Consider a survival analysis with time-constant coefficients, interval-censored, where the observation intervals are consistent across all individuals (e.g. each individual is observed at the end of every time period). I think I remember seeing it asserted somewhere that in this case a Cox proportional hazards model is equivalent to a binomial (Bernoulli) GLM with a complementary log-log link and a fixed effect of time for every observation period (this corresponds to the baseline hazard that is factored out of the Cox PH likelihood). Is this known to be true/false, and can someone provide a supporting argument or pointers to supporting references?

If true, this provides a very convenient way to avoid the computational/technical difficulties of fitting interval-censored Cox models in the specific case where the interval censoring dates are completely (or mostly) identical across individuals (e.g. see this question and more generally these questions ...)

  • $\begingroup$ it looks like this might be covered here ... also, I realized my question looks a lot like this one ... (I don't think it can be closed as duplicate until there's an answer to one or the other) (The link from that Q is broken, on wayback machine here: web.archive.org/web/20141121080306/http://www.ics.uci.edu/… looks to be substantially similar ... $\endgroup$
    – Ben Bolker
    Sep 30 '19 at 21:12

This particular situation is explained in Section 3.3.2 of Tutz and Schmidt, Modeling Discrete Time-to-Event Data, Springer, 2016, adapted slightly below.

Say that there is an underlying continuous-time proportional hazards (PH) model, but failures are only coded within intervals of time, e.g. the event is coded as happening at $T=t$ if is occurs within $[t-1,t)$. Then with covariate vector $x$ and regression coefficients $\beta$ the discrete hazard $\lambda(t|x)=P(T=t|T\ge t,x)$ can be written as

$$ \lambda(t|x)= 1- \exp(-\exp (\gamma_{0t}+ x^T \beta)),$$

where the fixed time coefficient $\gamma_{0t}=\log (\Lambda_0(t)-\Lambda_0(t-1))$ and $\Lambda_0(t)$ is the cumulative baseline hazard function of the underlying continuous-time PH process. This is the form of a complementary log-log (Gompertz) model, as so nicely described here. This is called a "grouped proportional hazards model."

How many "computational/technical difficulties" are avoided, however, isn't so clear. Unlike a continuous-time Cox model, in which the baseline hazard effectively disappears from the maximum-partial-likelihood calculations, this fixed-time-effect discrete approach requires estimation of the baseline hazard (as parameterized in $\gamma_{0t}$) at each interval end-point along with estimation of the regression coefficient values. That's not a big problem with the type of panel data envisioned in this question. Except for the link function, it's not different from standard discrete-time survival models that use logistic regression for panel data with time modeled as a fixed effect.

"Computational/technical difficulties" arise with this approach when each individual has her own interval-censoring times, as in Finkelstein's original adaptation of PH models to interval censoring (Biometrics 42: 845-854, 1986). This occurs e.g. with modeling of disease progression in cancer, when between-visit progression is detected at clinical appointment times that aren't uniform among patients. In that situation with $N$ patients you might need to estimate as many as $2N$ nuisance values of a baseline hazard function (one for both ends of each patient's event-containing interval) in addition to the regression coefficient values of primary interest. That's why tools like those in the icenReg package can be needed.

  • $\begingroup$ @CliffAB I wanted to clear up this long-standing question but parts of it are outside my specific expertise. When you get a chance, could you see if I fairly described why the approach of the OP in its simple situation doesn't avoid "computational/technical difficulties" with semi-parametric models in a broader context? $\endgroup$
    – EdM
    Mar 31 at 20:29
  • $\begingroup$ thanks, this seems like what I was looking for! $\endgroup$
    – Ben Bolker
    Mar 31 at 21:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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