The role of unobservable characteristics in friendship network formation

Inbreeding homophily is a prevalent feature of human social networks with important individual and group-level social, economic, and health consequences. The literature has proposed an overwhelming number of dimensions along which human relationships might sort, without proposing a unified empirically-grounded framework for their categorization. We exploit rich data on a sample of University freshmen with very similar characteristic - age, race and education- and contrast the relative importance of observable vs. unobservables characteristics in their friendship formation. We employ Bayesian Model Averaging, a methodology explicitly designed to target model uncertainty and to assess the robustness of each candidate attribute while predicting friendships. We show that, while observable features such as assignment of students to sections, gender, and smoking are robust key determinants of whether two individuals befriend each other, unobservable attributes, such as personality, cognitive abilities, economic preferences, or socio-economic aspects, are largely sensible to the model specification, and are not important predictors of friendships.