Comparison of tests for spatial heterogeneity on data with global clustering patterns and outliers

Monica Jackson, Lan Huang, Jun Luo, Mark Hachey, Eric J. FeuerInternational Journal of Health Geographics 2009 · DOI

TL;DR

This paper compares statistical methods for detecting spatial patterns in health data, specifically focusing on global clustering and local outliers. They found that SaTScan excels at identifying outliers, while Tango's MEET is superior for detecting global clustering patterns, offering a more reliable approach for surveillance and analysis.

What Problem Does This Paper Address?

Understanding how diseases are distributed geographically is crucial for public health surveillance and research. This involves identifying areas where disease rates are unusually high or low (clustering) and pinpointing specific locations that deviate significantly from their neighbors (outliers). While many statistical methods exist for these tasks, their performance can vary greatly depending on the type of spatial pattern present. This paper aims to evaluate and compare several popular methods under two extreme scenarios: strong global clustering across a region and the presence of isolated local outliers. The authors highlight a gap in existing research, which hasn't thoroughly investigated how these methods perform on these specific, challenging data configurations.

Methods

The researchers compared several statistical methods:

  • Global Clustering Evaluation:
    • Tango's MEET (Maximised Excess Events Test): This method is designed to detect overall spatial clustering. They used two versions: Tango_PDM (with a population density-adjusted weight function) and Tango_ADJ (with an adjacent neighbor weight function).
  • Moran's I: A classic measure of spatial autocorrelation, used here as a benchmark. They used the adjacent neighbor weight function (Moran's I_ADJ).
  • Oden's I*pop: An adaptation of Moran's I that accounts for population density. They used both population density (I*pop_PD) and adjacent neighbor (I*pop_ADJ) weight functions.
  • Local Cluster Detection (Outlier Detection):
    • SaTScan (elliptic version): A widely used software for cluster detection, capable of identifying both large clusters and local outliers. They used the elliptic version for its better performance.
  • Local Moran's I: A localized version of Moran's I that assesses spatial association for individual geographic units.

The study involved simulating count data for U.S. counties under different scenarios: global clustering (with linear and exponential trends in relative risk) and local outliers (counties with a significantly higher relative risk). They then calculated the statistical power (ability to correctly reject the null hypothesis of no clustering) and, for cluster detection methods, the chance of a county being identified as part of a cluster.

Datasets

The primary "datasets" used in this study were simulated data. The simulations were based on the geographic configuration of 3109 counties in the continental United States. They used a fixed population size per county (5,000) for simulations to control for population heterogeneity, which they acknowledge as a limitation.

For real-world examples, they used:

  • HIV mortality data: From 1987-2004, extracted from the Surveillance, Epidemiology, and End Results (SEER) Program of the National Cancer Institute.
  • Lung cancer mortality data: For white males from 1950-1969, obtained from the National Cancer Institute's atlas.

Key Insights and Ideas

  • SaTScan is a champion for outliers: The simulations clearly showed SaTScan's elliptic version significantly outperforms other methods in detecting local outliers, even when the outlier is surrounded by many neighbors. This is attributed to its design for local detection and the use of small search windows.
  • Tango's MEET reigns supreme for global patterns: For identifying broad, region-wide clustering, Tango's MEET (especially with the population density weight function) demonstrated the highest statistical power. This suggests it's a more reliable choice for understanding overall spatial trends.
  • Moran's I struggles with extremes: Traditional Moran's I showed low power in both outlier detection and global clustering scenarios, especially when compared to the specialized methods.
  • Weight functions matter: The choice of weight function (e.g., adjacent neighbors vs. population density) significantly impacts the performance of global clustering methods. Population density-adjusted weights generally improved power for detecting global patterns.
  • SaTScan can be misleading for global patterns: While SaTScan can detect clusters, its application for global clustering evaluation should be approached with caution. It might identify statistically significant clusters but potentially misrepresent the true global spatial pattern, especially if the "clusters" are not truly global in scope.

Key Results

  • Outlier Detection: SaTScan-E achieved powers above 0.95 for outlier detection, significantly higher than global clustering methods (power < 0.2). Local Moran's I showed some ability to detect outliers but with lower precision and chance of detection compared to SaTScan.
  • Global Clustering:
    • Tango_PDM showed very high power (close to 1) for global clustering patterns, especially with exponential trends.
  • I*pop_PD also performed well (power > 0.8) with a large population density weight parameter (λ=50%).
  • Methods using only adjacent neighbor weights (Tango_ADJ, Moran's I, I*pop_ADJ) had low power (0.1-0.2) for global clustering.
  • SaTScan-E had moderate power (around 0.75) for global clustering when using a large search window (50% of population).
  • Real Data: Tango_PDM confirmed global clustering in HIV mortality data. SaTScan identified significant clusters of both high and low HIV mortality rates, consistent with a global pattern. For lung cancer mortality, SaTScan detected both global clustering patterns and local outliers.

Limitations and Future Work

  • Homogeneous Population Assumption: The simulation study assumed homogeneous county populations. Real-world data often have heterogeneous populations, which can complicate statistical power analysis. The authors acknowledge this limitation but explain it was necessary to isolate the performance of the methods themselves.
  • Weight Function Selection: The choice of spatial weight function is crucial and can be subjective. The paper explored adjacent neighbor and population density weights but suggests further research into defining appropriate spatial weight functions.
  • Alternative Patterns: The study focused on global clustering and outliers. Future research could explore other spatial autocorrelation patterns (e.g., different cluster sizes, shapes).
  • Other Methods: The authors suggest evaluating other outlier detection methods (like MGPS) and comparing them with SaTScan and Local Moran's I.
  • Heterogeneous Population Data: Future work could also involve comparing method performance on real-world data with naturally heterogeneous populations.

Summary updated

Abstract

SaTScan elliptic version is more efficient for outlier detection compared with the other methods evaluated in this article. Tango's MEET and Oden's I*(pop) perform best in global clustering scenarios among the selected methods. The use of SaTScan for data with global clustering patterns should be used with caution since SatScan may reveal an incorrect spatial pattern even though it has enough power to reject a null hypothesis of homogeneous relative risk. Tango's method should be used for global clustering evaluation instead of SaTScan.
BibTeX
@article{jackson2009comparison,
  author = {Monica Jackson and Lan Huang and Jun Luo and Mark Hachey and Eric J. Feuer},
  title = {{Comparison of tests for spatial heterogeneity on data with global clustering patterns and outliers}},
  journal = {International Journal of Health Geographics},
  year = {2009},
  doi = {10.1186/1476-072x-8-55},
}
Figure