The “Broken Windows” theory of crime predicts that the incidence of low intensity crimes in a neighborhood leads to future increases in the frequency of more severe crimes. We present a simple, dynamic model of law enforcement and test this prediction using a detailed database comprised of every crime report filed with the Dallas Police Department from 2000-2007. Notably, we isolate the intertemporal causal effect of crime that is outside of the control of law enforcement from the total reduced-form causal effect of previous crimes on future crimes, which includes the future police response to previous crimes. Our identification strategy, which takes advantage of geographic variation in social learning and law enforcement, explicitly accounts for the facts that many determinants of crime are unobservable and produces parameter estimates that are robust to a variety of empirical concerns. We find no evidence that reductions in light crime lead to short run or long run reductions in more severe crimes. These results cast considerable doubt on the claim that the dramatic reduction in the violent crime rate in US urban areas over the past fifteen years is due to “zero-tolerance” law enforcement policies. Policymakers aiming to decrease the prevalence of severe crimes ought to implement policies narrowly targeted to severe crimes. JEL Codes: K42 , R23