A1000
Title: Stable likelihood-based parameter estimation for semi-parametric accelerated failure time models
Authors: Aishwarya Bhaskaran - Macquarie University (Australia) [presenting]
Jun Ma - Macquarie University (Australia)
Abstract: Cox proportional hazards and accelerated failure time (AFT) models are two principal frameworks for assessing covariate effects in survival analysis. The AFT model is particularly appealing as it offers a straightforward interpretation of how covariates impact the event time. While semiparametric AFT models allow for flexible modelling without specifying the baseline hazard, they pose significant computational challenges. Unlike Cox models, estimation in semiparametric AFT models is markedly more demanding, as the transformation of failure times used to estimate the baseline hazard is dependent on the regression coefficients and must be updated iteratively, often leading to instability or convergence issues. Moreover, incorporating interval censoring complicates the likelihood expression, introducing algebraic challenges in deriving gradients and Hessians, particularly when the intervals are narrow. To address these challenges, a maximum penalized likelihood approach is proposed using rescaling techniques for fitting semiparametric AFT models to data with partly interval-censored failure times. The method employs M-splines to approximate the nonparametric baseline hazard, a constrained optimization framework for stable coefficient estimation, and an automatic smoothing selection criterion that accounts for active constraints. Simulation studies demonstrate the robustness and accuracy of our approach across a range of censoring scenarios.
