One place where I have seen this come up is in discussions of using "intention to treat" analysis versus an analysis that tries to get at the "efficacy" of a treatment in experiments with imperfect compliance. See the Wikipedia article on "intention to treat" (link), which includes some references.
In a run-of-the-mill randomized control trial with noncompliance, the intention to treat estimate examines only the difference between those assigned to treatment and control. However, noncompliance means that some people who were assigned to treatment may not have actually taken it, and some assigned to be in the control group may have actually received the treatment. If so, the intention to treat estimate may understate the average treatment effect that would obtain were all members of the population under study to actually take up the treatment.
When this kind of noncompliance is present, the analyst has a decision to make. She could decide to simply do the intention to treat analysis, justifying it by saying that in the real world, we cannot control compliance, and so the intention to treat analysis is more "realistic" as an estimate of what would happen were this treatment approved for use clinically. I have seen this referred to as an analysis of a treatment's "effectiveness." Or, she may use some kind of adjustment method to try to get at how people who actually took up the treatment differed from those who didn't. She could justify this by saying that what we are really interested in knowing is the biological (in the case of a medical trial) "efficacy" of the treatment, and to do so, we need to make the comparison between those who actually took the treatment and those who didn't.
The issue for an analysis of biological efficacy is, what "kind of adjustment method" is valid? The current state of the art, as I understand, is to view an experiment with noncompliance as an instrumental variables problem, a la Angrist, Imbens and Rubin (1996) (gated link), or, more generally to view the problem in terms of "principle stratification", a la Frangakis and Rubin (2002) (gated link). As such, the randomization serves as an instrument that nonparametrically identifies "efficacy" effects for at least certain subpopulations---namely, those who would comply with their treatment or control assignment. Beyond this, one could impose a more stringent model in order to identify efficacy effects, but then one may wonder, why did you bother to do a randomized experiment in the first place?