|
Abstract:
|
An important task in survival analysis is correctly structuring the relationship between covariates of interest and the time-to-event outcome. For example, the Cox proportional hazards (PH) model structures each covariate effect as a constant multiplicative shift in the hazard across time. Likewise, the accelerated failure time (AFT) model structures each covariate effect as a common multiplicative shift in the survivor distribution, across all survival percentiles. The PH structure is often scrutinized in practice via graphical methods and hypothesis testing, and deviations can be addressed by incorporating time-dependent effects. However, comparatively less attention has been paid to the AFT structure, and the potential for deviations from the common effect across survival percentiles. We address this issue via an AFT model structure which allows covariate effects to vary across survival percentiles. We derive methods for selecting, estimating, and graphically presenting these effects. We also demonstrate the utility of this modeling structure in a real data application.
|
