A1435
Title: A fully Bayesian approach to structural nested failure time models for longitudinal causal inference
Authors: Liangyuan Hu - Rutgers University (United States) [presenting]
Abstract: Structural nested failure time models (SNFTMs) offer a robust alternative to marginal structural models for causal inference with time-varying treatments, notably avoiding the strict positivity assumption. However, their widespread adoption is hindered by the computational intractability of frequentist g-estimation, where rank-based objective functions create a non-differentiable staircase topology that precludes standard gradient-based optimization. A Bayesian framework is proposed that fundamentally resolves this instability. By specifying a parametric baseline counterfactual distribution, the discontinuous objective is transformed into a smooth posterior surface, enabling efficient estimation via standard MCMC algorithms without the need for artificial smoothing parameters. The posterior consistency and asymptotic normality of the resulting estimators are theoretically established. Simulation studies demonstrate that this Bayesian approach yields superior stability and coverage compared to frequentist g-estimation, particularly in small-sample and high-censoring regimes. Finally, the method is applied to the CARRA Registry to evaluate the impact of adalimumab withdrawal strategies on the timing of disease flare in juvenile idiopathic arthritis. By avoiding the data wastage associated with frequentist artificial recensoring, this framework efficiently leverages the full follow-up period to recover precise estimates of the protective effects of maintenance therapies.
