There is a long-standing debate on whether the decentralization of government services improves the quality of their delivery. This paper exploits a natural experiment in Colombia whereby the administration of the education sector was decentralized to municipalities with populations over 100,000. By comparing municipalities above and below this threshold through a regression discontinuity design, I am able to estimate the effect of administrative decentralization on learning outcomes in public schools. I find that for those municipalities close to the cutoff, decentralization has a negative impact on student learning measured through performance on the high school exit exam. In order to understand mechanisms, I undertake a case study of the decentralization process in La Guajira, a region characterized by low institutional capacity that due to mismanagement of its resources recently lost administrative control of its education sector. Through interviews corroborated by administrative data, I find that decentralization impacts education through various channels, including the ability and the incentives to appoint qualified teachers, assign resources, contract with private providers and monitor schools.