Descriptive statistics are presented in for the total sample and for each income group. Youth from a low-income background had higher means on several measures of delinquency, and on the early-onset sex variable, compared to youth from a middle-income background. Regarding alcohol use, the sole statistically significant difference showed a higher level of early alcohol use among those from middle- compared to low-income backgrounds. There were no statistically significant group differences on the young adult outcomes.
Means and standard deviations for the study variables by low- and middle-income background
Correlations among study variables are reported in separately by income group. There were expected positive associations between the childhood problem behaviors and the indicators of adolescent delinquency and alcohol use, and the young adult outcomes. A somewhat larger number of statistically significant associations were observed among low-income compared to middle-income youth.
Correlations for youth from middle- (n = 385; upper-diagonal) and low- (n = 423; lower-diagonal) income backgrounds
The primary analyses were conducted in 3 stages. First, separate LGMs for delinquency and alcohol use, along with a combined (dual process) model, examined patterns of change and interrelationships over time, from age 14 to age 18. Loadings of the intercept [1,1,1,1] and slope [0,1,2,4] factors were fixed at values that correspond to linear change, with intercepts set at age 14. Further analyses examined possible nonlinear change. Linear growth models were supported for both delinquency, χ2 (5, N = 800) = 10.24, p = .07, TLI = .98, RMSEA = .04, and alcohol use,χ2 (6, N = 800) = 9.69, p = .14, TLI = .99, RMSEA = .03. There was a significant average decrease in delinquency (M = −.009) and a significant average increase in alcohol use (M = .155), with significant variability about the means (delinquency slope variance = .004, p < .05; alcohol use slope variance = .037, p < .05). The dual process LGM displayed acceptable fit, χ2 (15, N = 800) = 30.40, p = .01, TLI = .99, RMSEA = .04. Neither the predictive relationship between the delinquency intercept and the alcohol use slope nor that between the alcohol use intercept and the delinquency slope was statistically significant. Preliminary LGMs excluded covariates, which resulted in a loss of 8 cases; subsequent analyses used the full sample.
The second stage added predictors and outcomes to the dual process model. A conceptual illustration of the full LGM is depicted in . Covariances among exogenous variables, and also among residuals of the outcomes, were freely estimated, as were covariances among residuals of the intercept factors and (separately) the slope factors. Model fit was acceptable, χ2 (31, N = 808) = 47.33, p = .03, TLI = .99, RMSEA = .03. Path estimates are reported in .
Conceptual illustration of the latent growth curve model examining childhood predictors and young adult outcomes of changes in adolescent delinquency and alcohol use. Covariates (not displayed) include gender and income background.
Standardized (unstandardized) path coefficients for the total sample latent growth curve model (N = 808)
Early delinquent involvement and sex onset were associated with higher levels of middle-adolescent delinquency and alcohol use. Early alcohol use positively predicted a higher level of alcohol use in middle adolescence and a faster rate of increase in drinking thereafter, but was not associated with either the subsequent level or change in delinquency. Early delinquent involvement positively predicted young adult crime. Income background was not related to the outcomes. Middle-adolescent delinquency was associated with increased risk for each outcome. Predictive associations for alcohol use were more specific. For example, level of alcohol use in middle adolescence was positively related only to AUDs in early adulthood; however, growth in adolescent alcohol use was associated with increased risk for young adult risky sex in addition to AUDs. Gender had expected associations with the outcomes, with one exception: Male gender was associated negatively with age 14 alcohol use, reflecting the fact that girls (M = .37, SD = .71) in this sample reported a somewhat higher level of alcohol use than boys (M = .35, SD = .68) at this time. Tests of mediation through adolescent problem behaviors showed small but statistically significant (p < .05) indirect effects of early delinquent involvement on crime (β = .09), AUDs (β = .08), and risky sex (β = .08); early alcohol use on AUDs (β = .08); and early sex onset on crime (β = .08).
In the third stage of analysis, the full LGM () was estimated as a multiple-group model to test for differences in the path coefficients across income groups. Analyses began with a model that allowed all paths to vary freely across the two groups, which is similar to what would be obtained if separate models were conducted for each group, except the two groups were examined simultaneously in a single model. The fit of this unconstrained multiple-group LGM was acceptable, χ2 (51, N = 808) = 78.33, p = .01, TLI = .98, RMSEA = .04. Next, a model was estimated that constrained to equality across groups all 44 structural (i.e., directional) paths. These constraints forced each estimated path coefficient (e.g., between early alcohol use and the alcohol use intercept) to take on the same value in both groups. A model comparison showed that these constraints contributed to a statistically significant decrease in fit compared to the unconstrained model, χ2 (30, N = 808) = 46.95, p = .03, which indicates that some of the constrained path coefficients should be allowed to be different across groups. To systematically investigate which paths show group differences, modification indices from the constrained model were examined. Modification indices estimate the degree to which freeing a constrained parameter estimate will improve model fit; larger values suggest greater improvements to the fit of the model. The constraint on the path coefficient with the largest modification index was released and the model was re-estimated, comparing the new model fit with that of the unconstrained model. This process was repeated until a modified model with acceptable fit in comparison to the unconstrained model was obtained.
Constraints on 4 paths were released, resulting in a final model with some path coefficients that were forced to be the same and some that were allowed to vary across the two groups. This ‘partially constrained’ model had acceptable global fit, χ2 (72, N = 808) = 108.82, p = .00, TLI = .98, RMSEA = .04, and it did not fit significantly worse than the unconstrained model, χ2 (29, N = 808) = 42.77, p = .05. Thus, the partially constrained model was preferred because it struck the best balance between fit and parsimony. Three of the released constraints showed significant group differences in the (unstandardized) path estimates between the delinquency slope factor and young adult crime (b = 2.69, p < .05 for middle-income; b = 3.44, p < .05 for low-income), risky sex (b = 2.54, p < .05 for middle-income; b = 3.89, p < .05 for low-income), and AUDs (b = 2.21, p < .05 for middle-income; b = 5.88, p < .05 for low-income). The fourth released constraint revealed that gender had a statistically significant positive association with the delinquency slope factor only for youth from a middle-income background (b = .04, p < .05 for middle-income; b = .01, p > .05 for low-income).
As a final consideration, additional multiple group analyses were conducted to explore gender differences in the associations depicted in . Compared to an unconstrained model with path coefficients freely estimated across gender groups, a constrained model that forced each path estimate to take on the same value for boys and girls had significantly worse fit, χ2 (22, N = 808) = 57.38, p = .0001, suggesting the presence of gender group differences. Modification indices revealed that the (unstandardized) paths of both early delinquent involvement (b = .07, p < .05 for girls; b = .02, p < .05 for boys) and early alcohol use (b = .04, p < .05 for girls; b = .01, p > .05 for boys) to the delinquency intercept were stronger for girls than boys. After the constraints on these two parameter estimates were released, the resultant model did not fit significantly worse than the unconstrained model, χ2 (21, N = 808) = 30.81, p = .08.