Accounts in conflict: Evidence of selection bias in empirical violence data
There is a growing debate on measurement error in empirical data on political violence. Violent events are at risk of remaining unobserved if, for example, they occur in remote locations, if their frequency exceeds available registration capacity, or if perpetrators successfully conceal their deeds. The possibility of challenges to direct observation is usually unknown and can hence not be accounted for empirically. The case of armed conflict in Kosovo (1998-2000) presents an exceptional research opportunity because of various available data sources on war victims. Among them is a recently completed census by the Humanitarian Law Centre (HLC) that constitutes a reliable representation of ‘ground truth.’ In this paper, I develop a theory of measurement which is concerned with how data on violent events comes to be. I then link six data sources on violent deaths to the HLC census. I investigate underlying selection processes with regard to victim, event, and overall conflict context characteristics to explain the capture of deaths into each data system. Statistical models show that capture dynamics are systematic to statistical significance. I discuss the findings with regard to the issue of selection bias in empirical research on violence and formulate recommendations for future research.