[BostonCard]

Excellent and helpful analysis, BC, thank you. One quick follow-up: the pros and cons of an 8-point ordinal rating scale vs. days to recovery would presumably have been known to the researchers prior to launching the study, so do you have any insights as to why they would change midway through? I suppose my skepticism is that if it doesn't change outcomes but reduces recovery time for those who fully recover, then changing the primary endpoint would yield a more straightforward benefit from remdesivir. That's not nothing, but it does strike me as somewhat strange to change horses mid-race when the pros and cons of those horses were known out of the starting gate.

It's a WAG on my part, but here goes:
The ordinal scale generally has more power than looking at any single dichotomous outcome or time to event analyses, because every patient is potentially informative (in time to event analyses, only patients who have the event are informative).  So, two possibilities.  The first and obvious one is that the trial was oversubscribed (I think I saw that the actual recruitment was on the order of 1000), so they had more power to detect an effect in time to recovery.  The second is that a larger share of patients reached the event than they expected because recovery was generally faster in the blinded data than they had figured when they made the initial power calculation.  Note also that power calculations for ordinal endpoints are very hard because they require guessing the distribution of outcomes in the placebo group.

BC

[BostonCard]

Excellent and helpful analysis, BC, thank you. One quick follow-up: the pros and cons of an 8-point ordinal rating scale vs. days to recovery would presumably have been known to the researchers prior to launching the study, so do you have any insights as to why they would change midway through? I suppose my skepticism is that if it doesn't change outcomes but reduces recovery time for those who fully recover, then changing the primary endpoint would yield a more straightforward benefit from remdesivir. That's not nothing, but it does strike me as somewhat strange to change horses mid-race when the pros and cons of those horses were known out of the starting gate.

It's a WAG on my part, but here goes:
The ordinal scale generally has more power than looking at any single dichotomous outcome or time to event analyses, because every patient is potentially informative (in time to event analyses, only patients who have the event are informative).  So, two possibilities.  The first and obvious one is that the trial was oversubscribed (I think I saw that the actual recruitment was on the order of 1000), so they had more power to detect an effect in time to recovery.  The second is that a larger share of patients reached the event than they expected because recovery was generally faster in the blinded data than they had figured when they made the initial power calculation.  Note also that power calculations for ordinal endpoints are very hard because they require guessing the distribution of outcomes in the placebo group.

BC