Abstract:
|
The lifetime risk measures the cumulative risk for developing a disease over the lifespan. Models for the lifetime risk must account for delayed entry, the semi-competing risk of death, and inference at a fixed timepoint. Covariates may be associated with the cumulative incidence function (CIF) but not the lifetime risk. One can model the CIF with a Fine-Gray (FG) or Royston-Parmar (RP) model and predict the lifetime risk at a fixed timepoint, but there are no regression models for the lifetime risk. We introduce a novel model for the lifetime risk using pseudo-observations of the Aalen-Johansen estimator at a fixed timepoint, allowing delayed entry. We regress the logit-transformed pseudo-observations on covariates with a generalized linear model. We compared the performance of our method with the FG and RP models in simulations with non-proportional subdistribution hazards. We estimated the bias, relative bias, mean squared error, and coverage. Our model outperformed the FG and RP models in scenarios with smaller sample size, greater delayed entry, and greater right censoring. We illustrate our model for the lifetime risk of atrial fibrillation in the Framingham Heart Study.
|