We enroll patients consecutive. If a patient presents multiple times at the ED, and is therefore enrolled multiple times, he can generate multiple events in a survival analysis (cox) (depending on whether he has an event or not).
Do we have to exclude multiple enrollments in any circumstance?
When you say events, do you mean multiple records (i.e. observations) or do you mean truly multiple events.
It sounds like the same person can contribute data multiple times, but the unit of analysis is the times they contribute data (i.e., the observations). This is similar to a repeat measurement situation in longitudinal studies, or any other clustering scenario (like if several observations come from the same hospital, and you have several hospitals in the study). I suspect you can address it by taking account of clustering by patient (assuming you have unique patient IDs). Important aspects of the decision about how to address it seem to be how many patients re-present and how many times.
You can try a shared frailty approach where observations coming from the same person “share” frailty.
On the Pavel’s last point see Therneau’s article. You can also use a cluster sandwich covariance estimator to correct for intra-patient correlation (e.g., see the R rms package robcov function). But full modeling using random effects for patients is preferred.
Yes I do. However, I would like to understand in general how a small percentage of multiple enrolled patients would affect a standard cox regression. I would highly appreciate a comment of yours.
If it is a small proportion, then anything you do will have little impact on the results.
Most studies set an eligibility criterion, at design time, that patients can only be enrolled once. If that was not the case in your study, I suspect excluding observations after the first might introduce selection bias because people who died would not have had the chance to enter the study again after their death.
If you treat each observation as independent (ignoring the fact that some come from the same patient), your standard errors/confidence intervals/ and p-values will be too small. If you make some allowance for this “clustering” of observations within patients, you can avoid this. The trouble is that most of your “clusters” only have one observation and a few have ?4, so I’m not sure how the robust variance estimator that corrects the standard errors will perform. It might inflate them too much and therefore increase your p-values (I.e. increase risk of erroneously declaring “no statistical significance”).
I think reasonable to do anyway, though.
In any case, if the interpretation of your result is affected by these choices with so few affected patients, it is probably sensitive to many other issues too, including common things like confounding and somewhat nonproportional hazards.
i guess then recruiting the same patient was not intended? i have seen this before in academic clinical trials, ie a patient reappears at another hospital in the district and is re-recruited into the study. it’s not intented and the proportion is small and they should be excluded from the analysis. maybe you choose which of their visits to include by random selection