Bias in epidemiological studies of conflict mortality

Cluster sampling has recently been used to estimate the mortality in various conflicts around the world. The Burnham et al. (2006) study on Iraq employs a new variant of this cluster sampling methodology. The stated methodology of Burnham et al. (2006) is to (1) select a random main street, (2) choose a random cross street to this main street, and (3) select a random household on the cross street to start the process. We show that this new variant of the cluster sampling methodology can introduce an unexpected, yet substantial, bias into the resulting estimates as such streets are a natural habitat for patrols, convoys, police stations, road-blocks, cafes and street-markets. This bias comes about because the residents of households on cross-streets to the main streets are more likely to be exposed to violence than those living further away. Here we develop a mathematical model to gauge the size of the bias and use the existing evidence to propose values for the parameters that underlie the model. Our research suggests that the Burnham et al. (2006) study of conflict mortality in Iraq may represent a substantial overestimate of mortality. We provide a sensitivity analysis to help readers to tune their own judgements on the extent of this bias by varying the parameter values. Future progress on this subject will benefit from the release of high-resolution data by the authors of Burnham et al. (2006). Introduction Recent studies of conflict mortality, such as the one on Iraq (Burnham et al., 2006) survey households using a cluster sampling methodology. Cluster sampling itself is not problematic but the micro-level details on how households are selected at the final stage of sampling are crucial and widely overlooked (see Appendix 1 for details). As described by the EPI sampling methodology (e.g., Spiegel and Salama, 2000, Depoortere et al., 2004 and Coghlan et al., 2006), these studies often initiate the sampling process from some easily accessible geographical feature, such as the centre of a village, in order to economize resources and ensure staff safety. The stated procedures in Burnham et al. (2006) call for selecting a “constituent administrative unit” and then selecting a main street from “a list of all main streets. A residential street was then randomly selected from a list of residential streets crossing the main streets.” (Figures 1 & 2). The field team would enumerate the