On the Effect of Ranger Patrols on Deterring Poaching: A Bayesian Approach for Causal Inference Using Field Tests as an Instrument

Abstract

Wildlife conservation relies on ranger patrols to detect and remove animal traps set out by poachers. While these efforts are important, conservation biologists seek to understand the impact of ranger patrols on future poaching activity, beyond the immediate impact of removing snares. Quantifying the effect of patrolling, such as understanding the effect of patrols on deterring poaching in the future, can inform governmental policy and increase funding for conservation efforts.

However, studies to date have not moved beyond association to study the causal effect of patrolling on deterrence. Leveraging a previous field study in Uganda that encouraged rangers to patrol randomly selected sites as an instrument, we assess the causal effect of patrolling on deterrence.

Yet, the imperfect detection of snares--failure to detect a snare when one is present--poses a major barrier to obtaining a causal estimate for our scientific question of interest. Using ground-truth labels on the detection of snares would lead to estimates that reflect the effect of patrols on the detection of poaching, rather than the true causal effect of patrols on poaching levels.

To model the probability of the true presence of a poaching event, we present a novel Bayesian approach that uses learned neural network predictions of detection probability to model imperfect detection, conditioning on patrol bias which may result in false negative labels. Using these adjusted illegal activity probabilities as our outcome variable for causal inference allows us to answer the open question of the effect of patrolling on poaching deterrence. To improve the quality of the detection predictions used for modeling the presence of poaching, we engineer and evaluate a system that automatically extracts diverse geospatial features from publicly available remote sensing data for any national park in the world. Beyond this causal study, these enriched feature sets benefit under-resourced parks that may not have the means to obtain or compute their own geospatial features for risk predictions.

Through our Bayesian approach to causal inference, we detect a significant deterrence effect. Moving beyond a causal estimate, we present new knowledge gleaned from analyzing the learned nonlinear causal relationships of causal effects of varying patrol effort across sites. For the first time, we demonstrate the causality of patrols on deterrence. The insight to leverage historical field studies as an instrument for causal inference, especially when they were not originally conducted for the purposes of causal inference, presents a contribution to the statistics community. Our novel methodology for modeling imperfect detection in negative labels using machine learning outputs under a Bayesian model to enable causal inference for questions that cannot be answered otherwise presents a major contribution to the computer science community.