Comparing analytical protocols for identifying causes of population changes

Abstract

Conservation decision-making requires accurate identification of causes of population changes. Ecologists often rely on analytical protocols that aggregate high-dimensional monitoring data. We hypothesise that compressing data - either spatially, as in conventional time series (TS) analysis, or temporally, as in static species distribution models (SDMs) - destroys covariance structures and obscures the identification of causal drivers. To quantify this aggregation cost, we conducted a rigorous simulation experiment using virtual species to establish a known ground truth of population drivers. We then employed a virtual ecologist approach to mimic a 20-year large-scale bird monitoring scheme, and generate realistic spatiotemporal datasets to evaluate the analytical pipelines. We benchmarked the causal attribution accuracy of aggregated TS and SDM protocols against a full-resolution spatiotemporal (FRST) framework, which retains native data dimensions and integrates mechanistic spatiotemporal covariance structures. Our simulations revealed that spatial compression severely compromises causal inference: unpenalised TS models failed to detect any true underlying drivers (accuracy = 0.50, sensitivity = 0.00). Temporal compression (SDMs) performed moderately better (accuracy = 0.68), while the FRST model achieved superior accuracy (0.88), sensitivity (0.84), and specificity (0.93). Furthermore, we identified a variable selection paradox: double penalty shrinkage marginally improved underpowered TS models, although it degraded the specificity of SDM and FRST frameworks by forcing spurious, correlated variables to absorb residual variance. Our findings demonstrate that protocols that involve data aggregation reduce the informational value of large-scale monitoring datasets. Full-resolution, mechanistically informed frameworks are essential for reliable causal attribution and robust biodiversity monitoring. HighlightsO_LISpatiotemporal data aggregation obscures causal drivers in biodiversity monitoring. C_LIO_LISpatial compression in time series models fails to detect true population drivers. C_LIO_LIFull-resolution spatiotemporal models accurately identify true drivers. C_LIO_LIAutomated variable selection introduces false positives in high-resolution models. C_LI