Participants
We used two sources to examine the relationship between anxiety in adolescence and psychiatric outcomes in late adolescence/young adulthood: the Child and Adolescent Twin Study in Sweden (CATSS) and the Swedish National Patient Register (NPR). In addition, we conducted replication analyses on a longitudinal dataset that was retrieved from the Netherlands Twin Register (NTR).
Child and Adolescent Twin Study in Sweden
CATSS is an ongoing, longitudinal, population-based study on somatic and mental health problems in twins during childhood and adolescence. Details can be found elsewhere [19]. Briefly, since 2004, parents of all Swedish twins born from July 1992 and onwards are contacted in connection with the twins’ ninth or 12th birthday (CATSS-9/12) and asked to participate in a telephone interview, which includes measures of, among other things, neurodevelopmental disorders. When the twins reach the age of 15 (CATSS-15), families are contacted again and asked to fill out a web-based questionnaire, targeting various mental health problems and social milieus. CATSS-15 includes twins born from the first of January 1994 up until 2002. In the present study, we included 12,324 twins who completed the self-reports at age 15, and parental report covering 11,133 twins, out of which a total of 85% (self-report) and 87% (parental report) had complete data at age 9/12. A total of 14,106 twins where covered with either self- or parental report.
Measures
Anxiety in adolescence
CATSS-15 includes both the parental and self-report version of the Strengths and Difficulties Questionnaire (SDQ; [20]). The SDQ is a brief behavioral screening questionnaire for children and adolescents between the ages of 3 to 16 years. In the present study, the emotional problems scale, parental and self-report, was used as an indicator of anxiety in adolescence. The scale has 5 items with 3 response options: 0 (not true), 1 (somewhat true), and 2 (certainly true). In addition to dimensional scoring, three categories [20] have been proposed so that roughly 80% of the children fall into a ‘normal’ (in the present study scores between 0 and 2 in the parental report, scores between 0 and 4 in the self-report), 10% into a ‘borderline’ (i.e. score 3 in the parental report, score 5 in the self-report) and 10% into an ‘abnormal’ (i.e. ≥ 4 in the parental report, ≥ 6 in the self-report) category. The emotional problems scale has been reported to have a sensitivity/specificity ranging between 29/96% (youth report) to 54/91% (parent report) for anxiety disorders [21] and correlates strongly with other measures of anxiety in youth [22]. In the present study, the Area Under the Receiving Characteristics Curve for the scale was 0.67/0.69 when comparing parental/self-report to the clinical diagnosis of anxiety disorder from the register mentioned below. Studies of the Swedish version of the SDQ symptom scales have confirmed the factor structure of the original English SDQ [23], and results from a validation study of the Swedish SDQ indicate that the parental version has a good discriminatory validity [24].
Neurodevelopmental disorders
The Autism-Tics, ADHD, and other Comorbidities inventory (A-TAC [25]; is included in CATSS-9/12. The A-TAC is a fully structured 96-item parent-report telephone interview designed for large-scale epidemiological purposes. To screen for neurodevelopmental disorders in childhood, the A-TAC is based on symptom criteria and common clinical features. Items are scored as 1 (yes), 0.5 (yes, to some extent), and 0 (no) and are divided into modules corresponding to diagnostic domains. Cross-sectional and longitudinal validation studies show good to excellent predictive validity for, amongst others, ADHD, ASD, and DCD [26, 27]. In this study, cut-offs ≥6 for ADHD, ≥ 4.5 for ASD, and ≥ 0.5 for DCD with a sensitivity of (0.91/0.91/0.63) and a specificity of (0.73/0.80/0.68) were used as a screening cut-off for ADHD, ASD, and DCD.
Psychiatric outcomes in late adolescence/young adulthood
A personal identification number, which is assigned to all individuals living in Sweden either at birth or when permanently moving to Sweden, enables linkage across health and service registers. The Swedish National Patient Register (NPR) includes all inpatient data from 1987, and outpatient data from 2001 and is coded according to ICD-9 and 10 codes. The validity of the NPR is continuously assessed. Several studies report high validity and reliability of several disorders, e.g., bipolar disorder [28] and psychotic disorders [29]. The following ICD-10 codes were retrieved for all participants and are commonly referred to as ‘psychiatric outcomes’ in this study: F10–19 (alcohol and drug misuse disorders), F20–29 and F30–31 (psychotic disorders and bipolar disorders were merged into one category), F32–39 (depressive disorders), F40–41 (anxiety disorders), and X60–84 (suicidal ideation and suicide attempts) including all subgroups. In our linkage, the NPR was updated until the 31st of December 2014. For each individual, we recorded ICD-10 codes and age at first observed event. Any individual could have multiple outcomes but was still analyzed separately for each outcome. The overlap can be found in Additional file 1: Table S1. Further, we calculated a sum of outcomes, yielding values between 0 and 5.
Sensitivity analysis and attrition
To assess whether missing values affected our estimates and inferences, we used multiple imputation with chained equations to handle missing values in variables, as implemented in the R-package ‘mice’ [30]. Values were imputed using a random forest approach. We analyzed the associations between self- and parental reported anxiety and the 5 separate outcomes, as well as the collapsed, for the Swedish sample. No significant changes were found (Additional file 1: Table S2). Prevalence of psychiatric outcomes in individuals whose parents responded only at age 9/12 but not at age 15, versus those who responded at both assessment waves are characterized in Additional file 1: Table S3. The prevalence differed significantly in responders and non-responders (4.2 and 8.5%, χ2 = 153.56, p < 0.001).
Statistical analyses
We excluded individuals with psychiatric outcomes assigned before the age of 15 (N = 181) from all analyses and used log-linear regression models with the total number of psychiatric diagnoses (alcohol and drug misuse disorders, anxiety disorders, bipolar/psychotic disorders, depressive disorders) and suicidal ideation as the dependent variable, and the total score of anxiety as the independent variable (ranging from 0 to 10). Results are presented as rate ratios (RR), i.e. the increase in rate of the outcome, per unit increase in the exposure. Next, we used Cox proportional hazard regression to regress each of the specific psychiatric outcomes on anxiety (unadjusted model) and excluded individuals who were assigned the respective psychiatric outcome before the age of 15. Results are presented as hazard ratios, i.e. comparing the risk of the outcome in the exposed and unexposed group while accounting for follow-up time. Parental report and self-report in the independent variable were analyzed separately. We entered ADHD, ASD, and DCD cut-offs, as well as sex into the models as covariates in order to test for confounding.
A cluster-robust sandwich estimator was applied to adjust the standard errors for the nested twin data. For our collapsed outcome “any psychiatric outcome”, p < 0.05 was considered statistically significant. For the individual psychiatric outcomes, i.e., alcohol and drug misuse disorders, anxiety disorders, bipolar/psychotic disorders, depressive disorders, and suicidal ideation (unadjusted and adjusted for ADHD, ASD, DCD, and sex, yielding 10 tests), we set the statistical significance threshold at p < 0.005 to account for multiple comparisons. Estimates for anxiety categories ‘borderline’ and ‘abnormal’ were calculated relative to the anxiety category ‘normal’, which served as the reference category.
To account for unmeasured confounders, (a) we conducted a within-twin analysis, which capitalizes on the complete genetic relatedness of monozygotic twin pairs. The design takes genetic and environmental confounding into account by comparing the risk of the outcome in twins differentially exposed (i.e., having different levels of anxiety [31]), (b) we calculated the E-value, which is the lowest point estimate of both the confounder associations and the measured covariates that needs to be present in order for an (unmeasured) confounder to fully explain the association [32].
Survival analyses and attributable fractions
We created survival curves based on the unadjusted model using the R function ‘survfit’ from the package ‘survival’ [33]. We then estimated the total amount, while adjusting for childhood ADHD, ASD, and DCD, as well as sex, of psychiatric outcomes in late adolescence/young adulthood that could be attributed to anxiety by calculating the attributable fraction (AF) using the ‘AFcoxph’ function from the package ‘AF’ [34] in R. The ‘AFcoxph’ function allows the AF to vary over time and thus renders it possible to estimate age-specific AFs. The function estimates the model-based adjusted attributable fraction from the Cox proportional hazard regression model and is commonly used for data from cohort sampling designs with time-to-event outcomes while adjusting for potential confounders. All analyses were performed in R version 3.5.1 [35].
Replication from the Netherlands Twin Register
The Young Netherlands Twin Register (YNTR) and the Adult Netherlands Twin Register (ANTR) have been described in detail elsewhere [36, 37]. The sample included twins who were born the 1st January 1985 to the 31st December 1999 and were part of the YNTR. Nine thousand two hundred eleven twins were assessed at age 14 and then followed up at age 16 and 18. The retention rate varied between 10 and 47%, depending on follow up, giving sample sizes between 914 and 4283. In order to mirror the analyses from CATSS, we used the empirically based anxious/depressed syndrome scale scored from the Youth Self-Report [38] to create the exposure measure. The distribution of anxiety levels was carried out in the same fashion as in the main analyses (80% ‘normal’, 10% ‘borderline’, 10% ‘abnormal’). As outcomes, we used the Achenbach DSM-oriented scales at age 16: anxiety (“anxiety problems”, 6 items), and depression (“affective problems”, 13 items) and 18: for bipolar/psychotic disorders, we used the Adult Self-Report (ASR) Thought Problems scale [39]. For alcohol and drug misuse disorders, a total score of 16 or higher on the World Health Organization’s Alcohol Use Disorders Identification Test [40], implying high-risk/harmful alcohol use, was used. For suicidal ideation as an outcome, we used an individual question from the survey in the ASR indicating suicidal ideation: “I deliberately try to hurt or kill myself”, answered with either “somewhat or sometimes” or “very much so or often”. A dichotomization was conducted at the 90th percentile for all scales to classify individuals as screen-positive for the respective outcome. As covariates, the ten items ASD scales at age 7, 10 and 12, developed by So et al. [41], and the ten items Attention Problems scales at age 7, 10, and 12, developed by Achenbach and Rescorla [38], both from the Child Behavior Checklist, were included. As the outcome variables in the NTR consisted of self-report and we had no access to follow-up time, we used logistic regression models to estimate the association between exposure and outcome and included ADHD and ASD cut-offs, as well as sex, as covariates into the model. Results are presented as odds ratios. A cluster-robust sandwich estimator was applied to adjust for standard errors for the nested twin data. P < .05 was considered statistically significant. For the individual psychiatric outcomes, i.e., high-risk/harmful alcohol use, anxiety, thought problems, depression, and suicidal ideation (unadjusted and adjusted for ADHD, ASD, and sex, yielding 10 tests), we set the statistical significance threshold at p < 0.005 to account for multiple comparisons. Estimates for anxiety categories ‘borderline’ and ‘abnormal’ were calculated relative to the anxiety category ‘normal’, which served as the reference category.