Presentation (CTI)

OBJECTIVES: IPD meta-analysis of competing risks data enables pooling of individual-level information across outcomes and trials to precisely estimate treatment effects, adjust for covariates, and support extrapolation beyond observed follow-up. Existing frameworks rely on fixed-effect only, semi-parametric models that limit plausibility. We propose a general frequentist and Bayesian framework for one-step IPD meta-analysis of competing risks data using flexible parametric cause-specific hazard models.
METHODS: We extend the one-step stratified IPD meta-analysis approach of Meddis et al. (Biometrical Journal, 2020) to accommodate flexible parametric cause-specific hazard models within a unified framework. By modelling each cause separately, we isolate treatment effects on each competing event. Study heterogeneity is captured through stratified baseline hazards with optional random treatment effects. Treatment-covariate interactions and non-proportional hazards are estimable within the same model. Bayesian estimation is also considered, enabling prior incorporation across the cause-specific hazard structure and full posterior uncertainty propagation. Standardised predictions allow population-level inference on cause-specific hazard and cumulative incidence functions, and marginal treatment comparisons. Claude Code (Anthropic) assisted in software development; all AI contributions were reviewed and approved by the author.
RESULTS: We illustrate the framework using data from 23 trials in patients with nasopharyngeal carcinoma, comparing the effect of chemotherapy added to radiotherapy across four scheduling modalities - adjuvant, concomitant, concomitant plus adjuvant, and induction. Loco-regional failure was the primary endpoint with distant relapse and death without failure as competing events. Flexible parametric models offer improved fit over semi-parametric alternatives, smooth cumulative incidence function estimation, and seamless extrapolation beyond trial follow-up.
CONCLUSIONS: A flexible parametric approach to IPD meta-analysis of competing risks data enables robust cause-specific estimation, clinically interpretable predictions, and Bayesian uncertainty quantification - addressing key limitations of existing methods. Open-source software is freely available at https://github.com/RedDoorAnalytics.