Age Validation in the Long Life Family Study Through a Linkage to Early-Life Census Records

Abstract

Objectives.

Studies of health and longevity require accurate age reporting. Age misreporting among older adults in the United States is common.

Methods.

Participants in the Long Life Family Study (LLFS) were matched to early-life census records. Age recorded in the census was used to evaluate age reporting in the LLFS. The study population was 99% non-Hispanic white.

Results.

About 88% of the participants were matched to 1910, 1920, or 1930 U.S. censuses. Match success depended on the participant’s education, place of birth, and the number of censuses available to be searched. Age at the time of the interview based on the reported date of birth and early-life census age were consistent for about 89% of the participants, and age consistency within 1 year was found for about 99% of the participants.

Discussion.

It is possible to match a high fraction of older study participants to their early-life census records when detailed information is available on participants’ family of origin. Such record linkage can provide an important source of information for evaluating age reporting among the oldest old participants. Our results are consistent with recent studies suggesting that age reporting among older whites in the United States appears to be quite good.

Accurate age reporting among the oldest old study participants is needed for demographic, epidemiologic, and genetic studies of factors that influence extreme longevity. Typical patterns of age misstatements bias mortality estimates downward and can adversely influence estimation of health trajectories at the oldest ages (Preston, Elo, & Stewart, 1999).

Numerous sources of evidence suggest that age reporting among the oldest old people in the United States is a problem (Kannisto, 1994; Rosenwaike & Stone, 2003). For example, estimates of age-specific death rates based on vital statistics and census data at the oldest ages have been suspect due to inconsistencies in the ages recorded on death certificates (numerators) and population estimates (denominators) used to calculate these rates (Dupre, Gu, & Vaupel, 2012; Himes, 2002; Preston, Elo, Rosenwaike, & Hill, 1996). One reason is that many older individuals do not have a birth certificate because the birth registration area in the United States was not established until 1915, and it did not cover all states until 1933 (Shapiro, 1950).

Because of possible age misreporting among the oldest old group, it is recommended that steps are taken to verify ages of the oldest old people in studies of health and mortality (Cawthon et al., 2002; Perls et al., 2007). Therefore, in the Long Life Family Study (LLFS), an international study of familial clustering for exceptional longevity for which recruitment began in August 2006, an effort was made to verify ages for U.S. study participants born before 1930. In this article, we describe the procedures taken to verify an individual’s age at the time of the survey and evaluate the quality of age reporting among the oldest U.S. participants in the LLFS. These steps consisted of collecting information about the source document used to verify the date of birth reported by the respondent and of matching study participants to their early-life census records when they were children or young adults. This record linkage enhanced the LLFS data by providing information on early-life conditions as well as information to evaluate consistency of age reporting (Preston, Hill, & Drevenstedt, 1998; Rosenwaike, Hill, Preston, & Elo, 1998).

In the absence of a birth certificate, which is considered the gold standard for verification of an individual’s date of birth, an individual’s early-life census record has been used as a proof of age in several prior studies (Preston, Elo, Rosenwaike, & Hill 1996; Robine & Vaupel, 2001; Rosenwaike & Stone, 2003), and it has been accepted by the Social Security Administration as a proof of age in the absence of a birth certificate when applying for social security benefits (Social Security Administration, 1988). Although not foolproof, evidence suggests that early-life census record can serve as a good, if not a perfect, proxy to verify age for older adults in the United States. For example, in a study of U.S. supercentenarians, the authors documented that out of 43 cases for whom both a birth certificate and an early-life census record was located, the birth certificate confirmed supercentenarian status for 40 individuals, and in the three cases of age discrepancy, the calculated age was only 1 year younger than 110 (Rosenwaike & Stone, 2003). In a study to investigate the quality of age reporting among African Americans aged more than 65, the authors found that 77% of the study participants who were born in Maryland, were matched to their 1920 census record, and for whom a birth certificate was located, had their ages correctly reported in the 1920 census. For an additional 18.2%, the age discrepancy was within 1 year such that for 95.2% of the participants, the census age was within 1 year of their true age. When ages disagreed, it was more common for the age to be overstated than understated in the census record (Preston et al., 1996; Rosenwaike & Hill, 1996). We were also able to evaluate the accuracy of the early-life census age among LLFS study participants whose reported date of birth was verified by a birth certificate (see Discussion section).

Methods

Participants

The participants in this analysis consist of the oldest U.S. participants in the LLFS (http://longlifefamilystudy.wustl.edu/), a unique international study of factors that contribute to familial clustering of exceptional longevity in the United States and Denmark. In the United States, eligible families were identified from mailings of study information to Medicare enrollees and through media events and newspaper articles. Potentially eligible individuals were encouraged to contact three U.S. field centers located in Boston (MA), New York (NY), and Pittsburgh (PA) (Yashin et al., 2010). To be eligible for participation, the family had to demonstrate exceptional longevity by scoring high on the Family Longevity Selection Score, a summary measure based on the survival experience of the sibship of the family’s oldest living individual compared with what would be expected based on U.S. birth cohort life tables (Sebastiani et al., 2009). In addition, at least two members of the oldest generation had to be alive. In comparison to families enrolled in the Framingham Heart Study, fewer than 1% of the families in that study would qualify for participation in the LLFS (Sebastiani et al., 2009). Enrollment was conducted between 2006 and 2009 (Newman et al., 2011).

Upon enrollment in the study, field center staff collected information from study participants using standardized questionnaires and in-home physical and cognitive functional status examinations. One of the questionnaires was designed to obtain sociodemographic information, including age, date of birth, a source document used to verify date of birth, place of birth, place of residence before age 16, and date of or age at immigration to the United States, if foreign born. In addition, at the time of enrollment, information was obtained on the names of parents and siblings as well as their dates of birth. This information provided the key data for matching the participants to their early-life census records. In this article, we used data on 443 probands (the first enrolled participant in the family), and their 699 siblings who had completed the sociodemographic questionnaire (N = 1,142). We excluded 10 probands because they immigrated to the United States in or after 1930, the year of the last publicly available census, and 50 siblings who immigrated to the United States in or after 1930 or whose reported date of birth was after April 1, 1930, the 1930 census date. Because of potential misreporting of the date of birth, we also tried to match 25 siblings who were born after April 1, 1930 but before December 31, 1932; one sibling who immigrated in 1930 and one proband who immigrated in 1931 to the 1930 census without success. The study population was 99% non-Hispanic white.

Match to Early Census Records

We followed previously established protocols in linking study participants to publicly available 1910, 1920, and/or 1930 census records (Preston et al., 1996; Rosenwaike et al., 1998). This linkage was enhanced by the rich information on the family collected from study participants, including names of parents and siblings. Our ability to identify the correct participant strongly depended on the following information: the participant’s name, sex, state of birth, and state of residence before age 16, immigration status, father’s first and last name, mother’s first and maiden name, and the names of the participant’s siblings. In particular, the information on the names of parents and siblings improved our ability to locate the correct family, especially when there were many families in the state with similar last names.

Project staff conducted all searches on publicly available census manuscript records through Ancestry.com (www.ancestry.com). Staff members were provided with search sheets that contained information useful for the match: the participant’s sex, date of birth, state of birth, state of residence before age 16, state of current residence, full name at the time of the study, and full names of parents and first names of siblings and their dates of birth. Because one of the purposes of matching participants to early-life census records was to examine accuracy of age reporting, we did not use date of birth in the match except to identify the census closest to the time when the participant would have been 10 years old if the reported date of birth on the sociodemographic questionnaire was accurate. This census was the first to be searched and allowed both age understatement and overstatement relative to the estimated age based on the reported date of birth. In this census, we began the search in the state where the participant was born, if born in the United States, followed by the state where the participant lived prior to age 16, and then the current state of residence if these differed from one another. Among immigrants, we began the search in the state where the participant lived prior to age 16 followed by the state of current residence. We also searched in more than one census when the participant was not located in the first census searched.

The staff member conducting the search first assessed the match quality. “Very confident” matches were those families and individuals where the name(s) of the parent(s) and the participant matched and which included at least one sibling who should have been born prior to the census date. In some cases, a participant was matched to more than one census. In such cases, we used the earlier census record to assess age reporting consistency. We emphasized the need to be conservative in choosing a match and a second staff member reviewed all matches that were assessed by a staff member to be less than “very confident” and all disagreements were adjudicated in a staff meeting. We accepted only those matches that were deemed to be “very confident” after full staff review. Unacceptable matches consisted mostly of cases with common names.

Predictors of Successful Census Match and Age Reporting Consistency

We began by examining whether characteristics of the LLFS participants, available from the sociodemographic questionnaire, predicted match success. These attributes included the respondent’s age at the time of the survey, sex, educational attainment, whether foreign born and state of birth, if born in the United States. Educational attainment has been associated with accurate age reporting in prior studies (Elo et al., 1996), and we hypothesized that it would also be related to the accuracy of the information provided by the respondent and thus influence match success. In addition, we speculated that we would be more successful in finding the participant if she/he could be searched in more than one census. Multiple census searches were possible for the older participants in the sample whose reported birth dates were around or before 1920. Finally, we examined the extent to which information was missing on parents’ names, a key piece of information for locating the family in the early censuses. We had complete information on parents’ names, except for mother’s maiden name, for about 98% of the participants. Mother’s maiden name was missing for 19% of the sample, but preliminary analysis showed that it did not influence match success (results not shown). Thus, we do not include this variable in subsequent analyses.

We then examined the consistency of age reporting among participants linked to early census records. We first compared the participant’s age at the time of the interview calculated from the interview date and the participant’s reported birth date and his/her expected age based on the respondent’s age on his/her last birthday recorded in the early-life census record, the date of census enumeration, and the interview date. Age consistency, age at last birthday, occurred when the two ages agreed. In models predicting age agreement, our predictor variables included whether the participant was a proband or a sibling, his/her age at the time of the interview, sex, educational attainment, whether foreign born, and whether a birth certificate was used to verify date of birth.

In the logistic regression models predicting match success, the dependent variable was coded 1 if the participant was matched to an early-life census record and 0 otherwise. Similarly in the analysis of age agreement, the dependent variable was coded 1 if the two ages in years at the time of the interview agreed and 0 otherwise. We used Wald test to assess the significance of individual coefficients. All models were estimated in STATA 11 and standard errors were adjusted for clustering within families based on the Huber White sandwich estimator.

Results

Table 1 provides characteristics for the sample used in this analysis. The mean age of the respondents at the time of the interview was 90.7 years, over half of the participants were more than 90 years old and close to a quarter were aged 95 and older. Close to a fifth of the respondents had not completed high school or passed a General Educational Development Test (GED), whereas close to 30% had at least a college degree. Close to three quarters of the sample was born in the New England, New Jersey, New York, or Pennsylvania, reflecting the location of the U.S. field centers. Less than 5% of the sample was foreign born. Because of the high mean age of the respondents, we were able to search the vast majority of the respondents (71.9%) in more than one census.

Table 1.

Sample Characteristics of Long Life Family Study (LLFS) Study Participants (N = 1,142)

Characteristics	N	% or M
Mean age at interview (SD)	1,142	90.7 (5.6)a
Age at interview (age group)
<85	172	15.1
85–89	238	20.8
90–94	450	39.4
95 and older	282	24.7
Sex
Male	557	48.8
Female	585	51.2
Educational attainment
Less than high school	215	18.8
High school graduate/had GED	291	25.5
Some college/technical training	303	26.5
4-year college degree or more	329	28.8
Missing	4	0.4
Region of birth
New England	303	26.5
New Jersey, New York, or Pennsylvania	537	47.0
Other U.S. states	250	21.9
Foreign born	51	4.5
Missing	1	0.1
Number of possible census searches
One	321	28.1
Two or three	821	71.9
Matched to 1910, 1920, or 1930 census	1,012	88.6

Note: ^a SD in parentheses.

Table 2 provides information on the source document used to verify the date of birth of the study participants. In the sociodemographic questionnaire, the respondent was asked to verify his/her date of birth and the source document used to do so. As seen in Table 2, only about 21% of the study participants used a birth certificate to verify the date of birth. An additional 9.6% used a passport and a small number used a military record (1.1%) or a census record (0.5%). The most common form of age verification was a driver’s license or another form of state issued ID (58.4%); for 7% of the respondents, no official source was used or the information was missing. Thus, a large percentage of the documents used to verify date of birth for most individuals cannot be considered to be highly reliable for this purpose.

Table 2.

Source Document Used to Verify Date of Birth Among Respondents in the Study Sample (N = 1,142)

Source document	N	%
Birth certificate	243	21.3
Church record/family bible	23	2.0
Driver’s license	667	58.4
Military record	13	1.1
Census record	6	0.5
Passport	110	9.6
No official source/other/missing	80	7.0
Total	1,142	100.0a

Note: ^aPercent may not add up to 100 due to rounding.

Match to Early-Life Census Records

We were able to match 88.9% of the sample to their early-life census records. The mean number of matched siblings for a matched proband was 1.5 (SD 1.0). At the time of the census enumeration, 82% of the matched cases were less than 15 years old, about 16% were between 15–19 years, and about 2% were aged 20–24 (results not shown).

Table 3 presents results of bivariate and multivariate logistic regression models predicting match success. We were significantly more likely to match those with higher levels of education, U.S.-born study participants, and those who could be searched in more than one census, that is, those who were born before the 1920 census. Match success did not depend on whether the participant was a proband or a sibling, male or female, or U.S. region of birth.

Table 3.

Odds Ratios (ORs) from Bivariate and Multivariate Logistic Regression Predicting a Successful Childhood Census Match (N = 1,137)^a

	Bivariate OR	Multivariate OR
Proband (sibling)	1.03 [0.80, 1.34]	0.84 [0.63, 1.12]
Age at interview	1.02 [0.98, 1.07]	0.98 [0.92, 1.04]
Sex (male)	0.94 [0.64, 1.38]	1.10 [0.73, 1.66]
Educational attainment (<high school)
High school graduate/GED	1.63 [0.93, 2.86]	1.52 [0.87, 2.65]
Some college/technical training	2.24 [1.22, 4.12]	2.17 [1.20, 3.95]
4-year college degree or above	3.31 [1.68, 6.49]	3.05 [1.55, 5.99]
Region of birth (NJ, NY, or PA)
New England	1.19 [0.61, 2.32]	1.24 [0.63, 2.44]
Other U.S. states	2.28 [1.01, 5.13]	2.26 [0.98, 5.22]
Foreign born	0.20 [0.08, 0.47]	0.19 [0.08, 0.45]
Number of possible censuses (one)
Two or more	1.86 [1.20, 2.88]	3.18 [1.58, 6.40]
Wald χ2 (10)		53.59

Notes: ^aFour cases with missing educational attainment and one case with missing region at birth were dropped from this analysis. The value in parentheses indicates the reference group. The 95% CI is given in parentheses.

Age Agreement

Table 4 presents results of age agreement calculated as the participant’s age in years at the time of the survey obtained from the reported age in the census record and time between the census enumeration and the date of the interview minus the participant’s age calculated from the reported date of birth and the date of the interview. A positive difference between the two ages indicates that the age at the time of the interview calculated from the reported date of birth and the date of the interview is understated, whereas a negative value indicates age overstatement. Overall, for 88.7% of the cases, the two ages agreed. In general, age understatement was more common than age overstatement. There was no clear pattern in age overstatement or age understatement by 5-year age group at the time of the interview. Furthermore, age inconsistencies were small with less than 1% of the cases having inconsistent ages of greater than 1 year. Age agreement was significantly more likely for probands and those with higher levels of education in multivariate models that also controlled for place of birth, age, nativity, and whether a birth certificate was used to verify date of birth. Although age inconsistency was also lower for the foreign born, the difference was not statistically significant (results not shown).

Table 4.

Percentage Distribution of Discrepancies Between Age Based on Census Record and Age at Interview,^a LLFS Siblings, and Probands (N = 1,005)

	Census-based age − Age based on reported birth date
Age at interviewa	−2+	−1	0	1	2+	N
<80	0.0	0.0	89.3	7.1	3.6	28
80–84	0.8	1.7	92.4	5.1	0.0	118
85–89	0.5	1.0	88.4	10.1	0.5	207
90–94	0.0	2.4	89.5	7.8	0.2	411
95–99	1.5	4.4	85.4	8.8	0.0	205
100 and more	0.0	0.0	88.4	7.0	4.7	43
Total	0.4	2.3	88.7	8.1	0.5	1,012

Notes: ^aAge at interview based on reported date of birth and date of interview. Percentages may not add up to 100 due to rounding. A positive difference indicates age understatement at the time of the interview based on the reported birth date compared with census-based age, whereas negative values indicate age overstatement.

Discussion

We have demonstrated that it is possible to match a high fraction of elderly study participants to their early-life census records when detailed information is available on participants’ family of origin. The match percentage attained for LLFS study participants is high compared with prior studies that have attempted to link death certificates of older decedents to the censuses of 1900, 1910, and/or 1920. For example, Rosenwaike and Logue (1983) linked 42% of 1968–1972 death certificates of white and non-white decedents aged 85 and older from Pennsylvania and New Jersey to the 1900U.S. census manuscript records, whereas Preston and colleagues (1996) matched 56.8% of a national sample of African American death certificates from 1985 for decedents aged 65 and older to 1900, 1910, or 1920 censuses. In a subsequent study of age misreporting on death certificates of elderly white Americans, Hill and colleagues (2000) successfully matched 69.3% of 984 death certificates of native-born white Americans ages 85 and older 1900, 1910, or 1920 census records. A match percentage more similar to ours was achieved in a study of early-life predictors of the development of Alzheimer’s disease, in which 82% of persons with Alzheimer’s disease and 87% of the controls were matched to an early-life census record (Moceri et al., 2001).

We believe that our high match percentage resulted from having information about the family of origin from the participant, himself or herself, including information on the full names of parents and siblings, as well as information on the participant’s place of birth and place of residence prior to age 16. Furthermore, census enumeration is likely to have been more complete for whites than for African Americans who were the participants in the Preston and colleagues (1996) study and who made up a large fraction of the Rosenwaike and Logue (1983) study population.

We further showed that age reporting consistency was high among the study participants. We found that the ages at interview based on the reported date of birth and the census age agreed for close to 90% of the participants. When ages disagreed, age understatement was more common than age overstatement, and these disagreements were small such that for about 99% of the participants the two ages were within 1 year of each other.

Among our explanatory variables, educational attainment was significantly associated with both match success and age agreement with the more highly educated participants being more likely to be matched to their early-life census record and to have age agreement. In addition, the ability to search the participants in more than one census improved match success. Perhaps, not surprisingly, we were less successful in matching the foreign-born than the U.S.-born study participants.

We should also note, however, some limitations of our study. First, age consistency does not guarantee that the date of birth is correctly reported in the LLFS. The birth certificate is considered the gold standard for the verification of an individual’s date of birth and birth certificate was used to verify the date of birth for less than a quarter of the study participants. Among the LLFS participants for whom the date of birth was verified by a birth certificate, the census age agreed with the calculated age for 91% of the cases. When ages disagreed, the discrepancies were within 1 year for an additional 8.5% of the cases. Age overstatement was more common than age understatement in the census record, a pattern similar to that found in other studies (Preston et al., 1996; Rosenwaike & Hill, 1996; Rosenwaike & Stone 2003), reflecting a tendency to round up the age of the child rather than reporting age as of last birthday.

Second, our relatively small sample size above age 100 is of some concern, and we should be careful not to generalize from our findings to the very oldest ages. Third, we were unable to match 11% of the study participants to an early-life census record. These individuals were significantly more likely to be foreign born and have lower levels of education than those who were successfully linked. Because higher levels of schooling predicted greater age consistency, it is possible that the quality of age reporting was lower among those whom we were unable to link to their early-life census records. Fourth, the findings cannot be generalized to race–ethnic groups other than whites. Age reporting among black and Hispanic elderly people is known to be worse than among whites in the United States (Hill et al., 2000; Kestenbaum, 1992; Preston et al., 1996). Finally, another aspect of the LLFS that could influence age reporting is that one of the recruitment criteria into the study was based on the age of the proband and thus the emphasis on age may have improved age reporting in this study, at least among the first person recruited for participation. That age agreement was somewhat more likely for probands than for siblings points to this possibility.

In this study, we have demonstrated that it is feasible to link a high fraction of older individuals to their early-life census records when detailed information on their families of origin is obtained at the time of recruitment. Such record linkage can provide important information to evaluate age reporting among the oldest study participants. We also demonstrated that age reporting in the LLFS appears to be of high quality. This result is likely related to the fact that our study population was 99% non-Hispanic white (Hill et al., 2000; Shrestha and Preston, 1995). It is recommended that, when feasible, detailed familial information is collected in studies of the oldest old participants that permit linkage to other data sources, such as early-life census records, that can be used to evaluate the quality of age reporting (Jeune & Vupel, 1999).

Funding

Sponsored by the National Institute on Aging (NIA cooperative agreements: (U01-AG023755: I. T. Elo, L. Mykyta, P. Sebastiani, and T. Perls; U01-AG023712: K. Christensen; U01-AG23744: N. W. Glynn; U01-AG023746 and U01-AG023749).