Modelling spatially autocorrelated detection probabilities in spatial capture-recapture using random effects

TL;DR

This paper evaluates Bayesian spatial capture-recapture models incorporating random effects to account for spatial heterogeneity in detection probabilities, demonstrating that spatially autocorrelated random effects yield less biased population estimates especially under high autocorrelation.

Contribution

It introduces and compares three Bayesian SCR models with different random effects structures, highlighting the effectiveness of spatially autocorrelated random effects in reducing bias.

Findings

SARE provided the least biased population size estimates.

SARE best predicted spatial heterogeneity when autocorrelation was high.

FM outperformed SARE and RE at intermediate autocorrelation levels.

Abstract

Spatial capture-recapture (SCR) models are now widely used for estimating density from repeated individual spatial encounters. SCR accounts for the inherent spatial autocorrelation in individual detections by modelling detection probabilities as a function of distance between the detectors and individual activity centres. However, additional spatial heterogeneity in detection probability may still creep in due to environmental or sampling characteristics. if unaccounted for, such variation can lead to pronounced bias in population size estimates. Using simulations, we describe and test three Bayesian SCR models that use generalized linear mixed models (GLMM) to account for latent heterogeneity in baseline detection probability across detectors using: independent random effects (RE), spatially autocorrelated random effects (SARE), and a two-group finite mixture model (FM). Overall, SARE provided the least biased population size estimates (median RB: -9 -- 6%). When spatial autocorrelation was high, SARE also performed best at predicting the spatial pattern of heterogeneity in detection probability. At intermediate levels of autocorrelation, spatially-explicit estimates of detection probability obtained with FM where more accurate than those generated by SARE and RE. In cases where the number of detections per detector is realistically low (at most 1), all GLMMs considered here may require dimension reduction of the random effects by pooling baseline detection probability parameters across neighboring detectors ("aggregation") to avoid over-parameterization. The added complexity and computational overhead associated with SCR-GLMMs may only be justified in extreme cases of spatial heterogeneity. However, even in less extreme cases, detecting and estimating spatially heterogeneous detection probability may assist in planning or adjusting monitoring schemes.