Predicting heterogeneous treatment effects of an Internet-based depression intervention for patients with chronic back pain: Secondary analysis of two randomized controlled trials

Depression is highly prevalent among individuals with chronic back pain. Internet-based interventions can be effective in treating and preventing depression in this patient group, but it is unclear who benefits most from this intervention format. In an analysis of two randomized trials (N = 504), we explored ways to predict heterogeneous treatment effects of an Internet-based depression intervention for patients with chronic back pain. Univariate treatment-moderator interactions were explored in a first step. Multilevel model-based recursive partitioning was then applied to develop a decision tree model predicting individualized treatment benefits. The average effect on depressive symptoms was d = −0.43 (95 % CI: −0.68 to –0.17; 9 weeks; PHQ-9). Using univariate models, only back pain medication intake was detected as an effect moderator, predicting higher effects. More complex interactions were found using recursive partitioning, resulting in a final decision tree with six terminal nodes. The model explained a large amount of variation (bootstrap-bias-corrected R2 = 45 %), with predicted subgroup-conditional effects ranging from di = 0.24 to −1.31. External validation in a pilot trial among patients on sick leave (N = 76; R2 = 33 %) pointed to the transportability of the model. The studied intervention is effective in reducing depressive symptoms, but not among all chronic back pain patients. Predictions of the multivariate tree learning model suggest a pattern in which patients with moderate depression and relatively low pain self-efficacy benefit most, while no benefits arise when patients' self-efficacy is already high. If corroborated in further studies, the developed tree algorithm could serve as a practical decision-making tool. •Effect moderators of a depression intervention for pain patients were explored.•Multilevel tree learning was used to create a precision treatment rule.•Predicted treatment effects varied considerably, from di = 0.24 to −1.31.•The resulting tree model could be useful as a practical decision-making tool.