Research on the relationship between police and crime, like many criminological topics, is subject to uncertain causal direction and omitted controls. We recommend procedures that mitigate these problems: the Granger causality test, proxies for missing variables, robustness checks, and making data available to other researchers. Because specification problems are common in the social sciences, this strategy has applicability beyond the issue of police and crime. We analyze yearly police data and UCR crime rates, at the state and city levels, pooled over two decades. We find Granger causation in both directions. The impact of crime on the number of police is slight, but the impact of police on most crime types is substantial. The latter result is more robust at the city level.