The marshmallow test was never about willpower

You have heard the story: the 4-year-olds who waited for two marshmallows did better at everything for the rest of their lives. When it was retested on a bigger, more representative sample, most of that link turned out to be something else entirely.

Published 5 June 20266 min read4 published sources, caveats printedFor parents of 6–11s
The short version
1The 2018 replication used 918 children rather than a few dozen. Each extra minute a 4-year-old waited predicted about 0.1 standard deviations of achievement at 15.
2That correlation was half the size of the original — and dropped by about two thirds once family background, early ability and home environment were accounted for.
3Most of the predictive power came from whether a child could wait 20 seconds at all, not from heroic patience.
4It is still argued over. But nobody now claims a marshmallow at 4 tells you who your child will be — which should be a relief.
01

What the retest found

The original studies were small — dozens of children, drawn largely from a university community — and the correlations they reported were strong. In 2018 researchers ran a conceptual replication on 918 children from a large, more representative longitudinal study, focusing on the group whose mothers had not completed college.

The link was still there. It was just much smaller than the story required, and most of it dissolved once you accounted for the circumstances a child was growing up in.

The same finding, at three levels of rigour
Original studies, small samplesstrong link
Replication, 918 childrenhalf the size
…adjusting for background and early abilityabout a third of that
Watts, Duncan & Quan, Psychological Science, 2018. Bars are our illustration of the reported reductions — the bivariate correlation was roughly half the original and fell by about two thirds with controls.
02

The 20-second detail nobody quotes

Buried in the replication is the most interesting finding of all: most of the variation in later achievement came from whether a child could wait at least 20 seconds. Beyond that, waiting longer added little.

That points at something much more ordinary than willpower — basic impulse control, plus whether a child had reason to believe a promised second marshmallow would actually arrive. A child from a chaotic environment who eats it immediately is not failing a character test. They are making a reasonable bet.

918
children in the replication, versus dozens originally
0.1 SD
achievement at 15, per extra minute waited at 4
20 s
where nearly all the predictive power sat
A child who eats the marshmallow immediately is not failing a character test. They are making a reasonable bet about whether adults deliver.
03

What to take from it as a parent

Stop
Reading destiny into one test

No single behaviour at 4, 6 or 9 forecasts a life. Anything sold to you on that premise is selling you the 1990 version of a story that did not survive.

Start
Making promises you keep

Waiting is partly a judgement about whether adults deliver. Predictability is something you can actually give a child.

Keep
Practising the pause, gently

Simple impulse control does seem to matter, and it responds to ordinary things: games with turns, plans made out loud, a countdown before acting.

And the honest coda: this is a live argument. Several researchers have published rebuttals arguing the replication’s statistical choices were too harsh and that the original finding partly survives. We are printing that too, because a Notebook that only reports the tidy version is not much use.

Questions parents actually search for

Was the marshmallow test debunked?

Not exactly. A 2018 replication with 918 children found the link between waiting at 4 and achievement at 15 was real but about half the original size, and fell by roughly two thirds once family background, early cognitive ability and home environment were taken into account.

Does self-control in childhood predict success?

Weakly, and mostly through circumstances that travel with it. The replication suggests interventions aimed narrowly at making children wait longer are unlikely to move later achievement much.

What actually predicted the outcome?

Being able to wait about 20 seconds — basic impulse control — carried most of the predictive power. Waiting several minutes added little beyond that.

Is the original research now considered wrong?

It is considered overinterpreted rather than fabricated, and the interpretation is still disputed. Commentaries published after 2018 argue the replication’s controls were too aggressive and that a smaller version of the original effect stands.

Everything this piece leans on
Watts, Duncan & Quan, Psychological Science, 2018
Conceptual replication with 918 children: one extra minute waited at age 4 predicted about 0.1 SD in achievement at 15; roughly half the original correlation, reduced by about two thirds with controls; most variation came from waiting at least 20 seconds.
Caveat: observational data; the choice of control variables is exactly what critics dispute.
Shoda, Mischel & Peake, Developmental Psychology, 1990
The original longitudinal finding linking preschool delay of gratification to adolescent outcomes.
Caveat: small, unrepresentative sample from a university community.
Falk, Kosse & Pinger, Psychological Science, 2020 (commentary)
Argued the replication’s correlations were biased toward zero by censoring of waiting times and that controlling for non-exogenous covariates is hard to interpret.
Caveat: a commentary rather than new data collection.
Michaelson & Munakata, and later reanalyses
Reanalyses of the same dataset found links between delay of gratification and several adolescent outcomes, strongest for problem behaviours at 15.
Caveat: same data, different analytical choices — a good illustration of why single studies should not settle questions.

One finding a week, in plain words, with its caveat attached

That is the whole newsletter. Two minutes on a Friday, nothing sold, unsubscribe in one click.

Get the weekly noteIt comes with a parent account — the only place we ask for an email, and it is covered by the privacy notice.
Keep reading