Survival analysis with withdrawn patients

[Abby Chew]

In STATA, for those participants who were withdrawn from the study, do we treat these participants as 0 or 1 for censoring. For example, we have 5000 patients. 300 died, 50 were withdrawn from the study and the rest still not developing any events. However, we have time to event and time to withdraw. I am not sure in state should I treat 50 who withdrew from the study as . or 0 or 1. Thank you.

[Paul Dickman]

I suggest you consult someone knowledgeable about survival analysis since devising an appropriate analysis strategy for your data requires considerably more information than you have provided. We would need to know full details of the study and, in particular, why people drop out.

I'll try and answer your question, but cannot be confident in my answer without more information. The short answer is that the withdrawals/dropouts have event indicator 0 and are treated in the analysis as censored observations in the exact same way as those who had not experienced the event at the close of the study. You do not want to exclude them (event .) because they have been at risk of the event and did not experience it. Excluding them would induce a bias because you are systematically excluding only participants who did not experience the event. Classifying them as having the event is also incorrect because they did not experience the event.

So the short answer is to classify the withdrawals as censored. Analysing and interpreting these data, however, requires strong assumptions. You need to assume that censoring is *non-informative*. The reason for withdrawal needs to be independent of the future probability of an event. I suggest you discuss with someone familiar with such matters.