There is a vigorous debate on whether arrests for domestic violence (DV) will deter future abuse or create a retaliatory backlash. We study how arrests affect the dynamics of DV using administrative data for over 124,000 DV emergency calls (999 calls) for West Midlands, the second most populous county in England. We take advantage of conditional random assignment of officers to a case by call handlers, combined with systematic differences across police officers in their propensity to arrest suspected batterers. We find that an arrest reduces future DV calls in the ensuing year by 51%. This reduction is not driven by reduced reporting due to fear of retaliation, but instead a decline in repeat victimization. We reach this conclusion based on a threshold reporting model and its testable implications regarding (i) the severity of repeat DV calls and (ii) victim versus third-party reporting. Exploring mechanisms, we find that arrest virtually eliminates the large spike in re-victimization which occurs in the 48 hours after a call, consistent with arrest facilitating a cooling off period during a volatile, at-risk time. In the longer run, we estimate a sizeable deterrence effect. Substantiating this, arrest increases the probability an offender is charged with a crime. Our findings argue against recent calls for a decriminalization of domestic violence and suggest the optimal police response is to lower the threshold for arrest.