Comparison of reproductive history gathered by interview and by vital records linkage after 40 years of follow-up: Bogalusa Babies

With the growing emphasis on use of existing data and cohorts [1, 2], as well as data harmonization to create large analyses across disparate datasets [3‐5], it becomes more important to understand the degree to which these study designs provide accurate, reliable, and consistent data. While linkage of existing datasets and databases can be powerful and cost-effective, it can also magnify errors [6]. If multiple recordings of data fundamentally derive from the same source, or if linking tends to bias systematically the group of participants that are included in large-scale analyses, such study designs run the risk of leading to a greater degree of confidence in fundamentally flawed or biased analyses.

For example, migration limits the possibility of linkage between datasets. Most data are stored as part of a study database or as clinical or administrative data, and are limited by jurisdiction. Thus, any factor that affects the likelihood of migration affects the probability of linkage across databases. Socioeconomic status (SES) is likely to be particularly important, as it affects mobility, health, and quality of reporting, and can lead to serious bias in the conclusions of studies based on these datasets. Previous studies of mortality linkages have found reduced linkage with Hispanic populations, for instance [7, 8].

In addition to the general issue of the quality and value of data linkages, a question that has recently become more prominent is that of the relationship between reproductive history and health during other parts of the life course. There is a growing recognition that pregnancy does not operate independently of health during other periods of life [9‐13]. While it has long been known that parity and age at pregnancy are risk factors for breast cancer [14], more recent research indicates a relationship between pregnancy complications and birth outcomes and later health, particularly cardiovascular disease and diabetes [15‐17]. Studies of chronic disease, which are usually conducted in middle-aged or older populations, are therefore likely to be interested in finding data on the reproductive years. While several previous studies have looked at the comparison between self-report and other sources of data for studying pregnancy health [18‐20], in most cases these were pregnancy or child cohorts, so the timing and usually location of the births was known precisely, and most often compared were medical records, rather than vital statistics data. In this analysis, we compare the results of a linkage with vital statistics data with women’s self-report of their pregnancy history in the context of a study designed to assess cardiovascular health, and in which the timing of pregnancies was not known and took place over a period of forty years.

There were 1026 women with data from both vital records and interview, with a total of 2658 births reported (Fig. 1). Of these, 1624 were exact matches. An additional 113 matched on year only. 32 of these had year only provided from vital statistics due to confidentiality restrictions (Texas). Of the remaining 81, the median difference in time between the birth certificate and interview data was 2.0 days, with a mode of 10 days, a minimum of − 300 days, and a maximum of 228 days (date from birth certificate – date from interview).

Of the remaining 958 births reported in the interview, 65 occurred prior to 1982 and 105 after 2010, and 51 births were reported to occur outside of Louisiana, Mississippi, and Texas, and thus would not have been eligible to be linked in the linkage. 734 births to 465 women (62 women with non-matching information in both sources, 38 only vital statistics data, and 365 only interview data) were not included in both sources, but had no obvious reason for a lack of match in the other. Of these, 16 births were exactly one year or one year and 1–2 days apart.

Overall, 3944 (66.7%) of participants were located by one or both sources. The strongest predictor of either linkage or interview was recent and/or frequent participation in the parent study (Table 1). Those who were interviewed had more study visits (median 5) than those who did not (median 2, p < 0.01), and were more likely to have participated in the study as an adult. The groups that were interviewed were also more likely to have ever smoked, even after the age distribution and years of the interviews were controlled for (aOR for smoking for those with interview and linkage, 1.32, 1.05–1.66; with interview only 1.45, 1.13–1.86). Parental education was more likely to be missing for those who were not located (this data was not collected at early visits); among those with data, those who were located were more likely to have higher parental education. Differences in BMI, cholesterol, and blood pressure were largely explained by the age distribution of participation in the groups, although mean childhood BMI was higher in those who only interviewed (absolute values provided in table; adjusted beta for difference = 0.80, p < 0.01).

When the matched pregnancies were compared, agreement between the two sources was generally quite good, with kappa statistics > 0.9 for birthweight and 0.8 for gestational age (Table 2). Mean and median differences were close to 0. 128 births (7.5%) were reported as LBW and 1523 (88.8%) as not LBW by both sources; 47 (3.0%) were reported as LBW by the interview and not the birth certificate, while 18 (1.1%) were reported as LBW by the birth certificate but not the interview. 106 births (6.8%) were reported as PTB and 1340 (96.1%) as not PTB by both sources; 54 (3.9%) were reported as PTB by the interview and not the birth certificate, while 49 (3.2%) were reported as PTB by the birth certificate but not the interview.

Few consistent predictors of discrepancy in reporting could be identified (Table 3, Additional file 1: Table S1). Black race was associated with an increased likelihood of discrepancy. First births had a higher likelihood of disparity in LBW and greater discrepancy in gestational age. Those with lower education were more likely to have a discrepancy in reporting LBW (though not birthweight) and in gestational age (though not PTB). Older age was generally associated with greater difference in gestational age, as was time since pregnancy. Cardiovascular risk factors did not show a consistent pattern of being associated with discrepancies in reporting, though occasionally there were statistically significant associations (childhood BMI and blood pressure for birthweight, adolescent cholesterol for PTB).

This analysis serves as background in assessing the likely degree of bias for the overall Bogalusa Babies study, which aims to determine the relationship between cardiovascular risk factors and pregnancy outcomes. Overall, there are two questions to be answered: when considering information about reproductive history in a long-term study with no original goal of assessing reproductive outcomes, does linkage to vital statistics or interview find more participants or more representative participants; and when both data sources are available, how do they compare? These questions are relevant not only to our own study, but to other studies who may be interested in studying the relationship of pregnancy outcomes with chronic disease, and those determining the best way to capture such information.

Generally, we found that consistent participation in the study was the best predictor of being located, via linkage, interview, or both. Black women were also more likely to be linked or interviewed, which differs from other analyses of loss to follow-up [28]. Previous studies of linkage to vital statistics indicate lower linkage of those living in deprived areas and rural areas [29], and that therefore, such studies may suffer from a bias in estimating social gradients of health. Studies also indicate increased attrition with lower SES [30]. To some extent, we found a small tendency for lower education to be associated with loss to follow-up, although in this case, those who seek higher education are likely to move from the area (a relatively small town with no university in the parish), at least temporarily. Other studies have also found that more frequent or more intense involvement in the study reduces attrition [31, 32]. Generally, clinical trials and longitudinal studies find those at increased medical risk, advanced-age, and young adult participants are more likely to drop out [30, 32, 33]. Smokers are also more likely to be lost to follow-up [28, 30, 33], which, again, was not the case in our study, although this is probably partly due to the fact that those lost at a young age might not have begun smoking at the time they participated in the study.

The major question of concern is whether use of one or both sources is likely to lead to biased estimation of the relationships. Overall, two-thirds of all participants were located by one or both sources. While 33% loss to follow-up is easily sufficient to bias an analysis, the sample size that remains is adequate for many research questions, so the concern is whether this sample is representative of the larger study. The analysis is generally reassuring on that point, as cardiovascular risk factors usually did not vary between those linked and those not, or those interviewed and those not. There was not a consistent profile indicating that those with worse or better health were systematically excluded, nor of exclusion of those with low or high socioeconomic status.

Agreement between sources for those included was generally quite good, although there was some indication that black race might have been associated with larger discrepancies in reporting, as well as time since the pregnancy. Several reasons for discrepancies can be imagined. They include 1. Poor memory; 2. Misassigning outcomes (i.e., mixing up birthweights of siblings); 3. Misunderstanding or lack of communication around medical issues (e.g., change in due date based on ultrasound not being communicated to or understood by a woman); 4. Approximation, particularly for full-term gestational ages and pregnancies occurring before the routine use of ultrasound; 5. Not regarding gestational age at birth as worth keeping track of, particularly for earlier births that were not ultrasound-dated and went to full term; 6. Data issues: incorrect linkage or data entry, although studies comparing medical records to vital statistics find that vital statistics data are accurate for birthweight and gestational age [34, 35]. Many of these factors are likely to be correlated with education and the effort and respect accorded a woman by medical providers, all of which are more likely to be provided to white women than black women. Black women also tended to have children earlier and thus had a longer time since pregnancy, although this did not fully explain the difference.

Overall, results are generally reassuring as to possible bias; the limited variation by cardiovascular predictors and the good quality of agreement about birth outcomes suggests that loss to follow-up or missed linkage is not likely to produce major bias for studies of those topics.

Our results are generally consistent with previous studies indicating that mothers remember the birthweight and gestational age of their infants quite well, even after many years [36, 37]. A few facts about self-report should be considered. In the U.S., women generally report birthweight in pounds and ounces, while vital statistics data are in grams; however, the conversion did not produce major issues. Perhaps more serious is that women often remember their babies’ gestational age in terms of weeks while medical records and vital statistics report in days; although we allowed for reporting in both weeks and days, most women reported only in weeks. We also began the interview asking whether the baby was early, late, or on time, and women generally reported the baby was on time if s/he was born within the week expected. The more precise recording in medical records and vital statistics is better for studies that treat gestational age as a continuous variable. Finally, many of the earlier births in this study occurred prior to routine ultrasound dating, so women may have had less exact dating available to them.

The question then arises as to whether these results apply to other studies. Some aspects of the study are unusual, though possibly relevant to other studies. Participants did not initially agree to be in a long-term study; particularly, the original waves of data collection were collected as cross-sectional studies rather than a planned longitudinal analysis. Therefore, the loss of participants who participated once, many years ago, as children, is not particularly surprising. This analysis also assesses only women, who are generally more likely to continue participation in studies [28, 38] but also more likely to change their last names. Any analysis addressing pregnancy will have this population. The geographic basis for the study also affects the follow-up; in this semirural area, higher-SES individuals are more likely to leave the area, which affects their loss to follow-up, not necessarily the case for more extensive studies or other types of areas.

Combining information from multiple sources to increase sample size and outcome ascertainment may be valid. We have demonstrated support for use of data harmonization across sources as a feasible and valid way to create analytic epidemiologic cohorts. Studies will generally consider consistently-collected data such as vital records as the preferred source, but can be augmented with maternal self-report for these outcomes. This is good news for population health sciences’ ability to leverage the many existing, large-scale sources of data on health and health determinants for research that expands their scope further by answering previously unexplored questions, even those that the data were not initially collected to answer.