J Interpers Violence. 2025 Jul 18:8862605251350127. doi: 10.1177/08862605251350127. Online ahead of print.
ABSTRACT
To align with emerging policies for adolescents, feasible, accurate, and equitable trauma-focused assessment protocols need to be developed. To date, most research on this topic has focused on whether traditional adverse childhood experiences (i.e., maltreatment, impaired caregiving) can adequately index mental health risk. Yet, there are noted clinical and statistical drawbacks to this approach. Instead, examining threat and reward biases, two subtypes of cognitive biases stemming from interpersonal trauma exposure, may provide a reasonable alternative to adversity screening. Thus, the aim of this study was to examine the accuracy and fairness of self-reported, trauma-informed cognitive vulnerabilities for classifying concurrent and prospective adolescent mental health risk relative to more commonly assessed childhood adversities. In a diverse adolescent sample (N = 584; MAge = 14.43; 48.9% female; 35% African American; 38.5% White; 40% Hispanic) youth completed measures for adversity exposure (family, dating, and community violence), threat biases (posttraumatic cognitions, hostility), and reward biases (anticipatory, consummatory) during an initial assessment, as well as symptoms of posttraumatic stress (PTS), depression, and violent behavior at baseline and 1 year later. Indices of statistical discrimination, calibration, and statistical fairness were examined using an evidence-based medicine analytic approach, which was subsequently compared to a machine learning approach. Overall, posttraumatic cognitions emerged as an accurate and statistically fair predictor of prospective PTS (area under the curve [AUC]95% CI = [0.63, 0.78]; diagnostic likelihood ratio [DLR]95% CI = [1.32, 3.52]), and to a lesser extent depression (AUC95% CI = [0.56, 0.70]; DLR95% CI = [1.25, 2.98]), and both models were well calibrated (i.e., p-value >05 for Spiegelhalter’s Z test). Meanwhile, community violence (CV) exposure best classified the risk for prospective violent behavior (AUC95% CI = [0.62, 0.73]; DLR95% CI = [2.68, 5.49]), especially in males, and was well calibrated. The machine learning algorithms added limited incremental validity to our predictions. Our study suggests that focusing on posttraumatic cognitions and less invasive adversity items (i.e., CV exposure) may lead to trauma screening and assessment protocols that are accurate, equitable, and feasible to implement within applied settings serving diverse youth.
PMID:40682318 | DOI:10.1177/08862605251350127
