Monitoring civilian mortality in conflicts can help target humanitarian assistance and minimize loss of life among those caught up in conflict. While in stable situations surveillance of mortality using vital registration systems is the gold standard, these systems are rarely functional during conflicts and are often nonexistent where major conflicts occur ‐. Alternative approaches, such as passive surveillance through news media or press reports have been shown to under-record deaths and may be distorted by political agendas . Consequently, epidemiologists are often limited to estimating mortality using retrospective population-based surveys at the household level . In these, a representative sample of consenting households is selected to assess mortality events over a given time period, and mortality rates, along with their upper and lower confidence intervals, are calculated . These calculations can also be compared to other time points to estimate excess mortality related to a conflict. However, simple random and systematic random sampling methods are difficult during conflict given data unavailability and logistical and security constraints [6, 7].
One alternative approach is to apply cluster sampling to estimate conflict-related mortality rates. Two-stage cluster sampling was standardized in 1978 by the World Health Organization's Expanded Programme on Immunization (WHO EPI) to assess vaccine coverage and has since been extended to estimate conflict-related mortality in Iraq [8, 9], Kosovo , the Democratic Republic of Congo , and Sudan . This approach is relatively fast, can be done with limited financial and human resources, and exposure to unsafe areas can be limited. Additionally, a complete sampling frame is not required. These are all very important considerations in conflict settings.
In conventional two-stage cluster sampling, the first sampling stage involves the selection of a predetermined number of clusters. Clusters are mutually exclusive subpopulations, most frequently constructed from recognized administrative boundaries . Clusters are selected from a list of primary sampling units (e.g. census areas, township boundaries) with the probability of selection proportional to population size (or estimated population size) . In the second stage, starting households are selected from each cluster. As complete and adequate listings of households rarely exist, households are not selected from a sampling frame. Rather, they are selected by the survey team in the field based on a random procedure . Most commonly, starting households are selected in the field based on the “random walk” method, which involves identifying the center of the cluster, or another easily distinguishable feature such as a main street, and selecting a random direction to walk, thus drawing a transect across the cluster. In practice, the random direction is often selected by “spinning a pen” . Among those households that lie along the transect, one household is randomly selected as the starting household and a predetermined number of next nearest households are surveyed. Ultimately, the data collected from each cluster are pooled to make inferences with respect to the target population and standard errors are adjusted for design effects of using a cluster sampling approach .
Despite the benefits of this WHO EPI-type cluster sampling, the validity of this approach has been questioned . Most criticisms are related to the potential for bias in the second stage when using the “random walk” approach, which has been shown to introduce bias if the household selection procedure is not in fact random [7, 14, 16, 18, 19]. In addition, this approach is subject to interviewer bias, whether conscious or unconscious, and can take a significant amount of time to implement in the field. Too much field time exposes survey team members to risk in conflict settings. It is also impossible to calculate the probability of selection at the household level, so the sample is not a true probability sample .
To date, a few variations on this conventional cluster sampling approach have been developed for application in nation-wide health studies. Relevant examples that have been used and show promise for certain settings include compact segment sampling [14, 20] and random spatial sampling using global positioning system (GPS) coordinates . However, these approaches may not be appropriate in conflict settings. Compact segment sampling requires a significant amount of field time exposure and two visits to each cluster  while the use of GPS units is often a security risk in the context of modern warfare . Variations on the conventional two-stage cluster sampling designed for nation-wide mortality estimates in conflict settings are needed to generate accurate and useful mortality estimates and to contribute to theoretical and practical advances in the field of conflict epidemiology .
This paper presents a two-stage cluster sampling method implemented in a retrospective mortality study in Iraq. Our goal was to develop a cluster-based sampling method while taking into consideration the specific challenges of conflict settings. Our cluster sampling uses a gridded population dataset and a spatial sampling algorithm in a geographic information system (GIS) to select clusters in the first stage. Starting households are selected in the second stage using imagery and a sampling grid in Google Earth TM.