With covariates $X$ and parameters $\beta$, the Cox proportional hazard model assumption is that the hazard function is $\lambda(t|X) = \lambda_0(t)\exp(X\beta)$, where $\lambda_0(t)$ is baseline hazard function.

Question is why $\exp(X\beta)$? Is this an arbitrary choice? Why not like $(X\beta)^2$ or some such thing?


1 Answer 1


Consider the log hazard, $\eta(t|X)=\eta_0(t) + X\beta$

and coding different groups using indicator variables with reference-cell coding.

Note then that $\eta_0(t)$ is the log-hazard function for the reference group and that the other groups have a log-hazard that is shifted up or down by some $\beta$ (or combination of $\beta$s).

A shift in log-hazard is the same as a scaling of the hazard function by a multiplicative constant, which is exactly what the name "proportional hazards" implies (that the groups' hazards are simply scaled versions of each other).

So this formulation using the $\exp$ is the natural way to get the log-hazards to be shifted (i.e. proportional hazard).

Now it is not necessarily the case that the model should be linear in $X$ on the log-hazard scale, that's an assumption, just as it would be in say a GLM (where it's not necessarily the case that the link-transformed mean $\eta=g(\mu)$ is linear in $X$, but it's an assumption that is often made).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.