When does a common group identity improve efficiency in coordination games? To answer this question, we propose a group-contingent social preference model and derive conditions under which social identity changes equilibrium selection. We test our predictions in the minimum-effort game in the laboratory under parameter configurations which lead to an inefficient low-effort equilibrium for subjects with no group identity. For those with a salient group identity, consistent with our theory, we find that learning leads to ingroup coordination to the efficient high-effort equilibrium. Additionally, our theoretical framework reconciles findings from a number of coordination game experiments. (JEL C71, C91, D71)