Rangers serve as a primary defence for wildlife conservation, patrolling 21.7 million sq. km of protected areas globally to protect from poaching and other threats. The high cost of patrols is justified assuming that rangers deter poachers from returning in the future.
However, patrol-induced deterrence has not been established causally, nor its effect size quantified, due to complex dynamics of ecosystem—human interactions and the difficulty of detecting poaching events.

Using 12 years of field-collected ranger data from Murchison Falls Conservation Area in Uganda, we present the first causal evidence of ranger-patrol deterrence. We leveraged a past field evaluation of a machine learning tool as a natural experiment, then used machine learning to produce temporally and spatially granular estimates of site properties for matching and Bayesian inference to impute unobserved outcome labels. We found that increasing patrol effort by 1 kilometre in a 1 sq. km area reduced the poaching probability in the next month by 45.0%. Our results also enable downstream analyses to study patrol efficacy: we estimated that deterrence is most effective in areas with greater accessibility and that optimizing patrol allocation can increase deterrence by up to 46.6%, creating possible cost savings of USD$300,685 annually in one park alone.