Sparse models for predicting psychosocial impairments in patients with PTSD: An empirical Bayes approach

Posttraumatic stress disorder (PTSD) is associated with psychosocial impairments, which represent a relevant focus for therapy. Previous results on the clinical predictors of these psychosocial impairments were inconsistent. The data analyzed in these contexts often suffer from a high number of correlated predictors and small sample sizes, entailing the risk of model overfitting. In Bayesian regression, the problem of overfitting can be mitigated by usage of specific zero-centered (regularizing) prior distributions. In this study, we used the 2 most common Bayesian regression models, the Bayesian Ridge and the Bayesian Lasso, to predict psychosocial impairments in 192 patients of a day clinic for the treatment of PTSD.

Predictions were based on specific dimensions of PTSD symptoms previously revealed by factor analyses, as well as posttraumatic cognitions, depressive symptoms, comorbid disorders, and demographics. The variance of the prior distribution was estimated through empirical Bayes (maximum marginal likelihood) and an approximation to the posterior distribution was obtained with stochastic variational inference and with a local approximation (Laplace approximation).

Severe psychosocial impairments were mainly related to depressive symptoms and symptoms from the amnesia and numbing dimension of PTSD, while gender, posttraumatic cognitions, and reexperience and avoidance symptoms had no impact. As expected, the model coefficients were shrunken to zero when regularizing prior distributions were used, particularly for the Bayesian Lasso.

Depressive and numbing symptoms are the main clinical correlates of psychosocial impairments in patients with PTSD. Usage of Bayesian and regularized regression can contribute to the generalizability and interpretability of research results. (PsycInfo Database Record (c) 2023 APA, all rights reserved).