A life-and-death problem in statistics

Drugs are accepted or denied by the FDA on the basis of their efficacy. Efficacy is defined as the excess of "success" over what might occur by chance. A drug which, for instance, resulted in no change in the average outcome for the disorder at which it was aimed could never be passed because it appears to be no better than a placebo.

However, basing success or failure on averages is a flawed procedure, as it assumes that the results lie on the normal, "bell-shaped" curve. Sometimes, however, they don't. Here's an example: Bill is dying of cancer, with a relatively short and painful time-span ahead of him. But suppose there is a drug which, in X percent of the cases, results in death immediately, but in the remaining percentage the patient is cured. Let us suppose that the drug, on average, does not change the average outcome, so it is not and will not be approved.

But for Bill, if he takes the drug, he will either die immediately or be greatly improved or cured. For Bill, taking the drug is a no-brainer. Of course he will chance dying, since he's dying painfully anyway and there is no escape. But if he takes the drug, he may be improved or cured. The flaw in FDA thinking is that there may be bi-modal or even trimodal results, and "averages" do not reflect the importance of this distribution of data.

I would appreciate any comments by someone knowledgeable about statistical analysis.