Does X (hazard) variable in Cox proportional hazard regression analysis always have to be time? If not, could you provide an example, please?

Can age of cancer patient be a hazard variable? If so, can it be interpreted as the risk of getting cancer at a certain age? Would Cox regression be a legitimate analysis to study the association between gene expression and age?


Usually, age at baseline is used as a covariate (because it is often associated to disease/death), but it can be used as your time scale as well (I think it is used in some longitudinal studies, because you need to have enough people at risk along the time scale, but I can't remember actually -- just found these slides about Analysing cohort studies assuming a continuous time scale which talk about cohort studies). In the interpretation, you should replace event time by age, and you might include age at diagnosis as a covariate. This would make sense when you study age-specific mortality of a particular disease (as illustrated in these slides).

Maybe this article is interesting since it contrasts the two approaches, time-on-study vs. chronological age: Time Scales in Cox Model: Effect of Variability Among Entry Ages on Coefficient Estimates. Here is another paper:

Cheung, YB, Gao, F, and Khoo, KS (2003). Age at diagnosis and the choice of survival analysis methods in cancer epidemiology. Journal of Clinical Epidemiology, 56(1), 38-43.

But there are certainly better papers.

  • $\begingroup$ @chi: Thanks a lot. I'll look the papers. Would you please comment the first question? Is hazard variable always time? $\endgroup$ – yuk Oct 14 '10 at 19:12
  • $\begingroup$ @yuk Not necessarily, as suggested by @whuber. I have in mind another application of Cox regression dealing with the treatment of systematic pattern of missing responses in educational testing, as it arises when a student has not enough time to complete the test (missing responses might then be considered as right-censored) -- in this case, this is item ordering that is considered as the time scale. I'll look at the original paper (although I think this was also the subject of a PhD). $\endgroup$ – chl Oct 14 '10 at 20:43
  • $\begingroup$ +1. There are other papers, but I'm not sure they're necessarily better; I think Chalise does a pretty good job summing up the situation. $\endgroup$ – ars Oct 15 '10 at 2:13

No, it doesn't always have to be time. Many censored responses can be modeled with survival analysis techniques. In his book Nondetects and Data Analysis, Dennis Helsel advocates using the negative of a concentration in place of time (in order to cope with nondetects, which when negated become right-censored values). A synopsis is available on the Web (pdf format) and an R package, NADA, implements this.

  • $\begingroup$ +1, thanks for pointing out the NADA package. I noticed it makes it easier to handle left-censored data through the survival package -- is left-censored a common scenario with environmental data? $\endgroup$ – ars Oct 15 '10 at 2:11
  • $\begingroup$ @whuber: Thank you for the comment, NADA package looks very interesting. $\endgroup$ – yuk Oct 15 '10 at 5:23
  • $\begingroup$ @Andy: Thanks for the links. I think its worth to be an answer. I'd upvote. $\endgroup$ – yuk Oct 15 '10 at 5:25
  • $\begingroup$ @Yuk, per your request I made my comment into an answer, and @whuber thanks for your example. $\endgroup$ – Andy W Oct 15 '10 at 12:03
  • $\begingroup$ @ars: Yes, left censoring is characteristic of environmental data (and is a key concern of chemometrics in general). It's a tricky and interesting problem. Among the reasons are (1) the censoring limits are themselves determined by statistical estimates (through a calibration process), (2) censoring can occur in multiple ways--as limits of detection, limits of quantification, or "reporting limits", (3) the thresholds often vary in response to covariates ("matrix interferences") that can be strongly correlated with the original censored values, (4) data often are lognormally distributed. $\endgroup$ – whuber Oct 15 '10 at 14:11

On the age-scale vs. time-scale issue, chl has some good references and captures the essentials -- in particular, the requirement that the at-risk set contain sufficient subjects from all ages as would arise in a longitudinal study.

I would only note that there is no general consensus around this yet, but there is some literature to suggest that age should be preferred as the time scale in certain cases. In particular, if you have a situation where time doesn't accumulate in the same way for all subjects, for example due to exposure to some toxic material, then age may be more appropriate.

On the other hand, you can handle that specific example on a time-scale Cox PH model by using age as a time varying covariate -- rather than a fixed covariate at start time. You need to think about the mechanism behind your object of study to figure out which time scale is more appropriate. Sometimes it's worth fitting both models to existing data to see if discrepancies arise and how they might be explained before designing your new study.

Finally, the obvious difference in analyzing the two is that on an age-scale, the interpretation of survival is with respect to an absolute scale (age), whereas on a time-scale, it's relative to the start/entry date of the study.


Per the OP's request, heres another application I have seen survival analysis used in a spatial context (although obviously different than measuring environmental substances mentioned by whuber) is modeling the distance between events in space. Heres one example in criminology and here is one in epidemiology.

The reasoning behind using survival analysis to measure the distance between events is not per say an issue of censoring (although censoring can definately occur in a spatial context), it is more so because of the similar distributions between time to event characteristics and distance between events characteristics (i.e. they both have similar types of error structures (frequently distance decay) that violate OLS and so the non-parametric solutions are ideal for both).

Because of my poor citation practices I had to spend and hour finding the correct link/reference to the link above.

For the example in criminology,

Kikuchi, George, Mamoru Amemiya, Tomonori Saito, Takahito Shimada & Yutaka Harada. 2010. A Spatio-Temporal analysis of near repeat victimization in Japan. 8th National Crime Mapping Conference. Jill Dando Institute of Crime Science. PDF currently available at referenced webpage.

In epidemiology,

Reader, Steven. 2000. Using survival analysis to study spatial point patterns in geographical epidemiology. Social Science & Medicine 50(7-8): 985-1000.


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.