Read on Twitter
Because it is all very confusing, let me once again set out my perspective on analyzing AEs in a trial (like the $FGEN roxadustat NDD trial) where there are many more placebo dropouts than drug dropouts, particularly among the sickest patients https://twitter.com/Biomaven/status/1360609990957031430
So there are two ways to handle this, both prospectively specified. The first is to examine events on an ITT basis - that is keep tracking events through end of trial even if people drop out. They did exactly this for mortality and MACE+ events.
But realistically, you will miss many more minor events (UTI's, hypertensive episodes, etc.) in subjects that have dropped than in the subjects that are still enrolled and being closely monitored. So for minor events, the reported ITT data will be somewhat biased against roxa.
Other approach is to report events adjusted for time-on-drug. Because there were many more patient-years for subjects on roxa than placebo, you have to adjust raw events to reflect this. But even this does not fully adjust correctly - the reason is that dropouts were not random.
Because the patients with worse kidney function differentially dropped from the placebo group, over time the roxa group consisted of comparatively sicker patients who would likely demonstrate more AEs. This is known as "depletion of susceptibles." It also impacts QoL measures.
So both adjustment measures in the context of this trial are somewhat biased against roxa for all but the most serious AEs measured on an ITT basis. Despite this, adjusted AEs were essentially balanced and there are no safety signals that I can discern.
