Population mortality before and during armed conflict in Yemen: geospatial and statistical analysis of cemetery data

Abstract

Since 2014, Yemen is affected by crisis conditions due to armed conflict. Evidence on the impact of this large-scale crisis on mortality is lacking. We analysed archive very high-resolution satellite imagery from a sample of Yemeni subdistricts to quantify changes in burial incidence attributable to the crisis. We identified possible cemeteries through remote and ground sources in 24 sampled subdistricts. After initial triage and extensive steps to improve the interpretability of archive imagery spanning the period 2011 to 2021, a pool of crowd workers, supervised by expert analysts and aided by an automated algorithm, annotated surface area and grave counts in sequential images from a set of analysis-eligible cemeteries. We complemented these longitudinal observations with data on different predictors including three crisis proxies (incidence of insecurity events, price of staple cereal, internal displacement), and fitted statistical models to compare predicted burials under observed and assumed counterfactual (no crisis) conditions. We identified 561 potential cemeteries within 24 sampled subdistricts, but excluded most due to inability to geolocate them or see the cemetery and/or graves in available imagery, yielding an effective sample of 110 image observations across 35 cemeteries in 10 subdistricts. Burial rate generally decreased between 2014-2018 and rose sharply thereafter. Alternative regression models suggested that most cemeteries would have experienced lower burial rate under non-crisis conditions, with a crisis to non-crisis ratio of about two overall. The incidence of insecurity events appeared positively associated with burial rate. This unprecedented-scale geospatial analysis of cemeteries suggests an increase in burial rates attributable to crisis conditions in a non-representative, disproportionately urban sample of Yemen. The study identifies key challenges of such an analysis. We discuss possible methodological ways forward to further explore the feasibility and validity of this option for mortality estimation in settings with insufficient vital events registration and limited ground access.