Four experiments assessed how highlighting multiple identities with divergent implications for task performance can affect stereotype threat effects. In Experiment 1, female undergraduates completed a math test after receiving information either highlighting their gender (stereotype threat), their status as a college student (a group that stereotypically is good at math), both group memberships (multiple identity), or no information about their social identities (control). Math performance was poorest in the condition in which gender was highlighted, but performance in the condition where multiple identities were highlighted was equivalent with the control condition. Experiment 2 used the same procedures as Experiment 1, but identity accessibility was measured directly. Female students in the stereotype threat condition performed most poorly on the math test, and these participants also applied gender labels more quickly to themselves than women in the other conditions. In addition, mediational analyses suggested that decrements in performance could be explained by the relative accessibility of students' gender identity; to the degree that the female identity was more accessible than the college identity, performance was negatively affected. Experiment 3 assessed working memory capacity across the different conditions. Again, women in the stereotype threat condition performed most poorly on the math test and also showed lowest working memory capacity. Lower working memory capacity appeared to account for the poorer performance of these women. In Experiment 4, a different manipulation of group identity salience was used with a group of female undergraduates before they completed a math task. Half of the participants were asked to indicate their gender on a demographic form before taking the test whereas the other participants were not asked about gender. In addition, half were asked to indicate their status as a college student and the other half were not. Those women who were asked only about their gender performed most poorly on the math test. In sum, these results show that having at least one identity suggestive of high performance in a domain can serve a protective function, if that identity is highlighted within the testing context.