Modeling the probability distribution of positional errors incurred by residential address geocoding.

Dale L. Zimmerman, Xiangming Fang, Soumya Mazumdar, Alexander M. RushInternational Journal of Health Geographics 2007 · DOI

Abstract

Mixtures of bivariate t distributions with few components appear to be flexible enough to fit many positional error datasets associated with geocoding, yet parsimonious enough to be feasible for nascent applications of measurement-error methodology to spatial epidemiology.

Introduction

Geocoding addresses into geographic coordinates is a fundamental step in spatial epidemiology, enabling researchers to link health outcomes to environmental exposures. However, the process of converting a street address into latitude and longitude introduces positional errors. If these errors are systematic or large, they can lead to spurious findings, such as false disease clusters or attenuated exposure estimates. To mitigate these effects, researchers require robust mathematical models that accurately describe the probability distribution of these positional offsets.

Methodology: Finite Mixture Models

Traditional models, such as a single bivariate normal distribution, often fail to capture the complexity and heavy-tailed nature of real-world geocoding errors. This study proposes the use of Finite Mixture Models based on bivariate t-distributions.

Unlike the normal distribution, the t-distribution possesses "heavy tails," allowing it to account for significant outliers without compromising the fit of the central data. By combining two or three of these distributions, the model can represent "multi-axial clustering"—the tendency for errors to align along specific geographic axes. The authors utilized the Expectation-Maximization (EM) algorithm to estimate parameters and the Bayesian Information Criterion (BIC) to ensure model parsimony.

The evolution of model fit from a single component to a multi-component mixture, capturing non-isotropic error patterns.

The "Greek-Cross" Phenomenon

A key finding of the study is the "Greek-cross" pattern observed in error plots. In rural areas like Carroll County, Iowa, where the road network follows a strict North-South/East-West grid, positional errors are not randomly scattered. Instead, they cluster along these primary axes due to the nature of street-segment interpolation. Errors in interpolation or street-segment data naturally "stretch" the error distribution along these axes, creating a cross-like shape rather than a circular cloud.

Empirical error distributions exhibiting the characteristic 'Greek-cross' pattern aligned with the street grid.

Comparative Performance and Systematic Bias

The study compared automated geocoding (using ArcGIS and TIGER files) against E911 geocoding (used for emergency services).

  • Accuracy Gap: E911 geocoding proved significantly more accurate, with a median error of 44 meters compared to 168 meters for automated geocoding. E911 errors tended to be aligned orthogonally to the street orientation, likely due to how house offsets are calculated in emergency databases.
  • Systematic Bias: Automated geocoding exhibited a consistent shift toward the East and South, suggesting underlying flaws in the TIGER street databases or interpolation algorithms.
  • The Match Rate Trap: While "100% matches" were relatively accurate, "60% matches" in automated geocoding introduced massive errors, with some outliers exceeding 15 kilometers.

Conclusion

The research demonstrates that a mixture of just three bivariate t-distributions is sufficient to model the complex, heavy-tailed, and axially-aligned errors inherent in geocoding. These parsimonious models provide a foundation for measurement-error frameworks, allowing spatial epidemiologists to statistically "de-noise" their data and improve the reliability of public health mapping and analysis.

Summary updated

Abstract

Mixtures of bivariate t distributions with few components appear to be flexible enough to fit many positional error datasets associated with geocoding, yet parsimonious enough to be feasible for nascent applications of measurement-error methodology to spatial epidemiology.
BibTeX
@article{zimmerman2007modeling,
  author = {Dale L. Zimmerman and Xiangming Fang and Soumya Mazumdar and Alexander M. Rush},
  title = {{Modeling the probability distribution of positional errors incurred by residential address geocoding.}},
  journal = {International Journal of Health Geographics},
  year = {2007},
  doi = {10.1186/1476-072x-6-1},
}
Figure