As part of the broader study, at baseline and annual follow-up visits, participants completed a comprehensive battery including neuropsychological assessment , self-reports of behavior, psychiatric symptoms and substance use, and a multimodal neuroimaging session . Exclusionary criteria for study entry included current use of psychoactive medication, current or persistent major Axis I psychiatric disorders, significant learning or developmental disorders and serious major medical conditions . While a majority of participants had limited drug and alcohol exposure at enrollment, a small proportion were recruited who exceeded age-specific alcohol and marijuana low-use thresholds . Additional recruitment, demographic and procedural details have been published previous . The NCANDA project employs an accelerated longitudinal design, and the current study used all available data across the first seven waves of data collection, from November 2012 to December 2020.To measure the Big Five personality dimensions, participants were administered the TenItem Personality Inventory , a brief measure shown to have convergence with longer Big Five measures as well as good test-retest reliability . The TIPI consists of 10 questions, with two questions for each subscale: extraversion, agreeableness, conscientiousness, emotional stability, and openness. Each question asked participants to rate on a scale of 1 to 7 how much a pair of words applied to them. The responses on the two questions for each subscale were averaged and served as the primary outcome measure for all analyses. Due to protocol changes during the fifth wave of data collection, pipp racks the study moved from administering the TIPI during all annual visits, to only administering it if participants were completing their age 24 or 27 visit .

Alcohol and marijuana use were measured using the Customary Drinking and Drug Use Record . At all visits, participants self-reported the number of days they drank and used marijuana during the past year. That is, participants were asked: “During the past year, , how many days did you drink alcohol?”, with an identical question asked regarding marijuana use. Due to non-normal distributions, past-year alcohol and marijuana use variables were log-transformed prior to all analyses.Previous studies investigating the development of personality across adolescence often used linear growth models with polynomial effects . However, when examining development across a broad age range, it is possible that data do not always conform to this restricted parametric growth model, and when examining group-level effects , different groups may not necessarily demonstrate similar developmental trajectories. Generalized Additive Mixed Models , an extension of generalized linear mixed models, do not assume the shape of developmental growth a priori, but instead allow for age-related non-linear smooth functions that best represent the relationship between predictor variables and outcomes . Similar to traditional linear mixedeffects models , GAMMs allow for appropriate modeling of the within-subject correlation of longitudinal data, as well as other important random effects . Here, we modeled changes in personality development as a function of sex using both GAMMs and LMEs, and present findings side-by-side in order to assessthe impact of modeling choice. All tested models can be found in Table S1. Analyses were conducted using R 4.1.1 .Generalized Additive Mixed-effects Models —To assess the effects of age and sex on personality development, we fit GAMMs using the ‘mgcv’ package in R and carried out a series of model comparisons, similar to the approach taken in recent neurodevelopmental studies . For each of the 5 TIPI scales, we fit three successive models that included age-related development across the whole sample , a main effect of sex , and differences in the age-related personality development by sex . All models included a random intercept per participant, family, and data collection sites.

When interpreting sex effects on personality development, it is important to note when first fitting Model 3, sex was included as a ‘factor’, resulting in the estimation of a separate smoothed age trajectory in male and female participants. While this has the benefit of producing interpretable smoothed terms for each group, a traditional interaction term, such as that seen in linear modeling, is not produced. Therefore, to test the statistical significance of sex-specific developmental trajectories, standard hypothesis testing was used to compare the log-likelihood values from each model . Then, to provide additional statistical support, Model 3 was refit with sex coded as an ‘ordered factor.’ Here, a smoothed age trajectory is calculated for the ‘reference’ group only , and a smooth term representing the difference between the developmental trajectories of the reference group and the other group was estimated. While this method provides less information regarding the shape of the age-related trajectory for each group, it produces an estimate and significance-testing for the difference between groups, akin to traditional linear interaction terms, and has also been used previously in developmental studies . Finally, to assess the association between substance use and personality, we modeled time invariant , linear time-varying , and quadratic time-varying associations of past year alcohol and marijuana use. These three potential associations could occur for alcohol use, marijuana use, or both, resulting in a combination of nine different models . Additionally, to capture potential sex-specific associations of alcohol use, marijuana use, our both, a total 27 additional models were necessary to exhaustively explore these relationships . These predictors were added to the best fitting developmental model compared using standard hypothesis testing. Linear Mixed-effects Models—To compare developmental GAMM results to models with more traditional polynomial growth parameters, we fit a series of LME models. Unlike GAMMs, which allow for the assessment of sex differences in non-linear personality development with only 3 models, the LME framework requires the iterative addition of consecutive higher-order polynomial age predictors to statistically assess the benefit of added model complexity. Here, we chose to assess the effect of 3 orders of polynomial effects , along with their potential interaction with sex, using the ‘nlme’ package in R . Starting with a linear age effect, we compared three successive models to assess the pattern of age-related development across the whole sample , a main effect of sex , and differences in the age-related effects by sex . This process was then repeated for quadratic , cubic polynomial age effects. For each interaction model, the effect of sex was assumed to interact with all lower-order polynomial age effects. Finally, the best fitting model for each polynomial age effect was then compared, to determine the final model. Identical to GAMMs, all models included a random intercept per participant, family, and data collection sites. Briefly, female and male participants identified as either white non-Hispanic , African American/Black , Hispanic/ Latino , Asian , multi-racial , Pacific Islander , or Native American/ American Indian . At baseline, 20% reported parents with education below a college degree, 27% with at least one parent attaining a college degree, and 53% with at least one parent with education greater than a college degree; annual family income ranged from below $12,000 to greater than $200,000. The TIPI was completed during at least one visit for 829 of the 831 subjects, with a total of 3,402 case observations across the 7 waves. However, 24 cases included incomplete reporting of substance use measures, and 4 cases included inconsistencies in reported substance use . To provide direct statistical comparison of nested models, only subjects’ timepoints with complete data were included in the final analyses.

Notably, all developmental findings remain unchanged when those timepoints with missing substance use values were included. In total 3,374 case observations across 829 subjects were included in final analyses; the breakdown by wave follows: Baseline , Year 1 , Year 2 , Year 3 , Year 4 , Year 5 , Year 6 . As expected alcohol and marijuana use both increased with age . Overall, 68% of the sample reporting drinking, and 48% of the sample reported using marijuana during at least one wave of data collection. Of those reporting substance use, vertical growing racks over the course of the study to-date, past year alcohol use ranged from 1 to 365 days with an average of 28.5 days per year, and past year marijuana use ranged from 1 to 365 days with an average of 49.1 days per year. Fit statistics and model comparisons for GAMMs examining age- and sex-related effects on personality development, and the association between personality and past year substance use can be found in Table S2. Parameter estimates of the final best-fitting GAMMs can be found in Table 1. All significant findings reported herein are from the final best fitting models, including the effects of past year alcohol and marijuana use. For models with sex-by-age and/or substance use-by-age interactions, models were refit with their intercepts adjusted to ages 13, 16, 19, 22 and 25 in order to provide added interpretation to underlying main-effects of sex and substance use. In the absence of standard parametric age coefficients in GAMMs, we report the effective degrees of freedom , which sheds light on the degree of nonlinearity for a given developmental trajectory . Effect sizes for all parametric coefficients are reported as standardized regression coefficients for continuous predictors and Cohen’s d for categorical predictors .The current study sought to flexibly model developmental trajectories of personality in adolescence and young adulthood as a function of sex and explore the association between substance use and personality across age. We report three general conclusions: 1) there were linear increases in agreeableness and conscientious and decreases in openness, across this age range, the slope of which did not differ developmentally by sex, and significant sex-specific non-linear developmental differences in extraversion and emotional stability; 2) male participants reported lower agreeableness, conscientiousness,and openness across the entire age range, less extraversion at all ages except during midadolescence , and more emotional stability in all but early adolescence ; 3) alcohol use was associated with greater extraversion and openness across the entire age range, and less conscientiousness in adolescence , while marijuana use was associated with less agreeableness throughout the entire age range, less conscientiousness in early adolescence and young adulthood , less extraversion in young adulthood , and less emotional stability throughout the entire developmental age range in female youth, and in young adulthood in male youth. Developmentally, our findings provide partial support for the maturity principle , as we found both conscientiousness and agreeableness to increase linearly from ages 12 to 25. This is consistent with at least one report that found agreeableness and conscientiousness increased consistently across adolescence and young adulthood , with non-linear effects occurring primarily in other traits . Meanwhile, another study found the lowest levels of agreeableness and conscientiousness occurred around ages 12–13 . Our data provide strong replication of these results, in a large multi-site cohort, and suggest that any “disruptions” seen in agreeableness and conscientiousness may take place during childhood, prior to their continued maturation in adolescence and young adulthood. Contrary to this effect, we note decreases in openness across the entire age range. While openness has been shown to decline in late childhood and early adolescence , there is no evidence, to our knowledge, of self-reported decreases in openness in late adolescence and young adulthood, though parent-reported adolescent personality findings suggest decreases in openness in this age range . Interestingly, out of all five personality traits, when assessed in adolescence, openness has been shown to have the lowest internal consistency, and replicability across multiple samples and cultures . Thus, it’s possible that our 10-item question of personality could be less sensitive to true mean-level changes in openness in this population. Our findings also provide partial support for the disruption hypothesis , as we found extraversion in both male and female youth, and emotional stability in female youth, decreased in early adolescence . However, unlike previous reports , we found extraversion never increased during late adolescence and young adulthood. Instead, male participants continued to show linear declines in extraversion, while female participants showed a leveling off of extraversion. This is where we believe our flexible analytic strategy helps clarify past results. For example, Borghuis, Denissen et al. tested only linear and quadratic growth parameters, and found U-shaped trajectories for extraversion. Similarly, when quadratic growth parameters were fit to our data , we replicated this previously observed effect in female participants; however, more flexible modeling suggest, this quadratic growth does not best fit the data.