Used 911 calls in neighboring beats as an instrumental variable for police presence in Dallas (cops would be moved to a neighboring precinct to deal with the emergency). Considered unusually complicated for an economics paper since it involved what they consider big data, joining multiple datasets, and dealing with geolocations. Effect sizes seem implausibly large to me. A 10% reduction in police presence increases violent crime by 7.5% on average. Thefts are unchanged.
Effect size is much larger than difference-of-difference studies that use exogenous effects like terrorism to measure the impact of the police presence. Sarit claims that's because removing police is different than adding police. And endogenously added police haven't had time to learn the new beat yet (and probably have different incentives). Similar effect as previous IV studies that used new police hires as an IV but with better confidence intervals.