A1353
Title: Causal inference for tapering strategies in deprescribing using target trial emulation: A simulation study
Authors: Ryan Muddiman - RCSI (Ireland)
Teresa Perez-Perez - Complutense University of Madrid (Spain) [presenting]
Elisa Medrano - Complutense University of Madrid (Spain)
Elliot McClenaghan - London School of Hygiene and Tropical Medicine (United Kingdom)
John Tazare - London School of Hygiene and Tropical Medicine (United Kingdom)
Fiona Boland - RCSI University of Medicine and Health Sciences (Ireland)
Emma Wallace - University College Cork (Ireland)
Tom Fahey - RCSI University of Medicine and Health Sciences (Ireland)
Li Wei - University College London (United Kingdom)
Carmel Hughes - Queen University Belfast (United Kingdom)
Ann Doherty - RCSI University of Medicine and Health Sciences (Ireland)
Frank Moriarty - RCSI University of Medicine and Health Sciences (Ireland)
Abstract: Dynamic treatment regimens for deprescribing (reducing or stopping medicines that may no longer be beneficial) are common in clinical practice but are difficult to evaluate using randomized or observational data. Such settings involve irregular visits, dose changes over time, and withdrawal events that depend on patient history. A target trial emulation framework was used to define and compare a small number of clinically relevant tapering strategies (gradual medication withdrawal) against continuation. Synthetic data were generated using a multi-state model to represent visit times and clinical events, where treatment changes are only allowed at visit times. The data generating process was divided into components for baseline covariates, visit processes, dose changes, withdrawal events, and the outcome. The estimand is the average causal effect of each tapering strategy compared with continuation over a fixed follow-up period. Performance was assessed in scenarios reflecting challenges commonly seen in real-world data, including limited overlap between groups (non-positivity) and misspecification of the underlying statistical model. The parametric g-formula and inverse probability weighting using marginal structural models were compared, with targeted maximum likelihood estimation as a sensitivity analysis. The evaluation focused on bias in the estimated treatment effect, empirical standard error, root mean squared error, and coverage.
