Run a trial, or barring that, correct for obvious confounders that affect both post length and post quality (I am sure we can both think of a few right now).
Run a trial, or barring that, correct for obvious confounders that affect both post length and post quality (I am sure we can both think of a few right now).
Why not be more specific about what confounders should be included or how a trial should look like?
You cannot use observed dependences in the data to suggest decision changes because p(y | x) is not in general equal to p(y | do(x)).
What should cleonid do instead (if anything)? And even if something is not true in general, could it still be used as an approximation?
Run a trial, or barring that, correct for obvious confounders that affect both post length and post quality (I am sure we can both think of a few right now).
Why not be more specific about what confounders should be included or how a trial should look like?
Are you going to pay me to seriously look into this?