will J.C.
      • Galuska D.A.
      • Ford E.S.
      • et al.
      Cigarette smoking and diabetes mellitus: Evidence of a positive association from a large prospective cohort study.
      ], and psychiatric distress can be disabling [
      • Eaton W.W.
      • Martins S.S.
      • Nestadt G.
      • et al.
      The burden of mental disorders.
      ], so it is important to understand their development. However, considering the prospective associations among these outcomes, there could be significant benefits to examining development holistically across all three. This may help provide insights as to when secondary prevention efforts might be most effective, and improve understanding of etiology [
      • Cerda M.
      • Sagdeo A.
      • Galea S.
      Comorbid forms of psychopathology: Key patterns and future research directions.
      ], because the processes that lead to one of these outcomes occurring in isolation may be different from those processes that lead to them occurring together [
      • Beard J.
      • Galea S.
      • Vlahov D.
      Longitudinal population-based studies of affective disorders: Where to from here?.
      ].
      One potentially important etiological factor is a person's socioeconomic position (SEP), which could influence each outcome via the stratification of social and economic resources or stressors. If SEP is a common cause then this may explain the associations among these outcomes, though an etiological role of SEP does not exclude further pathways linking the outcomes to each other such as those suggested above. Although adolescents in a disadvantaged SEP are more likely to smoke [
      • Hanson M.D.
      • Chen E.
      Socioeconomic status and health behaviors in adolescence: A review of the literature.
      ] and experience depressed mood [
      • Lemstra M.
      • Neudorf C.
      • D'Arcy C.
      • et al.
      A systematic review of depressed mood and anxiety by SES in youth aged 10–15 years.
      ], studies on SEP and adolescent drinking vary, showing associations in either direction or no relationship at all [
      • Hanson M.D.
      • Chen E.
      Socioeconomic status and health behaviors in adolescence: A review of the literature.
      ]. However, these studies have tended to treat each outcome individually, without accounting for the relationships among them. The role of SEP may be clearer if these outcomes are examined together.
      This paper aims to identify the most common patterns of adolescent development in smoking, drinking, and psychiatric distress and see whether a disadvantaged SEP is associated with all patterns of increased health risk, or only with specific developmental patterns. Latent class analysis [
      ] is employed to identify distinct groups of adolescents with similar patterns of development, and then relate membership in those groups to SEP. SEP is commonly measured using a variety of indicators, but each may emphasize particular characteristics [
      • Galobardes B.
      • Shaw M.
      • Lawlor D.A.
      • et al.
      Indicators of socioeconomic position (part 1).
      ]. A range of SEP measures are employed to assess whether the associations are robust to measurement differences. Gender is also adjusted for as an important adolescent correlate of these outcomes [
      • West P.
      • Sweeting H.
      Fifteen, female and stressed: Changing patterns of psychological distress over time.
      ,
      • Sweeting H.
      • West P.
      Young people's leisure and risk-taking behaviours: Changes in gender patterning in the West of Scotland during the 1990s.
      ].

      Methods

      Sample

      Data are from the Twenty-07 Study based in and around Glasgow in the West of Scotland [
      • Benzeval M.
      • Der G.
      • Ellaway A.
      • et al.
      The West of Scotland Twenty-07 Study: Health in the Community: Cohort Profile.
      ]. People in three age cohorts have been followed for 20 years. This paper involves the youngest cohort, who had a baseline response rate of 85%. Baseline interviews with the respondents and their parents were conducted in 1987 (n = 1,515), a postal survey was conducted approximately 1 year later (n = 1,250), and further follow-up interviews took place in 1990 (n = 1,343). The mean age of the respondents was 15.7, 17.1, and 18.6 years respectively at each of these time-points. Ethical approval was obtained for each wave of data collection from the National Health Service (NHS) and/or Glasgow University Ethics Committees. Written parental consent for respondent's participation was obtained at the start of the baseline interview and from the respondents themselves at the follow-up interview. Regarding the postal survey, consent was indicated by return of the questionnaire. Baseline respondents were representative of the general population of the sampled area [
      ].

      Measures

      Outcomes

      Respondents self-reported each outcome at each measurement point. Regarding smoking and drinking, respondents were asked about their current status and then for further detail on quantity/frequency if they were current smokers/drinkers. For smokers the number of cigarettes smoked daily was obtained (dividing by 7 where respondents had reported weekly amounts). At baseline drinkers reported the frequency of their drinking, while in the two follow-up surveys they reported their drinking in detail over the past 7 days. Psychiatric distress was assessed using the 12-item General Health Questionnaire (GHQ-12) [
      ].
      A four-category measure was constructed for each outcome to cover the range from no use or no symptoms to heavy use or severe symptom levels. Smoking was categorized at each survey into: not currently smoking, smoking fewer than 1-a-day, smoking regularly (1-a-day or more), and smoking heavily (10-a-day or more). At baseline, drinking was categorized according to the available information into: not currently drinking, drinking less than monthly, monthly drinking, and weekly drinking. At the two later surveys, drinking was categorized into: not currently drinking, drinking less than weekly, weekly drinking within UK recommended limits in the past week (14 units for females, 21 for males) [
      Royal College of Physicians,Royal College of Psychiatrists,Royal College of General Practitioners
      Alcohol and the heart in perspective: Sensible limits reaffirmed.
      ], and weekly drinking exceeding recommended limits in past week. Psychiatric distress was categorized using GHQ-12 scores into: no (0), mild (1–2), medium (3–4), and severe symptoms (5+). Across all measures, for convenience, the four categories will be referred to as: none, low, medium, and high.

      Covariates

      Gender was coded 1 for females, 0 for males. All SEP indicators came from the parental interview at baseline, and were based on parental or household characteristics. They are viewed as representing the SEP of the households in which the adolescents were being raised and are thus considered conceptually as antecedent to the outcomes. Household social class was coded according to the UK Registrar General's 1980 classification [
      ], using the higher status occupation from couple parents, dichotomized into manual and nonmanual categories. Lone parenthood differentiated between respondents who had a single parent and those whose parents were married or co-habiting, and is viewed as a marker for socioeconomic disadvantage. Housing tenure dichotomized those in owned or mortgaged accommodation and those in rented or other types of accommodation. Parental education (taking the higher value from couples) separated those with and without education beyond the age of 16 years. Parental employment status was coded in three categories for the most economically active parent in the household: full-time, part-time, or not employed. Parents reported whether their weekly household income after tax was less than £50, £50–99, £100–149, £150–199, £200–249, £250–299, £300–349, £350–399, £400–449, £450–499 or greater than £500. The mid-point of the chosen band was equivalized for household composition [
      • McClements L.D.
      Equivalence scales for children.
      ], and the equivalized household income variable was split into tertiles. Area deprivation was based on Carstairs scores for baseline postcode sectors (average population = 5,000) derived from the closest Census information (1991) [
      • Mcloone P.
      • Boddy F.
      Deprivation and mortality in Scotland, 1981 and 1991.
      ]. Carstairs scores provide an index of deprivation based on proportions of: households in the area that are overcrowded; heads of household in the area who are in social classes IV and V; male heads of household in the area who are unemployed; and households in the area that do not have access to a car. Scores are commonly split into seven groups referred to as deprivation categories. These were further grouped into: least deprived (1–2); middling (3–5); and most deprived (6–7).

      Analysis

      Analyses were performed using Mplus version 7 [
      ] and models were estimated using maximum likelihood under the missing-at-random (MAR) assumption (i.e., that missingness is random given the other variables in the model) [
      • Clarke P.
      • Hardy R.
      Methods for handling missing data.
      ]. The analysis proceeded in two stages. First, latent class analysis [
      ] was used to identify patterns of development across the three outcomes over the three measurement points. Latent classes represent the most common and distinct developmental patterns, with each latent class having a profile of response probabilities detailing the likelihood of each outcome at each measurement. The number of latent classes was determined by estimating a series of latent class models each with an incrementally greater number of classes and then comparing these models on the basis of various model-fit statistics. Models with greater interpretative value were chosen where fit-statistics did not point to a single optimal model (see Appendix 1 in the online edition of this article for details). Two respondents were excluded at this stage because they had missing data on all of the outcome variables at all measurements (n = 1,513). Males and females could potentially have exhibited substantively different developmental patterns, so this stage of modeling was also carried out on males and females separately. Similar groupings were identified but at different frequencies (results not shown). Including gender as a predictor of class membership in the next stage of modeling was therefore deemed adequate for capturing gender differences in developmental patterns.
      Associations between SEP and latent class membership were examined in the second stage of modeling. Latent class analysis provides for each respondent the probability of being in each class given their observed responses. A common practice is to assign respondents to the class where they have the highest probability of membership and then treat these modal class assignments as if observed in further analyses. This, however, does not take account of the uncertainty in class membership and therefore tends to underestimate the magnitude of associations with covariates [
      • Vermunt J.K.
      Latent class modeling with covariates: Two improved three-step approaches.
      ]. In order to account for such uncertainty, this paper uses the 3-step modal maximum-likelihood procedure described by Vermunt [
      • Vermunt J.K.
      Latent class modeling with covariates: Two improved three-step approaches.
      ]. This procedure performs well at identifying true relationships between latent class membership and covariates in simulation studies [
      • Vermunt J.K.
      Latent class modeling with covariates: Two improved three-step approaches.
      ,

      Asparouhov T, Muthén BO. Auxiliary variables in mixture modeling: A 3-step approach using Mplus. Available at: http://www.statmodel.com/examples/webnotes/webnote15.pdf.

      ]. Each SEP indicator was included in a separate multinomial regression of latent class membership. All models were adjusted for gender, and interactions between gender and SEP indicators were examined. This stage of modeling used only those respondents with full data on all SEP covariates (n = 1,383), but for consistency the response probability parameters of the latent class model were fixed to those values identified in the previous stage. Modal class assignments for those who were excluded because of missing covariate information did not differ significantly from the class assignments of those who were included (chi-square; p = .12). The analysis was also performed using modal class assignments with similar findings (see Appendix 2 in the online edition of this article for details), except that the odds ratios (ORs) for modal assignments tended to be closer to unity and have smaller standard errors than those from the Vermunt 3-step method.

      Results

      Table 1 shows descriptive statistics for the covariates, and the proportion of those with these baseline characteristics at the two follow-ups. Drop-out was somewhat greater among males and those in a disadvantaged SEP, but these differences were not large.
      Table 1Descriptive statistics for baseline covariates and attrition
      Summary statistics are based on valid responses. Item-missingness was generally lower than 5% except for baseline household income (6.4%, 6%, and 5.8% at ages 15, 17, and 18 years).
      Baseline interview: Age 15Postal follow-up: Age 17Follow-up interview: Age 18
      N (%)1,515(100)1,250(82.5)1,343(88.6)
      N%N%N%
      Baseline characteristics
       Gender
      Male73748.658146.563847.5
      Female77851.466953.570552.5
       Household social class
      Nonmanual89159.876962.382762.4
      Manual59840.246537.749837.6
       Lone parenthood
      Couple parents1,27386.31,07788.11,14387.1
      Single parent20213.714511.917012.9
       Housing tenure
      owned64143.157446.660745.8
      rented84756.965853.471754.2
       Parental education
      Post-1651934.945837.248937.0
      Left by 1696965.177462.883463.0
       Parental employment status
      Full-time1,05971.291174.197573.7
      Part-time1248.3977.91138.5
      Not employed30420.422118.023517.8
       Household income
      Top tertile47133.342536.245035.6
      Mid-tertile47333.438933.142733.8
      Bottom tertile47233.336130.738830.7
       Area deprivation
      Least deprived24216.022117.723317.4
      Middling64842.855044.059244.1
      Most deprived62441.247838.351738.5
      a Summary statistics are based on valid responses. Item-missingness was generally lower than 5% except for baseline household income (6.4%, 6%, and 5.8% at ages 15, 17, and 18 years).
      Table 2 shows the prevalence of different levels of smoking, drinking, and psychiatric distress over the three measurement points. For all three outcomes, changes between ages 15 and 18 years mainly reflected shifts toward higher prevalence and heavier consumption or more severe symptoms.
      Table 2Frequency of smoking, drinking and psychiatric distress at each measurement point
      Summary statistics are based on valid responses. Missingness was generally lower than 5% except for psychiatric distress at baseline (7.1%) and drinking at age 18 (7.4%).
      Baseline interview: Age 15Postal follow-up: Age 17Follow-up interview: Age 18
      N (%)1,515(100)1,250(82.5)1,343(88.6)
      N%N%N%
      Outcomes
       Smoking
      Smoking: None, Light, Medium, and Heavy equate respectively to 0, <1, ≥1, and ≥10 cigarettes daily.
      None1,22581.389572.188165.8
      Low483.2322.6241.8
      Medium17011.317213.91188.8
      High644.214211.431523.5
       Drinking
      Drinking: At baseline, None, Light, Medium, and Heavy equate respectively to no drinking, <monthly, ≥monthly, and ≥weekly. At the two follow-ups, None, Light, Medium, and Heavy equate respectively to no drinking, <weekly, ≥weekly and within limits (14/21 units), and ≥weekly and over limits (14/21 units).
      None17411.521217.01239.9
      Low1,04068.970456.536129.0
      Medium21013.927422.049740.0
      High865.7554.426221.1
       Psychiatric distress
      Psychiatric Distress: None, Light, Medium, and Heavy equate respectively to scores of 0, 1–2, 3–4, and 5 + on the GHQ-12.
      None77855.357346.736728.2
      Low41529.531525.739930.7
      Medium1329.416513.431924.5
      High835.917414.221616.6
      a Summary statistics are based on valid responses. Missingness was generally lower than 5% except for psychiatric distress at baseline (7.1%) and drinking at age 18 (7.4%).
      b Smoking: None, Light, Medium, and Heavy equate respectively to 0, <1, ≥1, and ≥10 cigarettes daily.
      c Drinking: At baseline, None, Light, Medium, and Heavy equate respectively to no drinking, <monthly, ≥monthly, and ≥weekly. At the two follow-ups, None, Light, Medium, and Heavy equate respectively to no drinking, <weekly, ≥weekly and within limits (14/21 units), and ≥weekly and over limits (14/21 units).
      d Psychiatric Distress: None, Light, Medium, and Heavy equate respectively to scores of 0, 1–2, 3–4, and 5 + on the GHQ-12.
      A model with five latent classes was selected as the optimal description of the developmental profiles within the smoking, drinking, and psychiatric distress data (see Appendix 1 online). Figure 1 displays the proportions at each level of smoking, drinking, and psychiatric distress within each of the five latent classes. Class 1 had the healthiest pattern of responses: they had the lowest levels of psychiatric distress, which increased modestly with age; mainly low drinking, with some progressing to medium drinking by age 18; and very little smoking. We label this group Low Risk. Class 2 is labeled High Drinking because they started drinking earlier and many were drinking heavily by age 18. This group contained very few smokers but had higher distress levels than in the Low Risk class. Class 3 is labeled Early Smokers because there were many medium smokers at age 15 years with the majority smoking 10-a-day or more by age 17. Early Smokers also had greater increases with age in both distress and earlier and heavier involvement with drinking than those in the Low Risk class. Class 4 had relatively high levels of distress and a similar drinking pattern to that of the Early Smokers, but tended to take up smoking later and to smoke less than 10-a-day, so they are labeled Late Smokers. In this group the three problems appeared to develop more or less concurrently, whereas smoking tended to precede the development of drinking and distress problems among the Early Smokers. Finally, Class 5 is labeled High Distress because they had persistent and severe psychiatric symptoms across the three surveys, but were otherwise similar to the Low Risk class, with low levels of smoking and drinking. The estimated proportions in each class were as follows: Low Risk (39.8%); High Drinking (20.9%); Early Smokers (21.8%); Late Smokers (8.6%); and High Distress (8.9%).
      Figure thumbnail gr1
      Figure 1Latent class response probability profiles.
      Table 3 shows the odds ratios (OR) for membership in each class relative to the Low Risk class, for gender and SEP. Females were more likely to be in the High Distress and Late Smokers classes and less likely to be in the High Drinking class than males. Four of the seven indicators of a disadvantaged SEP were associated with lower odds of membership in the High Drinking class (p < .05 for housing tenure and area deprivation; p ≤ .1 for social class and income). Associations between most of the other indicators of a disadvantaged SEP and being in the High Drinking class showed trends in the same direction, but did not reach statistical significance. There was also a gender interaction (not shown) such that females with unemployed parents were less likely to be in this group (p < .05). All indicators of a disadvantaged SEP (except those for area deprivation) were associated with increased odds of being Early Smokers. In contrast, all SEP indicators showed a trend toward lower odds of being Late Smokers for those in a disadvantaged SEP, but this only reached statistical significance for area deprivation. For the High Distress class, there were significant associations with SEP in opposite directions for different measures: adolescents from lone parent families were more likely to be in this group and those from more deprived areas were less likely to be in this group. Those whose parents had less education were also somewhat less likely to be in this group (p < .1). However, most of the SEP indicators did not show significant associations with membership in this class. No other interactions between gender and SEP were observed (p < .05).
      Table 3Odds ratios for latent class membership
      All ORs are adjusted for gender except those for gender, which are unadjusted.
      Latent class (ref: low risk)
      High drinkingpEarly smokerspLate smokerspHigh distressp
      OR95% CIOR95% CIOR95% CIOR95% CI
      Males1111
      Females.43.23–.81.008.78.58–1.06.1132.041.02–4.10.0452.941.30–6.65.009
      Nonmanual household1111
      Manual household.58.30–1.11.1001.891.39–2.57<.001.84.43–1.65.606.89.44–1.80.735
      Couple parents1111
      Single parents1.20.52–2.78.6662.041.34–3.11<.001.87.29–2.64.8072.311.08–4.95.032
      Owned home/mortgage1111
      Rented/other home.41.23–.75.0032.381.69–3.34<.001.76.41–1.41.385.92.48–1.73.786
      Parent(s) in school after age 16 years1111
      Parent(s) left school by age 16 years.71.40–1.27.2512.041.43–2.92<.001.63.34–1.15.130.57.30–1.08.086
      Parent(s) in full-time employment1111
      Parent(s) in part-time employment1.23.51–2.97.6481.911.14–3.20.014.50.09–2.87.4371.16.33–4.07.815
      Parent(s) not employed.45.16–1.26.1311.831.28–2.62.001.48.16–1.47.1991.80.89–3.62.101
      Top income tertile1111
      Middle income tertile.64.34–1.22.1741.651.10–2.49.016.65.32–1.32.236.57.24–1.32.186
      Bottom income tertile.50.24–1.05.0662.421.62–3.61<.001.65.30–1.41.2741.01.49–2.08.980
      Least deprived areas1111
      Middling area deprivation.93.43–2.02.8591.18.68–2.04.561.31.15–.61.001.27.11–.66.004
      Most deprived areas.29.10–.80.0171.51.88–2.59.137.19.08–.43<.001.38.17–.83.015
      a All ORs are adjusted for gender except those for gender, which are unadjusted.

      Discussion

      Distinct patterns of adolescent development in smoking, drinking, and psychiatric distress were identified and support previous evidence of inter-relationships between smoking, drinking and psychiatric distress [
      • Chaiton M.
      • Cohen J.
      • O'Loughlin J.
      • et al.
      A systematic review of longitudinal studies on the association between depression and smoking in adolescents.
      ,
      • Armstrong T.D.
      • Costello E.J.
      Community studies on adolescent substance use, abuse, or dependence and psychiatric comorbidity.
      ,
      • Mathers M.
      • Toumbourou J.W.
      • Catalano R.F.
      • et al.
      Consequences of youth tobacco use: A review of prospective behavioural studies.
      ]. A Low Risk class had low levels of smoking and drinking, and low but increasing levels of psychiatric symptoms. Compared with this group, smokers had raised risks for drinking and psychiatric distress, and the majority of smokers were in the Early Smokers class where drinking and distress tended to develop after smoking initiation. This supports previous research showing prospective relationships between adolescent smoking and later problematic alcohol use and mental health problems [
      • Mathers M.
      • Toumbourou J.W.
      • Catalano R.F.
      • et al.
      Consequences of youth tobacco use: A review of prospective behavioural studies.
      ]. On the other hand, patterns where drinking and distress developed without smoking were also relatively common.
      The findings were contrary to what would be expected if SEP were a simple, common cause of these outcomes; the Early Smokers were the only class for which a disadvantaged SEP was associated with a higher likelihood of membership. In the High Drinking and Late Smokers classes, which both included increased risks for drinking and distress, there was either no association with SEP or an association in the opposite direction. For the High Distress class associations with SEP were inconsistent, most showed no effect but some measures showed associations in opposite directions, and thus this probably represents the more specific characteristics of each SEP measure more than SEP in general, suggesting a weak relationship with SEP. Adolescents in more deprived areas stood out as unlikely to be in the Late Smokers and High Distress classes. Both of these classes had high levels of distress, suggesting there may be something particular about more deprived areas (e.g., solidarity, social cohesion) that is protective in terms of distress. On the other hand, this may represent a cultural bias against reporting such symptoms within more deprived areas.
      As smoking in the Early Smokers class tended to precede problems with drinking and distress, it may be that a disadvantaged SEP promotes early uptake of smoking only, and this then acts as a causal factor leading to later problems with drinking and psychiatric distress [
      • Mathers M.
      • Toumbourou J.W.
      • Catalano R.F.
      • et al.
      Consequences of youth tobacco use: A review of prospective behavioural studies.
      ]. This could mean that the obvious benefits of preventing early smoking uptake among disadvantaged adolescents would additionally include beneficial effects on inequalities in distress and drinking. Alternatively, early smoking might not be causal but may instead be a marker for individual psychiatric vulnerability or for particular experiences within a disadvantaged SEP, either of which could then also lead to drinking problems and psychiatric symptoms. Indeed, the findings may represent an interaction between SEP and vulnerability for substance use and distress. Vulnerability in a disadvantaged SEP could lead to the Early Smoking developmental pattern described, while vulnerability in a more advantaged SEP leads into the High Drinking pattern.
      Inconsistent associations between drinking and SEP have previously led some to suggest that two opposing processes link SEP and drinking; that is, a lower SEP is generally associated with poorer health including heavier drinking, while a higher SEP indicates more resources for obtaining alcohol [
      • Wiles N.J.
      • Lingford-Hughes A.
      • Daniel J.
      • et al.
      Socio-economic status in childhood and later alcohol use: A systematic review.
      ]. These opposing processes could also be linked to different motivations for drinking; while some use alcohol to enhance pleasure, others use it as a mechanism for coping with stress [
      • Kuntsche E.
      • Knibbe R.
      • Gmel G.
      • et al.
      Who drinks and why? A review of socio-demographic, personality, and contextual issues behind the drinking motives in young people.
      ,
      • Colder C.R.
      • Campbell R.T.
      • Ruel E.
      • et al.
      A finite mixture model of growth trajectories of adolescent alcohol use: Predictors and consequences.
      ]. The adverse stressors and lack of other coping resources associated with socioeconomic disadvantage could promote coping-motivated drinking, while those of higher SEP have more resources to enable drinking for pleasure. Given that smokers often view smoking as a coping mechanism for dealing with stress [
      • Jarvis M.J.
      • Wardle J.
      Social patterning of individual health behaviours: The case of cigarette smoking.
      ], smoking that begins early and is maintained at increasingly heavier levels across late adolescence, as seen in the Early Smokers class, may be a marker for stress-related processes within a disadvantaged SEP, which may then also promote coping-motivated drinking. If drinking in the High Drinking class represented more pleasure-motivated drinking then this might explain why this pattern was somewhat more likely for those in a more affluent SEP. Alternatively, there may be other processes of socioeconomic disadvantage that promote both early smoking and drinking, such as fewer alternative activities or lower quality parental monitoring [
      • Stock C.
      • Ejstrud B.
      • Vinther-Larsen M.
      • et al.
      Effects of school district factors on alcohol consumption: Results of a multi-level analysis among Danish adolescents.
      ,
      ].
      Opposing processes might also explain why previous research from the Twenty-07 Study has indicated late adolescence as a period of relative equality in psychiatric distress [
      • West P.
      • Macintyre S.
      • Annandale E.
      • et al.
      Social class and health in youth: Findings from the West of Scotland Twenty-07 study.
      ,
      • Green M.J.
      • Benzeval M.
      Ageing, social class, and common mental disorders: Longitudinal evidence from three cohorts in the West of Scotland.
      ]. Adolescents in more affluent areas, for example, may experience anxiety-promoting pressure to do well in education [
      • West P.
      • Sweeting H.
      Fifteen, female and stressed: Changing patterns of psychological distress over time.
      ], while adolescents in disadvantaged circumstances experience other kinds of stress or lower levels of coping resources, leading both to increased psychiatric symptoms and other problems such as early smoking. If adolescent distress in an affluent SEP is associated mainly with education and tends to dissipate thereafter, while adolescent distress in a disadvantaged SEP is prompted by stressful life conditions that persist into adulthood, this may create socioeconomic inequalities in distress that widen with age [
      • Green M.J.
      • Benzeval M.
      Ageing, social class, and common mental disorders: Longitudinal evidence from three cohorts in the West of Scotland.
      ].
      These findings are presented with some caveats. The drinking measurements combined quantity and frequency, which might not have adequately reflected the consumption of those who drank heavily but infrequently, though previous research suggests that only a minority of adolescents drink this way [
      • Colder C.R.
      • Campbell R.T.
      • Ruel E.
      • et al.
      A finite mixture model of growth trajectories of adolescent alcohol use: Predictors and consequences.
      ]. Similarly, the smoking measurements may not have captured heavy smoking that occurred infrequently (i.e., less than weekly). If drop-out was associated with particular response patterns then the prevalence of these patterns may have been somewhat underestimated. With respect to SEP, however, the clearest effects were in relation to the Early Smokers class, many of whom would have been identifiable from the baseline data due to their early smoking. Thus the small differences in drop-out by SEP are unlikely to have greatly influenced the results. Also, the data refer to the specific geographic and temporal context of the West of Scotland in the late 1980s and early 1990s. Different developmental patterns and associations with SEP might be evident in other contexts where outcomes are more or less prevalent. For example, more recent female cohorts from this region have higher prevalence rates for all outcomes [
      • West P.
      • Sweeting H.
      Fifteen, female and stressed: Changing patterns of psychological distress over time.
      ,
      • Sweeting H.
      • West P.
      Young people's leisure and risk-taking behaviours: Changes in gender patterning in the West of Scotland during the 1990s.
      ]. Nevertheless, studies of developmental trajectories for individual outcomes in other contexts have identified broadly similar trajectories to those evident here. For example, U.S. studies have, for the ages studied here, distinguished between early and late onset smoking [
      • Weden M.M.
      • Miles J.N.V.
      Intergenerational relationships between the smoking patterns of a population-representative sample of U.S. mothers and the smoking trajectories of their children.
      ], between light drinking and increasingly heavy drinking [
      • Colder C.R.
      • Campbell R.T.
      • Ruel E.
      • et al.
      A finite mixture model of growth trajectories of adolescent alcohol use: Predictors and consequences.
      ], and among very high, consistently low, or moderate but increasing levels of depressive symptoms [
      • Wickrama T.
      • Wickrama K.A.S.
      Heterogeneity in adolescent depressive symptom trajectories: Implications for young adult's risky lifestyle.
      ]. Our findings replicate most of these patterns, but also indicate how they co-occur, and how SEP is associated with particular combinations of trajectories.
      Examining adolescent development across all three outcomes—smoking, drinking, and psychiatric distress—suggests opposing processes linking drinking and distress to SEP contingent upon early smoking. Such opposing processes could be missed in research that focuses on only one outcome at a time, as the opposition would result in weak or null associations. A key area for further research seems to be in determining whether early smoking makes a causal contribution to later drinking and distress, or is merely a marker for other causal processes related to a disadvantaged SEP. If early smoking is causal, then intervening to prevent smoking in early adolescence may be especially important, whereas if it is a marker for other processes it is important to understand what those processes are so that appropriate interventions can be devised.

      Acknowledgments

      An earlier version of this analysis was presented at the 2012 annual meeting of the Society for Social Medicine. M.G. wrote the first draft of this manuscript, under support from a doctoral training fellowship supplied by the Chief Scientist Office (CSO) of the Scottish Government Health Directorates (DTF/11/16). The CSO had no involvement in the study design, the collection, analysis, or interpretation of data, or the writing of the manuscript. They have approved the decision to submit the manuscript for publication. A.L. is also funded by the CSO ( MC_A540_5TK30 ). H.S. and M.B. are funded by the UK Medical Research Council (MRC) ( MC_A540_5TK10 and MC_A540_5TK50 ). We are grateful to all of the participants in the Twenty-07 Study, and to the survey staff and research nurses who carried it out. The West of Scotland Twenty-07 Study is also funded by the MRC ( MC_A540_53462 ) and the data were originally collected by the MRC/CSO Social and Public Health Sciences Unit. Information on how to apply for access to the data can be found at: http://2007study.sphsu.mrc.ac.uk/.

      Appendix 1. Determining the Number of Latent Classes

      The number of latent classes was determined by estimating a series of latent class models each with an incrementally greater number of latent classes. These models were then compared on the basis of various model-fit statistics, aiming for an optimal balance of fit and parsimony, that is, the lowest number of classes that could adequately describe the data [
      • Muthén B.
      • Muthén L.K.
      Integrating person-centered and variable-centered analyses: Growth mixture modeling with latent trajectory classes.
      ]. The log likelihood, Akaike's Information Criterion (AIC) [
      • Akaike H.
      A new look at the statistical model identification.
      ] and Bayesian Information Criterion (BIC) [
      • Schwarz G.
      Estimating the dimension of a model.
      ] are all measures of how well the model fits the observed data. Higher values for the log likelihood and lower values for the AIC and BIC indicate better fit. The AIC and BIC also both take into account the parsimony of the model (i.e., the number of model parameters being estimated), with the BIC being more stringent in terms of parsimony, taking the sample size into account as well as the number of parameters. Standardized bivariate marginal residuals can also be used to assess model fit with values greater than four representing poor fit [
      • Jöreskog K.G.
      • Moustaki I.
      Factor analysis of ordinal variables: A comparison of three approaches.
      ,
      • Mejlgaard N.
      • Stares S.
      Participation and competence as joint components in a cross-national analysis of scientific citizenship.
      ]. Here the percentage of all standardized bivariate residuals exceeding four is employed as a summary statistic [
      • Mejlgaard N.
      • Stares S.
      Participation and competence as joint components in a cross-national analysis of scientific citizenship.
      ]. As a rule of thumb, fit is considered to be adequate if less than 10% of these residuals exceed 4. Entropy indicates how definitively respondents are being classified into latent classes (values range from 0 to 1 with 1 representing definitive classification) [
      • Ramaswamy V.
      • Desarbo W.S.
      • Reibstein D.J.
      • et al.
      An empirical pooling approach for estimating marketing mix elasticities with PIMS data.
      ]. Because, for example, a 1-class model would perfectly classify respondents, but probably have poor fit, entropy is only really considered in situations where models do not differ much in terms of fit (in which circumstances a more definitive classification is preferred). Identification is the percentage of different sets of starting values that produced the best-fitting model and gives an indication of whether one can be confident that the model represents a global rather than a local maximization of model fit. If 100% of the sets of starting values converge to the same solution, then one can be confident that this represents a global maximum. Where the best-fitting solution is hard to replicate (i.e., it is reproduced in only a small percentage of the sets of starting values) it may well represent a local maximum or the model may not be identified [
      ].
      Because these various model fit statistics often disagree, it is possible that more than one model would appear as a viable candidate for an optimal summary of the data, in which case additional criteria relating to the interpretive value of the latent classes were employed. More parsimonious models, that is, with fewer classes, were preferred a priori. The interpretive value of additional classes was considered to be related to their prevalence and the uniqueness of the response probability profile. Additional classes that only represented a very small proportion of the sample, or which did not have very different response probability profiles from other classes, would not have added much interpretive value and could be accepted as noise within a more restricted classification. In contrast, additional classes that represented a sizeable proportion of the sample and had very distinct response probability profiles would ideally have been included for their extra interpretive value.
      Table A1 shows the model fit statistics for latent class models with two through seven latent classes. Models with additional classes were not considered because it was becoming difficult to replicate the best-fitting solutions (meaning that they could represent local maxima). The BIC had its lowest value at three classes, but values for the log likelihood, AIC, and standardized bivariate marginal residuals continued to improve with higher numbers of classes. The standardized bivariate marginal residuals were only at the threshold of acceptability for the three-class model. Entropy statistics also indicated a preference for higher numbers of classes over the three-class model. Thus it was not immediately clear which model should be considered optimal as some indicators pointed toward a three-class model while others pointed toward models with additional classes.
      In order to resolve this ambiguity the response probability profiles were inspected, starting with the three-class model, and then comparing models with additional classes, until they no longer suggested a meaningful addition to the model. This process led to the selection of the five-class model as the optimal representation of the data. The three-class model indicated patterns similar to the Low Risk, High Drinking, and Early Smokers classes presented in the paper (i.e., the three most prevalent), but there was less differentiation between these classes in terms of psychiatric distress. A four-class model drew out the High Distress pattern, which seemed to be an informative addition, considering that it also resulted in greater differentiation between the distress levels of the other classes. The five-class model differentiated between the Early Smokers and Late Smokers, which seemed to be a theoretically valuable distinction. The six-class model identified a small subgroup (approximately 6% of the sample) within the Low Risk class, who did not really start drinking at all until age 18. Because this group was relatively small and was only clearly differentiated from the Low Risk class in terms of no versus light drinking at younger ages, this was not considered a valuable addition and the five-class model was chosen.

      Appendix 2. Odds Ratios for Modal Assignment Method

      Table A2 shows the Odds Ratios (ORs) for latent class membership based on modal assignment of respondents into classes. The results are similar to those presented using the Vermunt 3-step method [
      • Vermunt J.K.
      Latent class modeling with covariates: Two improved three-step approaches.
      ,

      Asparouhov T, Muthén BO. Auxiliary variables in mixture modeling: A 3-step approach using Mplus. Available at: http://www.statmodel.com/examples/webnotes/webnote15.pdf.

      ] except that the magnitude of associations is somewhat smaller and the confidence intervals somewhat narrower.
      Table A1Model fit statistics for determining number of latent classes
      Number of ClassesLog likelihoodAIC
      AIC = Akaike's Information Criterion.
      BIC
      BIC = Bayesian Information Criterion.
      % of Residuals
      2-way item-by-item standardized residuals.
      >4
      EntropyIdentification
      Identification represents the % of times the best-fitting solution was replicated out of 20 sets of starting values. These 20 sets of starting values were identified by following 250 sets of starting values for 20 iterations and selecting those with the best log likelihood values.
      2−11,763.7823,637.5623,930.2612.868100
      3−11,650.0723,466.1423,907.8510.708100
      4−11,559.1323,340.2623,930.987.750100
      5−11,502.4223,282.8424,022.585.72565
      6−11,449.7623,233.5224,122.273.74315
      7−11,413.1723,216.3524,254.112.7355
      a AIC = Akaike's Information Criterion.
      b BIC = Bayesian Information Criterion.
      c 2-way item-by-item standardized residuals.
      d Identification represents the % of times the best-fitting solution was replicated out of 20 sets of starting values. These 20 sets of starting values were identified by following 250 sets of starting values for 20 iterations and selecting those with the best log likelihood values.
      Table A2Odds ratios for latent class membership using modal assignment
      All ORs are adjusted for gender except those for gender which are unadjusted.
      Latent Class (ref: Low Risk)
      High drinkingpEarly smokerspLate smokerspHigh distressp
      OR95% CIOR95% CIOR95% CIOR95% CI
      Males1111
      Females.68.51–.92.012.83.63–1.10.1931.50.99–2.28.0551.991.30–3.05.002
      Nonmanual household1111
      Manual household.73.53–1.00.0531.851.40–2.44<.001.98.64–1.49.9151.04.69–1.58.834
      Couple parents1111
      Single parents1.19.76–1.88.4441.961.34–2.87.001.99.52–1.91.9861.741.01–3.00.047
      Owned home/mortgage1111
      Rented/other home.62.46–.83.0022.231.65–3.01<.001.92.61–1.39.697.86.58–1.30.480
      Parent(s) in school after age 16 years1111
      Parent(s) left school by age 16 years.79.58–1.08.1421.861.36–2.54<.001.82.54–1.25.353.63.42–.95.029
      Parent(s) in full-time employment1111
      Parent(s) in part-time employment1.06.62–1.82.8321.701.06–2.73.027.78.34–1.79.562.88.38–2.02.756
      Parent(s) not employed.70.46–1.06.0931.751.26–2.44.001.74.42–1.31.3021.45.90–2.33.130
      Top income tertile1111
      Middle income tertile.77.54–1.10.1571.591.11–2.27.012.79.48–1.28.335.67.40–1.11.121
      Bottom income tertile.68.47–.99.0442.261.59–3.22<.001.88.54–1.45.6271.02.63–1.65.930
      Least deprived areas1111
      Middling area deprivation.87.57–1.33.5171.15.72–1.82.565.42.25–.70.001.52.30–.92.024
      Most deprived areas.49.31–.77.0021.45.92–2.29.110.32.19–.56<.001.54.31–.95.032
      a All ORs are adjusted for gender except those for gender which are unadjusted.

      References

        • Kessler R.C.
        • Berglund P.
        • Demler O.
        • et al.
        Lifetime prevalence and age-of-onset distributions of DSM-IV disorders in the national comorbidity survey replication.
        Arch Gen Psychiatry. 2005; 62: 593-602
        • Kandel D.B.
        • Logan J.A.
        Patterns of drug use from adolescence to young adulthood: I. Periods of risk for initiation, continued use, and discontinuation.
        Am J Public Health. 1984; 74: 660-666
        • Chaiton M.
        • Cohen J.
        • O'Loughlin J.
        • et al.
        A systematic review of longitudinal studies on the association between depression and smoking in adolescents.
        BMC Public Health. 2009; 9: 356
        • Armstrong T.D.
        • Costello E.J.
        Community studies on adolescent substance use, abuse, or dependence and psychiatric comorbidity.
        J Consult Clin Psychol. 2002; 70: 1224-1239
        • Mathers M.
        • Toumbourou J.W.
        • Catalano R.F.
        • et al.
        Consequences of youth tobacco use: A review of prospective behavioural studies.
        Addiction. 2006; 101: 948-958
        • Mason A.W.
        • Kosterman R.
        • Haggerty K.P.
        • et al.
        Dimensions of adolescent alcohol involvement as predictors of young-adult major depression.
        J Stud Alcohol Drugs. 2008; 69: 275-285
        • Cerda M.
        • Sagdeo A.
        • Galea S.
        Comorbid forms of psychopathology: Key patterns and future research directions.
        Epidemiol Rev. 2008; 30: 155-177
        • Kuntsche E.
        • Knibbe R.
        • Gmel G.
        • et al.
        Who drinks and why? A review of socio-demographic, personality, and contextual issues behind the drinking motives in young people.
        Addict Behav. 2006; 31: 1844-1857
        • Hart C.L.
        • Smith G.D.
        • Hole D.J.
        • et al.
        Alcohol consumption and mortality from all causes, coronary heart disease, and stroke: Results from a prospective cohort study of Scottish men with 21 years of follow up.
        BMJ. 1999; 318: 1725-1729
        • Robinson K.L.
        • McBeth J.
        • Macfarlane G.J.
        Psychological distress and premature mortality in the general population: A prospective study.
        Ann Epidemiol. 2004; 14: 467-472
        • Jarvis M.J.
        • Wardle J.
        Social patterning of individual health behaviours: The case of cigarette smoking.
        in: Marmot M. Wilkinson R.G. Social Determinants of Health. Oxford University Press, New York2006: 224-237
        • Becker U.
        • Deis A.
        • Sorensen T.I.
        • et al.
        Prediction of risk of liver disease by alcohol intake, sex, and age: A prospective population study.
        Hepatology. 1996; 23: 1025-1029
        • Will J.C.
        • Galuska D.A.
        • Ford E.S.
        • et al.
        Cigarette smoking and diabetes mellitus: Evidence of a positive association from a large prospective cohort study.
        Int J Epidemiol. 2001; 30: 540-546
        • Eaton W.W.
        • Martins S.S.
        • Nestadt G.
        • et al.
        The burden of mental disorders.
        Epidemiol Rev. 2008; 30: 1-14
        • Beard J.
        • Galea S.
        • Vlahov D.
        Longitudinal population-based studies of affective disorders: Where to from here?.
        BMC Psychiatry. 2008; 8: 83
        • Hanson M.D.
        • Chen E.
        Socioeconomic status and health behaviors in adolescence: A review of the literature.
        J Behav Med. 2007; 30: 263-285
        • Lemstra M.
        • Neudorf C.
        • D'Arcy C.
        • et al.
        A systematic review of depressed mood and anxiety by SES in youth aged 10–15 years.
        Can J Public Health. 2008; 99: 125-129
      1. Collins L.M. Lanza S.T. Latent Class and Latent Transition Analysis with Applications in the Social, Behavioral, and Health Sciences. John Wiley & Sons Inc., Hoboken, NJ2010
        • Galobardes B.
        • Shaw M.
        • Lawlor D.A.
        • et al.
        Indicators of socioeconomic position (part 1).
        J Epidemiol Community Health. 2006; 60: 7-12
        • West P.
        • Sweeting H.
        Fifteen, female and stressed: Changing patterns of psychological distress over time.
        J Child Psychol Psychiatry. 2003; 44: 399-411
        • Sweeting H.
        • West P.
        Young people's leisure and risk-taking behaviours: Changes in gender patterning in the West of Scotland during the 1990s.
        J Youth Stud. 2003; 6: 391-412
        • Benzeval M.
        • Der G.
        • Ellaway A.
        • et al.
        The West of Scotland Twenty-07 Study: Health in the Community: Cohort Profile.
        Int J Epidemiol. 2009; 38: 1215-1223
      2. Der G. Report No. 60: A Comparison of the West of Scotland Twenty-07 Study Sample and the 1991 Census SARs. MRC Medical Sociology Unit, Glasgow1998
      3. Goldberg D. Williams P. A User's Guide to the General Health Questionnaire. NFER-Nelson, Windsor1988
        • Royal College of Physicians,
        • Royal College of Psychiatrists,
        • Royal College of General Practitioners
        Alcohol and the heart in perspective: Sensible limits reaffirmed.
        RCP, RCPsych, RCGP, London1995
      4. Office of Population Censuses and Surveys: Classification of Occupations. HMSO, London1980
        • McClements L.D.
        Equivalence scales for children.
        J Public Econ. 1977; 8: 191-210
        • Mcloone P.
        • Boddy F.
        Deprivation and mortality in Scotland, 1981 and 1991.
        BMJ. 1994; 309: 1465-1470
      5. Muthén L.K. Muthén B.O. Mplus User's Guide. 7th ed. Muthén & Muthén, Los Angeles, CA2012
        • Clarke P.
        • Hardy R.
        Methods for handling missing data.
        in: Pickles A. Maughan B. Wadsworth M. Epidemiological Methods in Life Course Research. Oxford University Press, Oxford2007: 157-180
        • Vermunt J.K.
        Latent class modeling with covariates: Two improved three-step approaches.
        Polit Anal. 2010; 18: 450-469
      6. Asparouhov T, Muthén BO. Auxiliary variables in mixture modeling: A 3-step approach using Mplus. Available at: http://www.statmodel.com/examples/webnotes/webnote15.pdf.

        • Wiles N.J.
        • Lingford-Hughes A.
        • Daniel J.
        • et al.
        Socio-economic status in childhood and later alcohol use: A systematic review.
        Addiction. 2007; 102: 1546-1563
        • Colder C.R.
        • Campbell R.T.
        • Ruel E.
        • et al.
        A finite mixture model of growth trajectories of adolescent alcohol use: Predictors and consequences.
        J Consult Clin Psychol. 2002; 70: 976-985
        • Stock C.
        • Ejstrud B.
        • Vinther-Larsen M.
        • et al.
        Effects of school district factors on alcohol consumption: Results of a multi-level analysis among Danish adolescents.
        Eur J Public Health. 2011; 21: 449-455
      7. Hayes L. Smart D. Toumbourou J.W. Parenting Influences on Adolescent Alcohol Use. Australian Institute of Family Studies, Melbourne2004
        • West P.
        • Macintyre S.
        • Annandale E.
        • et al.
        Social class and health in youth: Findings from the West of Scotland Twenty-07 study.
        Soc Sci Med. 1990; 30: 665-673
        • Green M.J.
        • Benzeval M.
        Ageing, social class, and common mental disorders: Longitudinal evidence from three cohorts in the West of Scotland.
        Psychol Med. 2011; 41: 565-574
        • Weden M.M.
        • Miles J.N.V.
        Intergenerational relationships between the smoking patterns of a population-representative sample of U.S. mothers and the smoking trajectories of their children.
        Am J Public Health. 2012; 102: 723-731
        • Wickrama T.
        • Wickrama K.A.S.
        Heterogeneity in adolescent depressive symptom trajectories: Implications for young adult's risky lifestyle.
        J Adolesc Health. 2010; 47: 407-413