A Monte Carlo analysis of false inference in spatial conflict event studies

Abstract

Spatial event data is heavily used in contemporary research on political violence. Such data are oftentimes mapped onto grid-cells or administrative regions to draw inference about the determinants of conflict intensity. This setup can identify geographic determinants of violence, but is also prone to methodological issues. Problems resulting from spatial aggregation and dependence have been raised in methodological studies, but are rarely accounted for in applied research. As a consequence, we know little about the empirical relevance of these general problems and the trustworthiness of a popular research design. We address these questions by simulating conflict events based on spatial covariates from seven high-profile conflicts. We find that standard designs fail to deliver reliable inference even under ideal conditions at alarming rates. We also test a set of statistical remedies which strongly improve the results: Controlling for the geographic area of spatial units eliminates an important source of spurious correlation. In time-series analyses, the same result can be achieved with unit-level fixed effects. Under outcome diffusion, spatial lag models with area controls produce most reliable inference. When those are computationally intractable, geographically larger aggregations lead to similar improvements. Generally, all analyses should be performed at two separate levels of geographic aggregation. To facilitate future research into geographic methods, we release the Simple Conflict Event Generator (SCEG) developed for this analysis.

Perceived Racial Threats Increase Demand for Conservative Media: Evidence from Black Lives Matter Protests and Fox News Ratings

“…Namely, BLM protests do not appear to have increased CNN viewership, and perhaps even decreased viewership. In sum, the results of the census-tract analysis provide us with confidence that our CBG-level findings are not biased due to the effects of aggregating social phenomena into specific spatial units (Schutte and Kelling, 2022).…”

“…We conduct our primary analysis with the CBG panel and a secondary analysis with the tract panel. Using both geographies helps guard against potential biases resulting from aggregating social phenomena into spatial units (Schutte and Kelling, 2022). Note, however, that checking the robustness of results with larger geographies -in our case, tracts relative to CBGs -poses a harder test of statistical relationships because there are fewer observations.…”