Evaluating linkage approaches for address-level socioenvironmental exposure assessment

This study demonstrates that while address tag fuzzy matching achieves perfect accuracy for linking National Address Database records to parcel data, geocoding-based methods perform significantly worse—particularly in densely populated and deprived neighborhoods—highlighting the critical need for precise, standardized linkage approaches to prevent exposure misclassification in health research.

The Gist

Imagine you are a detective trying to solve a mystery about how the neighborhood a person lives in affects their health. You have a list of people's home addresses, and you want to attach specific details about their house to their file: Is it a single-family home or a big apartment? Is it worth $200,000 or $2 million? Is the roof leaking?

To do this, you need to find the exact "fingerprint" (the parcel ID) of that specific house in a giant digital database. The problem is, addresses are messy. People write them differently ("123 Main St" vs. "123 Main Street"), and sometimes the house numbers are tricky.

This paper is a report card on four different detective tools used to match a person's address to their specific house record. The researchers tested these tools on over 850,000 addresses in Ohio to see which one gets the job done right without making mistakes.

Here is the breakdown of the tools and what they found, using some everyday analogies:

The Four Detective Tools

1. The "Street Range" Guess (The Old Map):

   • How it works: This tool looks at a street and guesses where a house is based on numbers. If a street goes from 100 to 200, and you are looking for 150, it assumes you are exactly halfway down the block.
   • The Problem: It's like trying to find a specific apartment in a massive skyscraper by just guessing which floor you're on. It often puts you in the wrong building entirely.
   • The Result: Terrible. It got the right house less than 10% of the time in some areas. It's too vague for detailed health research.

2. The "Street Range" with a Better Map (The Improved Guess):

   • How it works: Same as above, but using a slightly smarter algorithm.
   • The Result: Still poor. It did a little better (up to 59% accuracy), but it still confused single-family homes with large apartment complexes way too often.

3. The "Address Point" GPS (The Drone Drop):

   • How it works: This tool uses a specific GPS coordinate for the front door of the house (like a drone dropping a pin exactly on the driveway). It then checks which property boundary that pin lands in.
   • The Result: Okay, but risky. It got the right house about 65% to 76% of the time. It's better than guessing, but if the GPS is off by just a few feet (like a 20-meter shift), it might drop the pin in your neighbor's yard or a different apartment building.

4. The "Fuzzy Text" Match (The Smart Librarian):

   • How it works: Instead of using GPS coordinates, this tool reads the address like a human. It breaks the address down into parts (Street Number, Street Name, Zip Code) and compares them to the database. It's "fuzzy" because it can handle typos or slight differences (like "St." vs "Street").
   • The Result: Perfect. It got the right house 100% of the time. It didn't rely on GPS coordinates that could be slightly off; it relied on the actual text of the address.

The Big Discovery: The "Crowded Neighborhood" Problem

The researchers found something very important about where these tools fail.

• The Analogy: Imagine trying to find a specific house in a quiet, empty cul-de-sac versus a dense city block with 50 houses on one street.
• The Finding: The GPS tools (Tools 1, 2, and 3) struggled the most in dense, poorer neighborhoods. In these areas, houses are packed close together, and there are many apartments. A tiny GPS error could easily send the data to the wrong unit.
• Why it matters: This creates an unfair bias. If the data is wrong for poor neighborhoods but right for rich, quiet neighborhoods, health studies might accidentally blame the wrong things for health problems in disadvantaged communities. The "Smart Librarian" (Tool 4) didn't have this problem; it worked perfectly everywhere.

The Real-World Test: Hospital Records

The team also tested these tools using real patient addresses from a children's hospital.

• When they used the "Smart Librarian" method, they could link patients to their housing data perfectly.
• When they used the GPS methods, they missed the mark significantly, especially for people living in large apartment buildings where one address might belong to many different units.

The Takeaway

If you want to study how housing affects health (like lead poisoning or asthma), don't just rely on GPS coordinates. They are too imprecise for the tiny details of property lines.

Instead, use text-based matching (the "Smart Librarian" approach). It's like reading the address label rather than guessing where the house is on a map. It is faster, more accurate, and ensures that the data for people in crowded, disadvantaged neighborhoods is just as reliable as data for people in quiet suburbs.

In short: To get the right health answers, we need to stop guessing where the house is and start reading the address label carefully.

Technical Summary

1. Problem Statement

Accurate linkage of individual addresses to parcel-level data (e.g., property value, land use, building condition) is critical for hyperlocal environmental exposure assessment in public health and clinical research. However, current methods suffer from significant limitations:

• Positional Uncertainty: Traditional geocoding methods, particularly street-range geocoding (interpolating coordinates along a street centerline), lack the precision required for parcel-level analysis. A small displacement (e.g., 20 meters) can misclassify a single-family home as a multi-unit apartment, leading to severe exposure misclassification.
• Lack of Standardization: There is no consensus on the most accurate method for linking addresses to parcels. Studies often rely on commercial geocoding software or manual inspection without rigorous validation against a gold standard.
• Equity Concerns: Errors in linkage may not be random; they may disproportionately affect high-density neighborhoods and communities with high material deprivation, introducing differential bias into health equity studies.

2. Methodology

The study utilized a large-scale, empirical evaluation comparing different linkage strategies against a "gold standard."

• Data Sources:
  • Input: 853,255 residential addresses from the National Address Database (NAD) covering Hamilton County (Cincinnati) and Franklin County (Columbus), Ohio.
  • Gold Standard: Authoritative parcel data from County Auditor Geographic Information Systems (CAGIS), including parcel boundaries, centroids, land use codes, and market values.
  • Validation: Real-world electronic health record (EHR) data from Cincinnati Children's Hospital Medical Center (1.3M+ patient records) and housing code violation data.
• Linkage Approaches Evaluated:
  1. Address Tag Fuzzy Matching: A deterministic, text-based method using the addr R package to parse addresses into structured components (street number, name, type, ZIP). It employs fuzzy string matching (e.g., allowing 1 edit distance for street names, exact match for numbers) to link directly to CAGIS records.
  2. Geomatching (Geographic):
     • Address Point Geomatching: Using precise NAD coordinates (derived from the Location Based Response System) to intersect with parcel polygons or find the nearest centroid.
     • Street Range Geomatching: Using coordinates generated by a street-range geocoder (DeGAUSS) to perform the same spatial intersections.
• Evaluation Metrics:
  • Completeness: Percentage of addresses successfully matched.
  • Agreement: Concordance with the gold standard for Parcel ID, Land Use Type (exact match), and Market Value (within 20%).
  • Stratification: Performance was analyzed by hyperlocal address density (addresses per 100m) and community material deprivation.

3. Key Results

The study revealed stark differences in performance between text-based and spatial-based methods:

• Address Tag Fuzzy Matching (Superior Performance):

  • Achieved 100% agreement with the gold standard for both Parcel ID and Land Use/Value in both counties.
  • Successfully handled multi-unit addresses (e.g., apartments) by matching the base address, though it required a rule to select a single parcel when multiple matches occurred.
  • Demonstrated high computational efficiency (100,000 addresses processed in ~20 minutes).

• Geomatching (Variable Performance):

  • Address Point Geomatching: Performed moderately well.
    • Centroid approach: 74.1% – 76.1% agreement.
    • Polygon intersection: 65.1% – 71.5% agreement.
  • Street Range Geomatching (Poor Performance):
    • Centroid approach: 18.1% – 28.3% agreement.
    • Polygon intersection: 7.2% – 11.7% agreement (the lowest performing method).
  • Completeness: Geomatching generally yielded higher "match rates" (fewer missing values) but with significantly lower accuracy compared to tag matching.

• Impact of Density and Deprivation:

  • Hyperlocal Density: As address density increased (more addresses within 100m), the agreement rates for all geomatching methods dropped significantly. For example, in the highest density quartile, street-range polygon matching dropped to 6.8% agreement in Hamilton County.
  • Equity: High address density correlated strongly with higher community material deprivation. This suggests that geomatching errors are not random but systematically bias exposure assessments in disadvantaged, high-density urban areas.

• Real-World Application:

  • When weighted by patient volume in the EHR, the disparity in accuracy widened. Geomatching methods showed lower agreement for addresses associated with large multi-unit buildings where many patients reside, further highlighting the risk of bias in clinical research.

4. Key Contributions

• Empirical Validation: Provides the first large-scale, head-to-head comparison of address tag matching versus various geomatching techniques specifically for parcel-level linkage.
• Methodological Recommendation: Demonstrates that address tag fuzzy matching is superior to geocoding for parcel-level exposure assessment, offering 100% accuracy where geocoding fails significantly (especially in dense urban settings).
• Equity Analysis: Quantifies the "differential misclassification" risk, showing that standard geocoding methods systematically fail in socioeconomically disadvantaged, high-density neighborhoods, potentially invalidating health equity findings.
• Scalability: Proves that open-source, deterministic text-matching pipelines are computationally feasible for large-scale clinical and population health databases.

5. Significance and Implications

• Research Validity: The findings suggest that studies relying on street-range geocoding for parcel-level data (e.g., housing quality, lead risk, property value) are likely suffering from severe exposure misclassification and bias.
• Clinical Integration: Health systems can integrate address tag matching (e.g., via the addr package) as a preprocessing layer to standardize and link patient addresses to environmental data before geocoding, improving data quality and reproducibility.
• Policy and Data Governance: The study underscores the need for standardized address maintenance practices and the harmonization of parcel data across jurisdictions (e.g., via the National Address Database) to support valid, scalable health research.
• Future Directions: The authors call for future research to assess how these linkage errors propagate through epidemiologic models to affect effect size estimates and health outcome associations.

In conclusion, the paper argues that for high-resolution, address-level environmental health research, text-based fuzzy matching of address components is a more accurate, scalable, and equitable approach than traditional spatial geocoding methods.