Clerical Review

Clerical review targets record pairs whose composite match score falls between two predefined thresholds: an upper threshold above which pairs are automatically accepted as matches, and a lower threshold below which pairs are automatically rejected as non-matches. This interval, often derived from the Felligi-Sunter model, represents the zone of statistical ambiguity where the cost of algorithmic error—either a false match or a false non-match—is too high to automate.

Human reviewers examine pairs using a clerical review interface that presents the original, unencoded records side-by-side. Reviewers assess semantic agreement beyond what fuzzy string metrics can capture:

• Name variations: Nicknames, cultural name order differences, or transliteration inconsistencies
• Address discrepancies: Unit numbers, rural route formatting, or historical addresses
• Date of birth typos: Transposed digits that fall within an edit distance threshold
• Family structure changes: Marriage, divorce, or adoption affecting surname matching

In Privacy-Preserving Record Linkage (PPRL), clerical review presents a unique challenge: reviewers must access plaintext records to make informed decisions, temporarily breaking the cryptographic privacy guarantees. This requires strict access controls, audit logging, and often a trusted third-party or secure enclave to decrypt only the pairs in the uncertainty region. The goal is to minimize the number of pairs requiring manual inspection to reduce both cost and privacy exposure.

Clerical review serves as the ground truth generation mechanism for evaluating and calibrating automated linkage systems. Adjudicated pairs become labeled training data for:

• Threshold tuning: Adjusting match score cutoffs to balance precision and recall
• Weight recalibration: Refining the agreement and disagreement weights in the Felligi-Sunter model
• Rule refinement: Identifying systematic failure modes in deterministic or probabilistic matching logic Review outcomes are measured using inter-rater reliability metrics to ensure consistency across reviewers.

Manual review is the most expensive and time-consuming phase of record linkage. The number of pairs in the uncertainty region grows with dataset size, making blocking key selection and match score thresholding critical upstream optimizations. Best practices include:

• Keeping the clerical review rate below 5% of total pairs
• Using active learning to prioritize pairs with the highest information gain
• Implementing consensus voting across multiple reviewers for ambiguous cases
• Applying transitive closure post-review to propagate manual decisions across connected components

Deterministic linkage typically requires clerical review to resolve partial matches where only a subset of identifiers agree exactly. Probabilistic linkage uses the uncertainty region explicitly, but the volume of pairs requiring review is often higher due to the method's sensitivity to data quality. Hybrid approaches apply deterministic rules first to reduce the candidate pool, then use probabilistic scoring with a narrow uncertainty region to minimize manual effort while maintaining high recall.

The process of setting cutoff values on a composite similarity score to automatically classify record pairs. Pairs scoring above the upper threshold are automatic matches, those below the lower threshold are automatic non-matches, and those falling between the two thresholds populate the uncertainty region that requires clerical review. Threshold selection directly controls the trade-off between manual effort and linkage accuracy.

A record matching methodology that uses statistical likelihood ratios to calculate the probability that two records refer to the same entity. Unlike deterministic rules, probabilistic linkage accounts for data errors, missing values, and typographical variations. The output is a continuous score that naturally creates the uncertainty region requiring human clerical adjudication.

The evaluation of record linkage output using precision, recall, and F-measure against a ground-truth set. This assessment measures the rates of false matches (Type I errors) and false non-matches (Type II errors). Clerical review decisions directly impact these metrics, and quality assessment validates whether human adjudication is improving or degrading overall linkage accuracy.

The proportion of record pairs incorrectly classified as matches out of all true non-matching pairs. In the context of clerical review, this metric is critical because human adjudicators can introduce bias by incorrectly confirming false matches in the uncertainty region. Monitoring false match rate before and after clerical review ensures the manual process is not degrading linkage quality.

A record matching methodology that classifies a pair as a match only if a predefined set of identifiers agree exactly. This approach offers high precision but low recall in the presence of typographical errors or data inconsistencies. Deterministic linkage produces no uncertainty region, eliminating the need for clerical review but missing many true matches that probabilistic methods would flag for human adjudication.