|
Abstract:
|
The hazard ratio is commonly used for comparing survival distributions. While easily estimated in the presence of censored data, it does not allow for the clinical relevance of differences in survival across groups to be easily judged. We consider an approach to nonparametric inference for clinically meaningful functionals of a survivor distribution (e.g., restricted mean, quantiles). In this approach we use different models to borrow information across sparse data than to form contrasts. Linear contrasts are evaluated and compared on mean squared error between approaches using nonparametric recursive partitioning, Cox's proportional hazards, and Buckley-James' linear regression with censored data. The nonparametric approach was superior to the semiparametric approaches when semiparametric model assumptions were violated, and had a slight loss of efficiency when they do hold.
|
