The University of California Study of Outcomes in Mothers and Infants (a Population-Based Research Resource): Retrospective Cohort Study

Introduction

Background

The use of population-based studies for public health surveillance and research has proved highly valuable for perinatal and pediatric epidemiology [1]. These studies frequently use data from birth certificate records, which provide a census of all births occurring within a specific region. Linking birth certificate records to external databases (eg, hospital discharge, environmental pollutant, and census data) can improve the value of population-based studies because it allows researchers to study a wider set of exposures and outcomes that are not available in the vital statistics files. Linkage with hospital records also enhances the measurement of variables that have been underreported on birth certificates, such as comorbidities for the birthing person [2,3], and allows for postnatal follow-up of the mother-child dyad.

With the purpose of studying pregnancy and birth-related outcomes, the California Office of Statewide Health Planning and Development (OSHPD), now known as the California Department of Health Care Access and Information (HCAI), contracted for the creation of linked birth files covering the years 1991 through 2012. The contractor conducted linkages between birth and fetal death certificate records with the hospital discharge records of the birthing person (9 mo before delivery and 1 y after delivery) and infant (1 y after delivery) [4]. These OSHPD datasets were used for many important studies [5-8]. However, file creation was discontinued after 2012, and updated files were needed to examine current data.

In 2015, the University of California Study of Outcomes in Mothers and Infants (SOMI) was created by CDC and GB at the University of California San Diego (UCSD) and LJ-P at the University of California San Francisco (UCSF) [9]. SOMI was developed in collaboration with community and public health partners to predict and promote healthy outcomes in birthing people, infants, and children across California, with a special focus on San Diego County. San Diego County is the second most populous county in California with a diverse racial, ethnic, and economic makeup, including a large population of migrants and refugees. SOMI’s goal is to develop a better understanding of the complex combination of factors—from biological to social and from economic to environmental—that contribute to birthing and child health outcomes. To achieve this goal, the SOMI data platform was created as a foundation of the research in collaboration with the UCSF California Preterm Birth Initiative [10] and the Healthy Outcomes of Pregnancy for Everyone study [11].

In 2019, SOMI received approval from the Committee for the Protection of Human Subjects within California’s Health and Human Services Agency to create updated linked birth files after this work was discontinued by the OSHPD. Subsequently, SOMI gained approval to link additional state databases to enhance the data platform’s utility. Future projects include using the data platform to obtain additional California state data that will provide information about outpatient visits, medication use, and childhood outcomes beyond the first year of life. The SOMI data platform allows multidisciplinary researchers, policy makers, and other stakeholders to visualize and quantify maternal and child health outcomes and related disparities across San Diego County and California and to examine trends over time.

Objectives

In this paper, we provide an overview of SOMI. The mission of SOMI is to predict and promote healthy outcomes in birthing people, infants, and children in collaboration with community and public health partners. The SOMI platform was created to achieve this goal. This platform connects birth, death, and other health records; biobank specimens; and environmental exposures at the individual or regional level. These data may be used to understand disease etiology, inform current diagnoses and clinical care protocols for infants affected by a variety of birth defects and medical conditions, and inform prevention policy. The SOMI data platform also allows multidisciplinary researchers, policy makers, and other stakeholders to visualize and quantify disparities in child health outcomes across San Diego County and beyond. SOMI scholars have been recruited to use the data platform to advance clinical knowledge and provide important big data experience for emerging researchers.

Methods

Overview

Before 2019 (when SOMI received institutional review board approval to create updated linked birth files), projects used OSHPD-created linked birth files as the data source. In 2019, the SOMI data platform was created using the methodology described in the next subsections. These same methods have also been used for the creation of other similar datasets by partners at the UCSF California Preterm Birth Initiative [10,11].

Data Sources

The primary source data file for the SOMI data platform is the birth cohort file maintained by California Vital Statistics. This file contains live birth certificates, linked infant death certificates within 1 year of birth when a death occurred, and fetal death certificates [12]. Births occurring in California and resulting in a live birth must be registered within 10 days of the birth, regardless of whether it occurred in a hospital, home, or elsewhere. Fetal death certificates are issued for fetal deaths, defined as those where “the fetus does not breathe or show any other evidence of life such as the beating of the heart, pulsation of the umbilical cord, or definite movement of voluntary muscles,” occurring after 20 weeks of gestation [13,14]. Fetal deaths exclude induced abortions and should be registered within 8 days of delivery. Birth certificate records include all California live-born infants and contain demographic information as well as information about pregnancy, labor and delivery, and infant outcomes. Infant death, defined as death within 1 year of birth, is recorded on death certificates, which include information about the cause of death. Between 2005 and 2020, 420,000 to 570,000 live-born infants and 2500 fetal deaths (deaths occurring after 20 weeks of gestation) were recorded annually in California [15].

The second source of data for the SOMI dataset comes from the HCAI, which collects patient discharge, emergency department, and ambulatory surgery center data HCAI [16]. The linked birth file is created through a probabilistic linkage algorithm, developed by the UCSD- and UCSF-based SOMI team and is similar to other published algorithms (Figure 1) [17]. Details are provided in the following subsections.

‎

Principal File Linkage

Link California Vital Statistics Record to Birthing Person’s Hospital Discharge Record for Birth Admission

We identified variables that were available both in California Vital Statistics records and HCAI files. These included the birthing person’s date of birth, hospital where the birth occurred, infant’s date of birth, and sociodemographic and clinical characteristics associated with the birth (Table 1). We used a blocking strategy that involves 2 stages. The first stage identifies possible links for each record based on linking the birthing person’s date of birth and hospital of birth. The second stage scores the possible links within each block (refer to the scoring system presented in Table 1). Scores for each variable were assigned based on the strength of the variable (eg, linked date of birth had a higher score than linked infant sex). Accurate linkages were prioritized over high linkage rates; therefore, a pessimistic linkage method was used to increase the likelihood of true matches. Before performing the linkage, it was determined that a successful link required that (1) the dates of birth for both the birthing person (10 points) and the infant (10 points) must be linked, the hospital (10 points) must be linked, and at least 1 other variable must be linked; or (2) if the dates of birth or hospital did not link exactly, the month and day could be reversed, or the year could differ for the birthing person. After scoring, approximately a subset of matched records with the lowest link score were randomly reviewed by RB to determine whether this requirement resulted in links that seemed accurate based on in-person record review. All match variables were compared among these subsets. On the basis of the point values of these links and a review of individual records, it was determined that a link score of at least 40 was required, in addition to the required matching variables mentioned previously. Although records with a link score of <40 had matching strong variables such as birthing person’s date of birth, infant’s date of birth, and hospital of birth, without further matching information (eg, mode of delivery, race and ethnicity, and county of residence), we were not confident enough to proceed with the match, and these records were discarded. For singleton births, we selected the record with the highest link score (above the cutoff score required) as the final linked record and removed all other records. For multiple births, we linked multiple records to the same birthing person and required that the births be classified as multiple (eg, twin or triplet) births in both the California Vital Statistics and HCAI records. In the rare instance where ≥2 records had the highest link score, the linked record was randomly selected form all records with this link score.

As we allowed for the month and day to be reversed or for the year to differ in the birthing person’s date of birth, we continued linkage using multiple passes and different linking variables. We repeated the process for any unlinked records after the first pass using different variables in the blocking step of the linkage (pass B: infant’s date of birth and patient zip code, pass C: infant’s date of birth and hospital of birth, pass C: birthing person’s date of birth and patient zip code, pass E: patient zip code and hospital of birth, and pass F: preterm birth and infant’s date of birth). Before each pass, previously linked records were removed. Records without a link after pass F remained unlinked.

Link Birth Certificate to Infant’s Hospital Discharge Record for Birth Admission

Hospital discharge records were restricted to records with an indication of pregnancy or birth (major diagnostic category code 14; International Classification of Diseases [ICD; ICD-9] codes V23 or any 640-679; ICD-10 O1, O2, O3, O4, O6, O7, O8, and O9). We used the same method to link the birth certificate records to the infant’s hospital discharge record for the birth admission, with different variables for the first blocking step (pass A: hospital of birth and infant’s date of birth; passes B, C, and D repeated pass A to obtain duplicates for twins or multiples; pass E: infant’s date of birth and patient zip code; pass F: repeat pass E to obtain duplicates for twins or multiples; pass G: patient zip code and infant sex; pass H: patient zip code and payer category; and pass I: patient zip code and hospital of birth). Records without linking infant sex were not considered links regardless of link score. After scoring, a subset of records were randomly examined by RB, and a cutoff score of 31 was determined to balance sensitivity and specificity to ensure accurate linkage (of note, this was only linking infant to infant because the birthing person’s date of birth was not available for infant hospital discharge records, hence the lower match score threshold). As fewer variables were available for comparison between the 2 infant datasets, the link score was set lower than for the birthing person. We note that if twins or multiples shared the same sex, birthweight range, and mode of delivery, and all survived, it was impossible to decipher which record should be linked, and the records were randomly assigned.

Linkage to Additional Hospital Discharge, Emergency Department, and Ambulatory Surgery Records for Birthing Persons and Infants

This step enables multiple linkages over time because a person might be admitted multiple times. It allows an evaluation of hospital visits during pregnancy and readmissions before or after delivery. The methodology for this was similar to the prior linkages; however, now encrypted social security numbers (SSNs) available within the HCAI files were used in addition to the variables listed in Table 1 to strengthen the link. Encrypted SSNs are present for approximately 80% of people giving birth and <10% of infants. As an SSN can potentially be used by different individuals, matching SSNs alone was not considered a match. Other matching variables, in addition to matching SSNs, were required. Moreover, the algorithm considers variables in previous HCAI records (from birth admission) to strengthen the link.

Evaluation of Linkage Quality

Performance of Linkage

We evaluated the performance of the linkage by estimating the percentage of birth and fetal death certificate records that were successfully linked to (1) hospital records for the birthing person, (2) hospital records for the infant, and (3) hospital records for both individuals. We estimated these percentages for the full cohort, as well as by year, sociodemographic characteristics, and pregnancy outcomes.

Comparison With Previous Linkage Methodology

Next, we compared our dataset to the one created using the prior OSHPD methodology. Prior files were created using probabilistic linkage [18,19], although the code was not available for comparison. For the evaluation of linked files, we selected the 2 most recent years of the OSHPD linked birth files: 2011 and 2012. By merging on the California Vital Statistics birth state file number (SFN), we compared the individual records that were linked to determine whether we were linking to the same patient. We compared each variable in the record between the 2 files, and exact links were considered the same patient. We also compared the sociodemographic and clinical characteristics of those in the linked and unlinked populations because unequally distributed unlinked records could result in bias [20]. Finally, we reviewed records linked by the previously used OSHPD methodology and the SOMI methodology to determine whether our algorithm needed modification to improve the number and percentage of true matches.

Evaluation of Data Quality and Bias

To evaluate the potential for sampling bias due to the exclusion of unlinked records, we analyzed the prevalence of preterm birth among the live-born infant cohort and separately by race and ethnicity. We first evaluated the prevalence of preterm birth using only California Vital Statistics data and then in the cohort of linked records. We repeated this with stratification by race and ethnicity. Preterm live birth was defined as gestation of 20 to <37 completed weeks. Gestational age at birth and race and ethnicity data were obtained from birth certificate records.

To evaluate the gain in information from querying hospital discharge records in addition to California Vital Statistics, we estimated the prevalence of several characteristics in the linked cohort first from only the information provided on the birth record and then separately from the presence of the characteristics on the hospital discharge record or the birth record.

Analytic Software

All linkages and analyses were performed using SAS (version 9.4; SAS Institute Inc).

Data Available in the Principal Data Platform

From the California Vital Statistics records, the SOMI platform contains information from the vital statistics records for the birthing person and nonbirthing person, typically the father (race and ethnicity, education, and occupation); residence (address, zip code, and census tract); payer for prenatal care and expected payer for delivery; prepregnancy and delivery weight; height; number of prenatal care visits; the month prenatal care began; the mode of delivery; gestational age at delivery; singleton or twin or multiple; infant’s birth weight; infant’s date of birth; complications during pregnancy and labor and delivery; and abnormal conditions of the newborn. Birth certificate files contain a birth SFN (unique by year). Linked infant deaths have an ICD code for the cause of death and a death SFN (unique by year) [12].

The HCAI files (patient discharge, emergency department, and ambulatory surgery center data) provide date of admission and discharge, ICD diagnoses and procedure codes, principal language spoken, and discharge status [21].

Build Out of the Data Platform

To improve the SOMI platform’s utility, additional datasets have been merged on residential zip codes or census tracts. These datasets include CalEnviroScreen [22], US census data [23], and AirNow [24]. California birth certificates include a text field for occupation and business or industry for the person giving birth (typically the mother) and the person not giving birth (typically the father). As these fields include different ways to describe occupation (eg, teacher or elementary school teacher) and may include typos, misspellings, and abbreviations, the SOMI data platform has cleaned these variables using the Occupational Safety and Health Administration algorithm, allowing for analysis of birth outcomes by occupation (eg, firefighter, bartender, or dry cleaner). In addition, the platform has the potential to merge indicators of redlining, food deserts, traffic, walkability, and any other data sources with geographic features (eg, zip code, address, and geocodes; Figure 2).

‎

Biological Specimens

With the birth SFN provided in the California Vital Statistics file, SOMI data platform is able to identify individuals of interest to be able to obtain biospecimens from the California Biobank (biospecimens in the biobank have birth SFN linked) [25,26]; for example, the authors of this study were awarded a genomic project, funded by the National Institutes of Health, examining unexplained causes of infant deaths in San Diego County. Through the SOMI platform, all infants who died in the first year of life in San Diego County over a 10-year period were identified, and newborn blood spots stored in the California Biobank were requested. Whole-genome sequencing was performed and is now being used in combination with SOMI data on environmental, demographic, geographic, pregnancy, and infant characteristics to understand the proportion of cases that might be explained by some combination of factors and potentially prevented [25-27].

Ethical Considerations

Methods and protocols for the study were approved by the Committee for the Protection of Human Subjects within California’s Health and Human Services Agency (project 2019-025) and the institutional review board at the UCSD (project 190892). This protocol was first approved in 2019 and has received annual continuing approval to date. The protocol includes a waiver of informed consent. All identifying data are stored behind secure UCSD firewalls in accordance with the requirements of the Committee for the Protection of Human Subjects, California Vital Statistics, and the HCAI. As the data used in this study come from an administrative database, no compensation was provided to individuals whose data were included in the files.

Results

Principal File Linkage

There were 8,084,212 deliveries, including 8,040,000 (99.45%) live births and 44,212 (0.55%) fetal deaths, in California between 2005 and 2020. Using the SOMI linkage methodology, 92.38% (7,427,738/8,040,000) of the California Vital Statistics live birth records were linked to the HCAI records for the birthing person, 95.53% (7,6880,597/8,040,000) of the California Vital Statistics birth records were linked to the HCAI records for the infant, and 90.61% (7,285,345/8,040,000) of the California Vital Statistics birth records were linked to the HCAI records for both the birthing person and the infant. Of the 44,212 fetal deaths, 33,355 (75.44%) had HCAI records linked to the birthing person. There was some variation in the percentage linked by year (Table 2). The percentage of records linked did not vary by the age of the birthing person but was poor 399/1788, 22.3%) for individuals whose age was unknown. There were some differences by race and ethnicity. Records where the birthing person was Asian or Hispanic had a higher proportion of linkage. Records where the birthing person was American Indian or Alaska Native, Hawaiian or Pacific Islander, or did not have a race reported had a lower rate of linkage. Approximately 73.31% (31,356/42,770) of the preterm infants born at <28 weeks had records linked to both the infant and the birthing person, whereas 91.07% (6,529,658/7,169,856) of the term infants born at 37 to 42 weeks had both linkages (Table 2).

Performance of Linkage

The linked population was more similar to the total population than the unlinked population, as one would expect given the high linkage rate. The population without a link to a birthing person had a smaller percentage of people using Medi-Cal (California’s Medicaid program), individuals of Hispanic ethnicity, and singleton births compared to the population linked to the records of the birthing person, as well as a larger percentage of birthing people delivering at 20 to <37 weeks of gestation and pregnant with twins. When examining the differences in the births linked to an infant hospital record, the findings were similar (Table 3).

Comparison With Previous Linkage Methodology

During 2011 and 2012, California Vital Statistics recorded 1,009,197 births. SOMI methodology linked 92.74% (935,895/1,009,197) of these records to the birthing person and live-born infant compared to 96.04% (969,258/1,009,197) for the OSHPD methodology (determined from the SOMI files obtained from the HCAI). Of the 928,459 records that had a link using both linkage methodologies, 923,705 (99.49%) were linked to the same HCAI record for the birthing person, 907,745 (97.77%) were linked to the same HCAI record for the infant, and 903,183 (97.28%) were linked to the same HCAI record for the birthing person and the infant.

There were 40,773 records with an OSHPD link between the birthing person and an infant who did not have this same link in the SOMI file. Of these 40,773 records, the SOMI file had 23,001 (56.41%) infants with a link between California Vital Statistics and HCAI records but had no link to a birthing person. Similarly, SOMI had 32.33% (13,182/40,773) of the birthing persons having a California Vital Statistics and HCAI link but no link to an infant. In 4590 (11.26%) of these 40,773 records, SOMI had no link for birthing person or infant between California Vital Statistics and the HCAI and also had no link between the birthing person and infant.

Data Source of Ascertainment: HCAI Versus California Vital Statistics

The records from HCAI data had greater capture of all variables compared to California Vital Statistics, with the exception of infant death, where both sources had similar capture rates. When only considering periviable births of between 22 and 23 weeks of gestation, the HCAI percentage of infant deaths was smaller than that reported in California Vital Statistics. Querying both data sources resulted in percentages consistent with those expected in the general population (Table 4).

Use Case: Preterm Live Birth Prevalence in Linked Versus Unlinked Files

The preterm live birth rate in the California Vital Statistics records was 8.91% (716,331/8,040,000), and it was 8.49% (285,345/618,325) in the linked cohort (Table 5). The rate was most altered for records where the birthing person selected “not stated/unknown” for their race (21,582/188,987, 11.42% in the California Vital Statistics file vs 15,455/150,624, 10.26% in the linked file) and least altered for people who selected “White, non-Hispanic” (186,166/2,240,243, 8.31% in the California Vital Statistics file vs 159,191/1,987,431, 8.01% in the linked file).

Discussion

Principal Findings

Overview

In this paper, we described the creation of the SOMI data platform, an administrative cohort derived from California birth and fetal death records. Although our linkage is not identical to the previous linked files created by the OSHPD, we demonstrate that the linked records are similar, as seen by the 97.28% (903,183 /928,459) concurrence between both infant and birthing person. We have demonstrated quality linkages for 90.61% (7,285,345/8,040,000) of the population. We did not see evidence of selection bias for a number of measured variables, including infant’s year of birth, gestational age at birth, and birthing person’s education level.

Interpretations, Implications, and Comparisons to Existing Literature

Leveraging birth certificate data to answer research questions has several advantages. First, it makes use of a population-based cohort of existing data, which may be necessary when primary data collection is unrealistic or too costly [20,28]. Second, capturing the entire population improves the generalizability of the research findings. Finally, large sample sizes permit the study of rare health outcomes (eg, periviable birth and birth defects) [28]. Birth certificate files have limitations when used in isolation. It is widely accepted that recorded information, such as birth defects, preeclampsia, and premature rupture of the membranes, is profoundly underascertained [2,3,29,30]. By linking these data to external data sources, variable ascertainment can be improved, and additional factors can be measured; for example, linkage between birth records and hospital discharge records facilitates the measurement of additional health factors such as maternal substance use disorder, mental health illness, infant resuscitation, and complications of prematurity (eg, intraventricular hemorrhage and bronchopulmonary dysplasia). However, linkage creation may lead to selection bias if the linkage is differential across groups.

There are many different methods that can be used to create these linkages. Some are deterministic, requiring perfect links across key variables to link records; others are probabilistic, linking records based on the probability that the records belong to the same individual. Researchers must decide which variables to include in the linkage, what characteristics to prioritize, and how stringent to be in determining a link. These subjective decisions determine the proportion of people who end up in the linked database and who, ultimately, are included in research studies. Thus, linkage procedure can affect both the external validity (who is represented in the data and who is not) and the internal validity (measurement error, unmeasured confounding variables, and selection bias) of the study [20,31]. Our linkage rate is slightly lower than that of the previous OSHPD files (935,895/1,009,197, 92.74% vs 969,258/1,009,197, 96.04%). While the starting populations differed somewhat—the OSHPD files excluded fetal deaths and multiple births [18], whereas the SOMI files included these records—there were differences in the linkage rates observed that were not explained by differences in the starting samples. Importantly, our record-by-record review of the 4590 individuals who were in the OSHPD matched file but were not in the SOMI matched file suggested that these were likely false matches (given missing or highly discrepant information) [16]. As such, it is possible that some other identifier was available for the linkage of this previous OSHPD set, and, in turn, these records could in fact be true matches. Nevertheless, based on the data available to the SOMI team and our commitment to link quality over quantity, no adaptations to our linkage methodology were made as a result of these comparisons.

As there are not many publications on the creation of datasets such as this one, it is difficult to compare our methodology, linkage rates, and data utility to those of other works. In 2013, Salemi et al [32] published on a similar dataset created for births in Florida. Their linkage rate for the 2.3 million birth certificate records to the infant birth hospitalization record was 92.1%. Similar to our study, they reported possible bias in linkage rates. The US government maintains several databases that can be used for maternal and child health research, such as the Pregnancy Risk Assessment and Monitoring System [33] and the Healthcare Cost and Utilization Project [34], both of which contain national data; however, both databases have limitations, including restricted capacity for merging additional data.

Key Findings to Date

To date, in partnerships with investigators across the country, SOMI has published >50 manuscripts with important findings on maternal and child health (highlights are presented in Table 6). The large population within the data platform has enabled the examination of rare diagnoses such as alcohol use disorder during pregnancy and infant heart defects, cannabis use disorder and infant death, postpartum suicidal behaviors, and infant readmissions associated with air pollution exposure.

Limitations

A knowledge of the strengths and limitations of this study is crucial for the effective and appropriate use of these data. With respect to limitations, the lower proportions of linkages to infant hospital records among gestations of <37 weeks reported in our study was similar to the findings from prior population-based linkage studies [31,50]. Greater success of linkage to the records of the birthing person compared to the infant records in these early-gestation pregnancies may be driven by a higher incidence of fetal death in this group. In fact, a prior study recommends excluding births of <24 weeks of gestation from analyses due to the potential misclassification of stillbirths as live births [31]. In our analyses using the SOMI data, we exclude births with gestational age of <22 weeks in all our analyses. Second, Harron [20] defines 2 parameters for linkage accuracy: missed links (2 individual records link, but they are not linked) and false links (2 records from different individuals are linked). We were not able to estimate these linkage accuracy parameters in our study because we do not have access to the electronic medical records informing the HCAI and California Vital Statistics data, and we have no other data source to confirm the correct link. Although we strove to gain better ascertainment of variables such as diabetes and hypertension by linking California Vital Statistics to HCAI records, we are unable to validate the linkage; therefore, true ascertainment is unknown. However, we can state that the rates of these variables increase with the linkage. Third, as previously noted, if twins or multiples shared the same sex, birthweight range, and mode of delivery, and none died or all died, it was impossible to decipher which record should be linked with which infant. This is an unavoidable limitation of this linkage, which must be considered when conducting analyses that include twins or multiples.

The data platform is missing variables that are important for some investigations, including medication use, laboratory values, timing of diagnoses, diagnoses not specified in ICD codes, and outpatient visit information. Infants delivered outside of the hospital (eg, home births) could not be linked to a discharge record for the birth, nor could the birthing person.

The Future of SOMI

SOMI scholars have been recruited to use the data platform to advance clinical knowledge and provide analytic experience for emerging researchers. Scholars have included master’s students, PhD students, medical residents, and fellows. These scholars identify a research question, develop an analytic plan, participate in the data analysis, and present their data at meetings or as a manuscript. Some of these studies have served as thesis projects or part of dissertations [51].

Future projects include using the data platform to obtain additional California state data that will provide information about outpatient visits, medication use, and childhood outcomes beyond 1 year. We also expect that the SOMI data platform will be used to further additional work focused on understanding racial and ethnic as well as socioeconomic disparities in maternal and child health outcomes across San Diego County and California, while also enabling an examination of trends over time by multidisciplinary researchers, policy makers, and other stakeholders.