0
$\begingroup$

I have two groups of patients, treated at two time periods. I'm comparing survival. In the later group I have 3 years follow up time. In the early group I have of course access to longer follow up time.

What is correct?

Should I stop follow up in the early group at 3 years (1095 days), i.e. for a patient in the early group that has more than 1095 days survival I set survival to 1095 and treat them as not failed? Or shall I use the longer follow-up time in the early group when doing Cox regression?

$\endgroup$
1
$\begingroup$

Cox regression is semiparametric and is usually used to evaluate the effect of covariates on survival time, rather than the actual survival time. I suggest that, rather than running two separate survival analyses you run one and include "group" as a covariate. You would then need to adjust starting time, but this should not be difficult. The simplest case would be if everyone in each group started at the same time, then you could just subtract the difference in the two times.

$\endgroup$
  • $\begingroup$ Thankyou! 'group' is a covariate. My question mererly concerns the fact that the long time survivors in the older group have longer survival time than it is possible to have in the later group. What is the effect on the analysis/Hazard ratio. For example a long time survivor in the early group could have 5000 days in survival time while the patients in the later group that also are "cured"(=long time survivors) can only have maximum 1095 days (=3 years). What is the statsitical correct way to deal with this? $\endgroup$ – Richard Bernhoff Jun 8 '14 at 12:55
  • 2
    $\begingroup$ The issue of censoring is what survival analysis is mostly meant to deal with. As far as I know, having different censoring times in different groups doesn't violate any assumptions of Cox regression, but I am not 100% certain. $\endgroup$ – Peter Flom - Reinstate Monica Jun 8 '14 at 13:06
  • 1
    $\begingroup$ To answer the question about censoring times systematically differing censoring times: this is totally fine as long as the proportional hazards assumption holds. If the proportional hazards assumption does not hold, I have heard it argued that the Cox PH model still has utility that it gives the average proportional hazard between the groups. However, if you have significant deviation from proportional hazards (ie group one is at high risk compared to group 2 early on but lower risk later on), then uneven censoring becomes a very big problem; you're now comparing average hazards ratios... $\endgroup$ – Cliff AB Apr 27 '16 at 20:26
  • $\begingroup$ ...over different time periods, when this hazard ratio changes over time periods. So your estimate will really have no interpretation. $\endgroup$ – Cliff AB Apr 27 '16 at 20:27

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.