Sample
Between August 1992 and January 2008 we conducted a nine-wave cohort study of health in young people living in the state of Victoria, Australia. The Ethics in Human Research Committee of the Royal Melbourne Children’s Hospital approved all data collection protocols. Informed written parental consent was obtained before inviting students to participate.
A two-stage cluster sampling procedure, first by school and then by class, was used to select the study population with class recruitment at two entry points. At stage one, 45 schools were randomly chosen from a stratified frame of government, Catholic and independent private schools with the probability of selection in each stratum proportional to the number of students of that age in each school. At stage two of the sampling procedure there were two entry points, with one intact class selected from each participating school in wave 1, and a second class selected six months later (wave 2). These two classes from each participating school then formed the basis of the cohort which was followed up during adolescence a further 4 times at six monthly intervals (waves 3 to 6) and during young adulthood 3 times (waves 7 to 9) (Fig. 1). After wave 1, one school (n = 13 participants) declined continued participation, leaving 44 study schools.
The analysis in this paper is mainly based on data collected during the adolescent waves (waves 1 to 6), with adult wave data (waves 7 and 8) only used to identify parental divorce and separation occurring during adolescence that was reported retrospectively by study participants in adulthood. From a total sample of 2032 invited participants, 1943 (96 %) participated at least once during the six adolescent waves. Seventy six invited participants were either refused consent by their parents or were never available for interview (Fig. 1). Participants self-administered the questionnaire on laptop computers, with telephone follow-up for those who were absent from school after wave 2. The seventh to ninth waves were undertaken using computer-assisted telephone interviews.
Alcohol measures (at each wave)
Drinking contexts
In Australia, the legal age for purchase of alcohol is 18 years. Consumption of alcohol by those under 18 years is illegal, unless under the direct supervision of a parent or another adult acting in loco parentis. All participants except those who said they were a non-drinker were asked where they drank and whether they did so repeatedly in the last 6 months (never, once or twice, less than monthly, or more than monthly). The drinking contexts were: at home with family, at home alone, at a party with friends, at a pub or place to eat, in a park or on the street or at the beach, in a car, or at a nightclub.
Within each wave of data collection and for each drinking context, we identified adolescents who (a) never drank in that setting, and for those who did so, we categorised frequency of drinking in that context as occurring (b) 1–2 times for those reporting “once or twice” or (c) 3+ times for those reporting “less than monthly” or “more than monthly”. We then combined, firstly, “in a park or on the street or at the beach” with “in a car” into a single measure termed “in a park/car”, and, secondly, “at a pub or place to eat” with “at a nightclub” into a single measure “in a bar/club”. For the combined categories, drinking was categorised as 3+ times if adolescents reported drinking “less than monthly” in both contexts or “more than monthly” in at least one context. In early adolescence (waves 1 & 2), for each context we identified those adolescents who were drinking 3+ times at one or both waves, and termed this “repeated” use in that context in early adolescence; and identified those with non-repeated use as non-drinkers or those who were drinking less than 3 times in that context at both waves.
Alcohol consumption
Participants who reported that they drank alcohol in the week prior to interview were asked to complete a beverage- and quantity-specific one-week alcohol diary (for more details see [17]). We calculated the number of alcohol units (1 unit = 10 g of alcohol) consumed each day of the diary week. Risky drinking was defined as having drunk 5 or more units on at least one day during the week prior to survey. Very risky drinking was defined as having drunk >20 units for males and >11 units for females on any day over the diary week [16, 18]. Incident risky and very risky drinking were identified in late adolescence.
Time-varying measures assessed in each wave
Daily tobacco use
We identified participants using tobacco daily.
Regular cannabis use
We identified participants using cannabis ≥ weekly.
Antisocial behaviour was assessed using 10 items from the Moffitt and Silva self-report early delinquency scale [19], assessing property damage, interpersonal conflict and theft. Items concerning alcohol use or other substance use were not included. Participants were asked if they had engaged in any of these behaviours never, once, or more than once in the last 6 months. At each wave, antisocial behaviour was identified in participants who reported multiple behaviours at least once or one behaviour more than once.
Mental health was assessed using the Clinical Interview Schedule (CIS-R) [20], which assesses symptoms of depression and anxiety in non-clinical populations. We identified those participants with a score >11 as having a mixed depression-anxiety state for which clinical intervention would be appropriate.
Background measures
Geographic location
Rural and urban areas were defined using the location of the school at study inception.
Frequency of parental drinking and smoking was assessed in the course of the study. Regular parental smoking and drinking was defined as at least one parent engaging in these behaviours on most or every day.
Parental divorce or separation in adolescence (by wave 6) was identified in the course of the adolescent surveys or from responses to enquiry about parental marital status in later waves if the adolescent was absent at wave 6.
Statistical analysis
Prevalence and incidence of risky drinking and very risky drinking were estimated overall and separately for males and females at each wave within a multiple imputation framework (see below). The frequency of drinking in each of the drinking contexts was estimated for males and females in early adolescence (waves 1 & 2). The association between drinking with family at home and drinking in each other context in early adolescence was estimated using odds ratios.
The association between adolescent drinking contexts measured at waves 1 and 2 and incident risky drinking in late adolescence (waves 3–6) was assessed using repeated measures discrete time proportional hazards models [21]. Adjustment was initially for (i) wave of observation and then for (ii) background factors (sex, school location (urban or rural) and parental divorce/separation, parental frequent alcohol use and parental cigarette smoking) and time varying adolescent measures at the previous wave (daily cigarette smoking, weekly + cannabis use, antisocial behaviour and symptoms of anxiety and depression). Initially, the effect of drinking with family and each other drinking context (i.e. home alone, at pub/club, at a party, in a park/car) was estimated separately and then jointly to assess the effect of each drinking context on incident risky drinking. To estimate the effect of joint drinking locations, we generated new four level variables for each other context: not drinking in other location and not drinking at home with family (baseline category); not drinking in other location and drinking with family; drinking in other location and not drinking with family; drinking in other location and with family.
In these models, wave of observation was entered as three dummy variables with wave 3 as reference category in order to avoid constraining the effect of wave/time to be linear. Interactions between sex and drinking context, and sex and wave were assessed in the fully adjusted separate models. There was no evidence of interactions between any of these variables so they were not retained in the final models. All main effects and interactions were tested for statistical significance using Wald tests . All data analysis was conducted using Stata 13 [22].
Multiple imputation [23] was used to handle missing data. We imputed 20 complete datasets, separately for males and females, under a multivariate normal model that incorporated all variables used in the analysis and auxiliary variables that were thought to be related to missingness. The imputation model contained 36 key variables used in the analysis and 29 auxiliary variables not used in the analysis but thought to be related to the missingness. Auxiliary variables included in the imputation model were age at wave 2, context of drinking at waves 3–6 and adolescent measures (smoking, cannabis use, antisocial behavior and symptoms of anxiety and depression at waves 1 and 6. Of these 65 variables, five had <10 % missing values, 19 had 10–14.9 % missing and 16 had 15–19.9 % missing, 15 had 20-23 % missing and 10 wave 1 variables had 53-61 % missing (because about half the cohort was not recruited until wave 2). A three level maximum drinking variable was created at each wave: no risky drinking, risky drinking and very risky drinking. These variables and context of drinking variables were log transformed before imputation. All other variables were not transformed before imputation.
After imputation, any transformed variables were converted back to their original scale and categorised for analysis, with adaptive rounding for binary measures [24]. Simple diagnostics were used [25] to assess if the imputed datasets were reasonable. The imputed distributions were similar to the distributions of observed values for all variables. All estimates were obtained by averaging results across the twenty imputed datasets with inferences under multiple imputation obtained using Rubin’s rules [23].