Abstract
We develop a novel application of machine learning and apply it to the interview transcripts from the American Voices Project (N = 1,396), using discourse atom topic modeling to explore social class variation in the centrality of family in adults' lives. We take a two-phase approach, first analyzing transcripts at the person level and then at the line level. Our findings suggest that family, as represented by talk, is more central in the lives of those without a college degree than among the college educated. However, the degree of institutional overlap between family and other key institutions-health, work, religion, and criminal justice-does not vary by education. We interpret these findings in the context of debates about the deinstitutionalization of family in the contemporary United States. This demonstrates the value of a new method for analyzing qualitative interview data at scale. We address ways to expand the use of this method to shed light on educational disparities.