As virtually always the case when looking at incomplete survival data the reported estimated median is the 50% cross of the KM curve.
Start at 100%. Every time period drop in proportion to the number of events / number then at risk. Say the curve was at 30% at 4 years. The next month half of the remaining patients event. The curve would then be at 15%.
Even when the trial is over the reported median will be the KM number. Not the time of the 166th event (unless they find all the lost souls).
Back on the earlier threads, you can not treat medians like means and average the average. Might get you in the ballpark at best. And in this data set can be way off as seen with meirluc's post.
Exwannabe, is right... the weighted average formula I was using is a ball park estimate at best, and in fact could be misleading.
One way to find the true affect the 38 unclassifieds had on the mOS is to remove the 38 unclassified patients from the dataset and recalculate the KM data. This should be feasible by using Sentiment stocks high resolution curves to determine which patients were the 38 unclassifieds.