Observing violent behavior is problematic for generating valid inferences. Acts are imperfectly observed because monitors and other observers are not able to access all places within a specific political context at all times. Strategic choices of perpetrators and limited resources of monitors entail that complete and unbiased data of violent acts or victims are rarely if ever available. Scholars are aware of these issues and have developed a variety of solutions. However to date, we are unsure of which solutions provide the most valid estimates of the “true” number of events or victims. Armed conflict over Kosovo presents an ideal case to test a variety of methods against each other because of the availability of a complete and accurate census of human losses that occurred from 1998 through 2000. Additional to the census, six other data sources generated by a variety of different enumerators are available. These various data sources are representative of the type of incomplete and biased data scholars usually have at their disposal to empirically research violence. Using this set of “imperfect data,” we apply the variety of research methods that have to date been suggested to mitigate issues of underregistration and selection bias. These methods include using the data as is, pooling several sets of data, weighting estimates according to hypothesized selection effects, standards-based approaches, latent variable models, and multiple systems estimation. A comparison of our estimates to the census allows us to establish which methods appear most valid and reliable for recovering the “truth” and making inferences with data from direct observation. Our results provide important implications and insights for the future of empirical research on political violence.