The Stanford Marshmallow Replication: What the 2018 Study Actually Found

Few psychology experiments have shaped public conversation about child development as much as the marshmallow test. The finding seemed to offer a simple window into a child's future: watch a four-year-old alone in a room with a marshmallow, and you could read decades of their life ahead. The 2018 replication by Watts, Duncan, and Quan did not debunk that experiment with a dramatic reversal. What it did was more methodologically significant: it showed the original effect was mostly a reflection of where the children came from, not who they were destined to become.

The Original Experiment and What It Claimed

Walter Mischel began the marshmallow studies at Stanford University's Bing Nursery School in the late 1960s and early 1970s. The basic procedure was straightforward. A child between three and five years old was brought into a room and shown a treat — usually a marshmallow, though the original experiments also used pretzels and cookies depending on the child's preference. The experimenter explained that they needed to leave the room, and offered the child a choice: eat the treat now, or wait until the experimenter returned (roughly 15 minutes) and receive two treats.

A hidden observation window allowed researchers to watch how children coped with the wait. Some ate immediately. Some lasted a few minutes. A small number waited the full 15 minutes. Mischel and colleagues tracked the children's strategies — looking away, covering their eyes, inventing games, singing to themselves — and these became the basis for later research on self-regulation.

Source: Mischel, W., Ebbesen, E.B. & Zeiss, A.R. (1972), "Cognitive and attentional mechanisms in delay of gratification," Journal of Personality and Social Psychology 21(2), 204–218 — one of the central early papers from the Bing Nursery School studies.

The results that made the marshmallow test famous came from follow-up studies in the 1980s and 1990s. Yuichi Shoda, Mischel, and Philip Peake tracked down the original Bing participants in 1988 and found that those who had waited longer at age four had significantly higher SAT scores in adolescence, better educational attainment, and lower body mass index in adulthood. The correlation between wait time and SAT score was approximately 0.42 — a remarkably large effect for a behavioral measure taken in preschool.

Source: Shoda, Y., Mischel, W. & Peake, P.K. (1990), "Predicting adolescent cognitive and self-regulatory competencies from preschool delay of gratification," Developmental Psychology 26(6), 978–986 — the longitudinal follow-up linking preschool wait time to SAT scores.

The popular interpretation was clear: a child's ability to delay gratification at age four is a window into their character, and that character predicts real outcomes. Self-control became a celebrated trait in parenting books, education policy, and popular psychology writing, with the marshmallow test as the foundational evidence.

The Problem No One Talked About

The original Bing Nursery School studies had a sampling problem that was clear but consistently underemphasized in the way the research was communicated to the public. The Bing Nursery School at Stanford University served the children of Stanford faculty, staff, and graduate students. This is about as non-representative a sample as it is possible to construct.

Children at Bing grew up in households where college education was nearly universal, incomes were well above average, and the surrounding environment was stable, resource-rich, and predictable. These are exactly the children for whom waiting for a second marshmallow makes rational sense — their experience tells them that promises made by adults in institutional settings tend to be kept, that resources are generally reliable, and that the future is a reasonable thing to plan for.

Source: Watts, T.W., Duncan, G.J. & Quan, H. (2018), "Revisiting the marshmallow test: A conceptual replication investigating links between early delay of gratification and later outcomes," Psychological Science 29(7), 1159–1177 — the 2018 replication paper.

The original sample was also tiny by modern standards. Depending on which specific follow-up study you examine, the longitudinal analyses tracked between 35 and 185 children, with high attrition across the years. Effect sizes estimated from small, homogeneous samples tend to be inflated compared to what appears in larger, more diverse populations — a phenomenon called the winner's curse in statistics, or more formally, the small study effect.

Winner's curse — the tendency for early, small studies to report larger effect sizes than later, larger studies, because only dramatic results tend to get noticed and published in the first place.

The 2018 Replication Design

Tyler Watts (then at NYU, now at Ohio State), Greg Duncan (UC Irvine), and Haonan Quan used data from the National Institute of Child Health and Human Development (NICHD) Study of Early Child Care, which followed a nationally representative cohort of children from birth. The NICHD sample was far larger and more socioeconomically diverse than the Bing sample.

Key differences between original and replication:

Feature	Mischel Original	Watts et al. 2018
Sample size (longitudinal)	35–185 children	918 children
Sample source	Stanford Bing Nursery School	NICHD national cohort
Socioeconomic diversity	Very low (mostly faculty families)	High (full range of backgrounds)
Age at delay task	~3–5 years	~4.5 years
Age at outcome measurement	Adolescence	Age 15
Controls used	Minimal	Extensive (SES, cognitive ability, home environment)
Primary outcome	SAT scores	Achievement test scores, behavioral measures

Source: Watts et al. (2018), Psychological Science — Table 1 in the paper gives the full demographic breakdown of the NICHD sample.

The delay-of-gratification task used was adapted from the original marshmallow procedure. At age 54 months (approximately 4.5 years), children were offered a choice between a smaller immediate reward and a larger delayed reward, with a maximum wait time of seven minutes. The researchers measured how long each child waited before eating the treat.

At age 15, the researchers measured academic achievement using standardized test scores and also assessed behavioral outcomes including self-reported self-control and externalizing behavior problems.

What the Replication Found

The first finding confirmed that the marshmallow test does predict later outcomes in a bare correlation. Children who waited longer at age 4.5 scored higher on achievement tests at age 15. In that limited sense, the basic result held.

But the effect was smaller than in the original. Expressed in comparable terms, the bivariate correlation between wait time and achievement was roughly half the size of what the Bing studies had reported.

Then the researchers applied controls.

When they added statistical controls for family socioeconomic status, maternal education, home environment quality, and the child's cognitive ability at age four, the predictive power of wait time dropped by approximately two-thirds. For children whose mothers had college degrees — the group most similar to the original Bing sample — the effect remained statistically meaningful. For children whose mothers had not completed college, the effect became very small and statistically fragile.

Source: Watts et al. (2018), Psychological Science 29(7) — Table 3 shows the coefficient for wait time before and after adding control variables, broken down by maternal education.

In practical terms, for a child from a lower-SES family, waiting an additional minute in the delay task at age four predicted an improvement of about one-tenth of a standard deviation in achievement scores at age 15 — a real but modest effect. For the group with college-educated mothers, the effect was larger. But in both cases, most of the predictive work was being done by the background characteristics, not by the wait time itself.

What the Results Mean

The replication did not show that self-control is irrelevant to life outcomes. Self-control is a real trait with real consequences. What the replication showed is that a four-minute observation of a preschooler's behavior around a marshmallow is not a reliable measure of that trait when considered in isolation from the environment in which the behavior occurs.

A child who does not wait for the marshmallow may be making a rational choice. Children from lower-SES backgrounds have more experience with environments where promised rewards do not materialize, where adults are unreliable, and where the future is genuinely uncertain. In that context, eating the marshmallow immediately is not an expression of poor self-control — it is a reasonable inference from experience. Research by developmental psychologist Celeste Kidd and colleagues published in 2013 demonstrated exactly this: children who had been made to wait for a promised art supply (and either received it or did not) performed very differently on a subsequent marshmallow task, with the "unreliable experimenter" condition producing dramatically shorter wait times.

Source: Kidd, C., Palmeri, H. & Aslin, R.N. (2013), "Rational snacking: Young children's decision-making on the marshmallow task is moderated by beliefs about environmental reliability," Cognition 126(1), 109–114 — the study showing that the perceived reliability of the environment modulates marshmallow task behavior.

This is the key conceptual shift. The original marshmallow interpretation treated the child's behavior as a readout of an internal, stable trait. The replication evidence, combined with Kidd's work, suggests that the behavior is partly a readout of environmental experience — and that what predicts outcomes is primarily that environment, not the behavior per se.

A child who grows up in a stable, resource-rich household waits for marshmallows. They then go on to achieve better educational and career outcomes, not primarily because they waited for the marshmallow but because they grew up in a stable, resource-rich household.

The Reaction and the Debate

The 2018 paper generated significant attention and some pushback. Several researchers argued that the NICHD delay task was different from Mischel's original procedure in ways that could reduce its validity as a measure of self-control. The seven-minute maximum wait was shorter than the 15-minute wait in some original experiments. The reward structure was somewhat different.

Source: Michaelson, L.E. & Munakata, Y. (2020), "Same data set, different conclusions: Preschool delay of gratification predicts later behavioral outcomes in a preregistered study," Psychological Science 31(2), 193–201 — a reanalysis of the same NICHD dataset that found some predictive effects remained after different analytical choices, illustrating how analytic flexibility affects conclusions.

The Michaelson and Munakata reanalysis is worth noting because it used the same dataset and found that wait time did predict some behavioral outcomes (externalizing problems in particular) even after SES controls, using different model specifications. This illustrates that the data do not deliver a clean verdict — they deliver results that depend on how the question is framed and how the analysis is structured.

Mischel himself, before his death in 2018, acknowledged that the original findings had been overgeneralized in popular discourse. He had always maintained that self-control was teachable and environmentally shaped, and he expressed concern that the marshmallow test had been used to support a view of self-control as a fixed childhood trait — a view he did not hold.

Source: Mischel, W. (2014), The Marshmallow Test: Mastering Self-Control, Little, Brown — Mischel's own popular account, which emphasizes the role of strategies and context in delay behavior rather than fixed trait.

The Policy Consequences of the Original Claim

The original marshmallow finding did not stay in academic journals. It entered education policy through a specific pathway: the idea that self-control is a teachable trait with large, lifelong payoffs made it attractive as a target for school-based intervention programs. If a single brief behavioral observation in preschool could predict SAT scores a decade later, then improving that early behavior through deliberate training seemed like a high-leverage investment.

This reasoning drove real policy. The KIPP (Knowledge Is Power Program) network of charter schools, founded in 1994 and growing to hundreds of schools by the 2010s, built a significant portion of its ethos around character education and self-control — influenced partly by the positive psychology movement that drew heavily on the marshmallow literature. KIPP schools emphasized grit, self-discipline, and delayed gratification explicitly as school culture priorities.

Source: Tough, P. (2012), How Children Succeed: Grit, Curiosity, and the Hidden Power of Character, Houghton Mifflin Harcourt — the popular book that synthesized the marshmallow and related research for a policy audience, influencing a generation of education reformers.

The 2018 replication does not prove that KIPP's character-education approach is ineffective. But it does undermine the specific theoretical justification for placing self-control training at the center of an educational program for low-income children. If the marshmallow test's predictive power in low-SES populations largely reflects environmental instability rather than a trainable character trait, then the intervention that matters is not practicing waiting — it is reducing the environmental uncertainty that makes waiting irrational.

This is a consequential difference. An education program built on the premise that low-income children need to be trained to delay gratification implicitly locates the problem in the child. An approach built on the premise that those children are behaving rationally in response to an unstable environment locates the problem differently — and points toward different solutions, including housing stability, food security, and reliable adult figures rather than classroom exercises in self-control.

Source: Harden, K.P. (2021), The Genetic Lottery: Why DNA Matters for Social Equality, Princeton University Press — discusses how attributing outcome disparities to trainable character traits can obscure structural and environmental causes, with the marshmallow test used as an example.

What the Replication Literature Teaches About Big Effects

The marshmallow story is a clean illustration of a pattern that appears repeatedly in psychology replication work. A small, homogeneous sample produces a large, striking effect. The effect attracts enormous attention because it is large and because it has a compelling narrative attached to it. That narrative enters popular culture and educational policy. Years later, a larger and more rigorous study finds a smaller effect that is largely explained by variables the original study failed to control.

This sequence does not mean that the original researchers were dishonest or careless. It means that effect sizes estimated from small, selected samples tend to be inflated, that variables omitted from an analysis do not disappear — they simply flow through the measured variables and make them look more powerful than they are, and that a compelling narrative can delay rigorous scrutiny for decades.

The replication crisis in social and behavioral science is, in large part, a story about what happens when these three factors combine.

Source: Ioannidis, J.P.A. (2005), "Why most published research findings are false," PLoS Medicine 2(8), e124 — the statistical paper that laid out why small-sample, high-flexibility research tends to produce inflated and irreplicable effect sizes.

The Takeaway

The Stanford marshmallow replication by Watts, Duncan, and Quan is less a debunking than a clarification: the link between early self-control and later outcomes is real but substantially smaller than the original studies suggested, and it is largely carried by socioeconomic background rather than by something that a preschool marshmallow test measures independently. A child's decision to eat or wait is shaped by their environment, their past experiences with reliability, and the cognitive resources available to them at age four — most of which are products of the circumstances they were born into. That is a more complicated story than "self-control predicts success," but it is a more accurate one, and it has meaningfully different implications for where we should direct effort in education and early childhood policy.

---

Read next:

• The Marshmallow Experiment
• The False Belief Task
• The Rosenhan Experiment

---

Q: Did the 2018 replication find any predictive power in the marshmallow test at all?
A: Yes. Even after controlling for socioeconomic status and other background factors, wait time at age four predicted achievement at age fifteen for children whose mothers had college degrees. The effect was statistically significant in this subgroup. For children from lower-SES backgrounds, the effect was much smaller and less reliable. The replication weakened the original claim without eliminating it entirely.

Q: Is the marshmallow test still a useful research tool?
A: As a measure of delay of gratification under specific conditions, yes. As a predictor of individual life outcomes stripped of context, it is a poor tool. Researchers continue to use variants of the task to study self-regulation processes, cognitive strategies for managing temptation, and how environmental factors modulate willingness to delay — all legitimate scientific questions.

Q: Why did the original study find such large effects if they were inflated?
A: Three factors combined. The Bing sample was socioeconomically homogeneous, which reduced variance in background characteristics and made the wait-time measure look more predictive by default. The sample was small, which inflates effect-size estimates due to sampling error. And the follow-up measurements were conducted by researchers who already had a theory about what the data would show, which can introduce subtle biases in data collection and analysis.

Q: What is the "rational snacking" finding?
A: Celeste Kidd's 2013 study showed that children who experienced an unreliable experimenter (one who promised art supplies and did not deliver) subsequently waited much less time on the marshmallow task than children who had a reliable experimenter. This suggests that the decision to wait is partly a rational inference about environmental reliability, not purely a measure of impulse control.

Q: Does teaching children to wait for rewards improve life outcomes?
A: The evidence is mixed. Programs that specifically train self-control strategies in preschool settings (such as Tools of the Mind) have shown benefits in controlled trials, but the effects are generally modest and the mechanisms are unclear. The more consistent finding is that improving the stability and resources of a child's environment — reducing poverty, housing instability, and parental stress — produces lasting improvements in self-regulation as a byproduct.

Q: What did Walter Mischel say about the replication?
A: Mischel, who died in September 2018 shortly after the replication was published, had always maintained that self-control was shaped by learning and context, not a fixed trait. He expressed concern in later years that his work had been misread to support a deterministic view. He argued that the key lesson of the marshmallow research was that cognitive strategies for self-regulation can be taught — not that a four-year-old's performance on a single test reveals their destiny.