Technical Summary: Distinguishing True Spillover Trends from Surveillance Bias in Detected Cases

Problem Statement

The scientific consensus that zoonotic spillover rates are globally increasing has driven research priorities and policy responses for decades. However, this apparent rise may be a statistical artifact resulting from changing detection and reporting efforts rather than a genuine biological signal. The core challenge lies in distinguishing true epidemiological trends from surveillance bias. Existing datasets are often incomplete and derived from passive surveillance, where case identification depends on care-seeking behavior, diagnostic access, and reporting practices. While statistical modeling is frequently used to correct for these biases (e.g., controlling for covariates like GDP, population density, or publication counts), previous studies have not systematically validated whether these adjustments can recover true trends against a known ground truth. Consequently, claims of rising spillover may reflect methodological artifacts rather than biological reality.

Methodology

The authors employed a two-pronged approach: extensive simulation studies with known ground truths and an empirical application to real-world data.

Simulation Framework

The study generated synthetic datasets mimicking spillover surveillance under four distinct bias scenarios, crossed with three prevalence levels (high, medium, low) and nine combinations of true trend/detection trend directions (increasing, steady, decreasing):

1. Linear: Both true spillover and detection probability change linearly over time.
2. Oscillating: True spillover exhibits seasonal fluctuations on a linear trend, while detection changes linearly without seasonality.
3. Distance-based: True spillover changes linearly, but detection probability varies spatially based on distance to a testing site.
4. Shared-Drivers: True cases and detection probability are driven by the same underlying environmental and infrastructural factors (e.g., rainfall affecting both vector populations and surveillance funding).

Statistical Models

Three modeling approaches were evaluated for their ability to recover trend direction and magnitude:

1. Generalized Linear Mixed Models (GLMM): Negative binomial models with random intercepts, controlling for detection drivers via covariates.
2. Generalized Additive Mixed Models (GAMM): Similar to GLMMs but utilizing smooth functions (splines) to capture non-linear relationships between predictors and outcomes.
3. Double Machine Learning (DML): A causal inference approach using random forests to separate the estimation of the detection process from the outcome process, explicitly isolating the treatment effect (temporal trend) after adjusting for confounders.

Empirical Case Study

The methods were applied to 27 years (1993–2019) of bovine rabies surveillance data from South Africa. This dataset was chosen because bovine rabies represents spillover events (negligible cow-to-cow transmission) and is subject to known detection biases, such as distance to diagnostic laboratories and shared drivers (e.g., reservoir host prevalence influencing both risk and detection effort). A hierarchical Bayesian N-mixture model was also applied to this dataset as a fourth comparative method.

Key Results

Simulation Findings

• Scenario Dependency: The success of trend recovery was driven more by the structure of the bias scenario than by the choice of modeling framework. All three methods (GLMM, GAMM, DML) performed well (>50% correct classification) in simple linear and distance-based scenarios but failed (<50% correct) in oscillating and shared-driver scenarios.
• Systematic Positive Bias: When the true spillover trend was steady, all methods exhibited a strong directional bias toward false positives (incorrectly identifying an increasing trend). This was most pronounced in oscillating scenarios with changing detection, where misclassification rates for "increasing" trends reached 92–100% across methods.
• Shared-Driver Limitations: In shared-driver scenarios, methods struggled to distinguish true trends from detection artifacts. While they were better at identifying steady trends, they frequently misclassified true increasing/decreasing trends as steady (insensitivity) or, when they did identify a trend, showed low reliability (e.g., DML was correct only 55–71% of the time when identifying a directional trend).
• Magnitude vs. Direction: While methods could sometimes recover the correct direction of a trend, they frequently failed to capture the correct magnitude. Confidence intervals often failed to contain the true annual percent change, even when the direction was correct.
• Prevalence Sensitivity: Performance degraded significantly at low prevalence. Parametric models (GLMM/GAMM) suffered from widened confidence intervals (false negatives), while DML produced falsely precise null intervals due to regularization shrinking estimates toward zero.

Empirical Case Study (Bovine Rabies)

• Directional Agreement, Magnitude Disagreement: All four methods (GLMM, GAMM, DML, N-mixture) agreed on the direction of the trend (declining). However, the estimated magnitude of the decline varied drastically:
  • GLMM: -2.74% per year.
  • GAMM: -9.64% per year.
  • DML: -0.003 cases per district per year.
  • N-mixture: -2.5% per year.
• This divergence demonstrates that agreement on trend direction does not imply agreement on scale or the specific estimand being measured.

Key Contributions and Claims

The paper does not claim to prove that global spillover is not increasing; rather, it argues that current statistical methods cannot reliably distinguish a true increase from surveillance artifacts.

1. Methodological Limitation: Standard and state-of-the-art statistical adjustments (including causal inference methods like DML) cannot reliably recover true spillover trends from biased surveillance data, particularly under realistic conditions like oscillating detection or shared drivers.
2. The "False Positive" Risk: There is a systematic tendency for models to infer spurious increasing trends when the true trend is steady and detection effort is changing. This suggests that previous claims of rising spillover may be partially driven by these directional artifacts.
3. Reporting Guidelines: The authors propose a shift in how spillover trends are reported:
   • Prioritize Direction over Magnitude: Given the inability to accurately estimate magnitude, focus should be on establishing trend direction with appropriate uncertainty.
   • Disclose Reliability: Researchers must disclose the reliability of their methods under specific bias conditions.
   • Treat Disagreement as Signal: Cross-model disagreement (as seen in the rabies case study) should be treated as informative signal regarding uncertainty, not just noise.
   • Flag High-Risk Regimes: Analyses should explicitly flag when detection regimes are prone to generating spurious trends.

Significance

The significance of this work lies in its call for "constructive skepticism" regarding global spillover trends. The authors argue that methodological correction alone is insufficient; the gap between detection bias and true signal requires investment in surveillance infrastructure, particularly in low- and middle-income regions where infrastructure is weakest and spillover risk is hypothesized to be highest. Without improved data collection and metadata on surveillance effort (e.g., testing numbers, healthcare access), statistical models will continue to risk mislocalizing spillover risk and inflating apparent trends in improving surveillance systems while understating risk in contracting ones. The paper concludes that honest uncertainty is more valuable for public health decision-making than false precision derived from flawed data.