I've been trying to find a mathematical explanation for this. Note, I'm talking about the "classic" Cox-PH, not the extensions given after to incorporate interval/left censoring.

In this paper it is claimed that:

"Incorporation of interval censoring into the proportional hazards model does not enable canceling of the baseline hazard function".

Can anyone show this? Is this even correct?

From the definition of the partial likelihood as given by cox - it seems that left censored data will simply be ignored, and interval will be taken to be right censored with only the left part of the interval. So, it seems that Cox could still work, but will probably not be very accurate, as it will ignore some data.


1 Answer 1


The calculation of the Cox regression coefficients depends on knowing the risk set at each event time. For convenience, I'm copying some relevant parts of this answer here for the situation without tied event times; $dN_i(s)=1$ represents an event at observation time $s$:

Differentiating the log partial likelihood with respect to the parameter-value vector $\theta$ gives a score vector (Equation 3.4 of Therneau and Grambsch):

$$ U(\theta) = \sum_{i=1}^n \int_0^{\infty} \left[X_i(s) - \bar x(\theta,s)\right] dN_i(s) = \sum_{i=1}^n U_i(\theta)$$

where $X_i$ represents the covariate values for case $i$ and $\bar x$ is a risk-weighted mean of $X$ over observations at risk.* The maximum partial likelihood estimator $\hat \theta$ solves:

$$\sum_{i=1}^n U_i(\hat \theta) = 0. $$

The critical point is that you need to know the composition and characteristics of the risk set at each event time in the data. If one individual's event occurs in the middle of another individual's interval-censored event time, then you don't know whether that second individual should be included in the risk set for the first individual's event.

...it seems that Cox could still work, but will probably not be very accurate, as it will ignore some data.

Ignoring those data is the point. Your scenario treats an interval-censored observation as irrelevant at the beginning of the interval, and treats it as no event at the end of the interval. Yet for an interval-censored observation, we know that an event occurred during that time interval. Replacing true events with non-events will not lead to a proper model.

A common work-around with interval-censored data is to assign the event time to one or the other end of the interval, typically the end of the interval. With the latter choice you have a correct model of the time that the event was observed rather than when it occurred.

That's often good enough for a particular application. For example, cancer recurrence times are usually interval censored, with no recurrence seen at one clinical visit while a recurrence was seen at the subsequent visit. So the recurrence occurred within that time interval. Although there are ways to incorporate interval censoring into a true partial likelihood calculation for a Cox model (D. Finkelstein, Biometrics 42: 845-854, 1986), modeling the observation time instead of the occurrence time is frequent practice.

*The risk score for case $i$ in a regression without case weights is $r_i(\theta,s) =\exp[\theta' X_i(s)]$. Then the risk-weighted covariate average is:

$$\bar x(\theta,s) = \frac{\sum Y_i(s) r_i(s)X_i(s)}{\sum Y_i(s) r_i(s)}, $$

where $Y_i(s)$ is the at-risk indicator for time $s$.


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.