We study the distribution of food aid in Ethiopia between 1994 and 2004 using data from the Ethiopian Rural Household Survey. Over this period village leaders had considerable discretion in disbursing aid subject to official guidelines and periodic monitoring. We use a principal-agent model and household panel data for approximately 940 households to understand biases in the allocation of aid. The model shows that correlations between aid and observed measures of need are not a good measure of targeting because agents have incentives to distort allocations within targeted classes. Consistent with the model, we find that the aid recipients match official criteria but disbursements are negatively correlated with determinants of need that are not easily observable by monitoring agencies, namely pre-aid consumption, self-reported power and involvement in village-level organizations. Our results suggest informal structures of power within African villages influence the extent to which food aid insulates some of the world's poorest families from agricultural shocks but also that policy guidelines do constrain permissible deviations from need-based allocations.