A Model Selection Approach for Variable Selection with Censored Data

We consider the variable selection problem when response is subject to censoring. A main particularity of this context that information content sampled units varies depending on censoring times. Our approach based model where all 2k possible models are entertained and we adopt an objective Bayesian perspective choice prior distributions a delicate issue given well-known sensitivity Bayes factors these inputs. show borrowing priors from ‘uncensored’ literature may lead unsatisfactory results as default procedure implicitly assumes uniform contribution independently their In paper, develop specific methodology generalization g-priors, explicitly addressing particularities survival problems arguing it behaves comparatively better than standard approaches basis arguments (like e.g. predictive matching) in particular case accelerated failure time with lognormal errors. apply recent large epidemiological study about breast cancer rates Castellón, province Spain.