The marshmallow experiment, before and after replication

Walter Mischel and his team conducted from the late 1960s, at Stanford University, a series of experiments in which a preschool child could eat a sweet immediately or obtain two by waiting for the experimenter's return.

Later publications report associations between measured waiting time and school or behavioural outcomes in adolescence. Those analyses cover small samples, a few dozen children, from a campus school.

The original work was interested in the strategies children used to hold out: looking away, telling themselves a story, picturing the sweet differently. Mischel repeatedly stressed that those strategies could be learned.

The popularised reading retained something else, a stable and predictive capacity, closer to a character trait than to what the setup measured.

Tyler Watts, Greg Duncan and Haonan Quan published a replication on a much larger and more diverse sample, drawn from an American national study.

The sample numbered several hundred children, and the waiting measure used a shortened version of the original protocol. Later outcomes were assessed around age fifteen, using school achievement tests and behaviour questionnaires.

The association with later outcomes is found, but much weaker, and it shrinks further once social background and skills measured at the same age are taken into account. Part of what the test seemed to predict was predicted by family context.

A study by Celeste Kidd and colleagues in 2013 had shown that waiting time depended on the perceived reliability of the adult: after a broken promise, children waited markedly less. The study's sample was small, which calls for caution about the size of the effect. Its direction, on the other hand, is consistent with reading waiting as a decision taken in a more or less predictable environment.

That result suggests waiting is not only a matter of self-control but also a calculation about the probability that the reward will arrive. It has been replicated in several settings.

What remains is a measurable association between waiting at four and later indicators, of small size, which is already remarkable for a fifteen-minute measurement.

What has gone is the deterministic reading, and with it the idea that a single preschool test durably locates a person. The episode has become a textbook case of the difference between a result and its public uptake.

Public uptake had also slid from a statistical association towards an implicit promise, the idea that training children to wait would produce the associated outcomes. None of the studies cited tested that hypothesis, and the 2018 replication offers no argument in its favour.