Data preprocessing

Record linking is very sensitive to the quality of the data being linked, so all data sets considered (particularly their key identifier fields) should preferably undergo a data quality assessment prior to record linking. Many key identifiers of the same entity can be presented differently between (and even within) different data sets, which can make linking records considerably more difficult if not understood beforehand. For example, the key identifiers of a man named William J. Smith may appear in three different data sets as follows:.

In this example, the different formatting styles result in records that appear different but in fact all refer to the same entity with the same logical identifier values. Most record linking strategies would result in more accurate linking if these values ​​were first normalized or standardized into a consistent format (for example, all names are "Last Name, First Name", and all dates are "YYYY/MM/DD"). Standardization can be accomplished through simple rule-based data transformations or more complex procedures, such as lexicon-based tokenization and probabilistic Markov models.[8] Several of the packages listed in the Software Implementations section offer some of these features to simplify the data normalization process.

Entity resolution

Entity resolution is an operational intelligence process, typically driven by an entity resolution engine or middleware, through which organizations can connect diverse data sources to understand potential entity matches and non-obvious relationships between multiple data silos. It analyzes all information regarding individuals and/or entities from the various data sources and then applies a probability score to determine which identities match and what, if any, non-obvious relationships exist between those identities.

Entity resolution engines are typically used to uncover Risk (IT) risks, fraud, and conflicts of interest, but are also useful tools for use within customer data integration (CDI) and master data management (GDM) requirements. Typical uses of entity resolution engines include: terrorist detection, insurance fraud detection, USA Patriot Act compliance, detection of retail organized crime networks, and applicant screening.

For example: Across different data silos - employee records, supplier data, watch lists, etc. - an organization may have several variations of an entity called ABC, which may or may not be the same individual. These entries may, in fact, appear as ABC1, ABC2, or ABC3 within those data sources. By comparing similarities between underlying attributes, such as address, date of birth, or social security number, the user can eliminate some potential matches and confirm others as highly likely.

Entity resolution engines then apply rules, based on common sense logic, to identify hidden relationships in the data. In the example above, perhaps ABC1 and ABC2 are not the same individual, but rather two different people who share common attributes, such as an address or phone number.

Data matching

Although entity resolution solutions include data matching technology, many data matching offerings do not fit the definition of entity resolution. According to John Talburt, director of UALR's Center for Advanced Research in Entity Resolution and Information Quality, there are four factors that distinguish entity resolution from data matching:

Unlike data quality products, the most powerful identity resolution engines also include a rules engine and a workflow process, which apply business intelligence to resolved identities and their relationships. These advanced technologies make automated decisions and affect business processes in real time, limiting the need for human intervention.

The simplest type of record linking, called deterministic or rule-based record linking, generates links based on the number of individual identifiers that match between available data sets.[9] Two records are said to be matched using a deterministic record linking procedure if all or some identifiers (above a certain threshold) are identical. Deterministic record linking is a good option when entities in data sets are identified by a common identifier, or when there are multiple representative identifiers (for example, name, date of birth, and gender when identifying a person) whose data quality is relatively high.

As an example, consider two standardized data sets, set A and set B, that contain different data about patients in a hospital system. The two data sets identify patients using a series of identifiers: Social Security Number (SSN), name, date of birth (DOB), gender, and ZIP code (ZIP). Records from two data sets (identified by the "#" column) are shown below:

The simplest deterministic record linking strategy would be to choose a single identifier that is assumed to be unique, for example the SSN, and declare that records that share the same value identify the same person while records that do not share the same value identify different people. In this example, deterministic binding based on the SSN would create entities based on A1 and A2; A3 and B1; and A4. While A1, A2, and B2 appear to represent the same entity, B2 would not be included in the match because it is missing a value for the SSN.

Handling exceptions such as missing identifiers involves creating additional record linkage rules. One of these rules, in the case of a missing SSN, could be to compare the name, date of birth, gender and zip code with other records in the hope of finding a match. In the example above, this rule would still not match A1/A2 with B2 because the names are still slightly different: standardization put the names in the proper format (Last Name, First Name) but could not discern "Bill" as a nickname from "William". Running names through a phonetic algorithm (Soundex, NYSIIS), or Metaphone, can help solve these types of problems (although you can still run into last name changes as a result of a marriage or divorce), but then B2 would be paired only with A1 since the zip code in A2 is different. Therefore, another rule would need to be created to determine whether differences in certain identifiers are acceptable (such as zip code) and which are not (such as date of birth).

As this example demonstrates, even a small decrease in data quality or a small increase in data complexity can result in a very large increase in the number of rules needed to correctly link records. Ultimately, these linking rules will be too numerous and interrelated to build without the help of specialized software tools. Additionally, linking rules are typically specific to the nature of the data sets they are designed to link. One study was able to link the Social Security Death Master File to two hospital records in the Midwestern United States using SSN, NYSIIS-encoded name, month of birth, and sex, but these rules may not work as well with data sets from other geographic regions or with data collected in younger populations.[10] Therefore, ongoing maintenance testing of these rules is necessary to ensure that they continue to work as expected as new data are introduced into the system. system and need to be linked. New data that has different characteristics than initially expected could require a complete rebuild of the record linking rule set, which could be a very time-consuming and costly effort.

Probabilistic record merging, sometimes called fuzzy matching (also probabilistic merging or fuzzy merging in the context of database merging), takes a different approach to the record merging problem by considering a broader range of potential identifiers, calculating weights for each identifier based on its estimated ability to correctly identify a match or non-match, and using these weights to calculate the probability that two given records refer to the same entity. Pairs of records with probabilities greater than a certain limit are considered matching, while pairs with probabilities less than another limit are considered unmatched; pairs that fall between these two limits "Boundary (math)") are considered "potential matches" and may be treated accordingly (e.g., human-reviewed, linked or unlinked, depending on requirements). While deterministic record linking requires pre-programming a series of potentially complex rules, proballiistic record merging methods can be "trained" to perform well with much less human intervention.[1].

Many probabilistic record linkage algorithms assign match/non-match weights to identifiers using two probabilities called u and m. The probability u is the probability that an identifier in two unmatched records matches by pure chance. For example, the probability u for birth month (where there are twelve values ​​that are approximately uniformly distributed) is 1/12 0.083; identifiers with values ​​that are not uniformly distributed will have different probabilities u for different values ​​(possibly including missing values). The probability m is the probability that an identifier in matching pairs matches (or is sufficiently similar, such as strings with low Jaro-Winkler or Levenshtein distance). This value would be 1.0 in the case of perfect data, but since this is rarely (if ever) true, it can be estimated. This estimate can be made based on prior knowledge of the data sets, by manually identifying a large number of matching and non-matching pairs to "train" the probabilistic record fusion algorithm, or by iteratively running the algorithm to obtain closer estimates of the probability m. If a value of 0.95 were estimated for probability m, the match/non-match weights for the birth month identifier would be:.

The same calculations would be done for all other identifiers considered to find their match/non-match weight. Each identifier in one record would then be compared to the corresponding identifier in another record to calculate the total weight of the pair: the match weight is added to the running total whenever a pair of identifiers matches, while the non-match weight is added (i.e., the running total is decremented) whenever the pair of identifiers does not match. The resulting total weight is compared to the aforementioned limits to determine whether the pair should be bound, unbound, or reserved for special consideration (e.g., manual validation)[2][11]​.

Determining where to set the match/non-match boundaries is a balancing act between obtaining acceptable sensitivity (or completeness, the proportion of truly matching records that are linked by the algorithm) and positive predictive value (or precision, the proportion of records linked by the algorithm that actually match). There are several manual and automated methods for predicting the best limits, and some log merging software packages have built-in tools to help the user find the most acceptable values. Because this task can be computationally demanding, especially for large data sets, a technique known as blocking is often used to improve efficiency. Blocking attempts to restrict comparisons to only those records in which one or more particularly discriminating identifiers match, which has the effect of increasing positive predictive value (precision) at the expense of sensitivity (completeness).[11] For example, blocking based on a last name and a phonetically encoded zip code would reduce the total number of comparisons needed and improve the chances that the linked records would be correct (since the two identifiers already match), but would potentially miss records that were refer to the same person whose last name or zip code was different (due to a marriage or a move, for example). Blocking based on birth month, a more stable identifier that is expected to change only in the event of data error, would provide a more modest gain in positive predictive value and a loss in sensitivity, but would create only twelve distinct groups which, for extremely large data sets, may not provide a large net improvement in calculation speed. Therefore, extensive record merging systems often use multiple blocking passes to group data in various ways to obtain groups of records that must be compared with each other.[3].

Machine learning.

In recent years, a variety of machine learning techniques have been used in record linking. The classical Fellegi-Sunter algorithm for probabilistic record linkage described above has been recognized to be equivalent to the Naive Bayes algorithm in the field of machine learning and suffers from the same assumption of the independence of its features (an assumption that is usually not true). Greater accuracy can often be achieved by using other machine learning techniques, including a single-layer perceptron. Coupled with distributed technologies, the accuracy and scale for record linking can be further improved.

In an application with two files, A and B, denote the rows (records) by in file A, and in file B. Assigning characteristics to each record. The set of records representing identical entities is defined by.

and in the complement of the set that is, set representing different entities, it is defined as.

A vector is defined, which contains the agreements and disagreements encoded in each feature:

where is a subscript of the characteristics (sex, age, marital status, etc.) in the files. The conditional probabilities of observing a given specific vector are defined as.

and.

, respectively.

Master data management

Most master data management (MDM) products use a record linkage process to identify records from different sources that represent the same real-world entity. This link is used to create a "gold master record" that contains the clean and reconciled data about the entity. The techniques used in MDM are the same as for record linking in general. MDM extends this matching not only to create a "gold master record" but also to infer relationships. (i.e. a person has the same/similar surname and the same/similar address, this could imply that they share a family relationship).

Data Warehousing and Business Intelligence

Record linking plays a key role in data warehousing and business intelligence. Data warehouses serve to combine data from many different operational source systems into a logical data model, which can then be fed into a business intelligence system for reporting and analysis. Each operating source system may have its own method of identifying the same entities used in the logical data model, so it is necessary to link records between different sources to ensure that information about a particular entity in one source system can be seamlessly compared with information about the same entity in another source system. Data standardization and subsequent record linking often occur in the "transform" part of the extract, transform, and load (ETL) process.

Historical Research

Record linking is important for social history research, since most data sets, such as census records (Census (statistics)) and parish records, were recorded long before the invention of national identification numbers. When digitizing ancient sources, linking data sets is a prerequisite for longitudinal study. This process is often further complicated by the lack of standard spellings of names, surnames that change depending on place of residence, changing administrative boundaries, and problems of cross-checking data with other sources. Record linking was one of the most prominent topics in the field of history and computer science in the 1980s, but has since received less research attention.

Medical Practice and Research

Record linkage is an important tool for creating the data needed to examine the health of the public and the healthcare system itself. It can be used to improve data holdings, data collection, quality assessment and information dissemination. Data sources can be examined to eliminate duplicate records, identify missing cases and missing cases (e.g., population censuses), create people-oriented health statistics, and generate disease registries and health surveillance systems. Some cancer registries link multiple data sources (e.g., hospital admissions, pathology and clinical reports, and death records) to generate their records. Record linking is also used to create health indicators. For example, fetal and infant mortality is a general indicator of a country's socioeconomic development, public health and maternal and child services. If infant death records are compared to birth records, it is possible to use birth variables, such as birth weight and gestational age, along with mortality data, such as cause of death, when analyzing the data. Linkages can help in follow-up studies of cohorts or other groups to determine factors such as vital status, residential status, or health outcomes. Tracing is often necessary for monitoring industrial cohorts, clinical trials, and longitudinal surveys to obtain the cause of death and/or cancer. An example of a long-standing and successful record linkage system that enables population-based medical research is the Rochester, Minnesota-based Rochester Epidemiology Project ("Rochester, Minnesota").

The main reasons cited are:

The initial idea of ​​record linkage dates back to Halbert L. Dunn in his 1946 article titled "Record Linkage" published in the American Journal of Public Health.

Howard Borden Newcombe then laid the probabilistic foundations of modern record linkage theory in an article published in 1959 in Science.[4] These were formalized in 1969 by Ivan Fellegi") and Alan Sunter, in their pioneering work "A Theory For Record Linkage", where they demonstrated that the probabilistic decision rule they described was optimal when the comparison attributes were conditionally independent.[5] In their work they recognized the growing interest in apply advances in computing and automation to large collections of administrative data, and Fellegi-Sunter theory remains the mathematical foundation for many record linkage applications.

Since the late 1990s, several machine learning techniques have been developed that, under favorable conditions, can be used to estimate the conditional probabilities required by the Fellegi-Sunter theory. Several researchers have reported that the conditional independence assumption of the Fellegi-Sunter algorithm is often violated in practice; However, published efforts to explicitly model conditional dependencies between comparison attributes have not resulted in improved record linkage quality. [On the other hand, machine learning or neural network algorithms that do not rely on these assumptions typically provide much higher accuracy, when sufficient labeled training data is available.[6]​.

Record linkage can be performed completely without the aid of a computer, but the main reasons computers are often used to complete record linkage are to reduce or eliminate manual review and to make the results more easily reproducible. Computer matching has the advantages of allowing central supervision of processing, better quality control, speed, consistency and better reproducibility of results.[7]​.

Data preprocessing

Record linking is very sensitive to the quality of the data being linked, so all data sets considered (particularly their key identifier fields) should preferably undergo a data quality assessment prior to record linking. Many key identifiers of the same entity can be presented differently between (and even within) different data sets, which can make linking records considerably more difficult if not understood beforehand. For example, the key identifiers of a man named William J. Smith may appear in three different data sets as follows:.

In this example, the different formatting styles result in records that appear different but in fact all refer to the same entity with the same logical identifier values. Most record linking strategies would result in more accurate linking if these values ​​were first normalized or standardized into a consistent format (for example, all names are "Last Name, First Name", and all dates are "YYYY/MM/DD"). Standardization can be accomplished through simple rule-based data transformations or more complex procedures, such as lexicon-based tokenization and probabilistic Markov models.[8] Several of the packages listed in the Software Implementations section offer some of these features to simplify the data normalization process.

Entity resolution

Entity resolution is an operational intelligence process, typically driven by an entity resolution engine or middleware, through which organizations can connect diverse data sources to understand potential entity matches and non-obvious relationships between multiple data silos. It analyzes all information regarding individuals and/or entities from the various data sources and then applies a probability score to determine which identities match and what, if any, non-obvious relationships exist between those identities.

Entity resolution engines are typically used to uncover Risk (IT) risks, fraud, and conflicts of interest, but are also useful tools for use within customer data integration (CDI) and master data management (GDM) requirements. Typical uses of entity resolution engines include: terrorist detection, insurance fraud detection, USA Patriot Act compliance, detection of retail organized crime networks, and applicant screening.

For example: Across different data silos - employee records, supplier data, watch lists, etc. - an organization may have several variations of an entity called ABC, which may or may not be the same individual. These entries may, in fact, appear as ABC1, ABC2, or ABC3 within those data sources. By comparing similarities between underlying attributes, such as address, date of birth, or social security number, the user can eliminate some potential matches and confirm others as highly likely.

Entity resolution engines then apply rules, based on common sense logic, to identify hidden relationships in the data. In the example above, perhaps ABC1 and ABC2 are not the same individual, but rather two different people who share common attributes, such as an address or phone number.

Data matching

Although entity resolution solutions include data matching technology, many data matching offerings do not fit the definition of entity resolution. According to John Talburt, director of UALR's Center for Advanced Research in Entity Resolution and Information Quality, there are four factors that distinguish entity resolution from data matching:

Unlike data quality products, the most powerful identity resolution engines also include a rules engine and a workflow process, which apply business intelligence to resolved identities and their relationships. These advanced technologies make automated decisions and affect business processes in real time, limiting the need for human intervention.

The simplest type of record linking, called deterministic or rule-based record linking, generates links based on the number of individual identifiers that match between available data sets.[9] Two records are said to be matched using a deterministic record linking procedure if all or some identifiers (above a certain threshold) are identical. Deterministic record linking is a good option when entities in data sets are identified by a common identifier, or when there are multiple representative identifiers (for example, name, date of birth, and gender when identifying a person) whose data quality is relatively high.

As an example, consider two standardized data sets, set A and set B, that contain different data about patients in a hospital system. The two data sets identify patients using a series of identifiers: Social Security Number (SSN), name, date of birth (DOB), gender, and ZIP code (ZIP). Records from two data sets (identified by the "#" column) are shown below:

The simplest deterministic record linking strategy would be to choose a single identifier that is assumed to be unique, for example the SSN, and declare that records that share the same value identify the same person while records that do not share the same value identify different people. In this example, deterministic binding based on the SSN would create entities based on A1 and A2; A3 and B1; and A4. While A1, A2, and B2 appear to represent the same entity, B2 would not be included in the match because it is missing a value for the SSN.

Handling exceptions such as missing identifiers involves creating additional record linkage rules. One of these rules, in the case of a missing SSN, could be to compare the name, date of birth, gender and zip code with other records in the hope of finding a match. In the example above, this rule would still not match A1/A2 with B2 because the names are still slightly different: standardization put the names in the proper format (Last Name, First Name) but could not discern "Bill" as a nickname from "William". Running names through a phonetic algorithm (Soundex, NYSIIS), or Metaphone, can help solve these types of problems (although you can still run into last name changes as a result of a marriage or divorce), but then B2 would be paired only with A1 since the zip code in A2 is different. Therefore, another rule would need to be created to determine whether differences in certain identifiers are acceptable (such as zip code) and which are not (such as date of birth).

As this example demonstrates, even a small decrease in data quality or a small increase in data complexity can result in a very large increase in the number of rules needed to correctly link records. Ultimately, these linking rules will be too numerous and interrelated to build without the help of specialized software tools. Additionally, linking rules are typically specific to the nature of the data sets they are designed to link. One study was able to link the Social Security Death Master File to two hospital records in the Midwestern United States using SSN, NYSIIS-encoded name, month of birth, and sex, but these rules may not work as well with data sets from other geographic regions or with data collected in younger populations.[10] Therefore, ongoing maintenance testing of these rules is necessary to ensure that they continue to work as expected as new data are introduced into the system. system and need to be linked. New data that has different characteristics than initially expected could require a complete rebuild of the record linking rule set, which could be a very time-consuming and costly effort.

Probabilistic record merging, sometimes called fuzzy matching (also probabilistic merging or fuzzy merging in the context of database merging), takes a different approach to the record merging problem by considering a broader range of potential identifiers, calculating weights for each identifier based on its estimated ability to correctly identify a match or non-match, and using these weights to calculate the probability that two given records refer to the same entity. Pairs of records with probabilities greater than a certain limit are considered matching, while pairs with probabilities less than another limit are considered unmatched; pairs that fall between these two limits "Boundary (math)") are considered "potential matches" and may be treated accordingly (e.g., human-reviewed, linked or unlinked, depending on requirements). While deterministic record linking requires pre-programming a series of potentially complex rules, proballiistic record merging methods can be "trained" to perform well with much less human intervention.[1].

Many probabilistic record linkage algorithms assign match/non-match weights to identifiers using two probabilities called u and m. The probability u is the probability that an identifier in two unmatched records matches by pure chance. For example, the probability u for birth month (where there are twelve values ​​that are approximately uniformly distributed) is 1/12 0.083; identifiers with values ​​that are not uniformly distributed will have different probabilities u for different values ​​(possibly including missing values). The probability m is the probability that an identifier in matching pairs matches (or is sufficiently similar, such as strings with low Jaro-Winkler or Levenshtein distance). This value would be 1.0 in the case of perfect data, but since this is rarely (if ever) true, it can be estimated. This estimate can be made based on prior knowledge of the data sets, by manually identifying a large number of matching and non-matching pairs to "train" the probabilistic record fusion algorithm, or by iteratively running the algorithm to obtain closer estimates of the probability m. If a value of 0.95 were estimated for probability m, the match/non-match weights for the birth month identifier would be:.

The same calculations would be done for all other identifiers considered to find their match/non-match weight. Each identifier in one record would then be compared to the corresponding identifier in another record to calculate the total weight of the pair: the match weight is added to the running total whenever a pair of identifiers matches, while the non-match weight is added (i.e., the running total is decremented) whenever the pair of identifiers does not match. The resulting total weight is compared to the aforementioned limits to determine whether the pair should be bound, unbound, or reserved for special consideration (e.g., manual validation)[2][11]​.

Determining where to set the match/non-match boundaries is a balancing act between obtaining acceptable sensitivity (or completeness, the proportion of truly matching records that are linked by the algorithm) and positive predictive value (or precision, the proportion of records linked by the algorithm that actually match). There are several manual and automated methods for predicting the best limits, and some log merging software packages have built-in tools to help the user find the most acceptable values. Because this task can be computationally demanding, especially for large data sets, a technique known as blocking is often used to improve efficiency. Blocking attempts to restrict comparisons to only those records in which one or more particularly discriminating identifiers match, which has the effect of increasing positive predictive value (precision) at the expense of sensitivity (completeness).[11] For example, blocking based on a last name and a phonetically encoded zip code would reduce the total number of comparisons needed and improve the chances that the linked records would be correct (since the two identifiers already match), but would potentially miss records that were refer to the same person whose last name or zip code was different (due to a marriage or a move, for example). Blocking based on birth month, a more stable identifier that is expected to change only in the event of data error, would provide a more modest gain in positive predictive value and a loss in sensitivity, but would create only twelve distinct groups which, for extremely large data sets, may not provide a large net improvement in calculation speed. Therefore, extensive record merging systems often use multiple blocking passes to group data in various ways to obtain groups of records that must be compared with each other.[3].

Machine learning.

In recent years, a variety of machine learning techniques have been used in record linking. The classical Fellegi-Sunter algorithm for probabilistic record linkage described above has been recognized to be equivalent to the Naive Bayes algorithm in the field of machine learning and suffers from the same assumption of the independence of its features (an assumption that is usually not true). Greater accuracy can often be achieved by using other machine learning techniques, including a single-layer perceptron. Coupled with distributed technologies, the accuracy and scale for record linking can be further improved.

In an application with two files, A and B, denote the rows (records) by in file A, and in file B. Assigning characteristics to each record. The set of records representing identical entities is defined by.

and in the complement of the set that is, set representing different entities, it is defined as.

A vector is defined, which contains the agreements and disagreements encoded in each feature:

where is a subscript of the characteristics (sex, age, marital status, etc.) in the files. The conditional probabilities of observing a given specific vector are defined as.

and.

, respectively.

Master data management

Most master data management (MDM) products use a record linkage process to identify records from different sources that represent the same real-world entity. This link is used to create a "gold master record" that contains the clean and reconciled data about the entity. The techniques used in MDM are the same as for record linking in general. MDM extends this matching not only to create a "gold master record" but also to infer relationships. (i.e. a person has the same/similar surname and the same/similar address, this could imply that they share a family relationship).

Data Warehousing and Business Intelligence

Record linking plays a key role in data warehousing and business intelligence. Data warehouses serve to combine data from many different operational source systems into a logical data model, which can then be fed into a business intelligence system for reporting and analysis. Each operating source system may have its own method of identifying the same entities used in the logical data model, so it is necessary to link records between different sources to ensure that information about a particular entity in one source system can be seamlessly compared with information about the same entity in another source system. Data standardization and subsequent record linking often occur in the "transform" part of the extract, transform, and load (ETL) process.

Historical Research

Record linking is important for social history research, since most data sets, such as census records (Census (statistics)) and parish records, were recorded long before the invention of national identification numbers. When digitizing ancient sources, linking data sets is a prerequisite for longitudinal study. This process is often further complicated by the lack of standard spellings of names, surnames that change depending on place of residence, changing administrative boundaries, and problems of cross-checking data with other sources. Record linking was one of the most prominent topics in the field of history and computer science in the 1980s, but has since received less research attention.

Medical Practice and Research

Record linkage is an important tool for creating the data needed to examine the health of the public and the healthcare system itself. It can be used to improve data holdings, data collection, quality assessment and information dissemination. Data sources can be examined to eliminate duplicate records, identify missing cases and missing cases (e.g., population censuses), create people-oriented health statistics, and generate disease registries and health surveillance systems. Some cancer registries link multiple data sources (e.g., hospital admissions, pathology and clinical reports, and death records) to generate their records. Record linking is also used to create health indicators. For example, fetal and infant mortality is a general indicator of a country's socioeconomic development, public health and maternal and child services. If infant death records are compared to birth records, it is possible to use birth variables, such as birth weight and gestational age, along with mortality data, such as cause of death, when analyzing the data. Linkages can help in follow-up studies of cohorts or other groups to determine factors such as vital status, residential status, or health outcomes. Tracing is often necessary for monitoring industrial cohorts, clinical trials, and longitudinal surveys to obtain the cause of death and/or cancer. An example of a long-standing and successful record linkage system that enables population-based medical research is the Rochester, Minnesota-based Rochester Epidemiology Project ("Rochester, Minnesota").

The main reasons cited are: