[mistercow]

Well, take all the ways that it could be unsafe, and (by some measure) estimate some utilitarian cost of rolling the drug out early for each individual person for each unsafe scenario. Now multiply each of those costs by the probability of that unsafe outcome happening for a given individual, sum those up and place them in column A. Now take the probability that the drug is effective and multiply it by each individual's probability of getting Alzheimer's, and then multiply that by the utilitarian penalty for that person having Alzheimer's. Put that in column B. Subtract column B from column A and put that in column C. Any row who has a positive value in column C - that is for whom it's better to wait and vet the treatment's safety.

Practically speaking, these will be people who know they are genetically predisposed to Alzheimer's and get the vaccine before they show any signs.

Now, if we could actually run the calculations I mentioned above, the problem would disappear because we wouldn't simply say "approved as safe" or "rejected as unsafe". We'd just keep a running tally of the odds of safety and give the treatment to people according to the desperation of their individual case. But unfortunately, that's not the world we live in.