This is probably a basic question, but here goes: I have a data set with a group of at-risk patients (for illustrative purposes lets say for heart disease) who were enrolled in a databases, some were treated with a drug (lets say an anti-platelet drug) and they are followed to see if they die. I want to use Cox proportional hazards to model prognostic factors that affect survival time and see if drug exposure is one of them, how long do I follow them for before I censor?

A lot of the patients are enrolled at various times and have varying lengths of follow-up until either they are lost to follow-up or die (for example, some patients have 2 months of follow-up data because they were enrolled 2 months ago, others have 3 years of data because they have been followed that long but not died, some have died after 1 month of enrollment, and some are lost to follow-up after 1 month).

Do I just choose an arbitrary length of follow-up (such as 1 year since enrollment) before censoring anyone alive at that time point? Do I use the longest length of follow-up available and censor everyone before that time point who was lost to follow-up? Any advice would be really helpful, sorry if its a question that is answered elsewhere but I couldn't find the answer when I looked. Thanks!

  • 2
    $\begingroup$ Start each person when the relevant starting event happens to him/her and continue as long as possible. $\endgroup$
    – mdewey
    Jan 9, 2018 at 14:20

1 Answer 1


Always think like a trialist when running observational analyses like these:

  1. Minimum length of (potential) follow-up: If 6 months, exclude any patient who entered the study less than 6 months before your data cutoff data. This patient could not have provided you the minimum length of follow-up and is considered ineligible (i.e. outside of the "recruitment period" for a clinical trial)
  2. Maximum length of follow-up: Specify the maximum time since enrolment that you want to keep following up patients for. Let's say this is 2 years. Then anyone who hasn't had an event by year 2 and is still being followed simply gets censored at the 2-year mark. This would like a clinical trial interested in 2 years of follow up.

Both of these should be based on domain expertise (e.g. expectations about event rates and dropouts/missing data and so on). If no domain expertise is available, pick your best option and run some sensitivity analyses around different choices of these to make sure they don't affect your results very much.

  • $\begingroup$ Although the issues of domain expertise and sensitivity analysis are important, the specifics of your 2 recommendations seem unnecessarily to limit the number of cases and events. That could reduce power substantially. Could you please edit the answer to provide references to texts or to the research literature that support those suggestions? $\endgroup$
    – EdM
    Sep 13 at 20:55
  • $\begingroup$ @EdM I don't give much importance to power in observational studies (e.g. DOI: 10.1016/j.jclinepi.2021.08.028) but as I mentioned in my answer, considerations about cases/events can certainly feed into the approach I suggested. Observational studies framed as trials is a fairly common approach nowadays and both items I mentioned are commonly used in practice (e.g. doi: 10.1038/s41591-019-0597-x). $\endgroup$
    – elbord77
    Sep 14 at 0:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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