Revisiting the Marshmallow Test: A Conceptual Replication Investigating Links Between Early Delay of Gratification and Later Outcomes

918 full analytic sample; 552 children of mothers without a college degree (the preregistered primary subsample). Shoda, Mischel & Peake 1990 had 35-48.

NICHD Study of Early Child Care and Youth Development (SECCYD); delay of gratification measured at 54 months, outcomes at Grade 1 and age 15.

Conceptual replication of the most famous early-childhood predictive correlation in psychology, in a sample roughly 10x the original and far more demographically varied. Not causal: this is an observational longitudinal design. Its value is the covariate ladder — the same association is reported bivariate, then with family background + HOME environment, then with concurrent 54-month cognitive and behavioural measures added.

The marshmallow effect largely dissolves under controls. In the primary subsample an additional minute waited at age 4 predicted ~0.1 SD in age-15 achievement — HALF the original correlations — and that association fell by two thirds once family background, early cognitive ability and the home environment were controlled (beta 0.236 -> 0.081), and to a non-significant 0.050 (p=.140) with concurrent 54-month measures added. Almost all the predictive signal came from waiting at least 20 SECONDS, not from heroic self-control. Behavioural outcomes at 15 showed virtually no association even WITHOUT controls. Among children of college-educated mothers the bivariate association (0.167, p=.007) became statistically indistinguishable from zero once controls were added.

Directly on point. The controls that kill the association include early cognitive ability at 54 months and the home environment — both substantially heritable and both supplied by the same parents. This is a worked example of a famous "self-control" predictor being mostly a proxy for heritable ability and family background rather than a teachable trait.

Contested in both directions. Falk, Kosse & Pinger (2020, Psych Sci 31:100-104) argue that once the two studies are made design-comparable the direction is confirmed though the effect is less pronounced; Doebel, Michaelson & Munakata (2020) argue the controls over-adjust (construct clarity). Watts & Duncan (2020, Psych Sci 31:105-108) reply defending the specification. The undisputed core is that the original effect size was roughly double the replicated one.