1
$\begingroup$

I have a typical Survival analysis data-set - patients progressing through various stages of Alzheimer's Disease over 8 years of time, with some covariates, plenty of censored records.

However, I'm trying to ask a different question, that is somewhat sideways to Survival analysis. I'd like to take the group of patients that progressed (event) and those who did not (no event) and compare them against one another at a certain point in time, taking the censoring into account.

Perhaps I just can't get my head around this, but it doesn't seem to me that Cox Proportional Hazards or Accelerated Failure Time models can do that. Any advice, please?

$\endgroup$
  • 1
    $\begingroup$ What exactly do you mean by comparing them? If you want to assess survival conditional on having survived up to time $T$, plain ol' stratified Kaplan-Meier will do that for you. $\endgroup$ – Marc Claesen Sep 16 '14 at 19:55
  • 2
    $\begingroup$ I am 100% certain that your "censoring" is informative and you should first address that by using some combined endpoints for sensitivity analyses. Think about reasons why subjects with Alzheimer's would be censored, then think about the risk that that is related to cognitive decline. $\endgroup$ – AdamO Sep 16 '14 at 20:40
  • $\begingroup$ @MarcClaesen in several study designs, it may be difficult or expensive to ascertain certain exposures, like chart reviews, and when an outcome is rare, you can save a lot of money by matching cases to controls after follow-up and then measuring these exposures. $\endgroup$ – AdamO Sep 16 '14 at 20:48
  • $\begingroup$ @MarcClaesen I honestly do not see how K-M would cut that. I'd need to compare covariates, so I'm already looking at Cox PH models. Technically I can run CPH regression for the group of progressors, but for non-progressors it wouldn't make much sense, since there's no event... $\endgroup$ – Michal J. Figurski Sep 17 '14 at 15:53
2
$\begingroup$

I'm fairly certain the type of analysis you are interested in conducting is a case-cohort design. If censoring is informative, you have competing risks for censoring versus outcome, and you cannot use regular techniques to calculate the necessary weights for your Cox model. There is no method for which I'm aware to do this.

$\endgroup$
  • $\begingroup$ That is what I was thinking. $\endgroup$ – Michal J. Figurski Sep 17 '14 at 14:11

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.