What did the marshmallow test find, and did it replicate?

The original marshmallow studies reported that preschoolers who waited longer for a bigger reward later scored higher on achievement tests. Conceptual replications found an association that was roughly half the size and that mostly faded once family background and early ability were taken into account. So the finding partly held up, but the original strength and the causal story did not. [1] [2]

In the 1990 study by Shoda, Mischel and Peake, four-year-olds sat at a table with a bell and a pair of rewards, such as one marshmallow or two. If they waited until the experimenter returned, they got the preferred larger reward. If they rang the bell, they got only the smaller one. Waiting time was recorded up to a maximum of 15 to 20 minutes. Time waited correlated 0.57 with SAT math scores and 0.42 with SAT verbal scores. The sample was small, with 35 children in the baseline condition who had later academic results, and it came from the Stanford community. The original authors themselves warned that the coefficients could exaggerate the true association. [3]

Later work by Mischel's team reported further links: waiting longer was associated with higher adolescent SAT scores, better peer-rated social competence, and, in adulthood, lower BMI, a better sense of self-worth and better coping. [2]

Watts, Duncan and Quan studied children whose mothers had not completed college. An extra minute waited at age 4 predicted a gain of about one tenth of a standard deviation in achievement at age 15. That correlation was half the size of those in the original studies, and controls for family background, early cognitive ability and home environment cut it by two thirds. Most of the variation came from whether a child could wait at least 20 seconds. Links to behavior at age 15 were much smaller and rarely statistically significant. [1] [4]

The design differed from the original: a maximum wait of 7 minutes, waiting time grouped into categories, a different achievement measure, and a sample limited to children of parents without a college education. [3]

Researchers read the results differently. Watts and colleagues, and later Sperber and colleagues, emphasize that the predictive power is modest and largely explained by background factors. In a 2024 study of 702 participants followed to age 26, bivariate links with educational attainment (.17) and BMI (-.17) were detected, but almost all adjusted coefficients were nonsignificant. The authors concluded that the test does not reliably predict adult outcomes. [2]

Michaelson and Munakata reanalyzed the same data set as the 2018 study with a preregistered analysis. They found significant associations for three of five outcomes, including problem behavior, which the 2018 paper had not found. They argued that social support explained these relationships better than self-control, so the test may reflect a child's early environment. [5]

Blogger Jason Collins argued that, given the small original sample and the design differences, a smaller correlation is unsurprising and the original result held up fairly well as a simple diagnostic. The search results do not show a consensus among specialists beyond agreement that the original correlations were overstated. [3]