(PS1-80) A Prediction Model for Social Isolation During the Transition to College: Findings from a Prospective Cohort Study

Background: Mental health disorders are highly common among first-year college students (Auerbach et al., 2018), and social isolation appears to be an important contributor to poor emotional health (Cipolletta et al., 2025). The development of multivariate statistical models that accurately predict levels of social isolation could help identify students at greatest risk and improve allocation of limited prevention resources. This study aimed to develop and internally validate a longitudinal prediction model for social isolation among students making the transition to a new postsecondary institution.

Method: Data were drawn from a prospective cohort study evaluating mental health trajectories of incoming students, including first-time college students and transfer students, during their first semester at a large public university. Participants (N = 276) completed six assessments, one prior to the start of the fall semester and five during the semester (Sep through Dec). Student self-reported social isolation (outcome variable) was measured at all six assessments using the PROMIS Social Isolation Short Form (Hahn et al., 2014). We identified potential predictors (k=26) to include in our model based on prior literature and investigator hypotheses. The predictors were sorted into five thematic groups: structural challenges (e.g., commuting, learning disorder, financial distress), marginalized identity (e.g., race, ethnicity, LGBTQ+), pre-semester mental health (e.g., current/past treatment, depression and anxiety symptoms), early life trauma (maltreatment and household dysfunction), and academic orientation (e.g., grit, intrinsic motivation). To account for within-person dependence in the outcome variable, we fit a first-order proportional odds Markov model. An enhanced bootstrap internal validation procedure (5,000 random draws) was used to calculate optimism-corrected metrics of model discrimination and calibration (C-index, Nagelkerke’s R², and Somers’ D_xy). Penalized maximum likelihood estimation was used to avoid model overfitting. To avoid bias due to missing follow-up data, we fit models and aggregated results across 25 multiply imputed data sets.

Results: Model-based predictions performed reasonably well in ranking students by social isolation severity even after correcting for potential overfitting: C-index=0.75, R²=.40, D_xy=.49. Composite tests showed unique contributions of variable groups representing structural challenges (x2[df=23]=58.23, p< .001) and pre-semester mental health (x2[df=27]=73.74, p< .001). There was insufficient evidence to conclude that variable groups representing marginalized identity, early life trauma, and academic orientation made significant contributions.

Conclusion: Our prediction model shows promise in improving identification of students with highest levels of isolation during the college transition. Future validation and refinement in independent samples is needed before this prediction tool can be implemented in real-world settings.