We experimentally study punishment patterns across network structures, and their effect on cooperation. In a repeated public goods setting, subjects can only observe and punish their neighbors. Centralized structures (like the star network) outperform other incomplete networks and reach contribution levels like the ones observed in a complete network. Our results suggest that hierarchical network structures with a commonly observed player benefit more from sanctions not because central players punish more, but because they follow, and promote, different punishment patterns. While quasi-central players in other incomplete architectures (like the line network) retaliate, and get trapped in the vicious circle of antisocial punishment, central players in the star network do not punish back, increase their contributions when sanctioned by peripheral players, and sanction other participants in a prosocial manner. Our results illustrate recent field studies on the evolutionary prevalence of hierarchical networks. We document a network-based rationale for this positive effect in an identity-free, fully anonymous environment. (JEL C72, C91, C92, D90, H41).