[Paper Review] Spatial Multiresolution Cluster Detection Method

Lingsong Zhang, Zhengyuan Zhu|arXiv (Cornell University)|May 9, 2012

Data-Driven Disease Surveillance17 references3 citations

TL;DR

This paper proposes a spatial multiresolution cluster detection (MCD) method that identifies irregularly shaped clusters in spatial data using multi-scale likelihood ratio tests for Bernoulli, Poisson, and Normal distributions. By leveraging a neighborhood variability measure to select optimal test thresholds, the MCD method improves detection power for non-circular clusters while maintaining computational efficiency, outperforming both spatial scan statistics and FDR-controlled single-scale testing in simulations and fMRI data analysis.

ABSTRACT

A novel multi-resolution cluster detection (MCD) method is proposed to identify irregularly shaped clusters in space. Multi-scale test statistic on a single cell is derived based on likelihood ratio statistic for Bernoulli sequence, Poisson sequence and Normal sequence. A neighborhood variability measure is defined to select the optimal test threshold. The MCD method is compared with single scale testing methods controlling for false discovery rate and the spatial scan statistics using simulation and f-MRI data. The MCD method is shown to be more effective for discovering irregularly shaped clusters, and the implementation of this method does not require heavy computation, making it suitable for cluster detection for large spatial data.

Motivation & Objective

• To address the limitation of traditional spatial scan statistics in detecting irregularly shaped clusters due to fixed window shapes like circles or ellipses.
• To overcome the lack of spatial continuity in single-location multiple testing methods that control for false discovery rate (FDR).
• To develop a computationally efficient method that leverages spatial dependence to enhance detection power for complex cluster shapes.
• To introduce a global variability measure that balances sensitivity and specificity without relying on multiple comparison corrections.
• To provide a flexible, scale-space-based framework applicable to various exponential family distributions (Bernoulli, Poisson, Normal).

Proposed method

• Derives a multi-scale test statistic based on likelihood ratio statistics for three distributions: Bernoulli, Poisson, and Normal, applied across varying spatial window sizes.
• Uses a 5-point neighborhood (center and four adjacent cells) to compute local test statistics and assess spatial heterogeneity.
• Defines a local variability measure as the sum of squared deviations from the local mean across a 5-cell neighborhood to detect boundary effects.
• Applies a global variability comparison between boundary and non-boundary points to select an optimal test threshold, reducing false positives.
• Employs asymptotic distribution theory to derive the expected value and variance of the test statistic under different cluster configurations (noise, boundary, signal regions).
• Uses central limit theorem approximations to show that the average test statistic is stochastically separable across noise, boundary, and signal regions as sample size increases.

Experimental results

Research questions

• RQ1 Can a multi-resolution approach improve detection power for irregularly shaped spatial clusters compared to fixed-shape scan statistics?
• RQ2 Does incorporating spatial dependence via neighborhood variability enhance the detection of non-circular clusters while maintaining computational efficiency?
• RQ3 Can a global variability measure effectively replace multiple comparison corrections in spatial cluster detection without sacrificing specificity?
• RQ4 How does the MCD method perform relative to FDR-controlled single-location testing in identifying spatially contiguous, irregular clusters?
• RQ5 To what extent does the method maintain statistical power when the true cluster shape deviates from circular or elliptical forms?

Key findings

• The MCD method demonstrates superior performance in detecting irregularly shaped clusters compared to spatial scan statistics, especially when the true cluster shape is non-elliptical.
• The method maintains high statistical power even when cluster boundaries are complex, due to its multi-scale and spatially adaptive test statistics.
• Simulation results show that the average test statistic is stochastically separable across noise, boundary, and signal regions: $ Eig( ext{Ave}_{ ext{noise}} Tig) < Eig( ext{Ave}_{ ext{boundary}} Tig) < Eig( ext{Ave}_{ ext{signal}} Tig) $, with convergence in probability as sample size increases.
• The variability-based threshold selection effectively distinguishes signal from noise, with $ Pig( ext{Ave}_{ ext{boundary}} ilde{V} - ext{Ave}_{ ext{non-boundary}} ilde{V} > 0ig) o 1 $ when $ rac{ u^2 N p_B (1-p_B)}{8 + 4(1-p_B)} o u^2 $, ensuring high detection power.
• The method is computationally efficient and scalable to large spatial datasets, as it avoids exhaustive scanning over all possible window shapes and sizes.
• In fMRI data analysis, the MCD method successfully identified biologically plausible activation clusters that were missed by both spatial scan statistics and FDR-based methods.