Fig 1.
Illustration of the workflow in spatial event aggregation designs.
Table 1.
Thirteen influential SEA designs published largely in the last decade.
Fig 2.
Illustration of the consequences of varying cell sizes.
On the left, smaller cells capture local variation in causes of violence. On the right, larger cells overaggregate events. Also note the difference is the numbers of cells available for the regression analysis.
Fig 3.
Illustration of areal confounding.
On the left, example data is depicted featuring uniform spatial distributions for events and population (p) figures. As artificial cells of varying areas (a) are superimposed, higher event counts correlate with higher population figures (top right). When smaller border cells are excluded (or area is controlled for), this correlation disappears (bottom right).
Fig 4.
At the smallest level, the SCEG simulator represents a lattice of simulated locations orders of magnitude smaller than the areal units used for statistical analysis. In one time-step of the simulation, all simulated locations inside the country potentially generate a conflict event associated with their location.
Fig 5.
Diffusion and error correlation in SCEG.
On the left, the diffusion mechanism is illustrated: locally caused conflict events will experience “outcome diffusion” and relocate to a random simulated location within a predefined radius. On the right, the presence of a military base indicated by the triangle in the lower right corner causes more conflict events in its vicinity.
Table 2.
Overview of the eight event-generating processes run for all countries.
In each experiment, 100 datasets were generated.
Fig 6.
Main results for the basic scenario.
Experiments 1a and 1b: Simple event-generating process analyzed across 100 datasets and different units without spatial diffusion or error correlation. Numbers of false significant results for the population estimate are conveyed visually. False negative (FN) results come about when the analysis fails to produce positive and significant results in experiment 1a (true population effect). False positive (FP) results refer to positive and significant findings in experiment 1b (no population effect).
Fig 7.
Simple event-generating process estimated with ‘area’ controls.
Fig 8.
Experiments 2a and 2b (diffusion scenario, with and without causal effect) analyzed with a simple linear model as well as a spatial lag model.
All models include area controls.
Fig 9.
Experiments 3a and 3b (latent predictor, with and without causal effect) analyzed with spatial error and linear models.
All models include area controls.
Fig 10.
Experiments 4a and 4b (all problems combined, with and without causal effect) analyzed at two different levels of aggregation for both PG and ADM units.
All models include area controls. Results are only accepted if they align at both levels of spatial aggregation.