Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 293
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Chapter 10 Invited Session on More Record Linkage Applications in Epidemiology Chair: Patricia Nechodom, University of Utah Authors: Selma C.Kunitz, Clara Lee, and Rene C.Kozlojf, Kunitz and Associates, and Harvey Schwartz, Agency for Health Care Policy and Research Christian Houle, Jean-Marie Berthelot, Pierre David, and Michael Wolfson, Statistics Canada Cam Mustard and Leslie Roos, University of Manitoba Steve Kendrick, National Health Service, Scotland
OCR for page 294
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition This page in the original is blank.
OCR for page 295
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Record Linkage Methods Applied to Health Related Administrative DataSets Containing Racial and Ethnic Descriptors Selma C.Kunitz, Clara Lee, and Rene C.Kozloff, Kunitz and Associates, Inc. Harvey Schwartz, Agency for Health Care Policy and Research Abstract In response to the lack of easily retrievable clinical data to address health services and medical effectiveness questions, especially as they relate to racial/ethnic minorities, the Center for Information Technology (CIT), Agency for Health Care Policy and Research (AHCPR) recently sponsored a project on record linkage methodology applied to automated medical administrative datasets containing racial and ethnic identifiers (Contract 282–94–2005). The primary objectives of the project were to: link patient-level related datasets that contain racial and ethnics descriptors; and assess the value of the linked data to address medical effectiveness research questions that focus on the quality, effectiveness, and outcomes from care for minority populations. KAI, AHCPR's contractor, received approval from the State of New York's Department of Health to utilize the Statewide Planning and Research Cooperative System (SPARCS) files Discharge Data Abstract (DDA) and Uniform Billing files (UBF), which contain all acute hospital discharge and claims data, the SPARCS Ambulatory Surgical files, and the Cardiac Surgery Reporting System (CSRS) files, a research dataset. KAI received files for the 1991, 1992, and 1993 time periods. The files were successfully linked by patient “visits” across the time periods. While the linked data appear to be of high quality, the process of obtaining and linking the data is lengthy. Additionally, these administrative health care data sets contain millions of records that document all hospital stays, and thus, identifying appropriate subpopulations for a particular research question is a time- and resource-consuming effort. While the administrative health care datasets may be useful in answering questions about charges, length of stay, and other health service issues, their current utility may be less useful in answering clinical questions for minority populations. These datasets can be used to explore potential associations among diagnoses, treatment, and outcome variables. However, understanding the mediating factors and the decision-making variables that result in patient care may not be possible. For example, the results of diagnostic tests such as angiograms are not generally recorded in these datasets, thus limiting the ability to carefully subgroup patients by disease severity. With consideration for the potential utility of these datasets, however, there are several recommendations that emanate from the study. This talk will briefly describe the research questions posed, linkage process, findings, and recommendations for additional action and policy considerations.
OCR for page 296
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Introduction The ability to link automated health data records is of critical importance in our rapidly changing health care system. In a managed care and cost containment environment, researchers require reliable and valid data collected over time and across providers that describe patient characteristics and the location, process, cost, quality, and outcome of care to analyze which procedures are effective and produce satisfactory patient outcomes. Approaches and methods to linking records across time and providers are needed to provide information to policy makers, health plans, practitioners, consumers, and patients to make decisions about accessing, using, and paying for care, as well as the effectiveness of that care. Background In response to the lack of easily retrievable clinical data to answer medical effectiveness questions, especially as they relate to racial/ethnic minorities, the Agency for Health Care Policy and Research (AHCPR) sponsored a project on “Record Linkage Methodology Applied to Linking Automated Data Bases Containing Racial and Ethnic Identifiers to Medical Administrative Data Bases” under AHCPR contract number 282 –94–2005 (Kunitz and Associates, Inc., 1996). This linkage demonstration project contributed to AHCPR's research goals by reviewing and adding to record linkage methodology; illustrating the value of this methodology; assessing the need for further development; and providing guiding principles to developers. The primary objectives of this record linkage methodology project were to: link two patient- level related datasets that contain racial and ethnic descriptor; and assess the value of the linked data to address medical effectiveness research questions that focus on the quality, effectiveness, and outcomes from care for minority populations. Data Sets AHCPR's contractor, KAI, a health research firm, identified data sets to use for assessing the value of linking administrative health related data bases to support medical effectiveness research in minority populations. KAI received approval from New York State's Department of Health (NYSDOH) to utilize the Statewide Planning and Research Cooperative System (SPARCS) files Discharge Data Abstract (DDA) and Uniform Billing files (UBF), the SPARCS Ambulatory Surgery files, and the Cardiac Surgery Reporting System (CSRS) files. KAI received files for 1991, 1992, and 1993. The selected systems and data files are briefly described as follows: SPARCS (State Wide Planning and Research Cooperative System) is a system maintained by the NYSDOH. The Discharge Data Abstract files (DDA) contain all acute hospital discharge data and the Uniform Billing Files (UBF) contain all acute hospital billing records. Data about surgeries performed at hospital-based ambulatory care centers and certified diagnostic and treatment freestanding centers are maintained in the Ambulatory Surgery files. The data are used for planning and research. Three of the files extracted from SPARCS for this project were the DDA, UBF, and the Ambulatory Surgery file. NYSDOH staff combined the acute hospital DDA and UBF data files by individual hospital stay for this project. Thus, we received both matched and unmatched records from the DDA and UBF for 1991–1993. Because the files were selected based on DDA variables, unmatched records are those that are in the DDA file but do not have a corresponding match in the UBF. A completeness level of 95% is typically achieved in SPARCS files, a figure that is supported by our re-
OCR for page 297
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition search, as seen in Figure 1. Yet, those records which are unmatched may reflect not only missing UBF records, but also incorrect information which may have hindered the original matching process performed by the NYSDOH. Figure 1. —Matching Rates Between DDA and UBF Records Year Total Matched Unmatched N % 1991 626,222 594,302 31,290 5% 1992 699,246 663,323 35,923 5% 1993 714,583 677,778 36,805 5% KAI obtained 31,290 unmatched records out of a total of 626,222 for 1991, 35,923 unmatched out of a total of 699,246 for 1992, and 36,805 unmatched records out of 714,583 for 1993. The selection process did not enable KAI to receive data which was in the UBF file but absent in DDA. In addition, we received the Ambulatory Surgery files for these three years. Cardiac Surgery Reporting System (CSRS) is a voluntary reporting system of all in-hospital cardiac surgeries. It contains risk factors, clinical descriptors and procedure data and is used as a research data set. We received these files for 1991 –1993. Figure 2 summarizes the size of the original data files. The DDA/UBF files contain between 2.5 and 3 million records each year. The ambulatory surgery files were not segregated by year and contain slightly more than two million records. The CSRS data files are also summarized by year and contain a considerably smaller number of records because of the more narrow focus of the records on cardiac surgery. Figure 2. —Summary of Sizes of Complete Data Files Data Set Year(s) Number of Records SPARCS 1991 1,687,521 1992 1,677,948 1993 1,660,109 Ambulatory Surgery 1991–1993 2,121,542 Cardiac Surgery Reporting System (CSRS) 1991 19,783 1992 21,592 1993 22,491
OCR for page 298
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Research Question One of the primary goals of this project was to determine whether a medical effectiveness research question could be successfully addressed by the linked data. The selected research question for this project relates risk factors, treatment, and outcome of cardiovascular disease to minority status: Are the racial/ethnic differences in mortality and morbidity from coronary heart disease related to racial/ethnic differences in treatment? The working hypothesis stated that minorities are less likely to receive surgical treatment for coronary artery disease and, therefore as a group, experience higher incidence of cardiovascular morbidity and mortality than the majority U.S. population. The cohort was to be extracted from the linked SPARCS and CSRS datasets. The linked datasets were to contain records for 3 years, 1991–1993. Males and females aged 45–75 who were assigned a diagnosis of ischemic heart disease (ICD-9 codes 410–414) were to be included. Confidentiality Approach One of the primary issues in acquiring the New York State files was data confidentiality. Technically, the problems of confidentiality of data are often addressed by suppressing, encrypting, or compressing information. In these data sets primary identifiers such as name, address, and telephone number were removed or suppressed from the files and secondary identifiers such as Medical Record Number (MRN), Admission Number, and Physician License Numbers (PLNs) were encrypted consistently across files and years to aid the matching process. Typically, confidentiality restrictions hinder the matching of large data sets. Identifiers such as name, address, and medical record number are important in order to be confident that the correct linkages are being made. If only demographic data and broad geographic identifiers are available such as gender, race, age, and zip code, then a large group of people may have the same characteristics with the result that their records inaccurately matched. Cardiac Subset—Identification and Issues The original research plan specified the use of ICD-9 codes 410–414 to address the research question. The low yield on initial matches, however, indicated that we needed to expand these codes to obtain a more complete record match between the SPARCS and CSRS files. Therefore, for the linkage process, the codes were expanded to include: 390.xx —459.xx—disease of the circulatory system; 212.7x—benign neoplasm of the heart; 745.xx—bulbus cordis anomalies and other cardiac anomalies; 861.0x—injury to the heart without open wound to thorax; 861.1x—injury to the heart with open wound to thorax; 901.xx— injury to thoracic aorta; and 996.0x—mechanical complication of cardiac device. Figure 3 summarizes the number of potential patients on the DDA/UBF files using the ischemic heart disease codes (ICD-9 410–414) and an expanded set of codes. Figure 3. —Universe of Patient Records in DDA/UBF DDA/UBF File Year Initial Universe of ICD-9 Codes—410–414 Expanded Universe of ICD-9 Codes—390–459 1991 170,779 626,222 1992 189,198 699,246 1993 190,497 714,583
OCR for page 299
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Record Linkage The linkage software used for this project was MatchWare Technology Incorporated's (MTI) Automatch, developed by MTI's founder, Matthew Jaro (Jaro, 1997). MTI was KAI's subcontractor and its linkage experts collaborated with KAI's clinical researchers in conducting this project. Several steps were involved in the data preparation process prior to performing the record matching or linking process. Fields that are common to the files had to be identified and recoded, where necessary, for potential use in the linkage process. Common person and event fields included for all three data sets were MRN, sex, date of birth, patient county, hospital identification number, diagnosis, procedure code and date, and Physician License Number (PLN). Fields common to two of the three files included age, patient zip code and state, admit date, discharge date, and payor. As an example of recoding needs, race codes on the CSRS files were converted to correspond to SPARCS codes as shown in Figure 4. Figure 4. —Race Code Conversions Description SPARCS Race CSRS Race Asian or Pacific Islander 1 8 Black 2 2 Hispanic 3 8 Native American 4 8 Other 5 8 White 6 1 Linkage Objective The linkage objective was to build a longitudinal, comprehensive patient history that captured clinical encounters over time and across care settings. Thus, records for the same patient were linked in two ways: matches were performed within each of the three data sets; and matches were performed between the DDA/UBF files and CSRS and between the DDA/UBF and Ambulatory Surgery files. Steps in Record Linkage Steps in the linkage process included identifying duplicate records; running preliminary matches as an iterative process to determine which fields yielded the most appropriate matches; identifying appropriate cutoff weights; and running the final linkage. Duplicate records were identified on each of the files with no file having duplicates that exceeded 1% of the records. Automatch's method for determining most effective variables and probability weights to match across files was evaluated in preliminary iterative match runs. The process was iterative and consisted of selecting key variables for each match strategy, producing preliminary matched pairs, examining matched pairs with marginal match weights, and revising the parameters to better discriminate between apparent true and false matches. For the final matches specific probabilities of agreement were
OCR for page 300
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition determined based on the preliminary matches. The match cutoff weight was chosen so that the estimated absolute odds of a true match for record pairs with that match weight were 95:5. Linkage Data Quality Analysis The linkage results were reviewed for data reliability and validity. First, the same variables on linked and unlinked records were compared to assess internal consistency and reliability. Agreement was 99% or greater for all variables except for date of principal procedure (67%) and admission number (83%); MRN, zip code, county, and other procedure each exhibited an agreement rate of 93%. Principal procedure as well as other procedure differences may reflect differences in reimbursement categories that were changed on the UBF for payment advantages. Admission number and MRN are scrambled by computer and any clerical error such as a transposition of numbers in the original MRN yields an inconsistent scrambled MRN. Likewise, transposition of numbers in zip code and county can yield mismatches. The DDA and UBF responses for linked and unlinked records were then compared for the same patient. The responses are fairly consistent across DDA and UBF subfiles and between linked and unlinked records with slight differences in reimburser and diagnoses, which could be a function of the research question reflected in the linked files. The DDA variables were selected for matching and were compared for linked and unlinked patient records, because of their tendency to be more reliable in the clinical area. In the linked records, patients are older (age = 65–71% versus 59% for unlinked records), most likely reflecting the research question which focuses on cardiac diagnoses. Racial characteristics are similar as are ethnicity and gender. Linked and unlinked records for Ambulatory Surgery patients were also compared. Analysis showed a greater percentage of the linked records to have a higher proportion of angina as the primary diagnosis while in the unlinked files there was a higher proportion of arterial disease, perhaps reflecting procedures performed in ambulatory surgery, i.e., angiograms. There were more Medicare reimbursers in the linked records, which is consistent with differences in age groups. Other fields show no differences. Linked records compared with unlinked records for the CSRS patients showed a greater proportion of persons over 65, most likely reflecting the diagnostic groups of research interest. There were no gender, race, or ethnicity differences in the linked and unlinked records, reflecting similar patient populations. The general consistency between the DDA and UBF subgroups and the consistency between linked and unlinked records within each of the data sets demonstrate the reliability of the matching and indicates that the linked records generally reflect the file population. Racial Subsets Responses across racial subgroups for DDA variables were reviewed. As expected, more Blacks, Asians, and other minorities are treated in the New York City area (over 70%) than other parts of the state. Payment also differs, with a higher proportion of Whites on Medicare (69% versus 46% for Blacks and Others and 40% for Asian Americans). A higher proportion of Blacks and other minorities have Medicaid as the primary reimburser (Blacks—26%, Whites—5%, Asian Americans—25%, Other— 27%). Blacks have a higher proportion of diabetes (4% versus 1% for Whites, 2% for Asian Americans and Other) and hypertension diagnosis (5% versus 1% for Whites, and 2% for Asian Americans and Other), and a slightly lower proportion of myocardial infarctions (Whites—11%, Blacks—7%, Asian Americans—10%, Other—11%) as principal diagnosis. Responses for other variables for linked and
OCR for page 301
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition unlinked records by racial and ethnic categories are consistent, indicating that the linked file is a representative subset of the larger file. Research Subsets The research subsets, defined as the original diagnoses categories, 410.xx-414.xx, were examined next. Comparing the DDA and UBF records on the SPARCS data set for linked and unlinked records indicates that age is higher on the linked records (age = 65 = 75%) than on the unlinked records (= 65 = 59%), reflecting the cardiac procedure research question. Also reflecting the research question is the larger number of patients on Medicare in the linked data set (74% versus 59% in the unlinked data set). Comparisons between linked and unlinked records in the ambulatory surgery research files indicate no significant differences between the two subsets. A review of responses for racial and ethnic subgroups for the linked and unlinked subsets in the DDA research file indicates that in both, Whites are significantly older (78% Whites in the linked subset and 66% Whites in the unlinked subset are 65 or older). In the other racial categories, however, there is a larger proportion under 65 (Blacks—42%; Asian Americans—33%; and Other—56%) in both linked and unlinked subgroups. The age differences between White and minority racial subgroups are also reflected in the proportion of patients on Medicare. There do not appear to be other major differences between White and minority subgroups. These trends are also reflected in the differences between Hispanic and non-Hispanic subgroups. Linked Data Sets and the Research Question Preparing the data to answer the research question was a complex process despite the fact that record linkage had taken place. The primary reason for the complexity of the process is that the research question focuses on outcome while the linkage focused on diagnosis. The linkage focus on diagnosis appears logical because it is how patients are generally categorized for health services and clinical research. However, medical effectiveness questions often focus on outcomes and thus, within diagnoses, outcome is an important patient characteristic. The research question, while resulting in a complex subject identification procedure, was typical of many medical effectiveness questions. The amount of time, then, needed for progressing from a linked data set to analyses for outcomes research, is several months and should be built into the research planning process. Data and Linkage Issues Several issues related to health care data sets and application of linkage methodology were identified: Purpose. —The purpose of the primary data collection endeavor impacts on the quality of specific variables and on their utility for linkage and their relevance for addressing a medical effectiveness question. For example, primary diagnosis frequently differed between the DDA and UBF subfiles. The diagnoses in the DDA is driven by clinical practice while in the UBF it is driven by reimbursement. Variables such as age and date of birth, gender, county of residence, hospital identification number, MRN, admission date, and procedure date may not be consistent across billing and discharge administrative records as well as the research records for several reasons: accuracy is not important for billing, discharge, and some research; an individual 's high anxiety state; and family members reporting information under stress. Further, discharge abstracts generally reflect clinical diagnoses more accurately, while billing data typically reflect
OCR for page 302
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition charge justification. Encryption. —Encrypting the Medical Record Number (MRN), admission numbers, and physician license numbers degrades the efficiency of the matching software. The matching software used in this study can take into account slight differences among identifiers such as transposition of characters and adjust the match for them. However, since the encryption process scrambles identifiers or assigns a sequential number to records, the software is not dealing with actual numeric identifiers, which may have typographical errors. Thus this feature of the software is not useful for electronically encrypted or created numbers. The degradation was demonstrated in the first matching pass between the ambulatory surgery file and the DDA/UBF file. The MRN in the DDA/UBF file is defined as ten characters and was encrypted as such. The MRN in the Ambulatory Surgery file is defined as seventeen characters in which the first ten characters actually contain the MRN and the last seven characters are spaces. When the initial match between the DDA/UBF file and the Ambulatory Surgery file took place there were no matches. The resolution involved the recreation of the Ambulatory Surgery file using only the first ten characters of the MRN in the encryption process. If however, the MRNs had been provided without being encrypted, the software could have adjusted for the spaces at the end of the original MRN in the Ambulatory Surgery file. Race and Ethnicity Codes. —The race and ethnicity codes are not always accurate as demonstrated by all observations for a particular New York State hospital which contained a race code of 5 and ethnicity code of 2 for all patient records. Additionally, the state SPARCS programmer indicated that there were software problems for RACE and ETHNICITY for certain hospitals that affected accuracy. Dependent Relationships Among Variables. —Certain pairs of patient and provider variables are strongly dependent on each other. For example, MRN is frequently hospital-specific and physicians are generally associated with only a few hospitals, thus PLN (Physician License Number) and Hospital Identification Number are also strongly dependent as shown statistically by chi square and uncertainty coefficient tests. The Automatch software requires that only one member of each dependent pair is used as a match variable because of relative odds of a true match calculation. For example, if both date of birth and age were used in a matching process, the calculated match weight would overstate the relative odds of a true match by exactly the contribution of the second occurrence. While date of birth and age represent the same concept, hospital and physicians may be logically independent entities although statistically associated. The nature of association in health related records should be considered in the matching process and perhaps, a different statistical approach used for these data. Matching Variables. —A related issue is determining what variables provide the greatest yield during the blocking and matching procedures. Linking is generally dependent upon person identifiers such as name and address, and date of birth, as well as on procedure and diagnosis codes from health related records. Since name and address were omitted from the files used to preserve personal privacy, other variables assumed greater importance. The clinical research staff, experienced with clinical data, recommended the use of age and date of birth, gender, county of residence, hospital identification number, MRN, admission date, and procedure date. The researchers pointed out that procedure and diagnoses codes can vary between administrative and clinical data sets because of reimbursement interests and are more likely to be accurate in clinical files. Identification of the variables most appropriate for linking health related files is still an open research issue.
OCR for page 303
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Type and Number of Variables Utilized for Linking. —Personal identifiers such as name and address are frequently used in census and vital statistics linkage efforts. Since these variables are not present on the health files, other variables that appear in several files and have a high probability of accuracy must be identified. Some examples are hospital identification number, admission date, and zip code. Additionally, linkage software experts often argue for numerous variables upon which to link. We found that the health-related data sets were more frequently linked with fewer discrepancies in the matching records when fewer variables are used. Thus the percentage of “true” matches was higher with fewer variables or, conversely, the number of false positives was lower. Experience from Other Applications. —Experience and assumptions gathered from other applications of linkage methodology such as census data cannot necessarily be applied to health-related data. Thus, for health-related data, multidisciplinary teams of linkage software programmers and health researchers need to develop appropriate linkage algorithms and to identify variables pertinent for linking these files. Findings Despite time delays and other issues, the files were successfully linked and the data were used to address the above hypothesis that pertains to care among minority populations. General findings are as follows: Data quality in the administrative and research files generally appears high and the data are potentially useful for health services research. Both the linkage process and the analytic phase for large data sets are lengthy and resource consuming. The practicality of linking large health-related data sets needs to be balanced against the number of years the data will be useful. If data can be used to support research for three to five years, then the linkage overhead expense may be justifiable. Costs of linking large data sets, then, need to be balanced against the potential benefits. Linking is only the first step when the data are to be used to address research questions. The linkage process identifies a set of unique indexes for each of the patient records in each of the linked files. Depending upon the focus of the research question, it is necessary to carefully review the data files and the index files, which consumes both time and computer processing. Since the data files for large data sets must reside on mainframe computers, it also is a costly process. In this project, in which those subjects with the same diagnoses who received cardiac surgery are compared to those who did not, patients with relevant diagnoses had to be identified to form a subgroup from the SPARCS DDA/UBF files. The subgroup had to then be identified on the index files, determined whether linked or not linked to the CSRS file, and then found on the CSRS files. These steps precede any analytic procedures and represent the complexity of data management procedures that are associated with the analysis of the linked files. Utility of administrative data sets in answering medical effectiveness questions is variable. Clearly, identifying diagnoses, treatment, and outcome at a general level is possible and meaningful. The data set can be used to explore potential associations among diagnoses, treatment,
OCR for page 322
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition ology. Instead the development and refinement of linkage methods has taken place as a response to a wide variety of immediate operational demands. We have become to all intents and purposes a general purpose linkage facility at the heart of the Scottish Health Service operating to very tight deadlines often set in terms of weeks and in extreme cases, days. This has placed a high premium on developing quick, effective and accurate methods of linkage with an emphasis on fitness for purpose rather than straining for precision for its own sake. Despite the lack of time and resources available for background research and development in linkage methodology, these conditions have in fact fostered, especially in recent years, a rapidly changing and developing approach to linkage. Before describing the most significant developments involved, a brief overview of the main components will serve to set them in context. The Elements of Linkage For the purposes of this discussion, record linkage using probability matching can be regarded as having three phases or elements each involving a key question. Bringing pairs of records together for comparison. —How do we bring the most effective subset of pairs of records together for comparison? It is usually impossible to carry out probability matching on all pairs of records involved in a linkage. Usually only a subset are compared, those which share a minimum level of identifying information. This has been traditionally achieved by sorting the files into “blocks” or “pockets” within which paired comparisons are carried out (Gill and Baldwin, 1987). Calculating probability weights. —How do we assess the relative likelihood that pairs of records belong to the same person? This lies at the heart of probability matching and has probably been the main focus of much of the record linkage literature. (Newcombe, 1988). Making the linkage decision. —How do we convert the probability weights representing relative odds into absolute odds which will support the linkage decision? The wide variety of linkages undertaken has been particularly important in moving forward understanding in this area. It would probably be fair to say that of the three areas, it is the second, the calculation of probability weights which has received the most attention and is the best understood. Developments in Scotland over the last few years have occurred in the other two areas as the two subsequent sections will demonstrate. Before moving on to these developments, our approach to the calculation of probability weights has been relatively conventional and can be quickly summarised. A concern has been to avoid overelaboration and over complexity in the algorithms which calculate the weights. Beyond a certain level increasing refinement of the weight calculation routines tends to involve diminishing returns. This relatively basic approach has been facilitated by the relative richness of the identifying information available on most health related records in Scotland. To take an example, for the internal linking of hospital discharge (SMR1) records across Scotland we have available the patient's surname (plus sometimes maiden name), first initial, sex and date of birth. We also have postcode of residence. For records within the same hospital (or sometimes the same Health Board) the hospital assigned case reference number can be used. In addition positive weights can be assigned for correspondence of the date of discharge on one record with the date of admission on another. Surnames are compressed using the Soundex/NYSIIS name compression algorithms
OCR for page 323
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition (Newcombe, 1988) with additional scoring assigned for more detailed levels of agreement and disagreement. Wherever possible specific weights relating to degrees of agreement and disagreement are used. Bringing the Pairs of Records Together: One Pass Linkage The Limitations of Sort-and-Match By the time the largest linked data set covered several years of data and consisted of several millions of records, a particular challenge emerged. The linkage team began to be asked to link data sets consisting of relatively small numbers of “external” or “newcomer” records to the central catalog of identifiable records. The external or newcomer records might consist of respondents to a survey, a specialised disease register or a particular group of employees. In all cases the aim was to link the newcomer data set to the central catalog of records so that the experience of the individuals involved could be traced forward from the date of survey, the date last known to the disease register or the date of employment. As we have seen, in record linkage it is impossible to bring together and compare all the pairs of records involved in the linkage. The number of pairs which are brought together for comparison is normally reduced to manageable proportions by some form of blocking by which only those pairs of records which share common sets of attributes are compared. For example a common strategy is to compare only those pairs of records which share either the same first initial and NYSIIS/Soundex code or the same date of birth. The normal method of achieving such blocking is to sort the two files concerned on the basis of the blocking criteria. Thus, for a first pass of linkage, the files would be sorted by first initial and NYSIIS/Soundex code to bring together into the same “pocket” or “block” all records sharing the same NYSIIS/Soundex code and first initial. Records would only be compared within this block. Because a number of truly linked pairs of records would not be brought together on this basis (for example, because of a misrecording of first initial), a second pass could be carried out which blocks by date of birth. This second pass involves resorting the files on the basis of date of birth to create a second set of pockets or blocks within which comparison takes place. The results of the first and second passes need to be reconciled and this involves sorting the file yet again. The key point is that standard methods of blocking involve sorting all the records involved in the linkage at least twice and usually more often. When faced with the kind of linkage mentioned above, involving linking a small number of newcomer records to a central catalog holding several millions of records such a procedure is at best immensely wasteful and at worst impossible. No matter how few newcomer records are involved, it is still necessary to sort all the central catalog records for the years of interest. If only a few years are involved, and especially if linkage is restricted to a subset of the central records e.g., cancer registrations, the exercise is feasible but immensely inefficient. If it is desired to link newcomer records to the entire data set, the exercise becomes, in reality, impossible. One Pass Linkage: Blocking Without Sorting The question thus became: how can we link a relatively small number of newcomer records to the catalog without having to repeatedly sort the catalog? The solution adopted has been to store the newcomer records in memory and carry out blocking using indexes based on numerical elements of the blocking criteria. The catalog records can then be read in sequentially and compared with all the newcomer records which fit the chosen blocking criteria (Kendrick and Mcllroy, 1996).
OCR for page 324
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition The linkage is thus carried out in the course of “one pass” through the catalog data set. Before they are brought into contact with the catalog, all newcomer records are read into memory and stored in an array indexed by a unique numeric record identifier. Necessary pre-processing such as generation of NYSIIS/Soundex codes is also carried out. The next step is the creation of blocking index arrays. In this description we assume that two sets of blocking criteria are being used: first initial and NYSIIS/Soundex code on the one hand, and date of birth on the other. The blocking index arrays are indexed by numeric elements of the blocking criteria. Thus the first blocking array uses the numeric element of the NYSIIS/Soundex code as its index. All NYSIIS/Soundex codes consist of a letter followed by three figures e.g., A536 or B625. The first blocking index array has a row for each number from 001 to 999 which covers all possible numeric elements of NYSIIS/Soundex codes. In each row are stored the numeric identifiers of the newcomer records whose NYSIIS/Soundex code has the relevant numeric element. For example, the identifiers of newcomer records with surname FRAME (NYSIIS/Soundex code F650) and BROWN (B650) would be stored in the same row. The second blocking index array has three indices: year, month and day of birth. Along the fourth dimension of the array are stored the numeric identifiers of the newcomer records sharing that date of birth. Catalog records are then read in one by one. Suppose the first catalog record is for someone named BROWN (NYSIIS/Soundex code B650) with date of birth 1st March, 1922. Row 650 of the first blocking index array is inspected to see whether any newcomer records share the numeric element of the Soundex code. If any are found, then the first newcomer record is accessed via its numeric identifier in the newcomer record array. An immediate comparison is made of the first letter of the Soundex code and the first initial. If both match then we proceed to full probability matching between the catalog and newcomer records. If neither or only one match then no further action is taken. We then look at the next newcomer record (if any) indexed on the relevant row of the blocking index array. Blocking by date of birth is even easier to simulate in that the blocking criteria are entirely numeric. The catalog record can be directed to all newcomer records which share the same day (in this case 1); month (in this case 3); and year (in this case 22) by directly accessing the relevant array. How the results of the ensuing pair comparisons are stored and implemented depends upon the structure and purpose of the linkage. Whenever links above a certain weight occur they can be output and stored for implementation in a provisional linkage file. This provisional linkage file can itself be flexibly interrogated to implement a given structure of linkage e.g., we may be only interested in the best link (the link with the highest weight) achieved by each newcomer record (see below). One Pass Linkage: Practical Considerations The above strategy, whereby the newcomer records can be indexed only in terms of the numeric elements of any blocking criteria, is necessary when we are using a programming environment which only allows numeric indexing of arrays. If either the newcomer data or the catalog data is stored in a database
OCR for page 325
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition which allows direct access by any type of key then the logic of the exercise would be simplified. The file could be flexibly indexed by whatever blocking keys are felt appropriate. Our impression at present would be that using memory still has advantages in terms of speed of access. This of course is a practical issue and may well change quickly as relational databases and “search engines ” improve in speed and efficiency. The number of newcomer records which can be linked in one pass through the data is of course limited by the available memory. Memory is needed both for storing the elements of the newcomer records which are necessary for linkage and for storing the blocking index arrays. For most ad hoc linkages involving anything up to 15,000 newcomer records this has not tended to be a problem. Larger newcomer data sets can often be linked in sections without affecting the logic of the linkage. Of the three elements of linkage it is this which is most dependent upon the capabilities of available hardware and software. The implication is that as these capabilities develop, there will be immense potential for moving beyond current limitations. The Linkage Decision: Relative and Absolute Odds, Structuring the Linkage and the Best Link Principle At the heart of the record linkage enterprise is the decision as to whether two records are truly linked. Most often the question is one of whether the records involved relate to the same person. The calculation of probability weights aims to provide a mathematical grounding for this decision. However, it is a fundamental characteristic of the odds represented by probability weights that they are relative odds rather than absolute odds. They only serve to rank the pairs of records involved in a given linkage in order of the probability that they are truly linked. The relative odds do not represent fixed absolute or betting odds such that a probability weight of 25, for example, would always representing absolute or betting odds of 50/50. The conversion factor will vary from linkage to linkage. It is absolute odds which are needed to inform the linkage decision. In practical terms the issue of the determinants of the conversion from relative odds to absolute odds can often be bypassed in that the required threshold in terms of absolute odds can usually be identified empirically from inspection of a sample of pairs. However a broad understanding of the relationship between relative and absolute odds is useful in that it can help optimise the way a given linkage should be structured. Not all linkages require the same absolute odds. The absolute odds required depend upon the purpose of the linkage. They depend upon the costs associated with missing a true link compared with making a false link. For statistical purposes, the required absolute odds may be 50/50. If the linkage is to be used for administrative or patient contact purposes where a false linkage may have extremely damaging consequences, very high absolute odds may be required. Relative to Absolute Odds: A Priori Factors Two of the factors involving in converting relative to absolute odds take the form of relatively straightforward numerical principles. Newcombe has stated them in the context of a search file and a file being searched (Newcombe, 1988; Newcombe, 1995). The first principle is that the higher the proportion of records in the search file for which there exists a linked record in the file being searched, the more favorable
OCR for page 326
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition will be the conversion factor between relative and absolute odds. The second proposition is that the larger the file being searched, the less favourable will be the conversion factor between relative and absolute odds. These factors are given for any specific linkage. Relative to Absolute Odds: Structural Factors However the conversion factor between relative and absolute odds can be influenced by how the linkage is implemented. It is important to design the linkage in a way which takes maximum advantage of the structures of the files involved and the relationships between the records in the files. For example, are the relationships between the records in two files one-to-one, one-to-many or many-to-many? How much confidence do we have in previous linkages which may have been carried out on the files involved? How confident are we that a file to be linked already contains only one record per person? For example, if we want to link to each other a set of hospital discharge records, we have no a priori knowledge of how many records belong to each person. Our best bet is to do a conventional internal linkage and inspect all resulting pairs in setting a threshold. In this case we have relatively little leverage to improve the terms of conversion between relative odds and absolute odds. If however we are linking a file of hospital discharge records to a file of death records we can obtain some “structural leverage.” Death only occurs once and assuming that this is reflected in there being only one death record per person in the file of death records, the linkage becomes many-to-one. Each hospital discharge record should link to only one death record. The terms of conversion from relative to absolute odds can be improved by only retaining, for each hospital discharge record, the best (highest weight) link which is achieved to a death record (see also Winkler, 1994). Similarly, at the other end of the life cycle, if we are linking baby records to mothers records, assuming that the mothers records themselves have been correctly linked we should allow each baby to link to only one mother. This in fact was the first context in which the importance of structuring the linkage emerged in Scotland. Best Link and Structural Leverage: An Example The CHI/NHSCR Linkage These rather abstract considerations can be best understood in the context of a particular example. In common with the rest of the United Kingdom, Scotland is committed to the development of a unique patient identifier to help streamline the management of all patient contacts with the health service. Historically, Scotland has possessed two health-related registers of the Scottish population. It was felt that the combined strengths of the two registers would provide a firm basis for a new patient identifier. For the last twenty years the Community Health Index (CHI) has operated on a regional basis as a primary care patient register for such purposes as screening for breast and cervical cancer and childhood immunisation. It contains a wealth of operational information with high population coverage. However, the regional indexes were initially compiled on an opportunistic basis and there was a general perception that there were gaps in its coverage and that there was a high proportion of duplicate records for people who had moved from one area of Scotland to another.
OCR for page 327
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Scotland also possesses a National Health Service Central Register (NHSCR) which has been carefully maintained to contain one record for each resident of Scotland. The NHSCR however contains relatively little operational information. The initial plan was to carry out an internal linkage of the aggregated regional CHI indexes in order to remove duplicate records and then to link the resulting aggregated CHI data set to the NHSCR to form the basis of a national index. However, based on our early experience of structuring linkages to maximise the power of the linkage, it was felt that linking the CHI databases to each other via a linkage to the NHSCR would provide more “leverage” in the linkage. (Kendrick et al., 1997) The data which was available on both data sets to enable linkage was reasonable but not excessively rich. We had forenames, surnames, sex and date of birth but the only residential data was Health Board of residence (average size 300,000). A major bonus was that National Health Service number was available in a well formatted form on all NHSCR records. Because of its irregular format, the British NHS number has been notoriously difficult to use and was available on only a proportion of CHI records and with wide variations in accuracy and formatting. Although the linkage was primarily concerned with “current” CHI records, those reflecting the current residence and GP registration of the Scottish population, “redundant” CHI records for people who had died or moved to a new Health Board were included in the linkage as a possible basis for constructing historical traces. In order to find a correct NHSCR “home” for as many CHI records as possible the NHSCR file also contained deaths from 1981 as well as known emigrants. Since we were confident that the NHSCR did contain one record for every Scottish resident but there were suspicions that the CHI data set contained duplicate records (as well as legitimately multiple historical records), it was decided to structure the link as many-to-one. Each CHI record was allowed to link to only one NHSCR record—the one with which it achieved the highest probability weight. Each NHSCR record on the other hand was allowed to link to as many CHI records as necessary. Relative to Absolute Odds: Conversion Factors The linkage can be described in terms of the factors which were outlined in the previous section as determining the conversion factor between relative and absolute odds. Purpose of the linkage. —Any links accepted from the linkage would form the basis of patient contact. A very high level of confidence in the validity of any links was required. Missed links were regarded as less of a problem in that they would normally be picked up in the course of the running of the new index. Thus very high absolute odds for linkage were required. A priori probabilities. —Given that both sets of data represented a high level of coverage of the Scottish population, there was a very high probability that a person represented on the CHI file would also be represented on the NHSCR file. In terms of Newcombe's first rule, circumstances could not have been more favourable. File sizes. —Reflecting as they did the entire Scottish population as well as deaths and transfers these were large files: approximately 6.3 million NHSCR records against 7.8 million CHI records.
OCR for page 328
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition This gives a high coincidence factor and, according to Newcombe's second rule, would normally serve to push up the relative odds required for given absolute odds. Structuring the file. —Given the knowledge that all Scottish residents were likely to be represented by one NHSCR record and one or more CHI records, it made sense to structure the linkage as a best link many-to-one linkage i.e. allowing each CHI record to link only to the NHSCR record with which it achieved best link would be the most effective route and would maximise the conversion factor between relative and absolute odds. In broad terms then the linkage faced two difficult circumstances: the requirement for very high absolute odds and the large file sizes. These were more than outweighed however by the massive leverage contributed by the use of the best link principle in the context of a very high a priori probability that people were represented in both files. Linkage Results Of approximately 5,360,000 current registered CHI records, 4,600,000 or 86% linked deterministically to an NHSCR record. There was a match between the Soundex/NYSIIS code of surname, first initial, date of birth, sex and NHS number. For the remaining 750,000 CHI records, probability matching was carried out. Resources for clerical checking were limited and such checking was limited to a sample of best link pairs to determine a probability weight which would represent absolute odds for the correctness of a linkage which were sufficiently high for administrative purposes. Staff of Health Board Primary Care Teams and the National Health Service Central Register checked 2,500 pairs using existing search and confirmation systems. No incorrect links were found at a probability weight greater than 30 and this was chosen as the administratively acceptable threshold. To put this outcome into a broad comparative perspective we can compare the CHI/NHSCR linkage with previous linkages in Scotland which did not use the best link principle but which linked similar types of record using virtually the same agreement and disagreement weights for the main identifying items such as name and date of birth. In the linkage of the Scottish hospital discharge and death record data sets using probability matching, the fifty/fifty threshold (i.e., the weight at which it is equally likely that the two records belong or do not belong to the same person) has remained relatively constant at a probability weight of 25. The fifty/fifty threshold for the best links of CHI to NHSCR records is around 15. Similarly, the threshold below which links between Scottish Cancer Registrations and death records are clerically checked and above which they are accepted automatically is a weight of 40. In the CHI/NHSCR linkage as we have seen, this threshold is 30. In both cases the difference is ten units in the currency of binit weights or logs to the base 2. In terms of odds this is an improvement in the conversion factor from relative to absolute odds of 210 or around a thousandfold. Why the use of only best links in this context should contribute so much extra leverage compared with a pure threshold method is perhaps intuitively obvious but is much more difficult to explain in principle. The logic is perhaps best illustrated by a hypothetical example. Let us suppose that a CHI record on which is recorded the name Angus MacAllan with date of birth 25/01/1952 has achieved its best link with an NHSCR record on which is recorded the name Angus
OCR for page 329
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition McAllan born 24/01/1951. There is no NHS number on the CHI record and no other elements agree so that the link achieves what would be, in the context of an unstructured purely threshold linkage, only a moderate probability weight implying a less than fifty/fifty chance that the records belong to the same person. We can best assess the likelihood that these two records would belong to the same person in the CHI/NHSCR linkage context by an indirect route. Let us imagine what would have to be true for the two records not to belong to the same person. Either: there is no NHSCR record relating to the individual represented on the CHI record and in addition there exists on the NHSCR file a record relating to another Angus Mc/MacAllan with a highly similar date of birth; or there is an NHSCR record corresponding to the individual represented on the CHI record but there are sufficient discrepancies in the recording of the identifying information for this “true link” Angus MacAllan that an NHSCR record for another Angus MacAllan in fact achieves a higher probability weight with the CHI record. Neither of these scenarios are impossible but they are highly improbable and it is much more likely that the two records really do belong to the same person. The method used had two additional advantages. The file which was output from the linkage took the form of a copy of each CHI record to which was appended an extract from the NHSCR record to which it had achieved the best link and the weight at which the link was achieved. This file was used as a basis for generating pairs for inspection and links could be extracted at whatever weights were necessary. In essence this means that the threshold for linkage was set and could be varied retrospectively without having to rerun the linkage. The problem of twins has always bedevilled record linkage. The CHI/NHSCR linkage was able to take advantage of the fact that the NHS numbers for most pairs of twins are consecutive and a high negative weight was given for pairs of records with consecutive NHS numbers. Linkages using best link are in normal circumstances better than linkages using only a numeric threshold. In the presence of consecutive NHS numbers for twins the linkage was very successful in correctly allocating the records for twins. One Pass Linkage and the Structuring of Linkages Although one pass linkage and the structuring of linkages in terms of the best link have developed as separate responses to different challenges, they are not entirely independent. Given that one of the main aims of one pass linkage is to avoid having to repeatedly sort or restructure the larger or target file, it is natural to implement one pass linkage as a best link procedure i.e., each newcomer record is allowed to link only to the catalog or target record with which it achieves the highest probability weight. Thus, it is not possible for the linkage to bring together records in the target file by “bridging” between them—this would involve restructuring or resorting. As we saw earlier, as patient record sets in the main linked database grew larger, the false positive rate crept upwards, often because of illegitimate bridging by new records. As the main production linkages are adapted to one pass linkage, this problem will be minimised. Although the affinity between one pass linkage and the best link principle is one of practical convenience, as we have seen, depending upon the circumstances of the linkage, the best link principle often has highly beneficial effects. Practicality and best practice often go hand in hand.
OCR for page 330
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Linkage in Scotland: A Possible Future Another way of looking at the CHI/NHSCR linkage is to see the NHSCR file as a target file at which the regional CHI files were aimed for linkage. As we have seen finding the best link record in the target file for each CHI record proved to have a dramatic effect on the accuracy of the linkage. The much richer “national CHI” file which has resulted from the linkage and the introduction of national search and enquiry facilities provides an even better target for the linkage of other data sets. For example, in November 1996 Scotland experienced a severe outbreak of infection by the E-coli 0157 bacterium. Several different sets of records were generated in the course of the outbreak: a case register, community clinic contacts, laboratory records, known exposed cohorts and hospital patients. The quality of identifying information on many of these records was rather poor reflecting the circumstances in which they were collected. ISD Scotland was asked to link these records so that the records for each individual involved in the outbreak could be gathered together. Rather than attempt to link the different sets of records directly to each other, the records were “aimed” at the local Community Health Index and linked to it. Again this method paid off in terms of much more accurate linkage. It is likely that more and more linkages in Scotland will take the form of aiming data sets at the target of the national CHI. Ultimately the objective is to use such linkages, whereby for example laboratory data sets or hospital Master Patient Indexes are linked to the national CHI, to populate an increasing proportion of Scotland's health records with a unique patient identifier. It is intended that this will eventually reduce the need to record patient identification details such as names and dates of birth on operational records and communications. Instead identification will be via the national CHI number. Such a system is already in place in Tayside Health Board where the CHI number is implemented on a wide range of primary and acute health care records. In this context the role of probability matching in the Scottish Health Service and the methods used to carry it out are likely to change even more rapidly over the next few years than they have over the last ten years. As we have emphasised it has been the openness of record linkage in the Scottish Health Service to the demands of a wide range of customers which has driven the rapid development in our methods and this is likely to continue. In this context the common sense and pragmatic approach to record linkage championed by Howard Newcombe has been especially useful and appropriate as guidance. Working as we are in his footsteps we can summarise some of the most salient emphases. Record linkage is about being guided by the data and staying as close to the data as possible at all stages. The people who know the data best must be involved. Linkage is an evolutionary and recursive process at all levels. Linkage is a continual learning process and linkage is about what works, not what ought to work. Finally, record linkage is not about the mechanical application of complex and abstract rules. As circumstances change and data sets vary there is unlikely ever to be one definitive best method of carrying out record linkage using probability matching. Progress will come rather from the flexible and responsive application of what are, at heart, very simple principles.
OCR for page 331
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Acknowledgments The following people have contributed over the years to the collective enterprise which is described. In ISD Scotland, James Boyd, Dorothy Gardner, Lena Henderson, Kevin McInneny, Margaret MacLeod, Fiona O'Brien, Chris Povey, Jack Vize, David Walsh, and Bruce Whyte have contributed their insight and programming skills. John Clarke and May Sleigh provided invaluable continuity and guidance. The expertise of Angela Bailey, Eileen Carmichael, Janey Read, and Maggi Reid in checking output has helped keep the system on course. In the Data Centre of the Scottish Health Service's Common Services Agency Gary Donaldson, Alison Jones, Ruth McIlroy, and Debbie McKenzie-Betts have laboured to put the linkage systems on a production basis. References Arellano, M.G. ( 1992). Comment on Newcombe et al. 1(992). Journal of the American Statistical Association, 87, 1204–1206. Fellegi, I.P. and Sunter, A.B., ( 1969). A Theory of Record Linkage, Journal of the American Statistical Association, 40, 1183–1210. Gill, L.E. and Baldwin, J.A. ( 1987). Methods and Technology of Record Linkage: Some Practical Considerations in Textbook of Medical Record Linkage, Baldwin J.A. et al. (eds), Oxford: Oxford University Press. Gillespie, W.J.; Henry, D.A.; O'Connell, D.L.; Kendrick, S.W.; Juszczak, E.; McInneny, K.; and Derby, L. ( 1996). Development of Hematopoietic Cancers after Implantation of Total Joint Replacement, Clinical Orthopaedics and Related Research, 329S, S290–296. Heasman, M.A. ( 1968). The Use of Record Linkage in Long-term Prospective Studies, in Record Linkage in Medicine: Proceedings of the International Symposium, Oxford, July 1967, Oxford: Oxford University Press. Heasman, M.A. and Clarke, J.A. ( 1979). Medical Record Linkage in Scotland, Health Bulletin (Edinburgh), 37:97–103. Hole, D.J.; Clarke, J.A.; Hawthorne, V.M.; and Murdoch, R.M. ( 1981). Cohort Follow-Up Using Computer Linkage with Routinely Collected Data, Journal of Chronic Disease, 34, 291–297. Kendell, R.E.; Rennie, D.; Clarke, J.A.; and Dean, C. ( 1987). The Social and Obstetric Correlates of Psychiatric Admission in the Puerperium., in Textbook of Medical Record Linkage, Baldwin J.A. et al. (eds), Oxford: Oxford University Press. Kendrick, S.W. and Clarke, J.A. ( 1993). The Scottish Medical Record Linkage System., Health Bulletin (Edinburgh), 51, 72–79. Kendrick, S.W. and McIlroy, R. ( 1996). One Pass Linkage: The Rapid Creation of Patient-Based Data, in Proceedings of Healthcare Computing 1996: Current Perspectives in Healthcare Computing 1996, Weybridge, Surrey: British Journal of Healthcare Computing Books.
OCR for page 332
Record Linkage Techniques—1997: Proceedings of an International Workshop and Exposition Kendrick, S.W.; Douglas, M.M.; Gardner, D.; and Hucker, D. ( 1997). The Best-Link Principle in the Probability Matching of Population Data Sets: The Scottish Experience in Linking the Community Health Index to the National Health Service Central Register, Methods of Information in Medicine (in press). Newcombe, H.B. ( 1988). Handbook of Record Linkage, Oxford: Oxford University Press. Newcombe, H.B. ( 1995). Age-Related Bias in Probabilistic Death Searches Due to Neglect of the Prior Likelihoods, Computers and Biomedical Research, 28, 87–99. Newcombe, H.B.; Kennedy, J.M.; Axford, S.J.; and James, A.P. ( 1959). Automatic Linkage of Vital Records, Science, 130, 954–959. Newcombe, H.B.; Smith, M.E.; and Lalonde, P. ( 1986). Computerised Record Linkage in Health Research: An Overview, in Proceedings of the Workshop on Computerised Linkage in Health Research (Ottawa, Ontario, May 21–23, 1986), Howe, G.R. and Spasoff, R.A. (eds), Toronto: University of Toronto Express. Newcombe, H.B.; Fair, M.E.; and Lalonde, P. ( 1992). The Use of Names for Linking Personal Records, Journal of the American Statistical Association, 87, 1193–1204. West of Scotland Coronary Prevention Study Group ( 1995). Computerised Record Linkage Compared with Traditional Patient Follow-up Methods in Clinical Trials and Illustrated in a Prospective Epidemiological Study, Journal of Clinical Epidemiology, 48, 1441–1452. Winkler, W.E. ( 1994). Advanced Methods for Record Linkage, Statistical Research Division, Statistical Research Report Series No. RR94/05, Washington D.C.: U.S. Bureau of the Census.
Representative terms from entire chapter: