The identification of counties burdened by exceptionally high rates of mortality is a fundamental step in the development of state-based intervention and prevention strategies. However, the estimation of rates from small geographic areas presents special problems, especially for rare events. This paper compares the use of crude and age-standardized rates to the use of Poisson regression models and empirical Bayes models for analysing county-level mortality rates. The results demonstrate both practical and heuristic advantages of the empirical Bayes models. Age-standardized rates adjust for differences in age structure among countries but are vulnerable to extreme variability in county age-specific rates. In our example--an analysis of diabetes mortality rates--Poisson regression did not improve the variability of estimated county-level rates. Adjusted empirical Bayes estimates dramatically shrink the observed rates while preserving some separation of the counties with extreme rates. Also, empirical Bayes estimates of rates for countries with no observed deaths are shrunk close to the prior mean.