ABSTRACT
The outcome event of interest may be a clinically significant morbid event such as death, stroke, myocardial infarction, serious infection, a major organ failure, or tumor progression. The KM product-limit estimator of the survival function and the log-rank test are nonparametric and do not involve model assumptions such as proportional hazards. The graphical method does not work well for continuous covariate or categorical covariates that have many levels because the graph becomes too cluttered. The log-rank test is always a valid test of the null hypothesis that the survival functions of two cohorts are the same. As in ordinary or logistic regression analysis, model diagnostic and influential observation or outlier detection are essential parts of modeling. Some of the covariates used in the Cox regression model may need some transformation, or some regression coefficients may be time-varying to have a better fit or to satisfy the proportionality assumption.
