Within-Individual Changes in Impulsivity and Sensation Seeking from Childhood to Early Adulthood and Educational Attainment

Importantly, those participants who were between 22-23 years old in 2018 were included in the models examining trajectories impulsivity and sensation seeking in childhood and adolescence, but were omitted from those models examining trajectories of these same constructs during adulthood to preserve temporal order. The omission of these younger participants is reflected in the final sample size differences between both sets of models.

Given these internal reliability coefficients, the underlying factor structure of the items was investigated across all examined two-year intervals. First, principal-components analysis was performed for each two-year time interval. The results revealed that all three items loaded on a single higher-order construct with factor loadings ranging between 0.64 and 0.80. Second, two sets of confirmatory factor analyses (CFAs) were performed. The first set included a separate model for each two-year interval. While this approach provided input into the magnitude of the factor loadings, the models (with three indicators at each interval) were just identified, resulting in perfect model fit. Standardized factor loadings ranged between 0.43 and 0.74. To assess global fit, another CFA model that specified latent factors for all eight measurement occasions (ages 10–11 through ages 24–25) was estimated. The resulting model provided an acceptable fit to the data (CFI = 0.933; TLI = 0.915; RMSEA = 0.024; SRMR = 0.057) with resulting standardized factor loadings that closely resembled those from the individual CFA models. Importantly, these supplemental models accounted for the categorical nature of the indicators (i.e., a weighted least squares estimator with a probit link and robust standard errors), which may be one issue contributing to the observed differences between Cronbach α coefficients and the patterns observed in the factor analysis. Taken together, even in light of the internal reliability coefficients reported above, the construct validity checks performed in previous studies (Harden & Tucker-Drob, 2011) coupled with the results from supplemental analyses investigating the factors structure of these items suggest that they tap a single higher order construct that behaves similarly to other impulsivity measures. With that said, readers should interpret the results from the primary analyses with these limitations in mind.

Once again, the factor structure of these indicators was investigated in supplemental analyses. Principal-components analysis indicated that all sensation seeking items loaded on a single higher order latent construct at each two-year age interval, with factor loadings ranging between 0.70 and 0.84. A similar pattern was observed in the individual CFA models estimated for each age interval with standardized loadings ranging between 0.54 and 0.80. Finally, a CFA model with factors identified for each individual age interval provided a close fit to the data (CFI = 0.983; TLI = 0.979; RMSEA = 0.018; SRMR = 0.046).

In addition to these qualities, this nonlinear approach is better suited for the current study relative to alternative approaches (e.g., higher order polynomial functions). More specifically, the use of polynomial trajectories requires the estimation of a third latent term—in addition to the latent intercept and slope factors—that defines the overall “bend” in the linear trajectory over time. In this way, the resulting trajectory of the slope factor is defined with two latent factors—the linear slope factor and the polynomial trend factor. Since the current study aims to examine the resulting slope factor as a predictor in subsequent multivariable models, it would then be necessary to include both factors in such models, resulting in two problems. First, interpretation becomes increasingly difficult, as the overall impact of both included latent factors would (at least potentially) be differently associated with the examined outcome. Second, and perhaps even more importantly, the resulting slope and polynomial trend factors are often highly correlated, resulting in multicollinearity. The latent basis approach described above remedies these issues by limiting the estimated slope factor to a single term while still allowing for a nonlinear trajectory across observation periods (Bollen & Curran, 2006). The inclusion of a single slope factor in subsequent models not only eases interpretation, but it also eliminates issues surrounding multicollinearity. For these reasons, a latent basis approach for nonlinear trajectories was employed.

While the examined educational attainment outcome is technically an ordered categorical variable, the measure is comprised of a total of 14 categories, resulting in the approximation of a discrete quantitative variable. For this reason, a linear regression approach was employed in the current study. However, to ensure that this decision did not systematically bias results, a series of sensitivity analyses in which all the estimated models from the primary analysis were re-estimated using ordered logistic regression was performed. The results of the sensitivity analysis directly aligned with those from the primary analysis.