Caretakers provided informed consent and adolescents provided assent until age 17 and consent thereafter. We restricted analysis to adolescents at ages, as substance use by year was rare at younger ages: 93.9% and 84.5% did not use marijuana or alcohol, respectively, on any occasion between the ages of 7-12. Study procedures were approved by the Institutional Review Boards of the University of Pittsburgh School of Medicine and the Columbia University Mailman School of Public Health. Alcohol and marijuana use were assessed semi-annually by a 16-item Substance Use Scale adapted from the National Youth Survey. Adolescents were queried about timing, quantity, and frequency of alcohol and marijuana use. We defined “marijuana frequency” as the number of occasions of marijuana use in the past year. We defined “alcohol frequency” as the number of occasions of drinking in the past year. We defined “alcohol quantity” as the average number of drinks per occasion in the past year. For phases separated by only 6 months, past-year values were constructed by taking the average of the two semi-annual interviews. Affective, anxiety, and conduct problems were measured with items from the Child Behavior Checklist , Teacher Report Form , Youth Self-Report , and Young Adult Self Report from the Achenbach system of assessment.DSM-oriented problem domains were measured with items rated as very consistent with DSM-IV symptoms of affective disorders, anxiety disorders, and conduct disorder by a group of mental health professionals.The scales were administered to caregivers and teachers from age 7 to 16, and youth from age 10 to 19 .Items were scored as 0 = not true, 1 = somewhat or sometimes true, and 2 = very true or often true .In order to facilitate comparison across informants, total scores for each scale were converted to t-scores based on age- and gender-specific national norms .An average T-score was then calculated for years when multiple informants completed the scales. The average internal consistency coefficients for the caregiver, teacher, and youth depression scales were 0.82, 0.76, and 0.81, respectively. For the anxiety scales,rolling bench the internal consistency coefficients for caregiver, teacher and youth scales were 0.72, 0.73, and 0.67, respectively.
For the conduct disorder scale, the internal consistency coefficients were 0.91, 0.9, and 0.83 for caregiver, teacher, and youth scales, respectively.These scales have been shown to discriminate between clinic referred adolescents with depressive, anxiety, and conduct disorders and non-referred adolescents. All the scales used have previously shown acceptable concurrent and predictive validity in ROC analyses comparing the scales with official records of offense and delinquency or by assessing discrimination between adolescents referred to psychiatric clinics and non-referred adolescents.Several potential time-varying confounding factors were included in the current study to parse out the effect of psychiatric problems from the constellation of time-varying risk factors that could increase both psychiatric problems and substance use. The selection of confounders was based on theory and a review of the literature, as detailed below. “Family factors” included changes in socioeconomic status , assessed yearly by applying the Hollings head Index of Social Status to data provided by the primary caretaker or youth no longer living with family beginning at age 16; changes in parental supervision/involvement, a 43-question scale concerning caretakers’ knowledge of the youths’ whereabouts, the frequency of joint discussions, planning, and activities, and the amount of time that the youth is unsupervised; positive parenting, a scale measuring perception of frequency of positive responses to youth behavior; parental stress, a 14-item scale measuring perceived stress levels and caretakers’ abilities to cope with stress in the previous month 18; and parental use of physical punishment, drawn from a scale that measures parental discipline strategies. “Peer Variables” consisted of changes in youth peer delinquency and peer substance use, a 15-item scale that corresponds to a self-reported delinquency scale.Analyses were conducted in R version 3.0.2 and 3.0.3. Missing data in the covariates were imputed using R package ‘mice’ for “multivariate imputation by chained equations,” an implementation of fully conditional specified models for imputation. The fully conditional approach differs from the more traditional joint modeling approach by specifying a multivariate imputation model on a variable-by-variable basis. This fully conditional approach is used as an alternative to traditional joint modeling when no suitable multivariate distribution can be found. We imputed 20 datasets, and in subsequent analyses used the R package ‘mitools to pool the results of functions run on the 20 data sets using Rubin’s Rules.
We employed quasi-Poisson regression techniques to assess the fixed effects that one-yearlagged changes in psychiatric problems had on subsequent changes in alcohol use frequency/quantity and marijuana use frequency from ages 13 to 19. Quasi-Poisson models are an approach to dealing with over-dispersion, which was apparent in initial Poisson models. They use the mean regression function and the variance function from Poisson generalized linear models but leave the dispersion parameter unrestricted and estimate it from the data. Unlike negative binomial models, the variance is assumed to be a linear function of the mean.This strategy leads to the same coefficient estimates as a standard Poisson model but standard errors are adjusted for over dispersion. Following the “dummy variable method” for fixed effects in Poisson models 41 we included k – 1 dummy variables to represent the sample participants in each model. A series of models were fit sequentially to test the association of each one-year-lagged psychiatric problem domain with each substance use outcome. First, we regressed separately each one year-lagged shift in the average psychiatric problem T-scores on each substance use outcome. Within these models, age-related changes in substance use were controlled for using natural cubic splines. Natural cubic splines are a flexible smoothing approach for non-linear relationships, and are composed of piece wise polynomial functions that split the continuous age variable into separate line segments, each free to have its own shape. Segments are joined by “knots,” which we specified a priori to result in line segments for ages 13-14, 15-16, and 17-19. Slopes are constrained to converge at each knot. Second, we sequentially tested groups of potential confounders. All covariates were back-lagged two years, so that they would be modeled prior to the measurement of the exposure. This ensured that the estimated total effect of change in psychiatric problems on change in substance use included effects mediated through the covariates that occurred contemporaneous to changes in psychiatric problems. In our second set of models, we adjusted for age, SES, substance use variables that were not modeled as the outcome , and measures of psychiatric problems that were not the exposure of interest . In our third and fourth sets of models, we adjusted for age and parenting variables and age and peer variables, respectively.
In our fifth set of models,dry rack cannabis we adjusted for covariates that were significant in models. Third, we tested whether age modified the effect of our exposures by including a product term between exposure and each age spline. Significant effect measure modification was then probed to clarify how the association between psychiatric problems and substance use changed across the age splines. We conducted a sensitivity analysis to establish the directionality of the association between psychiatric problems and substance use. We thus estimated, with linear fixed effects models, the effect that changes in one-year-lagged substance use had on change in psychiatric problem domains in the following year. We followed the same modeling strategy for these models as we did with our primary models. We adjusted for groups of confounders as described above, first adjusting for SES, psychiatric problem domains that were not modeled as the outcome , and measures of substance use that were not the exposure of interest . Next we adjusted for parenting variables and peer variables, respectively. Finally, we adjusted for covariates that were significant in any of the previous groups of confounder models. Covariates were lagged one year prior to the exposure measure , to avoid blocking the causal Table 1 shows mean substance use and psychiatric problem counts over time, as well as demographic characteristics at baseline. The reports of particular informants in our psychiatric problem measures did not influence the associations between psychiatric problems and substance use . Table 2 displays the exponentiated coefficients and confidence intervals of quasiPoisson models, which can be interpreted as rate ratios. Table 2 shows the rate of substance use associated with a one-unit within-subject change in lagged psychiatric problems. Changes in lagged conduct problems were positively associated with changes in marijuana frequency. During years in which adolescents experienced a one-unit increase in conduct problems, the rate at which they smoked marijuana the following year increased 1.03 times : 1.01, 1.05. For a standard deviation change in conduct problems, this is equivalent to a 1.15 times higher rate of marijuana use frequency . The magnitude of this association did not change appreciably after adjusting for potential confounders, including alcohol quantity and frequency, SES, affective and anxiety problems, parenting, and peer deviance. Changes in lagged conduct problems were also associated with changes in alcohol quantity, only after adjusting for peer deviance. During years in which adolescents experienced a one-unit increase in conduct problems, the rate of their average alcohol consumption per occasion the following year increased by 1.01 . For a standard deviation change in conduct problems, this is equivalent to a 1.05 times higher rate of alcohol use . Associations of all covariates with substance use are presented in Appendix C, Table C1. Table 3 presents results for tests of effect measure modification of the association between conduct problems and marijuana frequency and alcohol quantity by age. Because splines are polynomial functions, there is no simple quantitative interpretation of individual effect modification terms; however, the significance of the coefficients implies that the associations between lagged conduct problems and marijuana frequency, and lagged conduct problems and alcohol quantity, differed by age. For ease of interpretation we present these results in Figure 1, which shows the predicted values of substance use outcomes associated with minimum, mean, and third-quartile levels of lagged conduct disorder T-scores, over time. Compared to minimal changes in lagged conduct problems, adolescents with mean or third-quartile levels of change in lagged conduct problems show markedly different marijuana frequency trajectories, which become the most disparate at ages 17-19. Compared to minimal changes in lagged conduct problems, adolescents with mean or third-quartile levels of change in lagged conduct problems show higher alcohol quantity in early adolescence but lower alcohol quantity in later adolescence. The results of our sensitivity analysis are presented in Table 4 and 5, and Figure 2. Table 4 displays the change in psychiatric problems associated with a one-unit change in lagged substance use in the prior year. There was one reverse association: while changes in lagged anxiety problems were not associated with changes in substance use, the opposite did occur: changes in lagged alcohol quantity in the past year were positively associated with changes in anxiety problems. During years in which adolescents experienced a one-unit increase in the average quantity of alcohol consumed when drinking, their anxiety problems T-score increased the following year by 0.12 . For a standard deviation change in average alcohol quantity, this is equivalent to an anxiety T-score increase of 0.3 .Associations of all lagged covariates with psychiatric problems are presented in Appendix C, Table C2. Table 5 presents results for tests of effect measure modification of the association between lagged alcohol quantity and anxiety problems by age, and Figure 2 shows the predicted values of anxiety problem T-scores associated with minimum, mean, and third-quartile levels of lagged alcohol quantity, over time. Adolescents show a decline in anxiety problems throughout adolescence, and little difference by the magnitude of fluctuations in lagged alcohol quantity. However, deviations arose at ages 13-14 and 17-19, where those who exhibited a mean or third-quartile level of increase in lagged alcohol quantity showed slower declines in anxiety problems compared to those who did not increase alcohol intake over time. This study focused on the longitudinal relationship between changes in psychiatric problems and changes in substance use one year later. However, the temporal resolution of this relationship may occur on a much shorter time frame – that is, changes in psychiatric problems may have immediate effects on substance use .