Background: Prior research has demonstrated associations between anabolic-androgenic steroid (AAS) use and features from several childhood and adolescent psychosocial domains including body image concerns, antisocial traits, and low levels of parental care. However, prior approaches have been limited by their focus on individual features and lack of consideration of the relevant causal structure.
Methods: We re-analyzed data from a previous cross-sectional cohort study of 232 male weightlifters aged 18-40, of whom 101 had used AAS. These men completed retrospective measures of features from their childhood and early adolescence, including body image concerns, eating disorder psychopathology, antisocial traits, substance use, and family relationships. Using an approach informed by principles of causal inference, we applied four machine-learning methods - lasso regression, elastic net regression, random forests, and gradient boosting - to predict AAS use.
Results: The four methods yielded similar receiver operating curves, mean area under the curve (range 0.66 to 0.72), and sets of highly important features. Features related to adolescent body image concerns (especially muscle dysmorphia symptoms) were the strongest predictors. Other important features were adolescent rebellious behaviors; adolescent feelings of ineffectiveness and lack of interoceptive awareness; and low levels of paternal care.
Conclusions: Applying machine learning within a causally informed approach to re-analyze data from a prior study of weightlifters, we identified six factors (most prominently those related to adolescent body image concerns) as proposed causal factors for the development of AAS use. Compared with the prior analyses, this approach achieved greater methodologic rigor and yielded stronger and broader findings.
Keywords: Anabolic-androgenic steroids; Body-image disorders; Causal inference; Eating disorders; Machine learning; Risk factors.
© 2023 The Author(s).