I have a dataset in which we follow a cohort of healthy individuals at recruitment, and when an individual experiences the event we assign four matched controls to it by age, sex, and recruitment center (this is how the study was designed, I have no control over it). I'm looking for biomarkers associated with the outcome. This is all the information I have.

My advisor who seems to think that we can't do a survival analysis because when a case is identified both the case and the matched controls are censored at the time of matching even though the controls are still in the study.

Is this right? in my opinion, his assumption is not correct.

I understand there may be biases, but please focus on whether or not the cases and controls are really both censored at the time of matching.


  • $\begingroup$ Some types of case-control studies can introduce bias, so please say more about the design and purpose of the study. Are you, for example, looking for a relationship between some biomarker and outcome? Are cases defined by their having had an event, or by some status at entry into the study? Why are the matching criteria limited to age, sex, and center and not including other characteristics potentially related to outcome? Is there some risk of a survivorship bias in identifying controls? The more details you can provide the better a response can be. $\endgroup$ – EdM Jan 28 '20 at 20:45
  • $\begingroup$ I've tried to address as much as possible but is the idea of censoring both cases and controls at the time the case experiences the event true? $\endgroup$ – raygozag Jan 28 '20 at 21:00

Quoting from Leung et al, Annual Review of Public Health 18: 83-104, 1997:

Censoring occurs when incomplete information is available about the survival time of some individuals.

So ask yourself: what does survival time mean in this study, and is that information incomplete about some individuals?

If survival time is being defined for an individual as the time between recruitment into the study and the time of an event, then any case (defined in your study as having had an event) clearly has complete information about survival time and should not be considered censored.

It's possible, however, for a study to be interested in some starting time for survival before recruitment into the study. For example, Therneau and Grambsch (page 75) describe a study at a tertiary referral center, for which survival was calculated from the date of original diagnosis. If an individual's diagnosis was made elsewhere before entry into the study at the referral center, then that individual was not at risk for an observable death until entry into the study, and the survival time would be considered left truncated at the time of study entry. If survival time is calculated from some unknown time before study entry, the survival time is left censored. This paper describes some of the issues that arise in these cases.

Whether controls are considered censored depends on their follow-up. Provided that survival time is calculated from study-entry date (putting aside the possibility of left censoring or truncation), if any control individual is followed until the time of an event, then that individual's survival time is by definition not censored. If the event hasn't happened by the time of last follow-up for some individual, then that individual's survival time is right censored at the last follow-up time. So if you do an early analysis of the data most or all controls might have right-censored survival times, but at later analysis times (assuming follow-up continues) some of those might have defined event times.


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.