What Is Stereotype Threat? A Finding That Did Not Survive Replication Intact

The idea was introduced in the mid 1990s with experiments reporting that participants performed worse on a test when it was described in terms that made a relevant stereotype salient, and better when the same test was described neutrally. The proposed mechanism was that concern about confirming a stereotype consumes working memory and increases anxiety, leaving less capacity for the task. The finding was attractive because it offered a partial explanation for measured gaps in test performance between groups that did not rest on ability, and because it suggested an intervention, since changing how a test is described is cheap. It was taught widely, cited enormously, and incorporated into policy discussions about testing and education within a few years of the first publications.

Systematic reassessment has substantially weakened the case:

• •Meta-analyses correcting for publication bias find effect sizes near zero in several applications, with the correction methods themselves debated
• •Large preregistered replications have repeatedly failed to reproduce the effect, including well-powered multi-laboratory attempts
• •Original studies were small, and the pattern of results in the literature shows the signature of selective publication
• •Effects reported in laboratory settings have not translated into effects in real testing situations
• •Some specific applications retain more support than others, and the literature is not uniform
• •The underlying mechanisms proposed, including working memory disruption, have their own mixed evidence

The trajectory of this finding is worth examining independently of whether it is true, because the pattern recurs. The claim was socially important, offering a non-deficit explanation for group differences that many people had strong reasons to want. It was simple to state and produced a memorable experiment. It appeared in a period when the statistical practices that permit small unreliable findings to accumulate were standard and unquestioned. It reached textbooks and popular books quickly, which put it beyond the reach of the specialist literature where doubts were raised. And interventions built on it were implemented in schools and universities on the strength of the early work. The correction has been slower than the spread, which is the usual asymmetry, and the finding is still taught in many places as established.

The process by which the evidence was reconsidered is worth describing, since it is the machinery that corrects such errors. Meta-analysis combines results across studies and can be adjusted to estimate how much the published record is distorted by selective reporting, using methods that examine whether small studies report systematically larger effects than large ones. Registered replication reports commit journals to publishing a result before it is known, which removes the incentive that kept failures unpublished. Multi-laboratory projects run the same protocol in many places at once, producing samples far larger than any single study and revealing whether an effect varies by context. Preregistration fixes the analysis in advance. None of these existed as standard practice when the original work was done, and their introduction is why a reassessment was possible at all.

The episode supports several conclusions that hold regardless of how the specific question resolves. A finding that is socially useful receives less scrutiny than one that is not, and awareness of that asymmetry is a defence against it. Effect sizes from small early studies are systematically inflated, so a striking result from a small sample should be treated as a hypothesis rather than a finding. Replication failures are informative and were historically almost unpublishable, which is why the reassessment came decades late. And the existence of measured group differences in test performance is not in dispute, so weakening one explanation does not establish another, since the alternatives include differences in schooling, resources, expectations and much else that the debate about this particular mechanism can obscure.

The claim was that awareness of a stereotype impairs performance on the relevant task, supported by small studies from the mid 1990s and taught widely within years. Meta-analyses correcting for publication bias and large preregistered replications find effects near zero. A socially useful finding receives less scrutiny, and weakening one explanation of measured gaps does not establish another.