False Match Rate

FMR is formally defined as the ratio of false positives (FP) to the total number of true non-matching pairs (TN + FP).

• Formula: FMR = FP / (FP + TN)
• Range: 0.0 (perfect) to 1.0 (total failure)
• Interpretation: An FMR of 0.001 means 1 in 1,000 true non-matches is erroneously linked
• Complement: FMR is the inverse of specificity; a low FMR indicates high specificity

In PPRL, this metric is often estimated against a ground-truth set where true match status is known, as the encoded identifiers prevent direct inspection of false positives.

A non-zero FMR represents a direct privacy breach in PPRL contexts. Each false match incorrectly associates two distinct individuals' records, potentially exposing sensitive attributes across organizational boundaries.

• Healthcare: A false match could merge Patient A's medical history with Patient B's insurance claims, violating HIPAA
• Financial Crime: Erroneously linking two separate entities could flag an innocent party for money laundering
• Census Data: False matches distort demographic statistics and resource allocation models

PPRL protocols are explicitly designed to minimize FMR even at the cost of higher False Non-Match Rates (FNMR), as privacy preservation takes precedence over recall.

FMR is controlled by adjusting the match score threshold in probabilistic linkage models like the Fellegi-Sunter framework. This creates an inherent trade-off with the False Non-Match Rate (FNMR).

• Lowering the threshold increases recall but raises FMR, risking privacy breaches
• Raising the threshold improves precision and lowers FMR but misses true matches
• Clerical review region: Pairs with scores between two thresholds are flagged for manual adjudication

In PPRL, thresholds are typically set conservatively to keep FMR below 0.1% or lower, accepting a higher FNMR to protect against erroneous linkages.

FMR is directly related to precision in information retrieval terminology, though the framing differs based on the denominator.

• Precision = TP / (TP + FP) — measures accuracy among predicted matches
• FMR = FP / (FP + TN) — measures error rate among true non-matches
• Key distinction: Precision depends on the prevalence of true matches in the dataset; FMR does not

A system with 99.9% precision can still have an unacceptable FMR if the number of true non-matches vastly exceeds true matches, a common scenario in large-scale record linkage where the search space is quadratic.

Measuring FMR in privacy-preserving settings is challenging because plaintext identifiers are never revealed. Common estimation approaches include:

• Ground-truth subset: A manually labeled sample of record pairs used to approximate error rates
• Synthetic data injection: Known non-matching pairs are embedded in the dataset to detect false linkages
• Cryptographic auditing: Secure multi-party computation protocols that compute aggregate error statistics without exposing individual matches
• Differential privacy bounds: Mathematical guarantees on the maximum FMR given specific noise parameters

Without these techniques, organizations must rely on post-linkage clerical review of a sample to validate match quality.

FMR is often confused with the False Discovery Rate (FDR), but they measure different things and are used in different contexts.

• FMR: FP / (FP + TN) — error rate over all true non-matches; used in linkage quality assessment
• FDR: FP / (FP + TP) — proportion of false positives among all positive predictions; used in multiple hypothesis testing
• Practical impact: In a dataset with 1M non-matches and 1K true matches, an FMR of 0.01% produces 100 false matches, while FDR would be 100/(100+900) = 10%

PPRL literature consistently uses FMR to emphasize the privacy risk to the non-matching population, not just the quality of predicted matches.

The systematic evaluation of record linkage output by comparing algorithmic classifications against a verified ground-truth set. This process quantifies performance using standard information retrieval metrics.

• Calculates precision, recall, and F-measure
• Requires a gold-standard dataset of known matches and non-matches
• Essential for tuning match score thresholding parameters

The complement to the False Match Rate, representing the proportion of true matching pairs that the system incorrectly classifies as non-matches. This metric directly impacts linkage recall.

• Also known as the Type II error or miss rate
• High FNMR leads to fragmented golden records
• Often traded off against FMR via match score thresholding

The foundational metrics for evaluating binary classification in record linkage. Precision measures the purity of the match set, while recall measures its completeness.

• Precision: True Matches / (True Matches + False Matches)
• Recall: True Matches / (True Matches + False Non-Matches)
• The F-measure provides a harmonic mean balancing both concerns

The process of setting cutoff values on a composite similarity score to automatically classify record pairs. Pairs scoring above the upper threshold are matches; below the lower threshold are non-matches.

• Pairs falling between thresholds enter a clerical review queue
• Adjusting thresholds directly tunes the False Match Rate and False Non-Match Rate
• A critical control point for operationalizing the Felligi-Sunter Model

The manual human adjudication of record pairs that fall into an uncertainty region where automated probabilistic scoring cannot confidently classify them. This process resolves edge cases that drive error rates.

• Targets pairs near the decision boundary of the match score threshold
• Adds a human-in-the-loop safeguard to reduce both false matches and false non-matches
• Often the most resource-intensive phase of a linkage project

The foundational statistical framework for probabilistic record linkage that computes match weights based on the agreement and disagreement patterns of individual record fields.

• Estimates the likelihood ratio for a true match using m-probabilities and u-probabilities
• Provides the theoretical basis for calculating expected False Match Rate
• Underpins modern match score thresholding strategies