Randomized controlled trials have become the gold standard for impact evaluation since they provide unbiased estimates of causal effects. This paper studies randomized settings where treatment is assigned over geographical units. We analyze how omitting spatial correlation in outcomes or unobservables affects treatment effect estimates. First, we study spatial dependence in Mexico's Progresa program. Second, we conduct Monte Carlo simulations to generalize our results. Findings reveal that spatial correlation is more relevant than the literature suggests, and may affect both the precision of the estimate and the estimate itself. Existing spatial econometric methods may provide solutions to mitigate the consequences of omitting spatial correlation.