Association of Genetic and Phenotypic Assessments With Onset of Disordered Eating Behaviors and Comorbid Mental Health Problems Among Adolescents

Key Points Question Do associations exist between disordered eating behaviors and other mental health disorders in adolescence, and if so, to what extent are those associations genetically predisposed? Findings Longitudinal assessments in this cohort study of a population-based sample of 1623 adolescents indicated that body mass index (BMI), neuroticism, impulse control, and addiction-related behaviors at 14 years of age were differentially associated with future disordered eating behaviors and symptoms of depression and generalized anxiety. Genetic analyses suggested etiologic overlaps between BMI, neuroticism, and attention-deficit/hyperactivity disorder with dieting, binge eating, and purging, respectively. Meaning Genetic and phenotypic assessments of BMI, impulse control problems, and personality may inform early and differential diagnoses of eating disorders.

Disorders, Specific Phobias and Hyperkinesis computer prediction scores were excluded due to their low prevalence in this population. All DAWBA prediction scores were available at all three time-points, aside from ADHD and Separation Anxiety which were reported only at baseline and FU1 assessments. We derived computer-predicted DAWBA diagnoses by combining DAWBA scores 0-2 (probability of having the relevant diagnosis <1%) as 'healthy controls' and levels 3-5 (probability of having the diagnosis >15%) as 'cases'.

Development of mental health problems:
We used longitudinal assessments of DAWBA diagnoses, as defined above, to identify groups of individuals likely to develop mental health disorders over time. We defined 'developers'' as individuals identified as a 'case' (e.g., with a predicted diagnosis of depression) at age 16 or 19, but not at baseline (14y).
Addiction-related behaviors: These were assessed with two questionnaires, the European School Survey Project on Alcohol and Drugs (ESPAD 6 ) and the Alcohol Use Disorders Identification Test (AUDIT 7 ). The primary questions of interest regarding the ESPAD were lifetime binge drinking (> 3 occasions), drug use ever and smoking (past month). Binge drinking was a categorical variable based on reporting 3 or more episodes of drunkenness in the past year, versus no episodes of drunkenness (On how many occasions (if any) have you been drunk from drinking alcoholic beverages?). Drug-use was a categorical variable and defined as any drug intake in the past year. Smoking was also a categorical variable with 'cases' defined as those who have smoked in the past month, and 'controls' those who have never smoked. In the case of categorical variables, drug-use, binge drinking and smoking 'developer' groups were defined as individuals who were 'cases' at age 16 or 19, but not as baseline (14y).
The AUDIT was used to assess alcohol dependence as a continuous variable using the total AUDIT score.

Longitudinal Trajectory analysis for personalities
We hypothesised that DEBs had an impact on how personalities change over time. To test this, the slopes of the trajectories for personalities were compared between the DEB developers and controls by using linear mixed-effect regression models, implemented in the lmer4 [36] and lmerTest [37] packages of the R software (v3.5.0).The age variable was coded as 0, 2 and 5 to reflect the gap between phases (14y, 16y and 19y). Both the fixed effect of the group (developers vs. controls) and the random effect of participant were modelled on the intercept and the slop of the trajectories, controlling for gender and sites. The effect of interest was the effect of group on the slope, namely, the group×age interaction.

Acquisition of Genetic Data
DNA purification and genotyping were performed by the Centre National de Génotypage in Paris. DNA was purified from whole blood samples (~10 ml) preserved in BD Vacutainer EDTA tubes (Becton, Dickinson and Company) using Gentra Puregene Blood Kit (QIAGEN) according to the manufacturer's instructions. Genotype information was collected at 582,982 markers using the Illumina HumanHap610 Genotyping BeadChip. Single nucleotide polymorphisms with call rates of < 98%, minor allele frequency < 2% or deviation from the Hardy-Weinberg equilibrium (p ≤ 1×10 -4 ) were excluded from the analyses. Individuals with an ambiguous sex code, excessive missing genotypes (failure rate > 2%), and outlying heterozygosity (heterozygosity rate 3 standard deviations from the mean) were also excluded.
Identity-by-state similarity was used to estimate cryptic relatedness for individual using PLINK software 8 . Closely related individuals with identity-by-descent (IBD > 0.1875) were eliminated from the subsequent analysis. Population stratification for the genomewide association study was examined using multidimensional scaling (MDS) analysis with HapMap populations as reference groups. Individuals with divergent ancestry (from Utah residents with ancestry from northern and western Europe) were excluded through visual inspection of the first 2 components.

Polygenic Risk Scores calculation
Polygenic Risk Scores (PRS) were calculated in the target genotypes to be a sum of risk alleles across SNPs with the p-values below a given threshold, weighted by the effect sizes obtained from the GWAS summary statistics. PRSice software (v1.2.5) 9 was used to obtain the PRSs.
Clumping was performed to remove linkage disequilibrium (LD) using default PRSice parameters, i.e., r 2 threshold=0.1, window size = 250kb. A regression model was performed to test the association between the PRSs and the target phenotype, involving sex, acquisition sites and population stratification (the first four MDS components of the genotype data) as covariates. Model fit (R 2 ) were estimated by subtracting the R 2 for the null model (involving the covariates only) from that for the full model. The p-value thresholds varied from 0.01 to 0.5 in a step of 0.01, in order to achieve the best model fit. PRSs were derived from the bestfit model.
We used mediation models to investigate whether neuroticism (BL), BMI (BL) and ADHD (SDQ hyperactivity FU1) phenotype mediated the relationship between PRS for neuroticism, BMI and ADHD and DEBs. Control variables included sex, acquisition site and population stratification. The continuous variables were transformed to z-scores. Confidence intervals for the indirect effect were estimated with 5000 bootstrap samples by using the PROCESS macro (v3.2, http://processmacro.org) in SPSS (v25, IBM Corporation).

Longitudinal analyses of predictors and outcomes of DEBs across adolescence: Generalised
Estimating Equations (GEE) using DEBs as time-varying predictors and logistic and linear regression models were used to investigate the predictors, correlates and outcomes of DEBs.
All continuous variables were standardised and their statistical output reported as z-scores. The odds of presenting with each of the outcomes were estimated using an unstructured working correlation structure and a robust estimation of standard errors. All analyses were adjusted for covariates including sex and study site in stepwise regression models. * Indicates that differences between DEB groups and healthy controls are statistically significant to p<0.05 when controlling for covariates including gender and study site.. Gender was not included as a covariate when gender differences were investigated. Dietary restraint, eating concern, weight concern and shape concern were included in the same linear regression model. EDEQ global score was investigated in isolation.