Deterministic Linkage

The core mechanism requires a Boolean AND operation across a predefined set of identifiers. If all selected fields match precisely, the pair is linked; if any field disagrees, the pair is rejected. There is no concept of partial similarity or scoring. This binary logic makes the process computationally efficient but brittle, as a single typographical error in a name or date of birth will cause a true match to be missed, directly impacting recall.

The primary trade-off of this methodology is its performance profile. Precision (the proportion of declared matches that are true matches) is typically very high, often approaching 100% if the identifiers are highly discriminative. However, recall (the proportion of true matches successfully identified) is low. This is because the method cannot tolerate any data entry errors, missing values, or temporal changes in identifiers. It is best suited for environments with strictly controlled, high-quality data entry.

The success of deterministic linkage hinges entirely on the choice of matching keys. Common identifiers include:

• Unique Identifiers: Social Security Number, National Health Index, or a composite enterprise ID.
• Composite Keys: A combination of fields like full name, date of birth, and gender. The selected keys must be highly discriminative to avoid false positives. Using a non-unique key like only a birth date would result in a catastrophic false match rate, linking records of different individuals who share the same birthday.

This method has zero tolerance for data heterogeneity. Common failure modes include:

• Typographical Errors: 'Jon Smith' vs 'John Smith'.
• Phonetic Variations: 'Catherine' vs 'Kathryn'.
• Missing Data: A blank middle initial field breaks the AND condition.
• Temporal Shifts: A changed surname after marriage. To mitigate this, deterministic linkage requires rigorous upstream data standardization and cleaning, often involving parsing, casing, and trimming of string fields before comparison.

Compared to probabilistic methods, deterministic linkage is computationally cheap. The comparison logic is a simple string equality check, which can be accelerated using database indexing and exact-match joins. This makes it suitable for real-time, operational systems where a quick, definitive answer is needed and the data quality is known to be high. It avoids the complex weight calculation and iterative expectation-maximization steps required by the Fellegi-Sunter model.

Deterministic linkage is a subset of the broader entity resolution spectrum. Unlike probabilistic linkage, which calculates a likelihood ratio based on the agreement and disagreement of multiple fields, deterministic rules do not handle uncertainty. A probabilistic model might still link 'Robert Smith' and 'Rob Smith' if other fields like address and date of birth agree strongly, whereas a deterministic rule would fail. The choice between them is a direct trade-off between computational simplicity and robustness to real-world data noise.

The foundational statistical framework for probabilistic record linkage that formalizes the decision problem of classifying record pairs as matches, non-matches, or requiring clerical review.

• Computes agreement weights and disagreement weights per field
• Assumes conditional independence between fields given match status
• Defines an optimal decision rule using likelihood ratios
• Introduces the concept of an uncertainty region for manual adjudication

A fuzzy matching parameter that defines the maximum allowable string transformation cost for two values to be considered equivalent. Commonly implemented using Levenshtein distance or Damerau-Levenshtein distance.

• Levenshtein counts insertions, deletions, and substitutions
• Damerau-Levenshtein adds transposition operations for adjacent characters
• Thresholds are typically tuned per field based on data quality
• Bridges the gap between strict deterministic matching and full probabilistic models

A string similarity metric optimized for short strings like personal names, giving higher scores to strings with matching prefixes. Widely used in fuzzy record linkage to catch common typographical errors.

• Produces a score between 0 (no similarity) and 1 (exact match)
• Applies a prefix bonus scaling factor for initial character agreement
• Particularly effective for matching names with transposed letters
• Often combined with phonetic encoding for robust name matching

A class of algorithms that index words by their pronunciation to match homophones, enabling the linkage of records despite spelling variations. Common implementations include Soundex and Double Metaphone.

• Soundex encodes names into a letter followed by three digits
• Double Metaphone produces both primary and alternate encodings
• Handles cultural and linguistic pronunciation variations
• Essential for linking records with phonetic spelling errors that deterministic exact matching would miss