The marshmallow test replication: what survived.

The set-up fits in a sentence. A child sits alone at a table. One marshmallow waits on a plate. If she can hold out — up to fifteen minutes, door closed, nobody watching — she gets two. The experiments behind that scene began at Stanford’s Bing Nursery School in the late 1960s, run by Walter Mischel and his students (Mischel, Ebbesen & Zeiss, 1972). Few studies in psychology have travelled further from their own findings.

The travel happened in the retelling. Decades on, follow-up papers reported that children who waited longer went on to higher SAT scores and better teenage competence (Shoda, Mischel & Peake, 1990). The wait became a window into the soul. Books, talks and school programmes treated fifteen minutes at age four as a forecast of the next forty years. Willpower looked like destiny, measurable with confectionery.

Then, in 2018, the test was re-run at ten times the scale, with modern controls (Watts, Duncan & Quan, 2018). The result did not kill the finding. It resized it — downward, by roughly two-thirds. This article follows the whole arc: the original experiments, the small follow-ups that built the legend, the replication that shrank it, and the larger truth about childhood self-control that survived. The last part is the useful one for teachers, parents and ministries.

The experiment that became a parable

Start with what the Stanford studies were actually about. They were not built to predict anyone’s future. They were built to find out what makes waiting possible — and the answer was: the situation, far more than the child (Mischel, Ebbesen & Zeiss, 1972).

The 1972 experiments make the point cleanly. When the treats sat exposed on the table, children lasted only a few minutes on average. When the same treats were covered, waits stretched several times longer (Mischel, Ebbesen & Zeiss, 1972). Thinking instructions moved the needle just as far. Children told to dwell on the marshmallow’s sweetness caved quickly. Children told to picture it as a round white cloud waited far longer. Same children, same reward, different framing — and waits swung by many minutes.

Mischel’s own summary of two decades of this work ran in Science in 1989, and its emphasis is striking to reread now. Delay of gratification is presented as a set of learnable moves: where you point your eyes, what you tell yourself, how you recast the thing you want (Mischel, Shoda & Rodriguez, 1989). The children who waited were not gritting their teeth. They sang, covered their eyes, turned their backs, invented games. In the original work, self-control looks less like a muscle and more like a technique — one an adult can teach in a minute.

The follow-ups that built the legend

The legend came later, and from a modest place. The Bing families were still reachable, so Mischel’s team sent follow-up surveys as the children reached adolescence. Longer preschool waits correlated with parent-rated competence and with SAT scores (Shoda, Mischel & Peake, 1990). The reported SAT correlations were startling — roughly .42 for the verbal section and .57 for the quantitative section.

The fine print carried more weight than the headline. The correlations appeared only in one condition — treats exposed, no coping strategy suggested — the one arrangement where waiting was hardest and most revealing (Shoda, Mischel & Peake, 1990). And the SAT figures rested on roughly 35 children, drawn from a single university nursery school whose families were mostly Stanford faculty, staff and students. The authors flagged the small sample themselves and asked for caution. The culture skipped the caution.

One more follow-up deepened the myth’s respectability. In 2011, researchers re-tested a group of the original participants, by then in their mid-forties. On a laboratory task, the preschoolers who had struggled to wait were still slightly worse at holding back responses to tempting social cues, with matching differences in brain activity in a small imaging subsample (Casey et al., 2011). So something dispositional is real and durable, at least at the extremes of one small group. The question the field had never answered was different: how much does the wait tell you about an ordinary child’s future? Answering that needed a bigger and fairer sample.

The re-run: 918 children, one number

Tyler Watts, Greg Duncan and Haonan Quan published the answer in 2018 (Watts, Duncan & Quan, 2018). They drew on a large national study in which 918 children had completed a delay task at age four and a half — shortened to seven minutes — and had been followed to age fifteen. The sample was more than ten times the original, and far more diverse. Their headline analyses focused on the children of mothers without a university degree: the group the Stanford campus sample had never represented.

The unadjusted result looked comfortingly familiar. Waiting longer at four predicted higher achievement at fifteen — an association of roughly r = .28 (Watts, Duncan & Quan, 2018). Had the analysis stopped there, the legend would have gained a fresh citation. It did not stop. The authors added the controls any modern study would demand: family background, the home environment, and the child’s cognitive skills measured at the same age as the wait itself.

The association fell by roughly two-thirds, to about a tenth of a standard deviation — statistically marginal, practically small (Watts, Duncan & Quan, 2018). Links with adolescent behaviour problems vanished altogether once controls entered. And one further detail rearranged the picture. Most of the predictive power lived in the first 20 seconds of the wait. Whether a child then lasted two minutes or the full seven added almost nothing (Watts, Duncan & Quan, 2018). The marathon of willpower carried no signal. The first moments — gather yourself, or grab — carried what little there was.

What the shrinkage means

Be precise about what replicated. The raw correlation did. Children who wait longer at four really do show somewhat better outcomes later — in 1990 and in 2018 alike. What failed was the story stacked on top of it: the wait as a clean read-out of a causal trait, one that programmes could train early and cash out decades later.

The controls tell you what the wait was actually measuring. A four-year-old’s delay time bundles many things at once: language and memory, practice at following instructions, the predictability of home, trust that adults deliver what they promise. Those same inputs shape school achievement through their own channels. Control for them, and the marshmallow’s unique contribution nearly vanishes (Watts, Duncan & Quan, 2018). The test was a mirror of a childhood at least as much as a gauge of a child.

Standard caveats run in both directions, and an honest reading keeps them. This was a conceptual replication, not an exact one: a shorter task, a different population, a different era (Watts, Duncan & Quan, 2018). A seven-minute cap may compress differences a fifteen-minute wait would reveal. And a residual tenth of a standard deviation is not nothing; across a school system, small associations describe real patterns. What no serious reading survives is the parable — the one where a nursery-school wait forecasts a life.

Rational waiting

Five years before the replication, a small experiment had already shown why the parable was too simple. Celeste Kidd and colleagues ran 28 children through the classic wait — after a rigged first act (Kidd, Palmeri & Aslin, 2013). Each child first did an art project with an adult who made a promise about better crayons and bigger stickers. For half the children, the adult kept the promise. For the other half, the adult returned empty-handed and apologised.

Then came the marshmallow, with the standard offer. Children who had met the reliable adult waited about twelve minutes on average. Children who had met the unreliable adult waited about three (Kidd, Palmeri & Aslin, 2013). Nine of the fourteen children in the reliable condition lasted the full fifteen minutes. One of fourteen did in the unreliable condition. A nine-minute swing, produced by a single broken promise a few minutes old.

The authors’ framing is the one worth keeping: waiting is a decision under uncertainty, not a willpower meter. Two treats later beats one treat now only if later can be trusted. A child whose world has taught her that promised things vanish is not failing the task when she eats the marshmallow. She is solving it (Kidd, Palmeri & Aslin, 2013). Read this beside the replication and the shrinkage stops being surprising. Delay time partly measures the reliability of a child’s environment — which is precisely what the family-background controls absorbed.

What childhood self-control does predict

None of this makes childhood self-regulation a myth. The strongest evidence in the field says the opposite — it simply measures self-control properly. The Dunedin study followed 1,037 children, an entire year’s births in one New Zealand city, from birth to age 32 (Moffitt et al., 2011). Childhood self-control was measured the slow way: ratings from parents, teachers, trained observers and the children themselves, collected repeatedly between ages three and eleven. Not one task on one afternoon — a composite built over eight years of watching.

That measure predicts adult life with uncomfortable force. Sliding down the childhood self-control gradient, outcomes worsen step by step: more physical-health problems, more substance dependence, worse finances, more crime (Moffitt et al., 2011). Roughly 43% of the lowest self-control fifth had a criminal conviction by age 32, against roughly 13% of the highest fifth. About 27% of the lowest fifth had multiple health problems, against about 11% at the top. The gradient survived controls for intelligence and social class. It even held within families: the sibling with lower self-control tended to fare worse than their own brother or sister decades later (Moffitt et al., 2011).

School-age evidence points the same way. In two cohorts of eighth-graders, self-discipline — measured by combining questionnaires, delay choices and teacher ratings — predicted final grades roughly twice as well as measured IQ did (Duckworth & Seligman, 2005). These are correlations, not causes; disciplined students differ from their peers in many ways at once. But the pattern is consistent wherever self-regulation is measured broadly rather than by stunt.

Put the two literatures side by side and the resolution is almost tidy. One fifteen-minute task, once, in one room: a weak signal, heavily confounded, useless for judging an individual child. The same capacity observed by many eyes, across many years and settings: one of the most robust predictors developmental science owns (Moffitt et al., 2011). The replication did not subtract self-control from the story of childhood. It subtracted the shortcut.

What the evidence doesn’t show

The marshmallow literature is one of the most instructive in child psychology precisely because of its limits. Anyone using it — in a classroom, a parenting book or a policy paper — should hold these as firmly as the findings.

• One task cannot diagnose a child. Delay tasks are noisy and situation-sensitive; a broken promise minutes earlier moved average waits by roughly nine minutes (Kidd, Palmeri & Aslin, 2013). Nothing in this literature licenses screening, labelling or grouping children by a wait.
• Causation was never tested. No trial has trained preschool delay and followed life outcomes. The strategy experiments changed behaviour within a session (Mischel, Ebbesen & Zeiss, 1972); whether trained waiting compounds into adult advantage is unknown.
• The construct was not debunked. The 2018 result shrank one task’s predictive claim (Watts, Duncan & Quan, 2018); the broad, multi-rater self-control gradient stands (Moffitt et al., 2011). Over-reading the replication is as careless as over-reading the original.
• The gradient is not destiny. Even in the lowest Dunedin quintile, most adults had no conviction and no health cluster (Moffitt et al., 2011). Group gradients describe probabilities, never individuals.
• Strategy gains are measured in minutes. Covering treats and cloud-thinking stretched waits within the laboratory hour (Mischel, Shoda & Rodriguez, 1989); durability beyond the session was not the question those studies asked.
• Grades evidence is correlational. Self-discipline out-predicting IQ for GPA (Duckworth & Seligman, 2005) does not show that discipline training raises grades; no random assignment was involved.

Where the evidence stops

1. 1One task cannot diagnose a child
2. 2Causation was never tested
3. 3The construct was not debunked
4. 4The gradient is not destiny
5. 5Strategy gains are measured in minutes
6. 6Grades evidence is correlational

© 2026 FUTURE PROOF™

Self-control, by the evidence

Read as one body of work, the literature converts into practical guidance — and most of it runs against how the legend was used in schools and homes.

Retire the task as a verdict. A delay task tells you about a child’s morning, mood and trust in the adult across the table as much as anything durable (Kidd, Palmeri & Aslin, 2013). Teachers and parents who recreate the test for insight are reading tea leaves. If you need a signal, use the slow kind: repeated observations, several settings, several observers (Moffitt et al., 2011).

Teach the moves, not the trait. The original research is a manual, not a measuring stick: redirect attention, restructure the situation, recast the temptation (Mischel, Ebbesen & Zeiss, 1972). Children wait longer when the treat is covered — so cover it. Phones in a box during homework is the same experiment at age fourteen. The environment does the willpower.

Be reliable before demanding patience. Waiting is a bet on adults keeping their word (Kidd, Palmeri & Aslin, 2013). A classroom where promised rewards actually arrive, on time, every time, is teaching delay more effectively than any character lesson. A home that keeps small promises is doing the same.

Watch trajectories, not moments. The predictive power lives in patterns across years (Moffitt et al., 2011). A term-long drift in a child’s self-regulation — across subjects, noticed by more than one adult — is worth acting on. One bad afternoon is not.

Design routines that spend less willpower. Self-discipline predicts grades (Duckworth & Seligman, 2005), but the practical lever is lowering the price of diligence: fixed study times, small finishable steps, distractions removed in advance. Strong routines make average self-control sufficient — which is the only kind most children will ever need.

Aim policy at the gradient. The Dunedin authors’ own conclusion was for governments: self-control’s gradient runs through the whole population, so early, universal support — stable care, predictable environments, taught strategies — could shift outcomes at national scale (Moffitt et al., 2011). That is a ministry’s lever, not a nursery’s party trick.