Controlling false match rates in record linkage using extreme value theory

Area of Science:

• Data Science
• Biostatistics
• Information Science

Background:

• Data cleansing is crucial for high-quality data in disease registries and medical research.
• Record linkage methods minimize errors like synonyms and homonyms, enhancing data integrity.
• Homonym errors (false matches) are critical, where distinct entities are incorrectly identified as identical.

Purpose of the Study:

• To present a new approach for estimating false match rates in record linkage.
• To leverage Extreme Value Theory (EVT) for a more efficient and cost-effective solution.
• To address the limitations of manual clerical review and existing statistical models requiring training data.

Main Methods:

• Utilizing Extreme Value Theory (EVT) within the Fellegi and Sunter framework.
• Applying the generalized Pareto distribution and mean excess plots for analysis.
• Developing a method that does not require training data for threshold determination.

Main Results:

• The proposed EVT-based approach effectively estimates false match rates.
• The method significantly reduces costs by eliminating the need for training data.
• Experimental results show comparable accuracy to methods with match status information.

Conclusions:

• Extreme Value Theory offers a viable and cost-effective alternative for estimating false match rates.
• This method enhances data quality in critical applications like medical research networks.
• The approach provides a significant advantage by removing the dependency on calibration data.