Much of the literature on missing data assumes the datum is unambiguously defined, relevant, and unobserved, so that the problem is to guess what the unobserved value really was. In clinical trials this may not be the problem at all. If the patient dies, there are no relevant unobserved outcomes: what prevents their observation also precludes their existence. Or if the outcome was a surrogate, other data on the patient's actual clinical condition may be more relevant than a fabricated value of the missing surrogate. And sometimes the patient's choosing an alternative treatment may constitute in itself the most relevant datum, in preference to what might have been observed if he had chosen otherwise.

Before we apply standard solutions we should ask whether we have the standard problem. Sometimes we do; sometimes we have an easier one; sometimes we have an impossible one.