The marshmallow test, redone with ten times as many children, found that a four-year-old's willpower mostly stopped predicting teenage success once family background was taken into account

Put a four-year-old at a table with a marshmallow. Tell them they can eat it now, or wait fifteen minutes and get two. Leave the room.

This is one of the most famous psychology experiments ever conducted. Walter Mischel ran the original at Stanford in the late 1960s, and when he followed the children years later, the ones who had waited were doing better on almost everything — grades, SAT scores, emotional regulation, body weight. A single act of childhood self-control, the story went, could predict how the rest of a life would go.

In 2018, a team of researchers rebuilt the experiment with ten times as many children. Most of what the original story claimed turned out to be wrong.

What the replication actually showed

The study was led by Tyler Watts of New York University, with Greg Duncan and Haonan Quan. Rather than run a new marshmallow test from scratch, they drew on the NICHD Study of Early Child Care and Youth Development — a long-running longitudinal study that had already measured delay of gratification in 918 four-year-olds and then tracked those children through adolescence.

Nine hundred and eighteen children is roughly ten times the sample size of Mischel’s original work. It was also far more diverse socioeconomically and demographically. The original Stanford study had drawn its participants from the Bing Nursery School, an elite preschool associated with the university and populated largely by the children of Stanford faculty and staff. The NICHD sample included children from across the United States, from a wide range of family backgrounds.

When Watts and colleagues ran the standard analysis — the one that had produced Mischel’s famous findings — they found a correlation, but a much weaker one. In their subgroup of 552 children whose mothers had not completed college, an additional minute of waiting at age four predicted only about a tenth of a standard deviation gain in academic achievement at age 15. The bivariate correlation was roughly half the size of the effect Mischel had reported.

Then they added statistical controls for family background, early cognitive ability, and the home environment. The correlation shrank by another two-thirds. Once these factors were accounted for, the four-year-old’s ability to wait for a second marshmallow explained almost none of the variation in the fifteen-year-old’s outcomes.

What predicted the teenage outcomes was not willpower. It was the specific set of resources, education, and environment the child had grown up inside. Willpower had appeared to predict success in the original study because it was correlated with those factors, not because it was independently causing them.

What the four-year-olds were probably actually doing

Here is where the story becomes more interesting than the standard “willpower doesn’t matter” reframe.

A child sitting at a table with a marshmallow is making a decision based on the world they have experienced so far. If that world has consistently delivered on its promises — if adults who say they will come back have always come back, if two-marshmallow offers have always resulted in two marshmallows — then waiting is a rational strategy. The child has evidence that patience will be rewarded.

If that world has been less reliable — if adults have made promises they didn’t keep, if scarcity has been a regular feature of the child’s life, if the food available now cannot be counted on to be available later — then eating the marshmallow immediately is also a rational strategy. In fact, it is arguably the more rational one. A bird in the hand.

Some research has directly tested this. A 2013 study by Celeste Kidd, Holly Palmeri, and Richard Aslin at the University of Rochester exposed children to either a “reliable” or “unreliable” experimenter before running a marshmallow test. Children who had just experienced an unreliable adult waited, on average, three minutes for the second marshmallow. Children who had experienced a reliable adult waited four times longer. The willpower being measured, in other words, was heavily contingent on whether the child had reason to trust the world in front of them.

Four-year-olds from stable, resource-rich, promise-keeping homes are more likely to wait. Four-year-olds from less stable environments are more likely to eat. Both are making sensible decisions given their evidence. And both are also, downstream, likely to have quite different life outcomes — but not because of what they did with the marshmallow. Because of the environments that shaped their expectations of what promises are worth.

What the debate is really about

The Watts replication has itself been contested by psychologists who argue that the statistical controls used by the team may have been too aggressive. A 2019 commentary by Angela Duckworth and others argued that some of the covariates Watts included — early cognitive ability, home environment — may have captured aspects of self-control itself, not just confounds surrounding it. If early cognitive ability partly reflects the same underlying mental capacities as delay of gratification, then controlling for one to isolate the other may have inadvertently removed the very thing the marshmallow test was supposed to measure.

This debate is genuinely open. What is not really open is the finding that the original marshmallow test’s predictive power was substantially overstated. Even Watts’s critics accept that the effect of childhood delay of gratification on later outcomes is much smaller than the popular version of Mischel’s work suggested. Whether it is small because willpower doesn’t matter much, or small because willpower is inseparable from the environment that produces it, is what the field is still working out.

What this changes

The specific policy implication of the Watts replication is that early-childhood interventions designed to teach self-control — building willpower directly through training programs — are unlikely to produce large effects on later life outcomes on their own. If most of what looked like willpower’s effect was actually the effect of stable environments, family resources, and reliable relationships, then trying to boost the willpower without changing those underlying conditions probably will not work.

The more useful intervention, on this account, is to change the conditions that produce willpower in the first place. Reliable relationships. Predictable environments. Adults who keep their promises. Enough resources that a child does not have to make emergency decisions about the food in front of them.

This is a substantially less romantic story than the original marshmallow narrative. The child who resists is not a small hero. The child who eats is not a small warning. Both are, mostly, reading the world they were handed.

What shapes them, over the following decade, is the world itself — not the fifteen minutes they spent looking at a marshmallow.