Title: Variable selection in the power-generalized Weibull distributional regression model
Authors: Laura McQuaid - University of Limerick (Ireland) [presenting]
Shirin Moghaddam - University of Limerick (Ireland)
Kevin Burke - University of Limerick (Ireland)
Abstract: In medical applications, non-proportional hazards are often encountered. In these scenarios, standard survival modeling techniques are not appropriate. Instead, we propose using distributional regression where covariates enter the hazard function via multiple distributional parameters (e.g., scale and shape) simultaneously. This allows for more complex covariate effects, such as time-varying hazard ratios, to be captured. We develop the adapted power-generalized Weibull (APGW) distributional regression model, which, with three parameters (one scale, two shapes), encompasses various common survival models (Weibull, log-logistic, Gompertz) and hazard shapes (constant, increasing, decreasing, up then down, down then up), making it a highly flexible model. Variable selection is challenging in this setting (and distributional regression more generally) since covariates can enter the model in various ways. Thus, we propose the use of a computationally feasible adaptive lasso penalized estimation procedure for variable selection and explore its performance using numerical studies and real-world data application.