Read on Twitter
Two sort of contradictory thoughts on the one vs two dose argument, with the obvious caveat that I know very little about vaccines, but still based on some of my work which is related in a more abstract way.
First, if someone who hasn't studied a question much thinks the answer is obvious, but those who have studied it a lot are divided/don't think the answer is obvious, then the answer is almost certainly not obvious.
In uncertain environments, knowing what questions are hard is a generally better marker of expertise than being able to predict outcomes. See this paper (with the caveat that we pick an information structure where this is unusually clear).
http://andrewtlittle.com/wp-content/uploads/2020/04/i_dont_know.pdf
http://andrewtlittle.com/wp-content/uploads/2020/04/i_dont_know.pdf
To me the argument for getting out first doses seems obvious, but given the above I'm very wary of that reaction. Even if we later learn that one dose was better, that doesn't mean it was ex ante clear.
Second, two dose proponents seem to be making a kind of argument that I think is problematic: setting up a false binary where we know two doses is 95% effective but don't know much if anything about one dose.
Extrapolation is needed to say what will happen if second doses are delayed. But going from any experiment to a wider population requires extrapolation, if nothing else to a different time period (with a different strain of the virus spreading!) @kmmunger https://osf.io/4utsk/
A good way to think about this in general is "how correlated is my prior belief about the efficacy in the trials and the efficacy as it will be implemented".
http://andrewtlittle.com/papers/learningfromobservation.pdf
If you refuse to say you have a prior belief about that, you are fooling yourself.
http://andrewtlittle.com/papers/learningfromobservation.pdf
If you refuse to say you have a prior belief about that, you are fooling yourself.
Changing the timing of doses decreases that correlation, but not to 0! And the correlation between the study efficacy and what will happen now in a wider roll out is not 1 (even setting aside the standard sampling issues).
In other words, it's not like we are going from a domain of perfect knowledge to total ignorance. And it's up to the real experts to hash out exactly how uncertain we should be, so I'll stop there.
