Across species and cultures, individuals partition themselves by forming non-overlapping groups. Animals divide into herds, children separate into groups during playtime, politicians create competing political parties, and countries join multilateral defensive treaties. Social network analysis became a central tool of computational social science and has proven instrumental in analyzing complex relational data. However, networks (i.e., sets of dyadic relations) are not good representations of social partitions (i.e., sets of non-overlapping groups). In this talk, I introduce a new statistical framework for social partitions–the Exponential Random Partition Model (ERPM)–inspired by stochastic network models and biological models used in population genetics. I will first describe the properties of the ERP model and show how it can be estimated for empirical data. Second, I will provide a game-theoretical interpretation of the model and explain how this interpretation can be used to formulate hypotheses related to the psychological preferences, individual strategies, or institutional constraints driving group formation. Finally, I will present two case studies employing this new method. The first will focus on the self-formation of teams during hackathons and the second will examine socio-economic segregation in adolescent interaction groups.