The Relation of Change in Hostility and Sociability During Childhood to Substance Use in Mid Adolescence

Abstract

In a cohort-sequential longitudinal study (N = 1,075), we related change in children's hostility and sociability assessed from 1^st-8^th grade to their use of cigarettes, alcohol, and marijuana assessed from 9^th-12^th grade. Children who were more hostile at 1^st grade, and had higher rates of growth of hostility, used more of all three substances at 9^th grade, and those with higher initial levels of hostility increased their use of cigarettes and marijuana from 9^th to 12^th grade. Children who were more sociable at 1^st grade used more alcohol at 9^th grade. These findings demonstrate the significance of individual differences in the development of personality traits for the prediction of later substance use and have implications for prevention.

Adolescents experiment with substances, most commonly alcohol, cigarettes, and marijuana. Over high school, life-time prevalence for these substances increases such that by 12^th grade, the majority of adolescents (71.9%) have tried alcohol, approximately one half (44.7%) have tried cigarettes, and only somewhat fewer (42.6%) have tried marijuana (Johnston, O'Malley, Bachman, & Schulenberg, 2009). For some, experimentation will lead to regular use or abuse with damaging consequences for their later health and well-being (Newcomb & Bentler, 1988). Therefore, these levels of adolescent substance use are a cause for concern and call for the development of substance-use prevention programs based on etiological research.

Etiological theories view the development of substance use as a person-environment process that unfolds over time involving the integration of cognitive, social, biological, and personality factors (e.g., Lerner, 2002; Stattin & Magnusson, 1996). Cognitive factors include intention and willingness to try substances, social images of substance users, and beliefs about the prevalence of peers' substance use. These beliefs develop prior to actual substance use and increase its likelihood (Andrews, Hampson, Barckley, Gerrard, & Gibbons, 2008; Gibbons, Gerard, Blanton, & Russell, 1998; Kandel, 1996). Social factors include negative peer influence, parental, school, and neighborhood factors (Andrews & Hops, in press; Ary, Duncan, Duncan, & Hops, 1999; Brook, Whiteman, Nomura, Gordon, & Cohen, 1988; Dishion, 1990; Eccles, Lord, & Buchanan, 1996). At the biological level, early pubertal maturation is a risk factor for substance use, partially through its relationship with increased stress and affiliation with older peers (Tschann et al., 1994; Westling, Andrews, Hampson, & Peterson, 2008).

Compared to such cognitive, social, and biological factors, childhood personality traits have received less attention but are potentially of particular importance. Personality variables are usually included in models of substance use as distal factors that have direct and indirect effects on the development of substance use (e.g., Jessor & Jessor, 1977). Personality traits affect how children experience, interpret, and respond to developmental tasks including ones that have implications for substance use, such as relationship formation and school engagement (Caspi, Roberts, & Shiner, 2005). Traits develop simultaneously with cognitive, social, and biological factors that affect susceptibility to substance use, and they both influence these factors and are influenced by them (Roberts, Wood, & Caspi, 2008). Hence, individual differences in trait development are expected to predict subsequent substance use. In the present study, we focused on the development of hostility and sociability from childhood to early adolescence (1^st through 8^th grade), and we related individual differences in the development of these traits to growth in smoking, alcohol, and marijuana use in high school (9^th through 12^th grade).

Childhood Traits of Hostility and Sociability as Predictors of Substance Use

Early temperament forms the basis for later personality traits. Temperament dimensions include positive emotionality, negative emotionality, impulsivity, and effortful control (e.g., Ahadi & Rothbart, 1994; Caspi, 1998). The childhood traits of hostility and sociability may be viewed as transitions between temperament and aspects of the Big Five personality dimensions in adulthood (Caspi et al., 2005). Hostility reflects negative emotionality and poor anger regulation, and relates primarily to Big Five low agreeableness, unconscientiousness, and emotional instability. Sociability reflects positive emotionality and approach temperament, and relates primarily to Big Five extraversion. Both traits have been associated with substance use in adolescents and young adults (Shedler & Block, 1990).

There are strong parallels between hostility and sociability and the behavior patterns of aggression and social withdrawal studied by developmental researchers. Social withdrawal describes the absence of sociability (Rubin, Coplan, & Bowker, 2009), and aggression is a behavioral expression of hostility (Pepler & Rubin, 1991). Teachers' ratings of these internalizing and externalizing behaviors are used to categorize children as having or not having conduct disorders that can be predictive of and comorbid with substance use (Pepler & Rubin, 1991; Rubin et al., 2009; Serbin, Moskowitz, Schwartzman, & Ledingham, 1991). Withdrawal predicts loneliness, depression, and anxiety (Prior, Smart, Sanson, & Oberklaid, 2000; Rubin, Chen, McDougall, Bowker, & McKinnon, 1995), whereas aggression predicts adolescent bullying and delinquency (Moffitt, 1993; Pepler et al., 2006).

The trait approach taken in the present study views these observable behavior patterns as indicators of underlying personality traits that all children possess to varying degrees (Shiner, 1998). Supporting the trait approach, latent constructs indicated by teacher ratings of elementary and middle school children on behavioral measures predicted later teacher ratings on dimensions of the Big Five personality dimensions when these same children were in high school (Hampson, Andrews, Barckley, & Peterson, 2007a). Behavior ratings of sociability, popularity, social withdrawal, and activity formed a unidimensional sociability construct that correlated primarily with Big Five extraversion. Ratings of negative influence, overt, and covert aggression formed a unidimensional hostility construct that correlated with Big Five emotional instability, disagreeableness, and unconscientiousness.

Of primary interest to the present study, hostility and sociability have been related to various forms of adolescent substance use. Hostility and related traits such as rebelliousness, anger, and negative affect, have been related to adolescent substance use defined as a combination of one or more substances (Wills, DuHamel, & Vaccaro, 1995). These traits also predict use of specific substances such as cigarette (Burt, Dinh, Peterson, & Sarason, 2000; Forgays, Forgays, Wrzesniewski, & Bonaiuto, 1993; Hampson, Andrews, & Barckley, 2007b; Tyas & Pederson, 1998; Weiss et al., 2005) and alcohol use (Block, Block, & Keyes, 1988; Boyle et al., 1993; Brook, Whiteman, Gordon, & Cohen, 1986; Hampson, Andrews, Barckley, & Severson, 2006; Kellam, Brown, & Fleming, 1982). Such studies indicate that the base level of childhood hostility is a risk factor for substance use. However, it remains to be determined whether increasing levels of hostility over childhood are also a risk factor.

Findings for sociability have been less consistent. Whereas some studies report an association between higher levels of sociability and higher levels of substance use (e.g., Brook et al., 1988; Chassin et al., 1993; Jessor & Jessor, 1977), others have found that less sociability, social competence, or social orientation is related to more substance use (e.g., Tarter, 1988; Tarter, Sambrano, & Dunn, 2002; Wills, Vaccaro, & McNamara, 1994). In these studies, substance use has often been evaluated as a combination of alcohol, tobacco, and marijuana, so it has not been possible to determine whether the association with sociability varies by substance. In our research, we have found that higher levels of sociability predicted intentions to use alcohol (Hampson et al., 2006) but did not predict intentions to smoke cigarettes (Hampson et al., 2007b). Alcohol use is more prevalent and begins earlier than cigarette use, so sociable children may see it as less deviant and as a more acceptable activity than using other substances. It appears that the base level of childhood sociability can serve as a risk factor for some forms of adolescent substance use. It remains to be determined whether changes in sociability over time also put a child at risk for certain kinds of substance use.

The Development of Childhood Hostility and Sociability as Predictors of the Development of Substance Use

Models of the influence of personality traits on substance use have treated childhood traits as stable dispositions. However, what is known about the development of personality traits prior to early adolescence suggests this assumption may not be warranted. Roberts and DelVecchio (2000) meta-analyzed studies assessing differential stability of temperament and personality from birth to age 70+. Rank-order correlations showed a steady increase with age, and were lowest during infancy, middle childhood (age 6-12), and adolescence (age 12-18), suggesting that childhood and adolescence are periods of greatest personality instability. In contrast, a meta-analysis of personality maturation evaluated by mean-level change by Roberts, Walton, and Viechtbauer (2006) suggested that ages 10-20 is a period of relative stability (mean trait levels neither increase nor decrease over time). Reconciling the conclusions from these meta-analyses points to childhood and early adolescence as a period when traits are in flux. There appears to be individual change in trait level over time (resulting in relatively low rank-order stability) in the absence of a consistent maturational trend (resulting in apparently stable mean levels over time).

Individual differences in the rate at which traits develop over time are novel and theoretically interesting for the prediction of substance use. Traits increase or decrease over time through interactions with the environment involving mechanisms such as niche selection, evocation, manipulation, and accumulation (Roberts et al., 2008). Children tend to gravitate to environments compatible with their traits, such as groups composed of others with similar traits, and to leave ones that are incompatible (Caspi et al., 2005; Donohew et al., 1999; Roberts & Robins, 2004). Alternatively, through manipulation, children will shape their environments to increase person/environment fit. For example, children who display hostile behaviors will evoke hostility in others, confirming the child's developing aggressive and distrustful tendencies. A sociable child may create opportunities to interact with others and so become more popular, and see themselves as more sociable. Through accumulation of selected, evoked, and manipulated trait-related experiences, trait tendencies are strengthened or weakened resulting in children becoming, for example, more hostile or more sociable over time. By these processes, we hypothesize that increasing hostility over childhood will increase the likelihood of exposure to and involvement with well-established social and cognitive risk factors for substance use such as negative peer influence, conflictual parent/child relationships, school disengagement, positive social images of substance users, and beliefs that peers are using substances. In contrast, increasing sociability over childhood may increase the likelihood of exposure to factors that develop social investment and protect against deviancy, such as positive peer influence and school engagement (Lodi-Smith & Roberts, 2007). However, as indicated by the inconsistencies in past research, in some contexts (e.g., when combined with high sensation seeking) or for some substances (e.g., ones more commonly used such as alcohol), sociability may result in greater substance use (Wills et al., 1994).

To fully describe change and individual differences in the rate of change over time, at least three data points are required. Multiple assessments allow for the use of Latent Growth Modeling (LGM; Muthén, 1991), which is a technique for studying change over time in a latent construct. It is an extension of structural equation modeling that estimates the mean trajectory for the group as a whole and the variance in individual trajectories. In LGM, a latent growth construct is represented by two parameters: the intercept (initial level) and slope (change over time). The means of the intercept and slope represent the mean trajectory for the sample, and the variances of the intercept and slope represent individual differences in that trajectory. If there is little or no maturational change at the group level, then the mean of the slope will not differ significantly from zero indicating no overall pattern of trait increase or decrease over time for the sample as a whole (i.e., no mean-level change). However, the variance associated with the slope can be significant even when the mean of the slope is not, indicating meaningful individual differences in the rate of trait growth (i.e., low rank-order stability). For prediction purposes, the variance of the intercept and/or slope must be significant. Further advantages of growth modeling with three time points include the evaluation of linear versus quadratic change and, with four time points, the evaluation of cubic change. Growth modeling also allows an estimation of the relation between the intercept and slope of latent constructs (Fraley & Roberts, 2005; Raudenbush & Bryk, 2002; Singer & Willett, 2003).

Studies that have used growth parameters of personality traits to predict outcomes are rare. In an example from research predicting mortality, Mroczek and Spiro (2003, 2007) modeled the growth-curve parameters (i.e., intercept and slope) of neuroticism and extraversion for participants in the Normative Aging Study. The parameters for neuroticism, which was assessed on up to six occasions over a 12-year period, were used to predict survival time in this older, male sample. Men with the combination of higher initial levels and greater increases in neuroticism over time were more likely to die sooner than men without that combination. This study demonstrated the value of individual differences in rate of trait change, assessed by parameters of trait trajectories, as predictors of a consequential outcome. However, to our knowledge, there have been no comparable studies of trait change in childhood and early adolescence in which individual differences in trait maturation have been used as predictors of substance use.

The Present Study

The present study was conducted on a community sample of children participating in an ongoing longitudinal, cohort sequential study of the predictors of substance use, the Oregon Youth Substance Use Project (OYSUP). The annual teacher assessments of hostility and sociability through elementary and middle school enabled the use of LGM to evaluate individual differences in the development of these childhood traits. Growth parameters in these traits were used to predict the use of alcohol, tobacco, and marijuana assessed annually over high school and modeled as latent growth constructs. Because adolescent substance use typically increases over time, it is more appropriate and informative to model it as a latent growth construct than to study it at a single time point (Singer & Willett, 2003). Relatively higher levels and faster rates of growth indicate more problematic substance use, and the identification of predictors of problematic substance use by adolescents is particularly valuable for early prevention interventions. Alcohol, tobacco, and marijuana were modeled separately to allow for differences in predictors among substances.

We expected that hostility would be related to later use of all three substances. It was predicted that those children with higher levels of hostility, and greater increases in hostility over time, would be more likely to be using substances at higher levels, and have greater increases in their substance use over time, during high school. Previous findings for sociability indicated that a clear pattern of results was less likely. Children with higher levels of sociability, and greater increases in sociability over time, may be less likely to be using substances at higher levels and have less increase in their substance use over time, during high school. However, given previous inconsistent findings for sociability, we anticipated that its relation to substance use may differ across substances. Because of somewhat higher rates of illicit substance use among male adolescents (Johnston et al., 2009), and gender difference in our previous findings for the association of hostility to alcohol and cigarettes use (Hampson et al., 2006; Hampson et al., 2007b), differences between boys and girls were examined in all the models.

Method

Design

The OYSUP used a cohort sequential design in which five grade cohorts (1^st through 5^th grade at the first assessment) were assessed annually over four years (T1-T4), until they were in 4^th through 8^th grade. The fifth assessment (T5) took place two years after T4 when participants were in 6^th through 10^th grades. Thereafter, assessments have occurred annually and are ongoing. This design serves as a proxy for a true longitudinal design where a single cohort is assessed multiple times. As shown in Figure 1, by collapsing across cohorts, the cohort sequential design enabled us to study trajectories of hostility and sociability from 1^st to 8^th grade, assessed on eight occasions, and to relate these trajectories to substance-use trajectories from 9^th through 12^th grade assessed on four occasions.

Figure 1.

The cohort sequential design of the Oregon Youth Substance Use Project demonstrating how collapsing across cohorts yielded a longitudinal study of 1^st – 12^th grade.

NOTE: Boxes with broken outlines indicate that there was no assessment for that cohort at that grade. Due to a delay in funding, there was a gap of two years between T4 and T5 resulting in one missing grade assessment for each cohort. The number of participants at each time of assessment (T) is show in parentheses.

Cohort differences were examined using one-way analyses of variance followed by post hoc (Scheffé test) comparisons (Cohort 1 = 1^st graders at T1, Cohort 2 = 2^nd graders at T1, Cohort 3 = 3^rd graders at T1, Cohort 4 = 4^th graders at T1, Cohort 5 = 5^th graders at T1). We examined differences in hostility and sociability as a function of cohort at each grade separately for boys and girls. Cohort differences in hostility were found for boys at both Grade 5, F (3, 404) = 3.86, p < .01, and Grade 8, F (3, 338) = 5.80, p < .001. Post hoc comparisons showed that at Grade 5, Cohort 3 boys had significantly higher hostility than Cohort 2 boys, Cohen's d = .52, and at Grade 8, Cohort 2 boys were significantly less hostile than either Cohort 1, Cohen's d = .55 or Cohort 3 boys, Cohen's d = .52. Cohort differences for hostility were also found for girls at Grade 8, F (3, 369) = 3.67, p < .05. Cohort 3 girls were more hostile than Cohort 5 girls, Cohen's d = .46. There were no significant cohort effects in sociability for either boys or girls. We also examined cohort effects for use of cigarettes, alcohol, and marijuana at 9^th through 12^th grade by gender and found only one effect: At Grade 12 for boys, F (1,177) = 4.43, p < .05, Cohort 5 boys reported significantly more cigarette use than Cohort 4 boys, Cohen's d = .31. Because there were only a few small to medium differences between cohorts, participants were collapsed across cohorts for the analyses.

Participants

The OYSUP initially recruited 1,075 children from 15 elementary schools within one school district in Western Oregon. An average of 215 students in each of 1^st through 5^th grade participated at T1 (50.3% female). The mean age of the sample at T1 was 9.0 years (SD = 1.45) across cohorts. The sample at T1 was 86% Caucasian, 7% Hispanic, 1% Afro-American, and 2% each of Asian/Pacific Islander, American Indian or Alaskan native, and other or mixed race/ethnicity. Approximately 7% of mothers and 11% of fathers had not obtained a high school diploma, and 71% of mothers and 66% of fathers had some form of post-high school education. The schools were located in a predominantly working class community; 40% of the sample was eligible for a free or reduced lunch, an indicator of low family income. At T1, participants were representative of children in the school district in terms of race/ethnicity and participation in the free or reduced lunch program, but the 3^rd and 5^th grade cohorts had slightly higher achievement test scores on reading and math (Andrews, Tildesley, Hops, Duncan, & Severson, 2003). The highest level of attrition (10%) was observed between the fourth and fifth assessments. The main reasons for attrition were parents withholding consent for future participation and inability to locate participants for follow-up.

The sample used in these analyses consisted of 1,074 children (540 girls and 534 boys, 99% of the entire sample) who participated in at least one assessment. The proportion who participated in all eight assessments was 78.1%; in seven, 8.3%; in six, 2.4%; in five, 1.2%; in four, 4.8%; in three, .8%; in two, 1.6%; and, in one, 2.7%.

Of the children who participated in two assessments including T1, 138 (13%) did not participate in the T8 assessment. In comparing children who participated at T1 and T8 with children who did not participate at T8, significantly more boys in the oldest cohort (Cohort 5) did not participate at T8 compared with boys in the other cohorts, X²(4, N = 532) = 10.57, p < .05. There were no differences for girls by cohort for those who participated at both T1 and T8 and those who did not. For hostility, boys' hostility at T1 was significantly higher for those who did not participate at T8 than for those who did, regardless of cohort. There were no differences in sociability as a function of attrition for either boys or girls. For substance use, we found no attrition effects by cohort for boys or girls between those who participated at T5 and T8 and those who did not participate at T8. However, boys' substance use was higher at T5 for those who did not participate at T8 compared with those who did: cigarettes, t(450) = 4.43, p < .001; alcohol, t(452) = 2.93, p <.01; marijuana, t(451) = 3.50, df = 451, p < .01. Girls' use of cigarettes, t(460) = 3.58, p < .001, and marijuana, t(464) =2.51, p < .05, was higher at T5 for those who did not participate at T8 compared to those who did.

Overall, these attrition effects are typical for longitudinal studies of substance use, with older participants, males, and substance users more likely to drop out (Andrews, Hops, & Duncan, 1997; Andrews, Hops, Tildesley, & Li, 2002; Kandel, 1984; Newcomb & Bentler, 1988). Importantly, participants were similar to non-participants in terms of grade, gender, race/ethnicity, and proportion who received a free or reduced lunch.

Measures and Procedures

Hostility and sociability

Teachers rated their students towards the end of the school year on items drawn from behavioral observation scales including the Walker-McConnell Test of Children's Social Skills (Walker & McConnell, 1995), the Harter Social Acceptance subscale of the Perceived Competence Scale for Children (Harter, 1985), the Teacher Report Form (Achenbach, 1991), and the relational aggression, overt aggression, and pro-social subscales of the Children's Social Behavior Scale Teacher Form (Crick, 1996). Teacher's ratings of students' behavior were collected until the student was in the 8^th grade. In 1^st through 7^th grade, core teachers (i.e., ones who spent the most time with their students) were asked for ratings. Most students had only one core teacher. Where they had two (e.g., because of job-sharing), the mean of their ratings was used unless one teacher said they did not know the child “at all well,” in which case only the ratings from the other teacher were used. In 8^th grade, students did not have core teachers so, to increase reliability, two different teachers who taught academic subjects were asked for ratings. As with the younger grades, the mean of their ratings was used unless one of the teachers said they did not know the child at all well, in which case only the other teacher's ratings were used. For 1^st through 7^th grades, mean teacher ratings were used infrequently (ranging from one in 1^st grade to 40 in 5^th grade). In 8^th grade, mean ratings were used in 584 (75%) of cases.

A detailed description of scale development was provided by Hampson et al. (2007a). In an exploratory factor analysis, a two-factor solution produced independent and interpretable factors that accounted for 52% of the variance. The highest loading, univocal marker items were selected for each factor to maximize their orthogonality. The ten indicators for hostility were six items from the Crick relational aggression scale (gossips, isolates, excludes, rejects, threatens other children, tells lies), three items from the Crick overt aggression scales (bullies, hits, fights), and one OYSUP item assessing how often the student exerted a negative influence on friends. The eight indicators for sociability were three items from the Walker-McConnell scale (maintaining conversation, initiating conversation, and extending play or conversation), three items from the Achenbach withdrawn scale (withdrawn, prefers to be alone, and underactive), one item from the Harter scale assessing number of friends, and one OYSUP item assessing popularity. Coefficient alphas for the ten-item hostility scale assessed for each of the eight grade groups ranged from .89-.92 for boys and .88-.92 for girls, and for the eight-item sociability scale ranged from .93-.96 for boys and from .92-.95 for girls.

Substance use

Students completed paper and pencil self-report surveys at school or, if they had moved out of the school district, at Oregon Research Institute or via a phone interview. Participants indicated their level of use of each of cigarettes, alcohol, and marijuana during the past 12 months, 0 = “Never,” to 5 = “Some each day.”

Statistical Approach

We used cohort sequential latent growth modeling (LGM; Muthén, 1991) to test models relating the growth of hostility and sociability across the 1^st through 8^th grade to growth in cigarettes, alcohol and marijuana use across the 9^th through 12^th grade. All models were analyzed with Mplus Version 5.1 (Muthén & Muthén, 1998/2008) with the maximum likelihood method, using the EM algorithm for missing data (Dempster, Laird, & Rubin, 1977).

Initially, we tested separate growth models for hostility and sociability to determine the growth model that best represented the change in each latent construct over the eight grades (e.g. linear vs. quadratic). Within each growth model, the intercept and slope (latent constructs) were based on eight indicators: the assessment of the variable at each of eight grades (1^st-8^th). To test a linear model, factor loadings of the intercept on all indicators were set to 1 and factor loadings of the slope on the eight indicators were set sequentially to 0, 1, 2, 3, 4, 5, 6, and 7. To test a quadratic model, the loadings of the intercept on all indicators were set to 1 and the loadings of the slope were set to equal the square of each of the linear loadings (e.g., 0, 1, 4, 9, etc.). Correlations between the intercepts and slopes of the latent constructs were included in all models if significant. The residuals of variables at the same or adjacent time points were allowed to be correlated if indicated by modification indices. Using similar procedures, we tested the fit of separate growth models for each substance across 9^th – 12^th grades using the factor loadings for the intercepts and slopes described in the linear model above. For all these initial growth models, the inclusion of the quadratic model did not significantly improve the fit. LGM provides mean initial levels (intercepts) and mean rates of change (slopes) based on the growth trajectories for all participants, and the variances for these means that indicate the extent to which there are individual differences in growth. Significant variances, but not necessarily significant means, are required when using intercepts and slopes as predictors or the criterion.

In the next step, we used a parallel process model with the LGMs for both hostility and sociability across elementary and middle school (i.e., 1^st 8^th grades) predicting use of each substance across the high school years (i.e., 9^th-12^th grades). The model included structural paths from the intercepts and slopes of hostility and sociability to the intercepts and slopes of the respective substance. Separate models were evaluated for cigarettes, alcohol, and marijuana.

We examined gender differences in model parameters using multiple-group analysis. Within this analysis, all paths were fixed to be equal across gender and then sequentially freed in a stepwise manner according to the modification indices to test for gender differences. If a structural path was significant (p < .05) for either gender, it was included in the final model.

Three criteria were used to assess goodness of fit of the models (Duncan, Duncan, & Strycker, 2006). The χ² test compares the predicted and actual models: small (i.e., non-significant) values indicate goodness of fit. However, the χ² test is more likely to be significant when samples are large, so additional criteria were also considered: a comparative fit index (CFI) of at least .95, and a root mean square error of approximation (RMSEA) of below .08 (Kline, 2006).

Results

Latent Growth Constructs for Each Trait and Each Substance

The observed mean levels of hostility and sociability at each grade are shown separately for boys and girls across 1^st to 8^th grade in Table 1. The observed mean levels of cigarette, alcohol, and marijuana use in the past 12 months are shown separately for boys and girls across 9^th to 12^th grades in Table 2. Fit indices for the latent growth models for each construct are provided in Table 3, and the means and variances for the intercepts and slopes of each construct separately for boys and girls are shown in Table 4.

Table 1. Means and Standard Deviations of Personality Variables by Gender.

	Boys		Girls
Variable	M	SD	M	SD
Sociability
1st grade	3.02	.79	3.23	.71
2nd grade	3.06	.91	3.19	.78
3rd grade	3.00	.87	3.16	.80
4th grade	3.08	.89	3.26	.74
5th grade	3.04	.91	3.23	.82
6th grade	2.85	.93	3.09	.82
7th grade	2.83	.95	3.11	.83
8th grade	2.87	.77	3.04	.74
Hostility
1st grade	.83	.75	.69	.68
2nd grade	.72	.70	.73	.77
3rd grade	.76	.74	.77	.73
4th grade	.65	.68	.69	.71
5th grade	.66	.66	.70	.74
6th grade	.82	.70	.77	.74
7th grade	.73	.71	.67	.71
8th grade	.71	.63	.69	.66

Note. Sociability and hostility were measured on 5-point scales; higher scores indicate higher trait levels.

Table 2. Means and Standard Deviations of Substance Use Variables by Gender.

	Boys		Girls
Variable	M	SD	M	SD
Cigarettes
9th grade	.44	1.07	.59	1.30
10th grade	.67	1.33	.91	1.61
11th grade	1.01	1.57	1.07	1.75
12th grade	1.39	1.78	1.30	1.86
Alcohol
9th grade	.91	1.15	1.09	1.10
10th grade	1.18	1.27	1.29	1.26
11th grade	1.35	1.23	1.49	1.27
12th grade	1.79	1.38	1.68	1.33
Marijuana
9th grade	.51	1.17	.53	1.12
10th grade	.71	1.35	.78	1.32
11th grade	.92	1.48	.82	.126
12th grade	1.11	1.55	.94	1.30

Note. Cigarette, alcohol, and marijuana use were measured on 6-point scales: 0 = never, 5 = some each day.

Table 3. Fit Indices for Univariate Latent Growth Models.

Factor	χ2	df	CFI	RMSEA	90% CI of RMSEA
Hostility	93.28**	64	.98	.03	.02 - .04
Sociability	102.21**	63	.97	.03	.02 - .05
Cigarettes	19.34*	13	.99	.03	.00 - .06
Alcohol	14.63*	13	.99	.04	.02 - .05
Marijuana	33.97**	13	.98	.06	.03 - .08

Note. N = 1,074 for hostility and sociability; N = 960 for cigarettes, alcohol, and marijuana. CFI = comparative fit index; RMSEA = root mean square of approximation; CI = confidence interval.

^*p < .05.

^**p < .01.

Table 4. Growth Factors for Univariate Latent Growth Models.

	Boys		Girls
Growth Factors	Mean	Variance	Mean	Variance
Hostility
Intercept	.70***	.32***	.70***	.32***
Slope	.002ns	.01***	.002ns	.01***
Sociability
Intercept	3.12***1	.49***2	3.33***	.38***
Slope	-.04***	.004***	-.04***	.004***
Cigarette Use
Intercept	.50***	.96***3	.50***	1.56***
Slope	.28***	.23***	.28***	.23***
Alcohol Use
Intercept	.93***4	.87***	1.07***	.87***
Slope	.22***	.08***	.22***	.08***
Marijuana Use
Intercept	.53***	1.11***	.53***	1.11***
Slope	.17***	.18***	.17***	.18***

Note: These coefficients are unstandardized. Chi-square difference tests showed gender differences:

¹X²_diff (1) = 26.20, p < .001

²X²_diff (1) = 8.01, p < .01

³X²_diff (1) = 17.20, p < .001

⁴X²_diff (1) = 4.16, p < .05.

^*p < .05.

^**p < .01.

^***p < .001.

Hostility

A linear growth model fitted the data well and there were no significant gender differences in the overall growth model. The mean of the intercept was significant, indicating it differed significantly from zero, but the mean of the slope was not, indicating no maturational trend to increase or decrease in hostility across the first eight grades for the sample as a whole. However, the variances of the intercept and slope were significant, indicating individual variability in both initial level and change in hostility across grades. Inspection of the distribution of the individual slopes indicated that 46% were negative (i.e., declined over time, slopes < 0), 2% showed no change (i.e., slopes = 0), and 52% were positive (i.e., increased over time, slopes > 0). There was a negative correlation between the intercept and slope, indicating that children with lower levels of hostility at 1^st grade increased in hostility as they grew older, and it was the same for boys and girls, r = -.57, p < .001.

Sociability

The linear model for sociability fit the data well, and the overall model did not differ significantly between genders. The mean and variance of the intercept (initial level) were significant, and the mean varied by gender, with boys rated as less sociable than girls at 1^st grade (see Table 4). The mean and variance of the slope were significant, with sociability showing modest decreases over time for both boys and girls. Inspection of the distribution of the individual slopes indicated that 85% were negative (i.e., slopes < 0), 1% showed no change (i.e., slope = 0), and 14% were positive (i.e., slopes > 0). The negative correlation between the intercept and the slope was significant and did not differ by gender, indicating that children with higher sociability at 1^st grade decreased their sociability more rapidly across grades, r = -.44, p < .001.

Cigarettes

The linear model for cigarette use fit the data well, and the overall model did not differ significantly between genders. The means and variances of the intercept and slope were all significant. There was one gender difference: the variance of the intercept was significantly greater for girls than boys (see Table 4). The correlation between the intercept and the slope was not significant and did not vary by gender, indicating that the initial level of cigarette use in the 9^th grade was not related to change in use across grades, r = -.07, ns. There was one significant correlation for both boys and girls between the residuals for 10^th and 11^th grade cigarette use that was retained in the final model, r = .46, p < .001.

Alcohol

The linear model for alcohol use fit the data well, and the overall model did not differ significantly between genders. The means and the variances of the intercept and slope were significant and there was only one gender difference: Girls reported slightly higher initial levels of alcohol use than did boys at 9^th grade (see Table 4). The correlation between the intercept and the slope was not significant and did not differ by gender, indicating that children's level of alcohol use in the 9^th grade was not significantly related to change in use across grades, r_is = -.01, ns. The significant residual correlation for boys between alcohol use at 10^th and 11^th grades was retained in the final model, r =.35, p < .001.

Marijuana

The linear model for marijuana use also fit the data well, and the overall model did not differ significantly between genders. The means and variances of the intercept and slope were significant and did not differ by gender. The negative correlation between the intercept and the slope was significantly smaller for boys (r = -.20, p < .05) than girls (r = -.38, p < .001) indicating that girls with lower marijuana use in the 9^th grade increased their use at a faster rate across grades than did boys. There also was one significant correlation for girls between the residuals for marijuana use at 10^th and 11^th grades that was retained in the final model, r =.49, p < .001.

We used multiple sample analysis to examine cohort differences in the means and variances of the intercept and slope within each of the above univariate growth models. Since not all cohorts were assessed at each grade, separate analyses were conducted to examine change across specific grades. We examined cohort differences in the growth parameters of hostility and sociability by comparing cohorts 1 and 2 across 2^nd-8^th grades, cohorts 1 through 3 across 3^rd-8^th grades, cohorts 1 through 4 across 4^th-8^th grades, and all five cohorts were compared across 5^th-8^th grades. For hostility, we found the residual for 8^th grade was significantly different for cohort 4 compared to all other cohorts. In addition, the variation in the slope of hostility for cohort 3 as compared to other cohorts across 5^th-8^th grades was significantly greater. For sociability, for cohort 1 at 5^th grade, the error variance was significantly greater compared to the other four cohorts at 5^th grade.

Since cigarette, alcohol and marijuana use were measured across 9^th-12^th grades, we compared cohorts 2 through 4 for differences in growth parameters across 9^th-10^th grades, cohorts 3 and 4 across 9^th-11^th grades, and cohorts 4 and 5 for 11^th -12^th grades. We found no significant differences as a function of cohort. Minimal differences as a function of cohort provided further support for the use of cohort sequential analyses.

Predicting Substance Use from Latent Growth Constructs of Hostility and Sociability

The next models evaluated our hypotheses about the prospective associations between childhood hostility and sociability, assessed during elementary and middle school, and later substance use assessed during high school. Higher initial level and greater growth of childhood hostility was expected to increase the likelihood of higher initial level and growth of all three substances. Effects for sociability were less confidently predicted, but protective effects for higher childhood sociability were a possibility. These hypotheses were tested in separate models for each substance that estimated paths from the intercept and the slope of hostility and sociability to the intercept and slope of the use of the substance.

The results for each model are shown in Figures 2 – 4. In these figures, and the text, standardized path coefficients representing significant simple regressions are reported. The significance of these path coefficients is evaluated by the critical ratio of the parameter estimate divided by the standard error, which gives the z statistic that is evaluated against the normal distribution (i.e., it must exceed +/− 1.96 to be significant). Where gender differences were significant, the coefficients for both genders are reported (coefficients for girls are in parenthesis). Only significant structural paths are shown. For the sake of simplicity, the modest but significant correlations among the intercepts and slopes of hostility and sociability are not shown in the figures, but they are reported here (there were no gender differences). The intercept of hostility correlated negatively with the intercept of sociability (r = -.21, p < .001) and positively with the slope of sociability (r = .23, p < .01). The remaining correlations between the slope of hostility and the intercept and slope of sociability were not significant and were fixed to zero. These correlations were the same across all three final models. Residuals for sociability and hostility at grade 5 were significantly correlated for both boys and girls (r = -.21, p < .001). For boys, residuals for sociability and hostility at 2^nd grade (r = -.30, p < .01) and 4^th grade (r = -.30, p < .001) were correlated. For girls, the residuals for hostility at 4^th and 5^th grades were significantly correlated (r = .25, p < .001). These residual correlations (not shown in the figures) were the same for all three final models predicting substance use.

Figure 2.

Hostility and sociability in elementary and middle school predicting cigarette use in high school.

NOTE: Hostility and sociability were measured at seven assessments (T1 – T7) and cigarette use was measured at four assessments (T5-T8). Therefore, using cohort-sequential analyses, sociability and hostility factors are represented by 8 grades in elementary and middle school (1^st-8^th) and cigarette use is represented by 4 grades in high school (9^th-12^th). Only significant standardized structural paths are shown. Where there were significant gender differences, estimates are shown separately for boys and girls (girls are in parentheses). Several residual correlations and correlations among the intercepts and slopes of hostility and sociability were estimated, but are not shown. I = intercept, S = slope.

* p = <. 05. ** p = < .01. *** p = < .001.

Figure 4.

Hostility and sociability and in elementary and middle school predicting marijuana use in high school.

NOTE: Hostility and sociability were measured at seven assessments (T1 – T7) and marijuana use was measured at four assessments (T5-T8) for 6^th-10^th grade at T5 through 9^th-12^th grade at T8. Therefore, using cohort-sequential analyses sociability and hostility factors are represented by 8 grades in elementary and middle school (1^st-8^th) and marijuana use is represented by 4 grades in high school (9^th-12^th). Only significant standardized structural paths are shown. Where there were significant gender differences, estimates are shown separately for boys and girls (girls are in parentheses). Several residual correlations and correlations among the intercepts and slopes of hostility and sociability were estimated, but are not shown. I = intercept, S = slope.

* p = <.05. **p = < .01. ***p = < .001.

Cigarettes

The model for cigarette use is shown in Figure 2. This model fit the data well, X² (352, N = 1,074) = 493.72, p < .001; CFI = .97; RMSEA = .027; 90% CI = .021, .033. The intercept of hostility predicted the intercept of cigarette use, and the effect was significantly stronger for girls (.42, p < .001) than boys (.39; p < .001). The hostility intercept predicted the slope (.40, p < .001) of cigarette use for both boys and girls. Thus, as hypothesized, children with higher initial levels of hostility (particularly girls) were smoking more at 9^th grade and were more likely to increase their cigarette use during high school. Growth (slope) in hostility across elementary and middle school predicted the intercept of cigarette use at 9^th grade for both boys and girls (.48, p < .001). That is, as hypothesized, children who became increasingly hostile during elementary and middle school were smoking more by 9^th grade. However, contrary to our hypothesis, the slope of hostility did not predict the slope of cigarette use during high school for either boys or girls. There were no significant effects of childhood sociability on adolescent cigarette use.

Alcohol

The model for alcohol use is shown in Figure 3, and demonstrated an adequate fit, X² (353, N = 1074) = 444.87, p <.001; CFI = .98; RMSEA = .022; 90% CI = .015, .028. In partial support of our hypotheses, the intercept of hostility predicted the intercept but not the slope of alcohol use for both boys and girls, and this association was significantly stronger for girls (.58, p < .001) than boys (.40, p < .001). That is, children, particularly girls, who were more hostile at 1^st grade, were drinking more alcohol by 9^th grade, but they were not more likely to increase their alcohol use during high school. The slope of hostility also predicted the intercept (.47, p < .001) but not the slope of alcohol use for boys and girls. That is, children who became more hostile during elementary and middle school had higher levels of alcohol use by 9^th grade but did not increase their alcohol use from 9^th to 12^th grades.

Figure 3.

Hostility and sociability in elementary and middle school predicting alcohol use in high school.

NOTE: Hostility and sociability were measured at seven assessments (T1 – T7) and alcohol use was measured at four assessments (T5-T8). Therefore, using cohort-sequential analyses sociability and hostility factors are represented by 8 grades in elementary and middle school (1^st-8^th) and alcohol use is represented by 4 grades in high school (9^th-12^th). Only significant standardized structural paths are shown. Where there were significant gender differences, estimates are shown separately for boys and girls (girls are in parentheses). Several residual correlations and correlations among the intercepts and slopes of hostility and sociability were estimated, but are not shown. I = intercept, S = slope.

* p = <.05. ** p = < .01. *** p = < .001.

One significant association for sociability was observed. For both boys and girls, the intercept of sociability predicted the intercept of alcohol use indicating that children who were more sociable at 1^st grade were more likely to have higher levels of alcohol use by 9^th grade (.20, p < .001).

Marijuana

The fit of this model, shown in Figure 4, was adequate, X² (353, N = 1074) = 486.47, p < .001; CFI = .97, RMSEA = .027; 90% CI = .020, .032. This model provided the most support for the hypothesized effects of level and growth of hostility. The intercept of hostility predicted both the intercept and the slope of marijuana use for both boys and girls. The association between the intercept of hostility and the intercept of marijuana use was higher for girls (.41, p < .001) than for boys (.31, p < .001), whereas the association between the intercept of hostility and the slope of marijuana use was higher for boys (.44, p < .001) than for girls (.26, p < .01). These findings indicated that children, particularly girls, who were more hostile at 1^st grade, were using more marijuana by 9^th grade, and that children, particularly boys, who were more hostile in 1^st grade increased their marijuana use significantly more during high school. The slope of hostility predicted the intercept of marijuana use for both boys and girls (.47, p <.001) but did not predict the slope of marijuana use. That is, children's increasing hostility over elementary and middle school was associated with higher levels of marijuana use by 9^th grade, but did not predict growth in marijuana use during high school. Sociability was not associated with marijuana use.

Discussion

This study extends prior research on the influence of personality on the development of substance use by relating growth of hostility and sociability in elementary school to growth of subsequent substance use in high school. By modeling trait predictors and substance use outcomes as latent growth constructs, it was possible to relate individual differences in the level and rate of growth of latent traits to individual differences in the level and rate of growth of subsequent smoking, alcohol, and marijuana. Consistent with our initial hypothesis, childhood hostility emerged as a more consistent predictor across all three substances than sociability, demonstrating the significance of the development of this particular trait for the etiology of substance use.

Growth of Adolescent Substance Use and Childhood Traits

Using LGM, developmental trends in levels and growth of substance use across the high school years were described, and differences between boys and girls examined. The levels of use of cigarettes and marijuana did not differ between boys and girls, but girls used more alcohol than boys. Consistent with national trends, use of cigarettes, alcohol, and marijuana increased linearly across the four high-school years for both boys and girls (Johnston et al., 2008). The correlations between the level and slope of substance use indicated that for marijuana, those who started off at relatively low levels of use in 9^th grade had more rapid rates of increase over the four years, and this was particularly true for girls. No such relation was observed for cigarettes and alcohol, suggesting no comparable catch-up process for these substances in high school, possibly because initiation of cigarettes and alcohol use occurs at an earlier age than marijuana (Andrews et al., 2003; Johnston et al., 2008).

Our findings add to what little is known about the development of personality traits over the elementary school years. There was no maturational trend to either increase or decrease in hostility, which is consistent with the conclusion of a prior meta-analysis that trait means show stability over time for this age range (Roberts et al., 2006). However, sociability showed a slight but significant decline. This finding is consistent with the modest reduction in childhood extraversion, assessed by mothers, observed by Lamb, Chuang, Wessels, Broberg, and Hwang (2002) over five assessments from age 2 to 15 years. The decline in sociability observed here, assessed predominantly by items referring to participation in social interaction, may reflect a maturational trend to develop social restraint during elementary and middle school, or it may reflect the decreasing opportunity for teachers to observer children's social interactions as classwork becomes more individual and there is less playtime.

There was significant variation in the level and the slope of both hostility and sociability. Variation in trait level confirmed that there were individual differences in the extent to which these children were hostile and sociable. Variation in slope was the more interesting and novel finding. It demonstrated that there were individual differences in trait change during childhood even when, as was the case for hostility, there was no evidence for a maturational trend. Individual differences in rate of change provide a new way of examining the influence of a personality trait on subsequent outcomes and here we showed that this variance predicted later substance use.

Hostility as a Predictor of Substance Use

Consistent with our initial hypothesis, childhood hostility emerged as an important predictor of later substance use. Higher initial levels of childhood hostility predicted higher levels of substance use in 9^th grade for cigarettes, alcohol, and marijuana. Children with higher initial levels of hostility were also more likely to increase their use of cigarettes and marijuana during high school. Furthermore, increasing hostility over childhood predicted higher levels of use of all three substances in 9^th grade. The findings for individual differences in rate of change for hostility are of particular interest. Increasing hostility across elementary and middle school emerged as a novel risk factor for greater substance use in high school.

Given that there was no maturational trend for children to increase in hostility, why some children become increasing hostile, and why this confers greater risk for later substance use, become important questions. Previous studies of the etiology of substance use point to possible explanatory mechanisms. Children's beliefs about substances, which include social images of substance users, subjective norms, and intentions and willingness to use, indicate cognitive susceptibility for later substance use (e.g., Andrews et al., 2008; Choi, Gilpin, Farkas, & Pierce, 2001; Gritz et al., 2003). There is evidence that more hostile children are more likely to develop greater cognitive susceptibility to substance use. For example, more hostile children in OYSUP were more likely to intend to use cigarettes and alcohol when older, and the effects of boys' hostility on intentions to smoke, and of both boys' and girls' hostility on intentions to drink alcohol, were mediated by subjective norms (Hampson et al., 2006; Hampson et al., 2007b). Subjective norms were operationalized as children's beliefs about the extent to which their friends and other kids in their neighborhood used the substance in question, and more hostile children believed that more of their peers were using substances.

The role of these peer-based norms as mediators of hostility suggests a peer-influence process, which is another well-established etiological factor in adolescent substance use (Andrews & Hops, in press). Hostile children may be more likely to join deviant peer groups and thus be at greater risk of substance use (Dishion, 1990; Hawkins, Catalano, & Miller, 1992; Patterson, Reid, & Dishion, 1992). For example, children from alcoholic families who were more impulsive and low in agreeableness (i.e., hostile), were more likely to become involved with deviant peers and at risk for substance dependence (Chassin, Flora, & King, 2004), and high sensation-seeking adolescents associated with deviant peers who were also higher sensation seekers (Donohew et al., 1999). Through the mechanism of niche selection, hostile children may seek out a deviant peer group composed of similar others, and then socialization within that group may contribute to the further development of their hostility by mechanisms such as evocation, manipulation, and accumulation, leading to the development of cognitive susceptibility (e.g., intentions, willingness) and eventual substance use (Roberts et al., 2008).

Sociability as a Predictor of Substance Use

As expected, sociability demonstrated a less consistent pattern of associations with substance use than hostility. The relatively few associations between childhood sociability and adolescent substance use may indicate that sociable children enjoy some degree of protection, benefitting from more prosocial early peer influences (Bukowski, Brendgen, & Vitaro, 2007). However, being more sociable in childhood did not predict less adolescent substance use. Childhood sociability was unrelated to adolescent cigarette or marijuana use but, contrary to our hypothesis, higher levels predicted initial levels of alcohol use in high school, suggesting sociable children may develop beliefs that alcohol use is prevalent and acceptable and hence are more susceptible to using alcohol in adolescence. Sociable children are likely to be more popular, and popularity has been associated with adolescent alcohol use, suggesting the importance of peer influence (Hops, Davis, & Lewin, 1999; Pearson et al., 2006). Trait mechanisms such as niche selection may result in sociable children seeking out environments (e.g., team sports) that support the development of cognitive susceptibility (e.g., willingness and intentions) for later use of alcohol (Eccles & Barber, 1999).

Of interest was the absence of any effects of negative growth in sociability (sociability showed a slight maturational trend to decline). Although there was significant variance in growth, individual differences in change in sociability did not predict later substance use. Past developmental research on the conduct disorder of social withdrawal suggests that children who are less social and who become increasing less social, would be more likely to use substances in adolescence (Rubin et al., 2009). Our findings do not support this prediction. Instead, this study supported the position that higher levels of sociability can be a risk factor for substance use (Brook et al., 1988; Chassin et al., 1993; Jessor & Jessor, 1977), and underscored the importance of examining different substances separately.

Strengths and Limitations

Although latent growth of traits has been studied in young and old adults (Mroczek & Spiro, 2007; Vaidya, Gray, Haig, Mroczek, & Watson, 2008), to our knowledge no study has examined trait growth from early childhood to adolescence using LGM. Prior studies have used parameters of latent trait growth to predict a single outcome such as mortality (Mroczek & Spiro, 2007), or have used traits measured on only one occasion to predict an outcome modeled as a latent growth construct, such as growth in academic achievement (Hart, Atkins, & Matsuba, 2008). Here, to address novel substantive questions, we related growth of personality traits to growth of substance use.

In addition to multiple assessments, another strength of this study was the use of teacher ratings of personality. Teachers are more informed raters than parents because they can judge a child relative to his or her peers, and can be more objective. Ratings were made by unique teachers at each assessment, which reduced the risk of biases from repeated assessments by the same teacher. The same raters across assessments (teachers, parents, or the self) may be biased towards perceiving greater personality stability over time. In addition, stability can be confounded with method bias (e.g., using the same portions of the scale in their ratings). Ratings from unique teachers across across multiple assessments showed more trait instability in childhood than has been reported hitherto.

This study has a number of limitations. As a non-experimental study, causal relations between childhood personality traits and later substance use cannot be inferred, although the prospective nature of this design makes a causal account more plausible. The OYSUP sample is almost exclusively White (European American), thus restricting the generalizability of the present findings. This study only examined whether childhood trait change was related to later substance use and did not test potential mechanisms to account for these associations, although some potential mechanisms were outlined. Personality traits are one of many risk factors for substance use and the effects of these other factors were not studied here. In previous papers, we showed that personality traits may influence cognitions regarding substance use which in turn influence intentions and use (Hampson et al., 2006, 2007b). Other etiological factors not studied here, including ones related so socioeconomic status such as parenting styles, may also influence both the development of personality traits and substance use. Moreover, this paper did not distinguish between increasing substance use and problematic use, such as binge drinking. It is possible that differentiating between qualitatively different levels of use would yield yet more interesting associations with early childhood personality. For example, sociability may predict experimental but not problematic substance use, whereas hostility may be more strongly associated with problematic use.

Conclusions

The present findings demonstrate that individual differences in the growth of childhood personality traits, in addition to the individual differences in the level of these traits, can predict subsequent substance use. In particular, increasing hostility over childhood as well as higher levels of early hostility, were identified as risk factors for later use of the most prevalent forms of high school substance use. This knowledge could be used to inform interventions to prevent the development of adolescent substance use. Although it remains for future research to investigate the processes by which developing traits influence later behavior, an implication of this study is that intervening to alter the developmental course of potentially harmful traits could result in prevention of later problem behavior.