Cloud Radio Access Network (C-RAN) is a promising technology to improve user quality of service and reduce network capital and operating costs. The key concept behind C-RAN is to break down the conventional base station into a Base Band Unit (BBU) and a Remote Radio Head (RRH), and to pool BBUs from multiple sites into a single geographical point. Moreover, to achieve statistical multiplexing gain, RRHs should be efficiently clustered: many RRHs may be mapped into a single BBU. In this article, RRH clustering is formulated as a coalition formation game where RRHs collaborate and organize themselves into disjoint independent clusters, in a way to optimize network throughput, power consumption, and handover frequency. An optimal centralized solution, based on exhaustive search, is presented. We also propose a distributed algorithm, based on the merge-and-split rule, to form RRH clusters. Simulation results show that our centralized solution adapts to network load conditions and outperforms the no-clustering method, where only one RRH is assigned to each BBU, and the grand coalition method, where all RRHs are assigned to a single BBU. More importantly, our distributed algorithm achieves very close performance to the optimal solution, with significantly lower computational complexity.