Introduction: A common issue in internet interventions for mental health conditions is that many patients do not respond to treatment. This has led to a line of research that focuses on finding predictive and moderating factors of treatment outcomes with the goal to improve our understanding of these heterogeneous treatment effects. However, this literature often finds inconsistent results between studies making the findings difficult to implement in clinical practice. Drawing on methods from the from the causal inference literature could improve consistency and impact.
Methods and results: Using causal directed acyclic graphs (DAGs), we show that results from common statistical analyses used in this line of research are sensitive to the underlying causal structure of the data, under some realistic conditions. We use collider bias as an illustrative example. Results from simulation will be presented that show how inconsistent results can appear depending on how covariates are chosen.
Conclusions: Research on predictors and moderators of outcomes in internet interventions would benefit from applying methods from causal inference. Recommendations for future research are to clarify the intended aim of the analyses, precisely define what is to be estimated (the estimand), and justify the chosen statistical analyses using causal reasoning.