Regardless of which method of deterministic linking is used, entry errors, typographical errors, aliases, and other data transmission errors can cause problems. For example, one incorrectly entered digit of a Social Security number will produce a nonmatch between two records for which all other identifying information is the same. Names that are spelled differently across different systems also cause a problem. A first name of James that is recorded in one system as Jim and in the other as James will produce a nonmatch when the two records, in fact, belong to the same individual. The data cleaning in the record linkage process often involves (1) using consistent value-states for the data fields used for linking, (2) parsing variables into components that need to be compared, (3) dealing with typographical errors, and (4) unduplicating each source file for linkage.
Because record linking typically involves data sets from different sources, the importance of standardizing the format and values of each variable is used for linking purposes cannot be overemphasized. The exact coding schemes of all the variables from different source files used in the matching process should be examined to make sure all the data fields have consistent values. For example, males coded as "M" in one file and "1" in another file should be standardized into a same value. In the process, missing and invalid data entries also should be identified and coded accordingly. For example, a birth year 9999 should be recognized as a missing value before the data set is put into the record-linking process. Otherwise, records with a birth year 9999 from the two data sets can be linked because they have the "same" birth year. We also find that standardization of names in the matching process is important because names are often spelled differently or misspelled altogether across agency information systems. For example, a first name of Bob, Rob, and Robert should be standardized into a same first name such as Robert to achieve better record-linking results.
The data cleaning and standardization in matching process often requires parsing variables into a common set of components that can be compared. Names may have to be split or parsed into first name, middle initial, and last name and suffix (e.g., Junior). In using geographic information, street names and the form of the addresses must be standardized. This may mean parsing the address into number (100), street prefix (West), street name (Oak), and street suffix (Boulevard).
Because of typographical errors, an exact character-by-character comparison for certain fields used in a record-linking process may miss many "true" matches. A good example is variant spellings of names. For example, character-by-character comparison of a last name spelled as "Goerge" in one data file to a misspelled name "George" in another file would cause disagreement in the last name comparison even though "George" in the second file was a misspelling. In some situations, these types of typological errors can be a serious problem in record linkage. Winkler and Thibaudeau (1991) and Jaro (1989) describe how researchers at the U.S. Bureau of the Census reported that about 20 percent of last names and 25 percent of first names disagreed character by character among true matches in the Post Enumeration Survey. In recent years researchers in the field of record linkage have made substantial progress in developing algorithms to deal with such problems in character-by-character comparisons. As a result, some complex string comparator algorithms also have been developed to determine how close two strings of letters or numbers are to each other that account for insertions, deletions, and transpositions (Jaro, 1985, 1989; Winkler, 1990; Winkler and Thibaudeau, 1991).
In the record linkage process, one critical data cleaning process is to "unduplicate" each source data set before any two data sets are linked. As discussed earlier, often individuals are associated with several IDs because of data entry errors or a lack of concerted effort to track individuals in agency information systems. Obviously, multiple records for the same individual in each data set being linked produce uncertain links because the process must deal with N to N link situations. Unduplication of the records in a single data set can be thought as "self-match" of the data set. Once a match has been determined, a unique number is assigned to the matched records so that each individual can be uniquely identified. The end result of the unduplication process is a "person file," which contains the unique number assigned during unduplication and the individual's identifying data (name, birth date, race/ethnicity, gender, and county of residence) with a "link file" that links the unique individual ID to all the IDs assigned by an agency. Once each data set is unduplicated in such a way, the unduplicated person files can be used for cross-system record links.
Accuracy of Record Linking
Regardless of which method is used, the ultimate concern is in the degree of validity and accuracy of the links made. Whether it is a deterministic or probabilistic record-linkage technique that is used, the linking process essentially involves making an educated guess about whether two records belong to the same individual. Because the decision is a guess, it might be wrong. These errors in record linkage can be viewed as making false-positive and false-negative errors. A false-positive error occurs when the match is made between the two records when the two records, in fact, do not belong to the same individual. This type of error is comparable to a Type I error in statistical hypothesis testing. A false-negative error occurs when the match is not made between the two records when they, in fact, belong to the same individual. The type of error is comparable to a Type II error in statistical hypothesis testing.
As with Type I and Type II errors, although the probability of making a false-positive error can be easily ascertained in the linking process, determining the probability of a false negative error is more complex. Because the "weights" calculated in the probabilistic record-linkage method are essentially relative measures of the probability of a match, the weights can be converted to an explicit probability that a record pair is a true match (i.e., 1-false positive error rate). Belin and Rubin (1995) introduced a method for estimating error rates for cutoff weight values in the probabilistic record-linkage process. Many developments also have been made in dealing with linkage errors in post-linkage analysis stages (such as a regression analysis using linked files) (see Scheuren and Winkler, 1993). In the case of deterministic record linkage, an audit check on the matched pairs could provide an estimate of false-positive errors. Estimating the false-negative error rate is much more complex because it conceptually requires knowing the true matches prior to the linking and comparing the linking results to the true matches.
Adding to the complexity, as one tries to reduce one type of error, the other type of error increases. For example, in an effort to reduce false-positive errors, one might use a stringent rule of labeling the compared matches as matched pairs only when they are "perfect" matches. In the process, a slight difference in identifying information (such as one character mismatch in the names) might cause a non link when, in fact, the two records belong to the same individual. Hence, false-negative error rates increase. In the opposite scenario, one might accept as many possible matches as true matches, thereby relaxing the comparison rule by reducing false-negative errors. In this case, false-positive errors increase.
In practice, it would be useful to consider false-positive and false-negative error rates as a means to compare different methods of record linkage. One practical issue researchers face is determining which linkage method to use, especially when an ID variable such as SSN is available in the two data sets to be linked. Although most experts agree that probabilistic record linkage is a more reliable method than deterministic linking, it requires extensive programming or the purchase of software, which can be quite expensive. If one does not have ready access to suitable commercial record-linkage software, it may be sufficient for a good programmer to write a quick deterministic linkage program that matches a good deal of the records. There are other situations where there is no apparent common ID and the quality of identifying information in the data is questionable (such as many typographical errors in certain data fields), so that only using probabilistic record-linkage methods will yield acceptable linking results.
We present some empirical data comparing the two methods in the following paragraphs and corresponding tables. The methods compared are a deterministic record link using SSN and a probabilistic link using SSN, full name, birth date, race/ethnicity, and county of residence. We use data from the Client Database and the Cornerstone Database from the Illinois Department of Human Services. The Client Database records receipt of AFDC/TANF and Food Stamps and documents all those who are registered as eligible for Medicaid from 1989 to the present. The Cornerstone database contains WIC and case management service receipt at the individual level. There is no common ID between the two systems, while SSN and other identifying information are available in both systems.
Because both systems serve mainly low-income populations and contain data for a long period of time, we expected a high degree of overlap between the two populations. When the existence of SSN in both systems is examined, we find that about 38 percent of the Cornerstone records have missing SSNs while the Client Database identifies nearly 100 percent of the SSNs. In our first analysis, we excluded the records with missing SSNs from the Cornerstone data. Table 7-1 compares the number of matched and unmatched Cornerstone data records to the Client Database records comparing the deterministic match using SSN and the probabilistic match using all other identifying information, including SSN. As shown in Table 7-1, the probabilistic match identified about 86 percent of non-SSN-missing Cornerstone record links to the Client Database. The SSN deterministic method identified about 84 percent of the matches.
|Probabilistic Matching Number||Probabilistic Matching Percent|
Although the percentage of overall matches is similar, the distribution of error types is quite different, as shown in Table 7-1. The false-negative error rate of using the SSN deterministic record-linking method when compared to the results from the probabilistic match is about 38 percent. On the other hand, the false positive error rate is about 1 percent. We checked the results of the probabilistic link from random samples of the disagreement cells (i.e., probabilistic match/SSN no match and probabilistic nonmatch/SSN match) to verify the validity of the probabilistic match. We found that the probabilistic match results are very reliable. For example, we found that most of the pairs in the probabilistic match/SSN no match cell involve typographical errors in SSN with the same full name and birth date. Also, we found that most of the pairs in the probabilistic nonmatch/SSN match involve entirely different names or birth dates. Although the findings might be somewhat different when applied to different data systems, our finding suggests that employing a probabilistic record-linkage method helps to reduce both false-negative and false-positive errors. The findings also show that the benefit of employing probabilistic record linkage is greater in reducing false-negative errors (Type II errors) than in reducing false-positive errors (Type I errors) when compared with a deterministic record-linkage method using SSN.
Next, we included the Cornerstone records with missing SSN in the analysis. The findings are presented in Table 7-2. As one might expect, the probabilistic record-linkage method significantly enhances the results of the match by linking many more records. Compared with the results presented in Table 7-1, the number of matches from the probabilistic match increases by about 210,000 records, representing about 62 percent of matches made among the records with missing SSNs. Again, most of the benefit of using the probabilistic linkage method is in reducing false-negative errors. With about 30 percent of the records showing missing SSNs, the false-negative error rate of the SSN deterministic link method is about 57 percent. From the above results, one can conclude that when SSN information is nearly complete in the two data sets, the added benefit of using probabilistic linking is relatively smaller (although quite significant) and the benefit comes largely from identifying false-negative errors. As the number of records with missing SSN increases, the benefit of employing a probabilistic record-linkage method increases.
|Probabilistic Matching Number||Probabilistic Matching Percent|
Very often in practice, being able to link different data sources involves many other issues than that of the linking method. A key issue is data confidentiality, especially when full names are needed for linking purposes in the absence of a common ID. One possible solution to the confidentiality issue is the use of Soundex codes. Even though Soundex is not a complete method to preserve confidentiality, it provides added protection compared to using actual full names. The Soundex system is a method of indexing names by eliminating some letters and substituting numbers for other letters based on a code. Although experts disagree on what should be the authoritative Soundex system, the most familiar use of Soundex is by the U.S. Bureau of the Census, which uses it to create an index for individuals listed in the Census. Because it is impossible to derive an exact name from a Soundex name, the system can be used to conceal the identity of an individual to some extent. (For example, similar sounding but different names are coded to a same Soundex name.)
The issue in probabilistic linking, however, is how valid a Soundex name is alone compared to using full names. We examine this issue by comparing the two methods involving the same data sets with the other identifying information fixed. The other identifying information variables are SSN, birth date, race/ethnicity, and county of residence. Table 7-3 presents the results of such an exercise. The agreement rate between the Soundex-only method and the full-name method is very high--close to 100 percent. The results suggest that Soundex coded names work equally as well as full names in a probabilistic match. In situations in which full names cannot be accessed for linking purposes, Soundex names might be a good alternative while providing a better means of protecting individual identities.(4)
|Full Name Matching Number||Full Name Matching Percent|
"01.pdf" (pdf, 472.92Kb)
"02.pdf" (pdf, 395.41Kb)
"03.pdf" (pdf, 379.04Kb)
"04.pdf" (pdf, 381.73Kb)
"05.pdf" (pdf, 393.7Kb)
"06.pdf" (pdf, 415.3Kb)
"07.pdf" (pdf, 375.49Kb)
"08.pdf" (pdf, 475.21Kb)
"09.pdf" (pdf, 425.17Kb)
"10.pdf" (pdf, 424.33Kb)
"11.pdf" (pdf, 392.39Kb)
"12.pdf" (pdf, 386.39Kb)
"13.pdf" (pdf, 449.86Kb)