Background: There are large health and economic costs associated with attrition from psychological services. However, predicting treatment dropout is a major challenge in psychotherapy research and recent findings on predictors are rather heterogeneous. Recent approaches relied on single measurements of patients’ pre-treatment characteristics. Instead of using cross-sectional data only, the prediction of attrition might be improved by intensive longitudinal data (ILD) measured before the onset of treatment. The recently emerged, innovative statistical tool of network analysis is able to identify the centrality of symptoms based on ILD that might have predictive power over and above intake variables.

Objectives: To compute individual dynamic symptom networks before treatment onset, to extract centrality measures for each symptom and to use them as additional predictor variables for treatment dropout.

Research question: In the present study, we examined whether centrality measures of nodes in complex symptom networks can improve the prediction of treatment dropout. The predictive power of dynamic network analyses based on ILD was compared to that of patients’ intake variables measured at only one point in time.

Method: Fifty-eight patients undergoing psychological treatment for mood or anxiety disorders were assessed using Ecological Momentary Assessments (EMA) four times a day for two weeks before treatment (3,248 measurements). Multilevel vector autoregressive models were employed to compute dynamic symptom networks. The network of patients who dropped out of treatment was compared to the network of treatment completers using bonferroni corrected two-tailed t-tests. Measures of centrality were extracted from the network of all patients. Using random forest and LASSO approaches (least absolute shrinkage and selection operator), the most promising predictors of dropout were selected out of seven intake variables and 61 centrality measures separately. Resulting predictors of both LASSO models were entered hierarchically into a logitic regression model to examine incremental variance explanation in dropout. Final prediction models were evaluated via area under the curve, explained variance and confusion matrices.

Results: Networks for patients differed significantly between completers and dropouts. Among intake variables, initial impairment and sex predicted dropout explaining 6% of the variance. Thrity-six patients were classified correctly. The LASSO identified four additional predictors from the network analysis: Expected force of being excited, outstrength of experiencing social support, betweenness of feeling nervous, and instrength of being active. The final model with the two intake and four network variables explained 32% of variance in dropout and identified 47 out of 58 patients correctly (AUC = 0.85).

Conclusions and Implications: The findings indicate that patients’ dynamic network structures may improve the prediction of dropout. Four centrality measures predicted dropout over and above patients’ intake variables. This kind of research can lead to direct applications to personalized and patient specific predictions of dropout in clinical practice. Implemented in an app or webbased application, a patient’s individual dropout probability could be estimated even before treatment onset, being highly valuable to clinicians. The findings of this study will be discussed in the context of the Trier Treatment Navigator, a recently developed comprehensive feedback and decision support system for psychotherapists.