Presentation (CTI)

OBJECTIVES: Real-world data from electronic health records, patient registries, or wearable devices are increasingly used alongside randomised controlled trials to estimate the cost-effectiveness of healthcare interventions in heterogeneous routine-care populations. However, treatment switching, often associated with time-varying confounding, and missing data, can bias estimates of costs and effects if not handled appropriately, and guidance on methods to address these challenges simultaneously remains limited. We compared causal inference approaches for handling treatment switching, combined with different approaches to handling missing data, when estimating incremental costs, incremental effectiveness, and net monetary benefit.
METHODS: We conducted a simulation study informed by a large empirical dataset of adults with a primary diagnosis of depression from a Dutch mental healthcare institution. Across 2,000 datasets (n=500 each), we simulated treatment switching and 10%-50% missing outcomes. We compared marginal structural models, g-computation, and longitudinal targeted maximum likelihood estimation (LTMLE), combined with missing-data approaches including multiple imputation using predictive mean matching, classification and regression trees, or random forest, and inverse probability of censoring weighting. Performance was evaluated using bias, root mean squared error, and confidence-interval coverage.
RESULTS: LTMLE with predictive mean matching consistently produced low bias (<5%) and coverage closest to nominal levels (0.88-0.94) across scenarios. LTMLE with classification and regression trees or random forest performed well in several scenarios up to 25% missingness, but bias increased at 50% missingness. G-computation and marginal structural models performed well in selected scenarios, particularly for incremental clinical effectiveness, but were less reliable for incremental costs and net monetary benefit.
CONCLUSIONS: In real-world cost-effectiveness studies affected by treatment switching and missing data, LTMLE combined with multiple imputation using predictive mean matching provided the most robust overall performance. Broader application of this strategy may improve the validity and reliability of applied economic evaluations using real-world data.