Improving the spatial measurement of conflict exposure for population health research: Comparing kernel density estimation against conventional approaches in Mali

Background
Precise measurement of conflict exposure is required to understand its health impacts. Common approaches to operationalizing conflicts, such as counting events within administrative boundaries or applying a fixed radius around events to capture their spatial influence, can obscure important spatial and temporal patterns of conflict. Using Mali as a case study, we constructed a conflict intensity measure based on kernel density estimation (KDE) and compared it with standard approaches by examining population exposure and associations with key reproductive, maternal, and child health (RMCH) outcomes.

Methods
We compiled georeferenced conflict events from global datasets. Using KDE, we produced gridded annual conflict intensity surfaces (10x10 km) from 2000 to 2024 and aggregated them to district and health-facility levels. We also measured conflict exposure as the number of events per district and within a 50-km radius of each health facility. For selected years (2013, 2018, and 2023), we compared the proportion of population subgroups exposed to conflict across approaches, using percent agreement and Cohen’s Kappa. We also examined associations between conflict measures and key RMCH indicators, using correlation coefficients and key model performance metrics (coefficient of determination (R²) and root mean square error (RMSE)).

Results
In 2013 and 2018, KDE and standard approaches yielded similar estimates of the proportion of the population exposed to conflict, around 30%. In 2023, the measures diverged substantially: standard approaches classified 88% of the population as exposed to conflict, while KDE classified 38%. Agreement in classifying conflict-affected population groups was on average moderate, with kappa values ranging from 0.37 in 2023 to 0.75 in 2018. KDE revealed a log-linear dose-response relationship between conflict intensity and selected RMCH indicators in 2013 and 2018, with correlation coefficients ranging from 0.31 to 0.75. Standard approaches showed weaker relationships, with correlations ranging from 0.03 to 0.64. Across selected years, KDE yielded higher R² values and lower RMSE than the standard approaches.

Conclusions
KDE outperformed standard methods, enabling the unmasking of health impacts at district and facility levels. Its potential at fine spatial scales remains to be explored.