Abstract
Spatial scan statistics are essential for identifying geographical clusters of events. While classic versions rely on Bernoulli or Poisson models, this paper introduces a Hypergeometric probability model to construct the likelihood function. Applied to Japanese Encephalitis (JE) data in Sichuan, China, the Hypergeometric (H) model identifies clusters identical to those found by Kulldorff’s method. However, simulation benchmarks reveal that the H model outperforms traditional statistics in scenarios involving high population density or large cluster sizes.
Introduction
In epidemiology, identifying "hotspots" or clusters of a disease is critical for resource allocation and understanding disease etiology. The most widely used tool for this is the spatial scan statistic. Traditionally, these statistics rely on Bernoulli or Poisson models to calculate the likelihood that a specific geographic window contains a cluster. However, these models often serve as approximations. For binary outcomes—where an individual is either a "case" or a "control"—the Hypergeometric distribution, which models sampling without replacement from a finite population, may provide a more precise fit.
Methods
The researchers proposed a shift from maximizing a likelihood ratio to minimizing a probability under the null hypothesis of complete spatial randomness.
- The H Model: Instead of the standard Poisson likelihood ratio, the Hypergeometric distribution calculates the probability of observing a specific number of cases within a window, given the total population and total cases in the entire study region.
- The "Most Strange Window": The test statistic identifies the window yielding the minimum probability, termed the "most strange window" as it is the least likely to occur by chance.
- Significance Testing: Monte Carlo simulations (99,999 random replications) determine the p-value to ensure detected clusters are statistically significant.
- Evaluation Metrics: Models were compared using Sensitivity (SEN)—the probability of detecting the true cluster—and Positive Predictive Value (PPV)—the proportion of the detected cluster that actually belongs to the true cluster.
Results and Application
The study utilized two primary datasets: 2009 Japanese Encephalitis (JE) cases in Sichuan, China, and a Northeastern US benchmark for simulation.
Real-World Application in Sichuan
When applied to the Sichuan JE data (596 cases across 181 counties), both the H model and the Poisson model identified identical primary and secondary clusters. The primary cluster was located in the southeast, while the secondary cluster appeared in the northeast.
Simulation Performance
Benchmarks indicated a performance trade-off between the models. The H model excels in urban areas and for large-scale clusters, whereas the Poisson model remains superior for rural areas or very small clusters.
As shown in the simulation results, the H model consistently showed higher SEN and PPV in urban settings. For instance, in one urban simulation, the H model achieved a SEN of 86.40% compared to the Poisson's 84.00%. The H model's advantage became more pronounced as the size of the "hot-spot" cluster increased.
Discussion
The Hypergeometric model is fundamentally about sampling without replacement, which is theoretically more accurate for fixed populations in geographic regions than the Poisson model. While not a replacement for Kulldorff’s method, the H model serves as a specialized tool. It is particularly powerful for monitoring infectious diseases in high-density urban environments or when large-scale outbreaks are suspected.
Future work may extend the H model to accommodate irregularly shaped scanning windows and account for spatial autocorrelation, where neighboring areas naturally exhibit similar characteristics.
Summary updated