Match Score Thresholding

Most production PPRL systems employ two distinct cutoff values rather than a single threshold to create three classification zones:

• Upper threshold (match zone): Pairs scoring above this value are automatically accepted as matches with high confidence
• Lower threshold (non-match zone): Pairs scoring below this value are automatically rejected
• Clerical review zone: Pairs falling between the two thresholds require manual human adjudication

This architecture minimizes both false positives and false negatives while keeping manual review costs manageable.

Thresholds are empirically derived from labeled ground-truth data using statistical analysis of score distributions:

• Histogram analysis: Plotting the frequency of match scores separately for true matches and true non-matches reveals natural separation points
• ROC curve optimization: The receiver operating characteristic curve identifies thresholds that balance sensitivity and specificity
• Cost-based tuning: Thresholds can be adjusted based on the relative business cost of false matches versus false non-matches

Calibration requires a representative sample of manually verified record pairs to ensure generalizability.

Raw similarity scores from different field comparators must be normalized to a common scale before thresholding can be applied:

• Min-max normalization scales all scores to a [0,1] range
• Z-score standardization centers scores around zero with unit variance
• Logit transformation maps probability estimates to an unbounded scale for better separation

Without proper normalization, thresholds become non-transferable across datasets with different characteristics, undermining the reproducibility of linkage results.

Static global thresholds often perform poorly when data quality varies across different population subgroups or data sources. Adaptive thresholding dynamically adjusts cutoffs:

• Stratified thresholds: Different cutoff values are applied to different blocking partitions based on local data quality metrics
• Confidence-based adjustment: Thresholds tighten when auxiliary metadata indicates higher data reliability
• Iterative refinement: Thresholds are recalibrated after each linkage pass using newly discovered match patterns

This approach is particularly valuable in healthcare informatics where data quality varies significantly across providers.

Threshold selection directly governs the precision-recall trade-off in record linkage:

• Raising the upper threshold increases precision (fewer false matches) but decreases recall (more true matches missed)
• Lowering the upper threshold captures more true matches but introduces more false positives
• Widening the clerical zone improves overall accuracy but increases manual review costs linearly

Formal evaluation using F-measure (the harmonic mean of precision and recall) provides a single metric for comparing threshold configurations across different linkage runs.

A robust threshold should produce consistent classification behavior across minor variations in input data. Stability analysis techniques include:

• Bootstrap resampling: Repeatedly sampling the score distribution to measure threshold variance
• Sensitivity analysis: Testing how small perturbations to threshold values affect final match counts
• Cross-validation: Training thresholds on one data partition and validating on held-out data

Unstable thresholds indicate overfitting to the calibration dataset and predict poor performance when the linkage system encounters novel data patterns in production.