Distinguishing True Spillover Trends from Surveillance Bias in Detected Cases

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Claims that zoonotic spillover is rising globally have shaped research priorities for decades, but this apparent rise can be a statistical artifact of changing detection effort, not a genuine biological signal. Methodological correction alone cannot resolve the ambiguity: closing the gap requires investment in surveillance infrastructure, not modeling. Testing three approaches—generalized linear models, generalized additive models, and double machine learning—against simulated data spanning detection-bias and prevalence conditions, we find trend direction is recoverable under simple bias but not under oscillating or shared-driver detection, and all methods risk false increases when true rates are stable. Applied to 27 years of bovine rabies surveillance in South Africa, independent methods agreed on the declining trend direction but produced magnitude estimates differing more than threefold, showing model disagreement is informative. We translate these findings into reporting guidelines: prioritize direction over magnitude, disclose method reliability, treat cross-model disagreement as signal, and flag high-risk detection regimes.

This publication has 38 references indexed in Scilit: