 |
Another way to account for possibly
non-proportional hazards is to represent the variation of the covariate effect with time
by a time-by-covariate interaction. As suggested by Cox in his original paper, the
disease-related mortality hazard function can be expressed as given here where f(t) is a
pre-specified function of time, z2 is the covariate involved in this
interaction and g 2 is the regression coefficient
for this interaction. If g 2 is set to 0, the model
reduces to the conventional relative survival model. We consider here only two classic
functions of time, linear or logarithmic. It should be noted that they both impose the
monotonicity constraint. |
