Global games with endogenous information often exhibit multiple equilibria. In this paper we show how one can nevertheless identify useful predictions that are robust across all equilibria and that could not have been delivered in the common-knowledge counterparts of these games. Our analysis is conducted within a flexible family of games of regime change, which have been used to model, inter alia, speculative currency attacks, debt crises, and political change. The endogeneity of information originates in the signaling role of policy choices. A novel procedure
of iterated elimination of non-equilibrium strategies is used to deliver probabilistic predictions that an outside observer—an econometrician—can form under arbitrary equilibrium selections. The sharpness of these predictions improves as the noise gets smaller, but disappears in the
complete-information version of the model.