Predicting completion and early dropout in routine-care internet-delivered cognitive behaviour therapy: A temporal validation study

Abstract

Background In routine-care Internet-delivered cognitive behaviour therapy (ICBT), engagement may be better understood as a staged process (e.g., early discontinuation vs full completion), yet many studies use a single dropout definition and seldom test whether prediction models remain useful as services evolve over time. Objective To identify baseline predictors of completion and early dropout in routine-care ICBT and evaluate temporal transportability from earlier to later service years. Methods We analyzed de-identified routine-care data (N = 9057) from a transdiagnostic ICBT program delivered from 2013 to 2024. Outcomes were completion (all five lessons vs fewer) and early dropout (0–1 lesson vs two or more). Multivariable logistic regression models used baseline demographic, clinical, and contextual intake variables. Models were developed in TRAIN (2013–2021) and applied without refitting to TEST (2022–2024). TEST performance was summarized using Brier score, area under the curve (AUC), and calibration. Results Demographic factors, including age, education, ethnicity, and relationship status, were more consistently associated with engagement than baseline symptom severity. Completion predictions showed modest discrimination and modest overprediction in TEST, whereas early-dropout predictions showed weaker discrimination, slight underprediction, and limited separation of higher- and lower-risk clients. Conclusions Models based on screening variables showed higher discrimination for completion than early dropout. Engagement prediction in routine-care ICBT should be temporally evaluated and monitored for calibration drift. Future research should examine whether early indicators during treatment, such as initial logins, lesson access, and early therapist contact, improve prediction of early dropout.