f the survival benefit was to be measured at year 3 (and not year 5), and the expectation to beat was to be 10% (again, at year 3), I'd think the survival benefit would decrease with time, not increase. At least, that's what I've found to usually be the case... with each passing year, the survival benefit goes down.
Am I missing something?
Yes you are missing something very big
This is ARMS comparison, test/control
You can expect survival to decrease over time
This measure is how much BETTER is the test arm doing than the control arm in that regard.
If you are doing better at year 3 then over time you had more people alive at year 3 than the control arm, so what do you think will happen by year 5 ?
So, to be more clear, a 5% difference at year 3 can be 15% difference at year 5.
From year 3 to year 5, the survival rate is going down respectively, but with multikine is going down slower than without, so the difference between two group survival rates is getting larger.