This article examined the role of collective identity in stereotype threat (i.e., when thoughts about one's social groups are prominent in one's thoughts). It was reasoned that if stereotype threat affects performance by increasing the prominence or accessibility of a group-based stereotype, then the accessibility of stereotype-relevant group memberships should be higher in settings in which a stereotype might apply, than in situations in which it could not. Experiment 1 tested this notion by leading male and female undergraduates to expect that they would complete a math test that was described either as "diagnostic of math ability" (stereotype threat for women) or as a "reasoning exercise" (control). Before taking the supposed math test, students completed a "reading comprehension" exercise in which they were asked to list group-based pronouns (e.g., "we," "us") that had been underlined in a written passage. Consistent with predictions, women under stereotype threat produced more group-based pronouns than did women in the control condition. Men listed fewer pronouns than women in the threat condition, and their number of listings did not vary by condition. These results confirmed the hypothesis that collective social identities become highlighted by stereotype threat. Experiment 2 used a similar method, but just with female undergraduates to assess the kinds of social identities that are highlighted by stereotype threat. Stereotype threat increased the accessibility of women's gender category but not other categories that were irrelevant to the stereotype (e.g., friend, family member, student). Moreover, the degree of gender identification was negatively related to performance expectations for the upcoming math test. In Experiment 3, female undergraduates in the stereotype threat or control condition read a fictitious newspaper article that described a female student who either excelled or did less well in math. One-third of the students did not receive any information about another student (control). When the test was described as diagnostic of ability (and gender identity was presumably most accessible), women were strongly affected by whether they had been exposed to a positive role model. Their expectations of academic success were higher, their concerns about others' evaluations were lower, and their actual math performance was better than when they had been exposed to a less impressive role model. Moreover, these effects were generally reversed in the nondiagnostic condition, where social identities would not have been accessible, showing that role models can actually reduce expectations and performance if a shared social identity is not obvious or highlighted. Experiment 4 extended this research by providing female students with a role model demonstrating excellence in math or an unrelated domain (sports). Role models only improved expectations of success and actual performance if they demonstrated excellence in the relevant domain of math. These studies point to the important role of accessible collective identities in stereotype threat, and highlight that role models can eliminate performance decrements under threat (and, in fact, can improve performance) when they are viewed as sharing a collective identity with a person under threat.