In this paper, we propose to compare different partial transparency regimes in order to determine the optimal diessemination policy by the central bank, using an experimental approach. A treatment dedicated to the benchmark situation (where information is fully released) is also available. Our experiment is based on subsequent framework of Morris and Shin (2002), Cornand and Heinemann (2008) and Trabelsi (2012). The predictive power of K-level reasoning is an issue that is addressed also in this paper. Our experiment indicates that -when fully disclosed- players overreact to public information and this overreaction is efficiently reduced when the degree of publicity decreases (i.e. when the fragmentation measure increases). The average weight assigned to common signal decreases over treatments, especially when we establish partial transparent strategy (i.e. fragmented information). The results provide support both for and against global games theoretical predictions. In fact, although players overreact to public signal, their behavior is inconsistent with theoretical equilibrium, which means that the destabilizing effect of public information is less pronounced experimentally than when it does in theory. This is not the case when public information is fragmented ; subjects' behavior does approach equilibrium. These observations coincide with both a collective and an individual analyses of behavior.