(PS9-75) Using Machine Learning to Predict Perseverative Thought in Daily Life

Abstract Text: Perseverative thinking (PT), or repetitive negative thinking, plays an important role in the onset and maintenance of many disorders (Ehring & Watkins, 2008). Understanding the naturalistic contexts in which PT emerges is critical for refining theoretical models and designing effective interventions for this impairing cognitive process. However, research on precipitants of PT in the natural environment has had two major limitations. First, studies have typically focused on a narrow form of PT, such as rumination or worry, instead of on the overarching PT construct that is increasingly recognized as the core pathological mechanism across disorders (McEvoy et al., 2013). Second, studies have typically focused on one or a few predictors in isolation, hampering the development of comprehensive theoretical models of PT. To address these gaps, we used machine learning to test a large set of theoretically relevant predictors of PT in daily life.

To recruit participants who varied widely in PT, we enrolled a mixed sample of 162 adults with generalized anxiety disorder, major depressive disorder, or no lifetime psychopathology. They were signaled to complete 9 daily assessments over a 7-day ecological momentary assessment (EMA) study, yielding 8,587 surveys. Each survey assessed momentary PT at the time of the signal (4-item measure; within-person ω = .77–.80) alongside a wide range of potential predictors. We used multilevel machine learning to model the relationships between these 150 predictors and momentary PT. Model performance was evaluated by calculating R² on out-of-sample data, and predictor importance was quantified using Shapley Additive Explanations (SHAP) values and R² decomposition based on Gelman’s Bayesian R² (Redell, 2019).

The machine learning model demonstrated strong predictive accuracy, achieving an out-of-sample R² of .61 [95% CI: .59, .62]. The participant’s average level of momentary PT, operationalized by the random effects portion of the model, accounted for the majority of explained variance (R² =.42 [.40, .45]). Nevertheless, EMA features explained a substantial portion of the variance in PT (R² = .18 [.17, .19]). SHAP analyses and R² decomposition identified a number of significant predictors of PT (category R² and top predictors shown in parentheses), including characteristics of recent events (R² = .12; event valence, stressfulness) and anticipated events (R² = .03; event stressfulness, threat), prior PT levels (R² =.04; rumination, worry), recent event-related reactions (R² = .01; event-related PT and sadness), prior negative affect (R² = .005; sad, stressed), and prior night’s sleep quality (R² = .002). Other features did not meaningfully predict PT beyond the aforementioned predictors: time of day, prior positive affect and cognitions, prior approach and avoidance behaviors, and prior night’s depression and anxiety symptoms.

Our findings reveal the relative importance of various contextual features in predicting PT. They also demonstrate that theoretically derived EMA features cumulatively explain a substantial portion of the variance in PT, opening the door to just-in-time interventions aimed at preventing PT in daily life. Implications for theory and intervention development will be discussed.