In many developed countries, alcohol is the most commonly used substance among adolescents. Considerable research has been carried out on the etiology and progression of adolescent alcohol use (Donovan 2004; Fisher et al. 2007; Kirisci et al. 2013; Wills et al. 2001), noting that alcohol involvement has biopsychosocial roots and manifests itself in different forms (e.g., initiation, regular use, problematic use) over time.

Early-onset alcohol use (EOA) is associated with an increased likelihood of engaging in frequent and problematic use of alcohol among adolescents (Fergusson et al. 1994; French and Maclean 2006; Gruber et al. 1996; Hawkins et al. 1997; Hingson et al. 2006; Mason and Spoth 2012; Swahn et al. 2008). Although the defining age of EOA varies across studies (e.g., before age 13, 14, or 15; Kuntsche et al. 2016), there is a general consensus that the younger the youths are when they initiate alcohol, the more likely they are to experience alcohol-related harms (e.g., externalizing problems) later in life (Liang and Chikritzhs 2015; Morean et al. 2014).

Despite a well-established link between EOA and later alcohol harms (Buchmann et al. 2009; DeWit et al. 2000; Hingson et al. 2006), it is not well understood if EOA is itself a direct cause of these harms or whether its effect is indirect through high-frequency drinking. Moreover, few studies have systematically examined predictors that underlie adolescents’ progressive use of alcohol (Stice et al. 1998). Comparative studies of family and peer influences are particularly important in that certain factors in these domains may emerge as salient predictors. For example, parents’ favorable attitudes toward the use of alcohol may be more strongly related to EOA than to heavy and harmful drinking in the mid-adolescent years (van der Vorst et al. 2006). Social risks, such as exposure to siblings or peers who use alcohol, may be less predictive of early-onset use but more predictive of problematic use among older youth (Mason and Spoth 2012; Windle 2000). Additionally, more research is needed to identify protective factors, such as youths’ strong attachments to parents (Oxford et al. 2001; Patock-Peckham and Morgan-Lopez 2010), parental involvement and recognition of positive behavior (Nash et al. 2005), high academic achievement (Bryant et al. 2003), and youths’ bonds to school (Resnick et al. 1997), that emerge from within family and school contexts, reducing the risk of early-onset, frequent, or problematic use of alcohol.

As suggested by social-developmental theories (Catalano and Hawkins 1996; Tarter 2002; Toumbourou and Catalano 2005), risk and protective factors are developmentally ordered and have both proximal and distal influences on alcohol use behaviors. Individual cognitions and peer-influence risk factors, such as attitudes favorable to alcohol use and peer drinking, are strong proximal predictors of adolescent alcohol use (Fisher et al. 2007). Importantly, some of the strongest protective factors, although more distal to alcohol involvement than attitudinal and peer-related risk factors, emerge from within the family (e.g., parental discipline and family bonding; Nash et al. 2005; Patock-Peckham and Morgan-Lopez 2010) and school contexts (e.g., academic commitment and achievement; Bryant et al. 2003; Hawkins et al. 1997). Thus, social-developmental theories suggest a progression leading from background factors representing contextual family and school influences to more proximal attitudinal and peer-related risk factors for alcohol involvement, including EOA as well as a more frequent and problematic pattern of use. However, there is a need to fully test this hypothesized progression over time to provide information useful for understanding when and how to intervene.

Although debate exists (Kuntsche et al. 2016), the public health burden of EOA underscores the need to implement intervention programs that delay the onset of drinking (DeWit et al. 2000; Ellickson et al. 2003). Prior research has demonstrated that substance misuse preventive interventions can be effective in slowing the rate of alcohol and other substance use initiation in youth (Park et al. 2000; Trudeau et al. 2003). Yet, little is known about whether the same risk and protective factors associated with EOA and later problem drinking found in the USA are applicable to other countries that might have different alcohol-related cultures or policies. This knowledge is essential for understanding the degree to which evidenced-based prevention programs that address these risk and protective factors to prevent EOA and alcohol use might have comparable effects in the USA and other countries.

The current study seeks to examine the progression of adolescent alcohol use from EOA to heavy drinking and alcohol harm through frequent alcohol use. To address the question of whether the same risk and protective factors are associated with different dimensions of alcohol use, we include a range of individual and social influence variables, both proximal and distal, that have been documented as correlates and predictors of adolescent substance use (Beyers et al. 2004; Hawkins et al. 1992; Hemphill et al. 2011; Mason et al. 2011). Based on social-developmental theory and existing research (Bahr et al. 1995; Nash et al. 2005; Oxford et al. 2001), we hypothesized that family and school protection indirectly influence early and more serious alcohol use by influencing youths’ choice of friends and attitudes toward alcohol use and antisocial behavior.

Given that the data for this study are part of a large international investigation that focuses on risks and behaviors of youth in Washington (WA) State, USA and Victoria (VIC), Australia, the present study also seeks to contribute to extending knowledge regarding the generalizability of direct and indirect effects of EOA on adolescent heavy and harmful drinking through frequent use and associated risk and protective factors across two states. Findings from this study could contribute to knowledge of cross-national applicability of prevention programs that address the same risk and protective factors to reduce the overall rates of alcohol use and problems for youth in WA, USA and VIC, Australia, which have different alcohol-related norms and policies to guide the prevention and control of alcohol use.

Methods

Study Procedures and Sample

Data were collected through the International Youth Development Study (IYDS), a binational study of youth development in WA, USA and VIC, Australia. The study uses the Communities That Care (CTC) Youth Survey, an instrument with good reliability and validity for multiple demographic groups in the USA (Glaser et al. 2005). Items of the CTC survey were cognitively pretested and pilot tested in both states prior to being finalized. The larger study used matched sampling, recruitment, and survey administration procedures to ensure the comparability of the data collected (see McMorris et al. 2007 for details on the study design).

Data for the current analyses are from the grade 7 cohort of the IYDS, which includes 961 students in WA (78.4 % of those eligible) and 984 students in VIC (75.6 % of those eligible) who participated in three annual survey administrations from 2002 to 2004. Surveys were group administered each year in classrooms from February to June in WA and from May to October in VIC to maintain seasonal equivalence. Retention rates for two consecutive follow-up years were 98 % in both states. Consistent with prior analyses (McMorris et al. 2011), tests of selective attrition indicated that attrited versus retained students were somewhat more likely to be from Victoria, to be slightly older, and to be from slightly lower income levels (results available on request). Study protocols were approved and are in compliance with the University of Washington Human Subjects Review Committee and the Royal Children’s Hospital Ethics in Human Research Committee in Melbourne, Australia.

The analysis sample consists of 1833 students, excluding 112 students classified as dishonest (reported use of a fictional drug or had improbably excessive illicit drug use—cumulative use >120 times in past 30 days) or who had not responded to the “honesty” questions across all three data collection periods. This sample is composed primarily of 13-year-olds in the first year of the study (WA M = 13.1, SD = 0.4; VIC M = 13.0, SD = 0.4) and is roughly gender balanced (49 % male and 51 % female overall and in each state sample). In WA, 65 % of students described themselves as white, 16 % as Latino(a), 6 % as Asian/Pacific Islander, 6 % as Native American, 4 % as African-American, and 3 % reported belonging to other ethnic groups. In VIC, the majority of students described themselves as Australian (91 %), 6 % as Asian/Pacific Islander, 1 % as Aboriginal or Torres Strait Islander, less than 1 % each as African or Spanish, and 1 % reported belonging to other ethnic groups. Because Australians have a tendency to identify as Australians if they were not foreign born regardless of race, racial/ethnic categories are not directly comparable across the states; thus, the present analyses do not control for race/ethnicity.

Measures

Alcohol Involvement Variables.

EOA was measured at age 13 by asking students how old they were when they first had more than just a sip or two of an alcoholic beverage. Response options were coded based on a five-point scale (from 0 “never had by age 13” to 4 “age 10 or under” so that earlier initiation is a higher score). Frequent alcohol use was assessed at age 14 by asking students on how many occasions they had more than just a sip or two of an alcoholic beverage in the past 30 days. Heavy drinking was measured at age 15 by asking students how many times they had five or more alcoholic drinks in a row in the past 2 weeks. Although these measures of frequent alcohol use and heavy drinking originally used an eight-point response option from 0 “never” to 7 “40 or more times,” responses were recorded to reduce skewness. The recoded items were scored on a four-point scale, ranging from 0 “never” to 3 “6 or more times.” Alcohol harm, measured at age 15, focused on eight harmful consequences of drinking, scored on a frequency scale of never to 40 or more times over the past year (Hibbert et al. 1996). Consequences included loss of control (e.g., “not able to stop drinking once you had started”) and social conflict (e.g., “become violent and get into a fight”). Because few responses exceeded a frequency of three or more times, responses were dichotomized to indicate 1 “ever experienced a particular alcohol-related harm in the past year” versus 0 never. Items were summed to produce a count of the number of harmful alcohol consequences experienced in the past year, ranging from 0 to 6 or more consequences.

Individual and Social Influences and Other Covariates.

Measures of the 15 individual and social influences assessed at age 13 are summarized in Table 1. For all measures, higher scores indicate more of the individual and social influences, as labeled. Except for information on the household, which is based on parent reports, variables in the model are based on youth reports.

Table 1 Construct information and standardized factor loadings from the unconstrained confirmatory factor analysis
Full size table

Analysis Strategy

There were three steps in the analyses. First, we examined differences in the prevalence and mean levels of alcohol use for students in WA and VIC. Mean differences were compared using t tests, and effect sizes were calculated with pooled standard deviations (Cohen 1988). Second, we used exploratory factor analysis (EFA) to identify a set of latent factors that combine the individual and social risk and protective variables. The GEOMIN oblique rotation was used to handle variables loaded on more than one latent factor (Browne 2001), and the maximum likelihood method was used for factor extraction (Fabrigar et al. 1999). Parallel analysis was performed to determine the optimum number of factors to retain from randomly generated correlation matrixes (Hayton et al. 2004), which provides more accurate numbers of components to retain than does Kaiser’s (1960) eigenvalue greater than 1 rule (Zwick and Velicer 1986). In the third step, multiple-group confirmatory factor analysis (MGCFA) and multiple-group structural equation modeling (MGSEM) were conducted to examine correlations among the variables and to test hypothesized structural paths.

The EFA, MGCFA, and MGSEM analyses were conducted using Mplus version 7.11 (Muthén and Muthén 1998–2010). In order to accommodate the modeling of ordered categorical dependent variables (e.g., EOA, frequent alcohol use), we used the weighed least squares mean- and variance-adjusted (WLSMV) estimator. As a robustness check, the primary analyses were re-run treating the dependent variables as continuous under maximum likelihood robust (MLR) estimation. In that the substantive findings were highly similar across the two approaches, we present only results for the WLSMV approach below. Model fit was assessed using the mean- and variance-adjusted chi-square statistic (Muthén et al. 1997), root mean square error of approximation index (RMSEA; Browne and Cudeck 1993), and the comparative fit index (CFI; Browne and Cudeck 1993). WLSMV estimation in Mplus implements a pairwise missing data strategy known to perform well when the data are missing at random after taking exogenous covariates into account. Because the WLSMV estimator was used, differences in the fit of nested models were estimated based on mean- and variance-adjusted chi-square statistics and degrees of freedom using the Mplus Difftest command (Satorra 2000). Tests of mediation hypotheses are based on an estimation of indirect effects generated with the Mplus model indirect command (Muthén and Muthén 1998–2010).

Results

Differences in Prevalence and Levels of Alcohol Involvement

Table 2 provides differences in prevalence and means, as well as Cohen’s effect sizes (d; Cohen 1988) for alcohol involvement variables for youth in both state samples. For all types of alcohol use, students in VIC showed higher prevalence and mean levels relative to those in WA.

Table 2 Differences in prevalence rates and means in alcohol involvement variables between Washington State and Victoria
Full size table

Exploratory Factor Analysis

The EFA produced four latent factors with an eigenvalue of 1.16 from the sample data and 1.08 from the parallel analysis (the five-factor model eigenvalues were 0.84 for the sample data and 1.05 for the parallel analysis). Table 3 shows the results of the EFA for the four-factor solution, including rotated factor loadings of the measured variables. The latent factors are labeled to correspond to the items grouped in each factor: family protection, school protection, favorable attitudes toward problem behavior, and exposure to sibling and peer problem behavior. These four latent factors are used for subsequent analyses to test the model in Fig. 1.

Table 3 Oblique GEOMIN-rotated factor loadings from exploratory factor analysis
Full size table
Fig. 1
figure 1

Hypothesized model

Full size image

Multiple-Group Structural Equation Modeling

Multiple-Group Confirmatory Factor Analysis

To evaluate the equivalence of the measures, an MGCFA was conducted. The unconstrained model in which factor loadings of the two state samples were allowed to be free (χ 2 (284) = 1109.39, p < .05, CFI = 0.92, RMSEA = 0.06) showed that all factor loadings are significant and of a similar magnitude (Table 1). A subsequent MGCFA in which loadings were constrained equal for the two samples also fit the data adequately (χ 2 (295) = 896.23, p < .05, CFI = 0.94, RMSEA = 0.05). In addition, there was a nonsignificant difference in fit between the constrained and unconstrained CFA models after using a Bonferroni correction to adjust for the multiple constraints tested (0.05/11 = 0.005), Δχ 2 (Δdf) = 25.32 (11), p < .01, and Bonferroni adjusted p = .08. Thus, we constrained all of the factor loadings across two states in subsequent MGSEM analyses. Standardized correlation coefficients among the variables from the constrained MGCFA are presented in Table 4.

Table 4 Standardized correlation coefficients for Washington State and Victoria from the constrained multiple-group confirmatory factor analysis
Full size table

Multiple-Group Structural Model

To test the cross-state equivalence in structural paths shown in Fig. 1 (gender and SES were included as controls, although not presented in Fig. 1), we first estimated an unconstrained MGSEM in which all structural paths were freely estimated. That model fit the data adequately (χ 2 (295) = 896.23, p < .05, CFI = 0.94, RMSEA = 0.05).

A second test of the model constrained all structural paths to equality. Although the fully constrained model fit the data well (χ 2 (322) = 914.58, p < .05, CFI = 0.94, RMSEA = 0.05), the chi-square difference test indicated a significant group difference (Δχ 2 (Δdf) = 65.96 (27), p < .001, Bonferroni adjusted p = .001). To understand which path contributed to the significant state difference, we used modification indices from the fully constrained model. The cross-equality constraints producing the largest modification index values were on the paths from family protection to exposure to sibling and peer problem behavior and from frequent alcohol use to heavy drinking; thus, we compared the unconstrained model with a partially constrained model in which the cross-state constraints on those two paths were released from the fully constrained model. This comparison yielded a nonsignificant chi-square statistic (Δχ 2 (Δdf) = 32.55 (25), p > .05), resulting in the final model in which all of the paths were fixed to equality except for the two freely estimated paths mentioned above (χ 2 (320) = 882.85, p < .05, CFI = 0.95, RMSEA = 0.04).

Figure 2 shows only statistically significant unstandardized coefficients from the partially constrained model. Standardized coefficients, estimated using group-specific standard deviations of the variables, are also presented in parentheses to assist in understanding the magnitude of an estimated effect within a group. Coefficients for gender and SES are separately presented in Table 5. Findings indicated that the hypothesized relationships are more similar than different for the two states, with two exceptions: the magnitude of the state difference in the association between family protection and the exposure variable was larger for WA relative to VIC, whereas the association between frequent alcohol use and heavy drinking was larger for VIC relative to WA. Together, these paths partially mediated the relationship between EOA and heavy drinking and fully mediated the association between EOA and alcohol harm. Direct effects of EOA on alcohol harm were not statistically significant. Additionally, the standardized indirect effects of EOA on heavy drinking (WA = 0.07, VIC = 0.12) and alcohol harm (WA = 0.10, VIC = 0.10) through frequent use were significant in both states (p < .05).

Fig. 2
figure 2

Statistically significant path coefficients from the final structural equation model for Washington State and Victoria students. WA Washington, VIC Victoria. Note: Unstandardized coefficients are presented with standardized coefficients in parentheses, first for Washington and then for Victoria; path coefficients that were freely estimated across two states are in brackets; analysis sample size for Washington State is 923 and 910 for Victoria; family and school protection, attitudes toward problem behavior, and exposure to risky environment; heavy drinking and alcohol harm variables were freely correlated; all variables were regressed on gender and SES, and results for control variables are presented in Table 5. *p < .05 or better

Full size image
Table 5 Unstandardized and standardized path coefficients on gender and SES for Washington State and Victoria from the final structural equation model
Full size table

Results showed that protection by the family was only indirectly related to alcohol use variables shown in the model. School protection had direct effects on EOA and heavy drinking, after accounting for other variables. In neither state did favorable attitudes toward problem behavior at age 13 predict heavy drinking or alcohol harm directly. The effects of the attitudes variable only indirectly increased the risk of heavy drinking and alcohol harm through EOA and frequent alcohol use (range in standardized indirect paths: 0.07∼0.12, p < .05). For both states, exposure to sibling and peer problem behavior predicted heavy drinking and alcohol harm directly and indirectly by increasing risk of EOA and frequent alcohol use.

Discussion

This study analyzed binational data to examine the developmental progression of adolescent alcohol use and associated risk and protective factors in WA, USA and VIC, Australia. For both state samples, EOA by age 13 increased risk of heavy drinking and alcohol harm at age 15 through frequent alcohol use at age 14. EOA also had a small but statistically significant direct effect on heavy drinking for students in both states. This study also found that attitudes toward problem behavior and exposure to sibling and peer problems had positive influences on EOA as well as frequent drinking. The exposure variable further predicted heavy drinking and alcohol harm over and above EOA and frequent alcohol use among students in both WA and VIC. Family protection showed more distal impacts on alcohol involvement for students in the two states, having direct negative associations with attitudes toward problem behavior and exposure to sibling and peer problem behavior. In addition, school protection had direct negative associations with EOA and heavy drinking for students in both states. Results of the MGSEM suggest that these patterns are largely similar in the two state samples, consistent with previous findings from the larger IYDS study (Mason et al. 2011; McMorris et al. 2011).

Consistent with earlier published research addressing a robust association between EOA and later heavy drinking (DeWit et al. 2000; Liang and Chikritzhs 2015), this study found a persistent direct effect on heavy drinking when controlling for risk and protective factors. This suggests that early alcohol onset directly increases risk for later heavy drinking, perhaps because early-onset drinkers have more time to escalate their alcohol consumption and transition into heavier patterns of use than late-onset drinkers. The effect of frequent alcohol use on heavy drinking was two times greater in VIC. However, we found no direct impact of EOA on alcohol harm, which includes alcohol-related externalizing problem behavior (e.g., fighting) as well as loss of control (e.g., not able to stop drinking). This suggests that those negative consequences of drinking may be more vulnerable to alcohol-favorable attitudes and behavior within family, peer, and school compared to heavy alcohol consumption per se, resulting in no direct association between EOA and later alcohol harm when these environmental influences are held constant.

Consistent with previous research, this study found that contextual family protective factors predict alcohol involvement only indirectly through individual attitudinal and peer exposure variables (Bahr et al. 1995; Nash et al. 2005). Family protection had a stronger negative association with exposure to problem behavior for students in WA compared to those in VIC. For both WA and VIC students, positive influences and experiences within the school context further had direct negative predictive associations with early onset of alcohol and heavy drinking, but not later, more harmful drinking patterns. In contrast, parents’ and students’ favorable attitudes toward problem behavior and sibling and peer substance use predicted not only EOA but also alcohol frequency, and in the case of sibling and peer substance use, heavy drinking and alcohol harm as well. These findings illustrate both common (exposure to substance use models) and unique (e.g., school protection, attitudes toward problem behavior) predictors of the different dimensions of alcohol involvement (e.g., Mason and Spoth 2012), which has implications for understanding what intervention targets to address for specific outcomes.

Several study limitations are worth noting. First, although we treated risk and protective variables as predictors of EOA, these variables were measured contemporaneously, leaving the possibility of bidirectional effects. Second, this study relies predominantly on youth self-reports. Although, most studies have found these reports to be valid (Johnston et al. 2007), it is possible that the results are influenced somewhat from relying on data from a single source. Third, heavy drinking was measured in the past 2-week time frame, which may underestimate the prevalence of infrequent heavy drinking. However, a recent study found that the measure of past 2-week heavy drinking had about 78 % concordance with heavy drinking in the past year (Cranford et al. 2006). Fourth, generalizability of study results is limited to youths (ages 13–15) of the two states.

The current investigation extends previous findings (Hemphill et al. 2011; Mason et al. 2011; McMorris et al. 2011) by indicating that the direction and magnitude of associations between alcohol involvement variables in the pathways from EOA to heavy and harmful drinking and the risk and protective factor variables related to different dimensions of alcohol use were largely the same for youths in the USA (WA) and Australia (VIC). Differences were found between the two states in the magnitude but not the statistical significance of paths between EOA and frequent alcohol use and between family protection and problem behavior exposure. Thus, despite country context differences, students with higher protection from the family and school appear less likely to initiate drinking at an early age. Also, students in both countries who are exposed to peers and siblings who use alcohol are at higher risk themselves for drinking earlier and more frequently and for eventually experiencing alcohol-related problems. Therefore, the same prevention programs targeting these protective and risk factors are likely to reduce problems from alcohol whether they are implemented in the USA or in Australia. As noted elsewhere (see Hawkins et al. 1997; Spoth et al. 2009), delaying the age of alcohol use onset may help lower the risk of alcohol problems for most adolescents, particularly in Australia where the impact of EOA on frequency of use is higher.