The Marshmallow Test Explained

A four-year-old sits alone at a table with one marshmallow on a plate. Eat it now, or wait about 15 minutes for the researcher to come back and get two. Psychologist Walter Mischel built this task at Stanford University in the late 1960s, and it became one of the most cited findings in developmental psychology, one that most people know only through its oversimplified pop-science version. The original correlation was real but small; a far larger 2018 replication cut it roughly in half and then mostly erased it once researchers accounted for family income and background. This is the honest, fully sourced version of what the marshmallow test actually showed, what it did not, and why the story changed.

What Mischel actually built and measured

Delay of gratification — a child's capacity to forgo a smaller, immediate reward in favor of a larger reward later — is the trait Mischel set out to study, not willpower in the abstract. Between roughly 1968 and 1974, children at the Bing Nursery School, a preschool on Stanford's campus serving mostly faculty, staff, and graduate-student families, took part in a series of delay tasks designed by Mischel and colleagues. Six hundred fifty-three preschoolers participated across these studies, according to later accounts summarized by the University of Washington psychology department.

The setup varied study to study. In some versions the treat sat on the table in plain view; in others it was covered, or the child was coached to think about something other than the treat, or told to imagine the marshmallow as a picture rather than something to eat. Researchers timed how long each child resisted before either eating the treat, ringing a bell to summon the experimenter, or successfully waiting out the full interval. Mischel's team was not testing a single blunt idea of self-control; they were testing which mental strategies, distraction, reframing, physical covering, made waiting easier, a line of research summarized in his 1972 paper with Ebbeson and Zeiss.

The famous long-tail finding came later. In 1990, Mischel, Yuichi Shoda, and Philip Peake published a follow-up using a subset of the original Bing preschoolers, now teenagers, whose parents had supplied SAT scores. Of the original pool, 404 families received follow-up questionnaires, 185 responded, and only 94 parents actually supplied SAT data, an attrition funnel that alone should temper how much weight the finding can bear. Within that small remaining group, children who waited longer as four-year-olds tended to have meaningfully higher SAT scores as teenagers, and Shoda, Mischel, and Peake themselves urged caution given how small and skewed the sample had become, according to a summary from EBSCO's research overview.

The 1990 SAT follow-up rested on data from just 94 children, all drawn from a single Stanford-affiliated preschool population. The 2018 replication that reexamined the claim analyzed 918 children from a nationally representative U.S. sample, nearly ten times larger and vastly more diverse in income, education, and geography.

Why the sample size matters this much

A sample of 94 children, almost all from highly educated, financially secure Stanford-community families, cannot tell you how delay of gratification behaves across the income and education range of an entire country. Statisticians call this a restricted range problem — when every subject clusters near the same value on a variable like household income, you cannot detect how that variable might interact with your finding, because there is barely any variation to detect it with. The Bing preschool sample was, almost by design, a study of children who already had unusually stable, resourced home lives.

The 2018 replication that rewrote the story

In 2018, Tyler Watts of New York University, along with Greg Duncan and Haonan Quan of the University of California, Irvine, published "Revisiting the Marshmallow Test: A Conceptual Replication Investigating Links Between Early Delay of Gratification and Later Outcomes" in the journal Psychological Science. Rather than running a brand-new preschool study, they drew on an existing, large, federally funded longitudinal dataset: the National Institute of Child Health and Human Development's Study of Early Child Care and Youth Development (NICHD SECCYD), which had already given a delay-of-gratification task to children around 54 months of age and tracked them for years afterward.

Their working sample included 918 children with complete data. Because Mischel's original claim was strongest for the idea that self-control matters most for children facing real-world adversity, Watts, Duncan, and Quan paid special attention to a subgroup of 552 children whose mothers had not completed college, roughly ten times the size of the original 1990 SAT follow-up sample.

The raw, unadjusted correlation they found between waiting longer at age four and academic achievement at age 15 was about half the size of what the original small studies had reported. That alone was a meaningful downgrade. But the more consequential finding came next: once the researchers statistically controlled for family background, including household income, maternal education, and the child's early cognitive ability, the association between delay time and later achievement shrank by roughly two-thirds. Concretely, the bivariate correlation of 0.24 for the non-college-educated subgroup dropped to about 0.08 with background controls, and to a statistically insignificant 0.05 once concurrent cognitive and behavioral measures were added, per the study as summarized in coverage from PMC's review of the replication.

What "conceptual replication" means here

A conceptual replication tests the same underlying idea with a different, often improved, methodology rather than reproducing the original procedure exactly. Watts, Duncan, and Quan did not restage Mischel's exact lab task; they used an existing delay-of-gratification measure embedded in a much larger, nationally representative study, which let them test the same claim with statistical power the original researchers never had.

The cognitive machinery Mischel actually proposed

The headline "waiting time" number obscures what Mischel spent much of his career actually studying: not whether children can resist temptation, but how. Long before the 2018 replication controversy, Mischel and cognitive scientist Janet Metcalfe published a 1999 theoretical model in Psychological Review that tried to explain, mechanistically, why some strategies made waiting easier and others made it nearly impossible. That paper, "A Hot/Cool-System Analysis of Delay of Gratification: Dynamics of Willpower," is arguably a more durable scientific contribution than the marshmallow test itself, and it rarely makes it into the popular retellings, according to Columbia University's psychology department archive of the paper.

Metcalfe and Mischel proposed that self-control runs on two competing mental systems. The hot system is fast, emotional, and reflexive, wired to react to a tempting stimulus, the smell of a marshmallow, the sight of it sitting on a plate, the way a smoke detector reacts to smoke rather than reasoning about it. The cool system is slow, deliberate, and unemotional, capable of holding an abstract goal in mind ("I get two marshmallows if I wait") without being hijacked by the immediate pull of the object in front of you. Waiting successfully, in this model, is not a single trait a child either has or lacks; it is a contest between which system gets to run the show in a given moment.

The single most striking finding buried inside Mischel's original research program was not the SAT correlation. It was that simply telling children to imagine the marshmallow as "just a picture," a flat, non-edible image rather than a real food, dramatically increased how long they could wait, sometimes by several multiples over children given no such instruction.

Which specific strategies actually worked

Mischel's team tested this empirically well before the hot/cool model formalized it. Across the original Bing studies, children who physically or mentally distracted themselves, singing, playing with their hands, looking away, thinking about something unrelated, consistently outlasted children who stared directly at the reward. Children coached to reframe the marshmallow's sensory qualities (thinking of it as a cotton ball or a cloud, rather than something sweet and chewy) also waited substantially longer than children left to their own devices, a pattern documented in coverage of Metcalfe and Mischel's framework from the American Psychological Association.

This matters for how the marshmallow test should actually be interpreted. If waiting time were a fixed trait, like height, coaching would not move the number much. But it did, substantially, which suggests that what the original task captured was partly a set of learnable cognitive strategies, not a stable, unteachable disposition a child either brought to the table or didn't. That distinction complicates both the deterministic pop-science reading of the test and, to some extent, the purely environmental reading suggested by the 2018 replication: strategy use and family background are not mutually exclusive explanations, and Mischel's own hot/cool framework was, in part, an argument that self-control is a skill shaped by both instruction and circumstance rather than a fixed trait revealed once at age four.

How the marshmallow test escaped the lab and reshaped parenting advice

By the 2000s, the marshmallow test had migrated far from developmental psychology journals into a fixture of parenting books, TED talks, and education policy debates, almost always stripped of its confidence intervals and sample-size caveats. The basic pitch was seductive in its simplicity: teach your child to delay gratification early, and you are investing in their future SAT scores, career success, even physical health, a claim traceable to popular treatments of Mischel's work that outran what his own data, honestly read, supported.

That framing had real downstream consequences. Some early-childhood curricula and "grit"-focused education programs built explicit self-control training modules around the marshmallow paradigm, treating delay of gratification as a discrete, trainable skill worth dedicated classroom time, on the theory that instilling it early would pay compounding dividends decades later. Parenting advice columns, meanwhile, sometimes used a child's marshmallow-style behavior, refusing to wait for a promised reward, as an informal diagnostic for future character, an interpretation that both Mischel's own cautionary notes in the 1990 paper and the entire thrust of the 2018 replication argue against.

The deeper problem with this migration is that it treated a correlational finding from 94 privileged children as if it were a causal lever any parent could pull, regardless of a family's material circumstances. If Watts, Duncan, and Quan are right that most of the marshmallow test's apparent predictive power traces back to income, maternal education, and household stability, then a self-control curriculum aimed narrowly at a child's willpower, without addressing the material conditions the 2018 data suggest actually drive later outcomes, is solving the wrong problem. It is closer to treating a fever with a cold compress while ignoring the infection underneath: it may look like progress on the thermometer, but it is not addressing the mechanism the original researchers themselves were unsure they had isolated.

What the SECCYD dataset actually contains

The dataset Watts, Duncan, and Quan used, the NICHD Study of Early Child Care and Youth Development, was not built to replicate Mischel's work; it was a much larger federal research effort recruiting 1,364 children and families across ten U.S. locations starting in 1991, specifically designed to capture socioeconomic, racial, and family-structure diversity that a single university preschool never could, according to ICPSR's archived study documentation. Because SECCYD researchers had already collected detailed measures of family income, maternal education, and cognitive development at multiple points in each child's life, Watts, Duncan, and Quan could statistically hold those variables constant and isolate what, if anything, delay-of-gratification behavior at age four added beyond what family background alone already predicted. That is a methodological capability the original 1990 Stanford follow-up, working with 94 demographically similar families, never had the statistical range to attempt.

Why family background explains so much of the effect

Think of the marshmallow test's predictive power like a thermometer reading during a fever: the number is real and measurable, but it does not by itself tell you what's causing the illness. A child's capacity to wait for a second treat reflects trust, hunger, temperament, and the reliability of adults in that child's life, all bundled into one waiting-time score, not a pure readout of future self-discipline.

Consider two four-year-olds. One has grown up in a household where promises reliably come true, where "we'll get ice cream after dinner" actually happens. The other has experienced broken promises, whether from financial instability, unpredictable caregiving schedules, or simple bad luck. For the second child, refusing the marshmallow in hand on the bet that a researcher, a stranger, will really return with two more is not obviously the smarter choice. Researchers working in this area have proposed that trust in whether a promised reward will actually materialize plays a substantial role in delay behavior, independent of any innate capacity for self-restraint.

Family income and maternal education also correlate strongly, on their own, with future academic achievement, independent of anything a four-year-old does with a marshmallow. A wealthier, more educated household is more likely to provide stable housing, consistent nutrition, quality early schooling, and lower chronic stress, all factors with well-documented independent links to later test scores. Once Watts, Duncan, and Quan's statistical models accounted for those factors, there was comparatively little left for delay-of-gratification behavior to explain on its own.

Why might poorer children wait less on average?

Children from lower-income households, on average, waited less in the delay task, and the replication data suggest this is not a deficit in willpower but a rational response to environments where waiting does not reliably pay off. Broken promises and unpredictable resources make a bird in hand a defensible bet, not a character flaw.

A useful, if overstated, insight

None of this means willpower is meaningless or that self-regulation never matters for children's outcomes. It means the original marshmallow test, run on a narrow, privileged sample and popularized without its confidence intervals, became a much bigger cultural claim than the data ever supported. It is closer to a single elegant experiment run under ideal laboratory conditions, like a classic titration performed with perfectly calibrated reagents in a controlled room, than a rule that holds once you test it under the messier, more representative conditions of real families and real neighborhoods.

Later research, including a 2018 direct comparison of the original Shoda, Mischel, and Peake data with the Watts, Duncan, and Quan replication, has generally converged on this more modest picture: delay of gratification at age four correlates weakly with later outcomes, and most of the correlation that does exist likely runs through shared causes like family income and early cognitive skill, rather than delay behavior independently shaping a child's future.

The original correlation between age-four delay time and teenage SAT scores was calculated from just 94 children. No study of that size, however striking its results, can responsibly generalize to a national population, a limitation the original researchers themselves acknowledged in print.

Comparing the two studies directly

Feature	Mischel, Shoda & Peake (1990)	Watts, Duncan & Quan (2018)
Sample size	94 children with SAT data (from 653 original participants)	918 children (552 in focal subgroup)
Population	Stanford Bing Nursery School, mostly faculty/staff families	Nationally representative U.S. sample (NICHD SECCYD)
Age at delay task	Approximately 4 years old	54 months (approx. 4.5 years old)
Outcome measured	Parent-reported SAT scores as teenagers	Academic achievement and behavior at age 15
Raw correlation strength	Strong (widely reported, though from a small sample)	About half as strong as the original
After controlling for family background	Not tested for this variable in the original design	Shrank by roughly two-thirds, largely non-significant
Publication	Developmental Psychology, 1990	Psychological Science, 2018

The core difference is not that the 2018 team disproved Mischel outright; it is that a sample nearly ten times larger and demographically representative showed the effect was smaller than believed and mostly explained by factors the original small study could not control for.

Common misconceptions

The most persistent misconception is that the marshmallow test predicts destiny, that a single four-year-old's choice about candy forecasts a life trajectory. It predicts a modest correlation at best, one that weakens substantially once researchers account for the child's environment, and the popular self-help framing routinely strips out that nuance entirely.

A second misconception is that Mischel's team engaged in bad science or fraud. They did not. They worked with the sample sizes, statistical tools, and reporting norms typical of developmental psychology in the 1970s and 1990s, and Shoda, Mischel, and Peake explicitly cautioned readers about the small sample in their own 1990 paper. The distortion happened later, in decades of media coverage, bestselling books, and parenting advice that flattened a tentative, small-sample correlation into a confident law of child development.

A third misconception is that the 2018 replication "debunked" the marshmallow test entirely. It did not. Watts, Duncan, and Quan still found a small, real, positive correlation between delay behavior and achievement in their raw data. What their study did was reveal how much of that correlation traces back to family circumstances rather than an isolated trait of self-control residing purely inside the child.

The takeaway

The marshmallow test is a genuine experiment with a genuine, modest correlation between early delay behavior and later outcomes, but the version most people know, four-year-old willpower as a crystal ball for lifelong success, was built on a sample of 94 privileged children and never held up to scale. When Tyler Watts, Greg Duncan, and Haonan Quan tested the same basic claim on 918 children representative of the broader U.S. population in 2018, the effect shrank by half before controls and by roughly two-thirds after accounting for family income and background. The 2018 replication does not erase Mischel's original observation; it puts a much-needed decimal point back into a claim that popular science had rounded up to a whole number.

---

Read next:

• The Little Albert experiment explained carefully
• The Bobo doll experiment explained
• The Pavlov dog experiment explained

FAQ

Q: Who created the marshmallow test?

Psychologist Walter Mischel and colleagues at Stanford University designed the delay-of-gratification studies in the late 1960s and early 1970s, testing roughly 653 preschool-age children at the university's on-campus Bing Nursery School.

Q: Does the marshmallow test predict success later in life?

The original small study of 94 children found a correlation between longer waiting times at age four and higher SAT scores as teenagers. A 2018 replication with 918 children, a nationally representative sample nearly ten times larger, found the correlation was about half as strong, and most of it disappeared once researchers controlled for family income, maternal education, and early cognitive ability.

Q: What did the Watts, Duncan, and Quan 2018 study find?

Published in Psychological Science, the study used data from the NICHD Study of Early Child Care and Youth Development to reexamine the delay-of-gratification claim. It found the bivariate correlation between age-four waiting time and age-15 achievement dropped from 0.24 to about 0.08 once family background was controlled for, and became statistically insignificant once concurrent cognitive and behavioral measures were added.

Q: Why do some children wait longer than others in the marshmallow test?

Waiting behavior reflects more than raw willpower. Trust that the promised second reward will actually appear, household stability, hunger, temperament, and general cognitive development all shape how long a given child is willing to wait for it.

Q: Was the original marshmallow test flawed?

It was limited rather than fraudulent: a small sample of about 94 children with SAT data, drawn almost entirely from one financially secure, Stanford-affiliated preschool community, which made its conclusions far less generalizable than decades of subsequent popular coverage suggested.

Q: Did the 2018 replication completely disprove the marshmallow test?

No. Watts, Duncan, and Quan still found a small, real, positive correlation between delay behavior and later achievement. Their contribution was showing that most of that correlation is explained by family background rather than by an isolated childhood trait of self-control.

Q: What is Mischel's "hot and cool system" theory?

Published with Janet Metcalfe in 1999, it proposes that self-control is a contest between a fast, emotional "hot system" that reacts to tempting stimuli and a slow, deliberate "cool system" that can hold an abstract goal in mind. Waiting successfully depends on which system dominates in the moment, not on a single fixed willpower trait.

Q: Can children be taught to wait longer in the marshmallow test?

Yes. Mischel's own studies found that coaching children to distract themselves or to mentally reframe the treat, imagining a marshmallow as a flat picture or a cotton ball rather than food, substantially increased how long they waited, suggesting delay behavior reflects learnable strategy as much as fixed temperament.