Social identity threat refers to the process through which an individual underperforms in some domain due to their concern with confirming a negative stereotype held about their group. Psychological research has identified this as one contributor to the underperformance and underrepresentation of women, Blacks, and Latinos in STEM fields. Over the last decade, a brief writing intervention known as a values affirmation, has been demonstrated to reduce these performance deficits. Presenting a novel dataset of affirmation essays, we address two questions. First, what linguistic features discriminate gender and race? Second, can topic models highlight distinguishing patterns of interest between these groups? Our data suggest that participants who have different identities tend to write about some values (e.g., social groups) in fundamentally different ways. These results hold promise for future investigations addressing the linguistic mechanism responsible for the effectiveness of values affirmation interventions. [For complete proceedings, see ED560503.]