Show a four-year-old child some marshmallows and a bell. Tell them that you're going to leave the room for a while (fifteen or twenty minutes). Say that if they ring the bell, you'll come back and give them a marshmallow. However, say that if they don't ring the bell, but wait until you come back without ringing it, then you will give them two marshmallows. Record what happens.
Ten years later, assess the child's personality and general success in life by means of a questionnaire sent to their parents. What you discover is that the ones who rang the bell, or who rang it earlier, score relatively poorly on questions that measure social adjustment, "emotional intelligence," and so on. The ones who didn't ring the bell, or rang it later, score much better on those measures, and also score better on the SAT. Shoda, Y., Mischel, W., Peake, P. K. (1990). Predicting adolescent cognitive and self-regulatory competencies from preschool delay of gratification: Identifying diagnostic conditions. Developmental Psychology, 26(6) [PDF]
The part I think is really interesting is what the authors of that paper don't say about the experimental protocol.
What they don't say is whether, upon returning to the room, the experimenters actually did give the children one or two marshmallows as promised. That particular paper I mentioned is a meta-analysis grouping data from several previous papers, so this detail might have gotten lost in the summarization, but I read one of the original papers and it didn't say either. From context I think it's reasonable to assume the experimenters did keep their promises. Nonetheless, the fact that this wasn't mentioned says something about the values of the people who made the decision on what to mention (which might well include the journal editors as well as the authors).
Obviously, if you're only going to do the experiment once, giving the child the marshmallows or not doesn't matter to the immediate result; the data has already been collected. But if you do it repeatedly per child, then there are many more interesting questions to ask. Will a child who waits the first time, also wait the second time? Will one who doesn't wait the first time, also not wait the second time? How will a child's waiting time change with number of iterations?
And those questions exist even if you keep the promise every time. What if you don't? What if in some or all of the iterations, you tell the child "wait, and I'll give you two marshmallows; don't wait and I'll only give you one," but then if the child doesn't ring the bell, you don't give them anything?
Divide the children into groups. One group is promised two marshmallows if they wait, one if they don't, and that's what they actually get. One group is told they'll get two if they wait, one if they don't, but really they get zero if they wait, one if they don't. Other groups get one or the other of those policies determined at random per trial, with different probabilities per group. Run the experiment many times per child. What happens to their behaviour over successive trials? How, if at all, does membership in the different groups correlate with the "success in later life" measures examined in the earlier study?
Since we're already well outside the range of experiments we could actually do in real life (try getting the above through an ethics review) we might as well take it even further: subject a whole lot of the four-year-old population, like 51% of them, to this kind of thing and then see what kind of world you get ten years later.
Give me a thousand orphans, a hedge maze, and enough cheese...
Bonus: write an essay about privileged children who think they're entitled to marshmallows.