An efficient validation method of probabilistic record linkage including readmissions and twins

Area of Science:

• Data Science
• Biostatistics
• Public Health Informatics

Background:

• Probabilistic record linkage is crucial for integrating health data.
• Validating linkage accuracy is essential for reliable data analysis.
• Existing validation methods can be resource-intensive.

Purpose of the Study:

• To develop and apply an efficient, generalizable approach for validating probabilistic record linkage.
• To utilize model-guided error detection for enhanced linkage accuracy.
• To specifically validate the linkage of newborn admission records.

Main Methods:

• A double-blind validation procedure involving sample selection, data collection, and analysis.
• Oversampling uncertain links based on model-derived weights from a Dutch national newborn admission registry (30,082 records).
• External validation using faxed patient histories collected from pediatric units.

Main Results:

• A 97% response rate was achieved for the external validation faxes.
• Linkage accuracy for singleton admissions was high, with minimal errors (0.02%).
• Multiple birth readmissions showed a higher error rate (37%) due to data quality issues, but overall file error remained below 1%.

Conclusions:

• The developed external validation procedure is feasible, efficient, and effective.
• The approach successfully identified the sources of record linkage errors.
• The method provides valuable insights for improving data quality in linked registries.