Using Electronic Health Records to understand the population of local children captured in a large health system in Durham County, NC, USA, and implications for population health research

Abstract

Although local policies aimed at reducing childhood health inequities can benefit from local data, sample size constraints in population representative health surveys often prevent rigorous evaluations of child health disparities and health care patterns at local levels. Electronic Health Records (EHRs) offer a possible solution as they contain large amounts of information on pediatric patients within a health system. In this paper, we consider the suitability of using EHRs from a large health system to study local children’s health by evaluating the extent to which the EHRs capture the county’s child population. First, we compare the demographic characteristics of Duke University Health System pediatric patients who live in Durham County, NC (USA) to the child population estimates in the 2015–2019 American Community Survey. We then examine geographic variation in census tract rates of children captured in the EHR data and estimate negative binomial models to assess how tract characteristics are associated with these rates. We also perform these analyses for the subset of pediatric patients who have a well-child encounter. We find that the demographic characteristics of pediatric patients captured by the EHRs are similar to those of the county’s child population. Although the county rate of children captured in the EHRs is high, there is variation across census tracts. On average, census tracts with higher concentrations of non-Hispanic Black residents have lower capture rates and tracts with higher concentrations of poverty have higher capture rates, with the poorest tracts showing the largest racial gap in rates of children captured by EHRs. Our findings suggest that EHRs from a large health system can be used to assess children’s population health, but that EHR-based evaluations of children’s health disparities and health care patterns should account for differences in who is captured by the EHRs based on census tract characteristics.

Understanding the geographic distribution of childhood illness and receipt of care is critical for improving population health and reducing health inequities among children. Neighborhood disparities in health outcomes can reflect unequal access to social and economic resources—often resulting from differences in neighborhood socioeconomic characteristics or structural forms of racism—that shape both individuals’ risk of illness and access to adequate health care services (Aysola, Orav, and Ayanian 2011; Beck et al. 2017, 2020). Public health policies and programs that leverage local data are essential for targeting the underlying causes of disease to improve population health (Acevedo-Garcia et al. 2008; Beck et al. 2017). Health surveys that describe individuals’ health profiles allow for population representative analyses at the national or state levels but often constrain accurate descriptions at smaller levels of geography because of small sample size (Birkhead 2017; Tomasallo et al. 2014).

Electronic Health Records (EHRs) offer a possible solution to the problem of providing accurate descriptions of children’s health statuses and disparities across census tracts (Beck et al. 2017) and cities (Klompas et al. 2017) due to the large number of individuals included in these data. EHRs are administrative data maintained by a health care provider on patient medical histories over time (CMS 2012). They contain large amounts of information on patients’ interactions within a health system, allowing for the characterization of individuals’ health profiles. Access to EHRs is significantly less costly than health surveys in terms of labor, time, and response rates that are often low in surveys (Casey et al. 2016) and can provide real-time descriptions for surveillance purposes. EHRs can also be linked to contextual data to examine connections between social and environmental factors and health outcomes (Casey et al. 2016), thus making them an important resource for understanding the multiple dimensions of individual and community health status and vulnerabilities (Miranda et al. 2013).

As EHRs become a more common data source in social and health science research (see examples: Angier et al. 2020; Beck et al. 2014; Carlson et al. 2021 for EHR-based analyses of associations between census tract characteristics and health outcomes), there is a need to understand whom is captured in a health system and if the EHRs of the pediatric patient population sufficiently capture the local child population so that researchers may be able to use the data to draw conclusions about children’s population health and health disparities. In this paper, we consider the suitability of using EHRs from a large health system in Durham County, NC to study child health in the county and across census tracts by evaluating the extent to which the county’s child population is captured in the DUHS EHRs. We also examine geographic variation by census tract in the rates of children captured in the EHRs and assess how tract characteristics are associated with variation in these rates.

Using EHRs to Study of Population Health

As EHR databases have become widely implemented over the last decade, there has been a widespread call for social scientists and practitioners to use this data source to inform efforts aimed at improving population health and identifying health disparities (Calman et al. 2012; Perlman et al. 2017). Recent studies have relied on EHRs to examine differences in children’s health across small geographic areas, underscoring the relevance of using these data for understanding and addressing local child health disparities. These studies have used EHRs to estimate children’s illness-specific hospitalizations rates across census tracts in Cincinnati, OH and to examine how census tract characteristics help explain the geographic differences in those rates (Beck et al. 2014, 2016). EHRs were also used to examine the relationship between the built environment and children’s weight status across block groups in the Kansas City (MO) metropolitan area (Carlson et al. 2021).

While EHRs present a unique data source for the study of population health, their primary administrative purpose complicates efforts to accurately measure health outcomes in target populations (Casey et al. 2016; Phelan, Bhavsar, and Goldstein 2017). Only patients who receive care in a health system are included in EHRs, making the inclusion of individuals in this data source non-random (Phelan et al. 2017). Furthermore, the completeness of medical information depends on the type and consistency of care received within the health system (Bower et al. 2017; Casey et al. 2016; Phelan et al. 2017). Thus, individual characteristics that predict who is in a health system and health care utilization may introduce bias into EHR-based analyses.

A large literature considers how individual-level selection processes bias EHR-based analyses (Birkhead 2017; Bower et al. 2017; Casey et al. 2016; Phelan et al. 2017). Informed presence bias results from an individual’s health status predicting entry into a health system, so that those who experience illness or symptoms are more likely to seek care and systematically differ from those who remain outside of the system (Bower et al. 2017; Phelan et al. 2017). Individuals’ perceptions of a health system—including perceptions of cost, service availability, and cultural competency—and their ability to access its services also shape entry into that health system such that the population included in EHRs is often not representative of the target population to whom reference is made (Bower et al. 2017). Prior work suggests that individual characteristics such as gender, age, race, ethnicity, socioeconomic status and health insurance coverage predict health system entry and utilization (Birkhead 2017; Bower et al. 2017; Romo et al. 2016). These selection patterns can bias EHR-based estimates of health outcomes at the population level (Bower et al. 2017) and confound the understanding of the relationship between individual characteristics and health outcomes.

Characteristics of neighborhood are also associated with health system entry in ways that may bias EHR-based analyses. Measures of social cohesion, safety, and the built environment are associated with the use of primary care services (Kuang et al. 2017; Sabounchi et al. 2018) and access to patient-centered medical homes (Aysola et al. 2011; Kuang et al. 2017) among pediatric patients. Low-income neighborhoods also have higher rates of child mortality in emergency departments (Rees et al. 2020), which may be a marker of inadequate upstream support for health. By shaping children’s access to a health system and the type of care they receive, these community resources likely determine both children’s presence in a health system’s EHRs and the types of patient data available.

Because it inherently dictates access to resources, structural racism shapes individuals’ interactions with health systems in similar ways (Feagin and Bennefield 2014; White, Haas, and Williams 2012). Neighborhood racial composition—a marker of residential racial segregation and a primary institutional mechanism of racism (Gaskin et al. 2012; Massey 2001)—is associated with health care use across communities. For example, persons living in predominantly Black or Latino zip codes generally have lower rates of most types of outpatient visits and are less likely to have a usual source of care compared to those living in predominantly White zip codes, though results are mixed when comparing rates of hospital-based outpatient visits (Gaskin et al. 2012). In Phoenix, Arizona, high levels of Latino and Native American residential segregation are associated with increased odds of children using community health care clinics instead of a physician’s office, partly due to the unequal distribution of health care organizations across the city (Anderson 2020).

The potential for neighborhood characteristics to introduce bias in EHR-based analyses is exacerbated by the relationship between those same characteristics and health outcomes. Both neighborhood socioeconomic characteristics (Carroll-Scott et al. 2013; Rees et al. 2020) and racial composition (Bell et al. 2006; Kotecki et al. 2019) are associated with the prevalence of a wide-range of children’s health outcomes, which in-turn drive entry into a health system. These characteristics also likely combine to uniquely shape health risks and access to care (Beck et al. 2020).

These observed associations underscore the importance of considering whether and how neighborhood characteristics bias EHR-based estimates of children’s health across neighborhoods by shaping who is captured in a health system’s EHRs.

Despite these concerns, recent comparisons of the demographic characteristics of adult patients captured in EHRs with those of population representative health survey samples have highlighted the suitability of EHRs for providing accurate descriptions of population health at the state-level. These comparisons have found that, in terms of demographic characteristics, adult patients in a large state health system are compositionally similar to state- and county-level population representative samples (Tomasallo et al. 2014). Similar conclusions were drawn for state-level rates of adult health conditions from EHRs compared with health surveys (Klompas et al. 2017; McVeigh et al. 2016; Tomasallo et al. 2014). However, when comparisons are narrowed to smaller geographic areas, the accuracy of EHR estimates vary by factors such as the fractions of local populations captured in a health system (Klompas et al. 2017).

Although few studies draw similar comparisons between pediatric patients in EHRs and population representative samples of children (for an exception, see Flood et al. 2015), the threat of individual-level selection bias may be less severe in this population group. Most children receive some form of health care. Children are also more likely than adults to have some form of health insurance (Artiga and Ubri 2017) and they require more routine visits (e.g., well-child encounters) with health care providers for developmental surveillance and immunizations. These factors suggest that EHRs might represent a unique data source for examining children’s population health at the local level and across small geographies.

Current Study

Durham County in North Carolina is a prime example of a community in which most health care has been consolidated under a single system, in this case the Duke University Health System (DUHS). DUHS serves as the primary provider in the County and consists of a network of over 100 outpatient primary care and specialist clinics and three hospitals, including the only two in the county–a large referral center (Duke University Hospital) and a community hospital (Durham Regional Hospital) (Miranda et al. 2013). Since 2014, DUHS has utilized an integrated EPIC-based system, enhancing the capture and organization of health data for all DUHS-affiliated providers and allowing for the tracking of patients across multiple modalities of care. Thus, the consolidation of care in Durham County under the DUHS presents a unique opportunity to evaluate if the EHRs of pediatric patients in a large health system can be used to study local children’s population health and health disparities.

Here we combine DUHS EHRs and American Community Survey (ACS) data to: (1) evaluate the extent to which DUHS EHRs capture the county’s population of children by comparing the demographic characteristics of the pediatric patient population to the county child population; (2) examine geographic variation by census tract in the rates of children captured in the EHRs; and (3) assess the relationship between census tract characteristics—namely poverty and racial composition—and variation in capture rates. Our analyses underscore the importance of identifying potential biases across individual and census tract characteristics when using EHRs to examine children’s health at local levels.

We perform the analyses for pediatric patients with any DUHS encounter, as well as for a subset of pediatric patients with at least one well-child encounter. The any encounter group represents the total population of pediatric patients in the DUHS EHRs. Because detailed information on patients’ health histories and profiles are collected at every encounter, using the EHRs for all pediatric patients maximizes the amount of health information researchers can use to estimate local children’s health.

The well-child encounter group represents the subset of pediatric patients with regular contact within the DUHS. Well-child encounters are recommended pediatric visits that take place on a regularly scheduled basis for the routine evaluation of growth and development, provision of anticipatory guidance, and provision of preventive care, including vaccine administration. Since all children in the county are expected to attend well-child visits, this group may be less selected on health status, potentially reducing informed presence biases, though selections based on other individual and census tract characteristics may still affect who enters the DUHS. Thus, the well-child group can provide an additional angle from which to assess demographic and census tract patterns related to whom is regularly captured by EHR data.

The setting of our study, Durham County, encompasses a growing population that is geographically, racially-ethnically, and socioeconomically heterogeneous. It is composed of the major demographic groups that constitute the US population: 43% non-Hispanic White, 37% Black, 14% Hispanic, and 6% Asian (Census, 2019 population estimates). The county includes both urban and rural neighborhoods, and its per capita income (~$35,000) approximates the national figure ($34,000) with significant socioeconomic disparities within each population group (2015–2019 5-year ACS). Historically rooted and systemic forms of racism are prominent across the county, mirroring those that exist in many other US cities and counties. Redlining and the construction of a major freeway that physically separates Black neighborhoods from downtown Durham have contributed, together with a rapid process of gentrification, to sharp neighborhood economic and racial segregation and unequal access to resources, including health care (De Marco and Hunt 2018). The county’s heterogeneity and social settings make it ideal for taking a neighborhood-centric perspective to assess which children are captured in DUHS and how capture rates vary by census tract to identify gaps in care useful to social scientists, health practitioners, and policymakers.

METHODS

Data

American Community Survey Data

We use aggregated estimates from the 2015–2019 5-year American Community Survey (ACS) data tables, which are publicly available on the US Census Bureau website, to examine characteristics of the children living in Durham County. The data provide estimates of county and census tract child population totals, as well as county and census tract distributions of children’s age, sex, and race and ethnicity.

In these tables, race and ethnicity are defined as: White, Black or African American, and Hispanic or Latino origin (of any race). We combine the remaining categories (American Indian and Alaska Native; Asian; Native Hawaiian and Other Pacific Islander; Some other race; and Two or more races) due to low estimated proportions of each group in Durham County census tracts. The 5-year ACS estimates for Durham County include the proportion of children identifying as non-Hispanic (NH) White, but no other commonly used joint race and ethnicity categories.

In order to better understand if comparisons between EHRs and ACS data using separate or joint race and ethnicity designations yield different takeaways, and because joint racial-ethnic identities are commonly used, we supplement our comparisons with estimates for the joint racial-ethnic composition of children in Durham County available from the Annie E. Casey Foundation’s Kids Count Data Center (Kids Count). These estimates are produced through a collaboration between the Center for Disease Control’s National Center for Health Statistics and the U.S. Census Bureau according to the following designations: NH White, NH African American, NH Other, and Hispanic (Annie E. Casey Foundation 2020).

To assess the relationship between census tract characteristics and rates of Durham County children captured in the EHRs, we rely on ACS estimates that pertain to the entire census tract community, because community characteristics are indicators of social context and structural barriers to care. We use data for all but three of Durham County’s 60 census tracts, which mostly consist of commercial businesses and very low counts of children (<35 per tract).

The main tract-level variables are the total population racial-ethnic composition—an indicator of residential racial segregation—and percentages of poverty (families in the same household whose 12-month total income falls below the poverty-level). We also include covariates that measure census tract levels of health insurance (individuals reporting at least one type of public, private, or other health insurance), employment (residents aged 16 or older in the labor force), and household overcrowding (households with more than one person per room). We include these covariates as they are linked to poverty rates and may independently shape access to care and health outcomes (Beck et al. 2016; Yousey-Hindes and Hadler 2011).

DUHS Electronic Health Records

We abstracted data from an EHR-based datamart (Hurst et al. 2020) to create a dataset of patient characteristics for all children with at least one encounter between January 1, 2016 and December 31, 2017. This two-year period centers the cohort around the 2015–2019 5-year ACS estimates. We use a two-year period to evaluate the suitability to use EHR data to examine a snapshot of the local population’s health (Klompas et al. 2017), while also maximizing the completeness of EHR data on pediatric patients.

The dataset was further limited to pediatric patients who had a Durham County residential address—geocoded to counties and census tracts—that was associated with an encounter occurring during the study period. Patients with more than one Durham County address recorded during the period were assigned the first recorded address. Our final EHR dataset includes all patients aged under 18 at the end of the study period, with 56,627 pediatric patients in the any encounter group and 40,247 in the well-child encounter group.

Using the EHRs, patient sex is reported as female or male and age is measured on the last day of the study period. DUHS patients’ race and ethnicity are matched to ACS categories. We do not know whether race and ethnicity reported in the EHRs is self-identified or assessed by DUHS staff. Although this a common problem shared by many large health systems, systematic discrepancies between self-identified and staff-assessed race-ethnicity are possible and would prevent accurate comparisons between the race-ethnic composition of children in the EHRs and of the county’s child population in the ACS. However, these concerns are somewhat alleviated by prior studies which find high levels of agreement between patients’ self-reported race-ethnicity and EHR-documented race-ethnicity (Bergdall et al. 2012).

Only 4.44% of patients with a DUHS encounter during the study period do not have an associated geocodable address. In our analytic dataset, few pediatric patients are missing racial (7.09%), ethnic (0.16%), or joint racial-ethnic (3.59%) data and no patients are missing birthdate or sex.

Patient information is aggregated at the county and census tract level for the analyses.

Analytical Methods

We first evaluate the extent to which the DUHS EHRs capture the child population of Durham County. Following prior studies (Casey et al. 2016; Tomasallo et al. 2014), we compare ACS-estimated characteristics of Durham County’s child population with the characteristics of children in the EHRs, by sex, age, and race and ethnicity. We consider the EHR pediatric patient population similar to the county child population if the patient characteristics fall within the 95 percent confidence intervals of the corresponding ACS estimate. EHR figures are exact percentages for the population of DUHS pediatric patients.

Next, we examine the rate at which the DUHS EHRs capture the county population of children. We estimate capture rates as:

$$CaptureRat{e}_c=\frac{EHRcoun{t}_c}{ACSestimat{e}_c},$$

where c is the geographic unit (county or census tract), EHR count is the number of children in the EHR cohort who have at least one relevant encounter and reside in the geographic unit, and ACS estimate is the ACS-estimated number of children residing in the geographic unit. Thus, the capture rates are measured as the estimated proportion of children with at least one relevant DUHS encounter during the study period in a defined area. We map census tract capture rates in ArcGIS to geographically examine variance across Durham County.

Finally, we use aggregated census tract data to assess the relationship between census tract characteristics and DUHS capture rates. Our outcome variable is the number of children in each census tract who entered DUHS during the study period. We estimate a negative binomial model to account for the over-dispersion of the number of children across census tracts with DUHS encounters and we include a population offset to effectively define the outcome as the capture rate.

We conduct two separate sets of analyses: one using census tract characteristics to predict any encounter capture rates across census tracts, and another predicting well-child encounter capture rates. The first model in each set of analyses estimates the independent relationship between DUHS capture rates (either any or well-child) and ACS estimates of poverty, racial-ethnic composition, and a set of socio-environmental covariates. The second model in each set of analyses considers the multiplicative effects of poverty and racial composition, since census tracts with multiple social disadvantages related to experience of historical and persistent racial discrimination may uniquely shape engagement with the DUHS.

All predictor variables are standardized, and coefficients are presented as capture rate ratios. A capture rate ratio of 1.13 is interpreted as a one standard deviation higher value in the predictor variable being associated with a 13% higher census tract capture rate.

All analyses were conducted in R 3.6.1 and all maps were created in ArcMap 10.7. This study was approved by the Duke Health Institutional Review Board.

RESULTS

Demographic Comparisons

Table 1 presents the estimated demographic composition of children living in Durham County (ACS) and the composition of pediatric patients in the DUHS EHRs. Overall, the total pediatric population (any encounter) reflects the estimated demographic composition of Durham County children. Both groups are about 49% female and follow similar age distributions. But, the ACS and EHR groups qualitatively and statistically differ when race and ethnicity are analyzed separately. Compared to the ACS, the EHR any encounter group has higher percentages of children identified as Black (EHR Any: 42.4%; ACS: 38.1 [95% CI = 36.9, 39.3]) and Other (EHR Any: 25.3%; ACS: 14.7 [12.8, 16.6]) and lower percentages of children identified as White (EHR Any: 32.3%; ACS: 47.4 [45.7, 49.1]) and Hispanic (EHR Any: 20.0%; ACS: 24.8 [24.7, 24.9]). However, when race and ethnicity are analyzed jointly, the racial composition of the EHR cohort closely aligns with the ACS and Kids Counts estimates. The percentage of NH White patients in the EHR does not differ from the ACS estimate (EHR Any = 29.2%; ACS: 29.4 [29.3, 29.5]) and closely aligns with the Kids Count estimate (30.9%). The percentage of NH Black patients (40.6%) is also similar to the Kids Counts estimates for Durham County children (40.5%), though the percentage of NH Other patients (9.6%) is higher than the Kids Count estimate (4.3%).

Table 1.

Durham County, NC Comparison of Children’s Demographic Characteristics in the ACS (5-year, 2015–2019), Kids Count (2017), and DUHS EHRs (2016–2017)

Characteristic	ACS (N=65,628)		Kids Count (N=65,919)	EHRs: Any Encounter (N=56,627)	EHRs: Well-child Encounter (N=40,247)
	%	95% CI		%	%
Sex
Female	49.3	(48.0, 50.6)	---	49.3	49.6
Age
<5	31.2	(31.1, 31.3)	---	31.5	36.5
5 - <10	28.5	(27.3, 29.8)	---	27.8	27.1
10 - <15	25.5	(24.3, 26.8)	---	26.1	24.4
15 - <18	15.1	(15.0, 15.2)	---	14.8	12.0
Race alone
White	47.4	(45.7, 49.1)	---	32.3	34.5
Black	38.1	(36.9, 39.3)	---	42.4	41.5
Other	14.7	(12.8, 16.7)	---	25.3	24.0
Ethnicity alone
Hispanic	24.8	(24.7, 24.9)	25.0	20.0	16.8
Race-Ethnicity
NH White	29.4	(29.3, 29.5)	30.9	29.2	32.0
NH Black	---	---	39.7	40.6	40.2
NH Other	---	---	4.4	9.6	10.4
Any insurance	93.1	(91.7, 94.5)	---	--	100.0
DUHS capture rate	--	--	---	0.86	0.61

Note: Descriptive characteristics for the ACS-estimated child population and the DUHS EHR groups are considered similar if the EHR percentage falls within the corresponding 95% confidence interval for the ACS sample. Non-Hispanic White is the only joint race-ethnicity measure in the ACS for Durham County, so we include joint estimates from the Annie E. Casey Foundation Kids Count Data Center. All EHRs were extracted for children who were less than 18 years-old as of December 31, 2017 and had an encounter in the DUHS between January 1, 2016 - December 31, 2017 (centered around the 2015–2019 ACS). Insurance in the EHRs measures use of any type of insurance to pay for at least one well-child visit during the period. EHR missingness is low: race (7.09%), ethnicity (0.16%), and race-ethnicity (3.59%).

The demographic composition of the well-child encounter group differs more substantially from the ACS sample. In addition to a slightly higher percentage of NH White patients, children with a well-child encounter are younger than the ACS sample.

Table 1 also includes overall DUHS capture rates. An estimated 86% of Durham County children are captured in DUHS during the study period (any encounter), with an estimated 61% captured in DUHS for a well-child encounter.

Census Tract Analysis

Figure 1 demonstrates the variation in any encounter capture rates across census tracts. In most tracts, EHRs capture more than 80% of the child population. Census tracts with the lowest rates lie along the Southeast and Southwest borders of Durham County. This pattern holds for the well-child capture rates (Figure 2), which are above 50% in most tracts.

Figure 1. Percentages of Durham County Residents Under the Age of 18 Captured in the Duke University Health System’s EHRs with Any Encounter in 2016–2017, by Census Tract.

Percentages of children captured in DUHS are calculated using EHR data from encounters occurring between January 1, 2016 and December 31, 2017. Pediatric patients are included in numerator counts if they a Durham County address associated with an encounter during the study period and are < 18 years-old at the end of the period. Denominator counts are 2015–2019 ACS-estimated counts of children <18 years-old residing in each census tract. Percentages are not estimated for three census tracts (colored white) that have a low number of ACS-estimated child residents (<35). Some tracts have an estimated >100% capture rate.

Figure 2. Percentages of Durham County Residents Under the Age of 18 Captured in in the DUHS EHRs with a Well-Child Encounter in 2016–2017, by Census Tract.

Percentages of children captured in DUHS with a well-child encounter are calculated using EHR well-child encounters occurring between January 1, 2016 and December 31, 2017. Pediatric patients are included in numerator counts if they a Durham County address associated with a well-child encounter during the study period and are less than 18 years-old at the end of the period. Denominator counts are 2015–2019 ACS-estimated counts of children <18 years-old residing in each census tract. Percentages are not estimated for three census tracts (colored white) that have a low number of ACS-estimated child residents (<35).

Next, we assess how tract characteristics are associated with the geographic variation in capture rates across Durham County. Table 2 presents the estimates from the negative binomial models predicting capture rates for the any encounter and well-child encounter groups separately. In the first model for any encounter, the percent of the population identifying as NH Black and the percent of families living below the poverty level are associated with census tract capture rates, although in opposite directions. Compared to census tracts with the average percentage of NH Black residents, a one standard deviation higher percent NH Black is associated with an 8% lower DUHS capture rate. Compared to census tracts with the average percentage of poverty, a one standard deviation higher percent poverty is associated with a 16% higher DUHS capture rate. Other census tract characteristics are not associated with tract capture rates.

Table 2.

Capture Rate Ratios from Multivariate Models Estimating the Association Between Census Tract Characteristics and Capture Rates of Children within DUHS: Durham County, NC, 2016–2017

Census tract-level population characteristics	Any Encounter				Well-child encounter
	CRR Est.	95% CI	CRR Est.	95% CI	CRR Est.	95% CI	CRR Est.	95% CI
Race-Ethnicity
NH White (%)	Ref.		Ref.		Ref.		Ref.
NH Black	0.92	(0.85, 0.98)	0.90	(0.84, 0.97)	0.87	(0.81, 0.94)	0.86	(0.79, 0.92)
NH Other	0.95	(0.89, 1.01)	0.91	(0.85, 0.98)	0.93	(0.87, 0.99)	0.89	(0.83, 0.96)
Hispanic	0.99	(0.88, 1.12)	0.92	(0.80, 1.06)	0.99	(0.87, 1.13)	0.93	(0.80, 1.07)
Poverty (%)	1.16	(1.06, 1.28)	1.22	(1.11, 1.36)	1.17	(1.06, 1.30)	1.23	(1.12, 1.38)
Interaction: Race-Ethnicity x Poverty
NH White x Poverty			Ref.				Ref.
NH Black x Poverty			0.90	(0.84, 0.98)			0.90	(0.84, 0.98)
NH Other x Poverty			1.00	(0.95, 1.05)			1.00	(0.95, 1.05)
Hispanic x Poverty			0.99	(0.95, 1.04)			0.99	(0.94, 1.04)
Other Socio-Environment (%)
Insurance	0.91	(0.81, 1.02)	0.90	(0.79, 1.01)	0.94	(0.83, 1.07)	0.93	(0.82, 1.06)
Employment	0.98	(0.93, 1.04)	0.97	(0.91, 1.03)	0.98	(0.93, 1.05)	0.97	(0.91, 1.03)
Household overcrowding	0.91	(0.83, 1.00)	0.91	(0.84, 1.00)	0.90	(0.81, 1.00)	0.91	(0.83, 1.00)
Intercept	0.89	(0.85, 0.94)	0.93	(0.88, 1.00)	0.63	(0.60, 0.67)	0.67	(0.63, 0.72)

Notes: Models in this table estimate negative binomial regressions. Each model includes a child population offset, effectively estimating DUHS capture rates among children. Coefficients are presented as capture rate ratios (CRR). All population characteristics are measured as z-scores. Any encounter counts the number of children residing in Durham County, NC who have at least one encounter of any type in DUHS within the period (2016–2017). Well-child encounter counts the number of children with at least one well-child visit within the period.

The second any encounter model indicates an interaction between percent NH Black and percent poverty. Higher percent poverty remains associated with higher capture rates. However, the effect is moderated by percent NH Black: the positive relationship between percent poverty and capture rate is larger in tracts with lower concentrations of NH Black residents. Figure 3 graphically presents this interaction. Among census tracts with the lowest percentages of poverty, capture rates do not differ by racial composition. Among census tracts with the highest percentages of poverty, there is a large, estimated gap in capture rates between census tracts with high and low concentrations of NH Black residents, which is driven by a steeper slope among tracts with lower concentrations of NH Black residents. While the coefficients vary slightly, the patterns hold for well-child capture rates.

Figure 3. Predicted Capture Rate of Any Encounter Given Census Tract Characteristics: Percentages of the Population Identifying as non-Hispanic Black and of Families Living Below the Poverty Level.

Predicted capture rates are estimated using coefficients from Table 2, holding all values at their means while varying percentages for poverty and NH Black. Because all other racial-ethnic measures are held at their means while percent NH Black varies, total racial-ethnic composition can fall below and above 100% and CRRs can rise above 1. The graph displays the diverging capture rates across measures of poverty and concentrations of NH Black residents.

DISCUSSION

Our analyses combine ACS and EHR data to assess the potential of using EHRs to examine children’s population health and health disparities at local levels. We find that most children (86%) living in Durham County are captured in the EHR database because they have received some form of care from DUHS during the study period (2016–2017). This high capture rate may be partly attributable to most Durham County children having health insurance (93%, ACS for ages <19).

The demographic characteristics of pediatric patients in the EHRs compare well with those in the ACS. Although characteristics of children with any encounter in the DUHS qualitatively and statistically differ from ACS estimates when race and ethnic compositions are analyzed separately, these differences are largely attenuated when comparisons are made with the joint race-ethnicity composition in the Kids Count estimates. This results from differences in how race is reported among Hispanic children in the EHRs, with race for Hispanic pediatric patients more likely to be reported as “Other” or missing in the EHRs compared to the estimates for children in Kids Count.

The age composition of children with a well-child encounter has a higher percentage of children aged <5 and a smaller percentage of children aged 15-<18 compared to the ACS estimates. This is not surprising, since the youngest children require more frequent well-child encounters, while adolescents, especially those covered by Medicaid, are less likely to have regular annual visits (Dempsey and Freed 2010).

In our analysis of DUHS capture rates across Durham County census tracts, we find important geographic patterns. The tracts with the lowest capture rates are situated along the Southeast and Southwest borders of the county. These census tracts are respectively closest to Wake Medical Center and UNC Health, which may serve some Durham County children who live closer to these other large health systems.

The evaluation of how census tract characteristics shape capture rates across Durham County reveals a negative association between percent NH Black and capture rates and a positive association between percent poverty and capture rates. However, the positive association between poverty and capture rate is moderated by percent NH Black, such that the association is smallest in tracts with the highest concentrations of NH Black residents.

Several explanations are possible. First, areas abutting Wake Medical Center and UNC Health are more affluent and the “loss” of children from more affluent tracts to other health systems may help explain the positive association between tract capture rates and poverty. Second, Duke Health collaborates with community organizations to improve access to health care, particularly for children in low-income communities (DukeHealth 2020). These collaborations could lead to higher DUHS capture in communities with higher poverty rates relative to wealthier communities. Third, and as an alternative interpretation, higher capture rates in tracts with more concentrated poverty may point to higher health care needs that require pediatric subspecialty appointments or emergency department care, which, in Durham County, are only provided by DUHS. In fact, supplemental analyses show that tract-specific percentages of poverty are positively and significantly associated with visits to the DUHS emergency departments. However, this is unlikely the only driver of the differences in capture rates across poverty percentages since tract poverty is also positively associated with capture rates for well-child encounters. Fourth, children may be undercounted in census tracts with higher percentages of poverty, inflating capture rates in these areas. It should be noted however that the relationship between capture rates and poverty percentages persists even when ACS margins of error for tracts’ estimated child population are included in the model, suggesting that systemic differences in the accuracy of ACS counts do not fully explain the relationship between tract poverty and capture rates.

Lower EHR capture rates in census tracts with higher concentrations of NH Black residents likely reflect the history and persistence of multiple forms of racism within Durham County. The county has a well-documented history of forced residential segregation that has systematically reduced access to resources, including health care, in predominantly Black neighborhoods. In addition to redlining, the construction of a major highway in the 1960s cut-off Hayti, a thriving Black community, from the main downtown area and forced thousands of Black-owned homes and businesses to relocate (De Marco and Hunt 2018). Reverberations of these policies are still evident today, and gentrification has exacerbated concerns related to health inequities and unequal access to resources, including health care, across the county. Furthermore, a history of inferior medical treatment provision to NH Black patients within the DUHS has fueled mistrust (Tweedy 2015), likely leading residents of predominantly Black communities to seek care outside of the DUHS or refrain from regular care overall.

Census tracts with high percentages of both NH Black populations and poverty may be uniquely disadvantaged, yielding the observed interaction effects. The positive relationship between poverty and capture rates is weaker in tracts with high concentrations of NH Black residents, where the legacy of racism, including current structural barriers, works against efforts to improve access to health services in high poverty neighborhoods. The health care needs of children in these communities may be best addressed when the first points of contact are health care facilities outside of the DUHS EHR umbrella that specifically target the needs of historically underserved communities. In Durham County, one such facility is the Lincoln Community Health Center (LCHC), which is the largest outpatient primary and preventive care facility for low-income, underserved community members in the county. Eligible patients receive their primary care services through LCHC and their secondary and tertiary care at DUHS. Patients who receive their primary care services through LCHC and do not receive secondary or tertiary care are not included in the DUHS EHRs we analyze. Thus, the longstanding relationships that LCHC has developed with underserved communities may partially explain the disparate DUHS capture rates across concentrations of NH Black residents in high poverty areas.

Public Health Implications and Future Work

Our analyses demonstrate that the DUHS sufficiently captures the county-wide child population and that researchers may be able to use the DUHS EHRs to draw conclusions about local children’s population health. Although our findings of high EHR capture rates do not imply health or health care equity, they demonstrate that EHRs are an important source of information for identifying disparities in health conditions across neighborhoods and subgroups and for supporting policymakers and community organizers in designing targeted resource deployment and interventions.

Access to EHR data can facilitate community health assessments and the creation of community-based cohorts, a cost-effective alternative to population surveys. EHR databases also offer real-time data on children’s health compared to survey data which often lag by one or more years. Social scientists, health practitioners, and policymakers can use EHR data to evaluate the prevalence of chronic diseases among local children, to compare prevalence rates across subgroups and small geographic areas, and to examine changes in rates over time. The utilization of EHR data can have multiple applications, including making aggregated census tract data on children’s health publicly available on an interactive webpage that allows policymakers and community organizers to examine current health patterns when developing public health strategies or community-based programs. EHRs can also be combined with contextual data accessed from data sources such as the ACS or through partnerships with community organizations to provide a multidimensional understanding of the relationship between individual and neighborhood characteristics and health outcomes.

The large number of county children captured by the EHRs provide a unique opportunity to study small population subgroups and allow for meaningful population health comparisons across racial-ethnic, socioeconomic, and age groups. EHR data may be a particularly powerful when evaluating the health of children in the youngest age group (<5 years old), who are in a sensitive period of development when early life health risks and outcomes can have long-lasting effects. Although basic health information for older age groups is available through schools, localities often lack data for children not yet in school. Our findings demonstrate that children below age five are disproportionately represented in the well-child encounter group, suggesting they have relatively complete EHR data that can advantageously inform local health interventions aimed at improving young children’s health.

Our finding that census tracts with higher concentrations of NH Black residents have, on average, lower capture rates, and that the racial gap is largest in the highest poverty tracts, underscore the reality that historical and ongoing forms of discrimination shape children’s interactions with local health systems. This aligns with previous findings, in Phoenix, Arizona, that segregation is associated with the types of health care facilities from which children seek care (Anderson 2020). When developing public health interventions, practitioners and policymakers should consider differences in where local children primarily access their health care to ensure that programs can equitably provide support to local children. Gaining access to EHR data from community health clinics alongside data from large health systems will improve our understanding of how exposure to different types of health systems and care vary across neighborhoods and how large health systems and community clinics can work together to meet the diverse needs of local children. Using EHR data to evaluate patterns of specialty and emergency department use across subgroups and census tracts can also inform our understanding of inequities in health and health care use or quality (Beck et al. 2014, 2016; Rees et al. 2020).

Our study also highlights the importance of understanding differential spatial patterns related to who is captured in a health system and how these patterns relate to neighborhood characteristics before estimating disease prevalence across local levels. For example, when using EHRs to identify hot spots of disease or evaluate relationships between neighborhood characteristics and health outcomes, researchers and practitioners must account for selection biases related to who is captured in EHR data based on neighborhood characteristics, such as spatial proximity to adjacent health systems, racial-ethnic composition of neighborhoods, and poverty rates. To address selection bias based on individual characteristics, previous work suggests implementing statistical approaches—including post-stratification adjustments to standardize crude estimate, analytic sample weights, and propensity score matching based on the variables that drive selection (Bower et al. 2017; Weiskopf et al. 2013). Similar approaches can be implemented to address neighborhood-level selection biases.

While our study assesses the suitability of using EHRs to study population health with a focus on the extent to which they capture the child population of a county, other challenges must be considered when using EHRs for this type of research. First, researchers must consider how patients’ race-ethnicity is assigned and if there is systemic misclassification of patients’ race-ethnicity in the EHRs that may bias results. Our findings suggest that joint (v. separate) race-ethnicity designations may more accurately reflect the demographic characteristics of the county population and should be used for EHR-based analyses. Health systems should also continue to develop and implement standardized procedures for collecting patient demographic data.

Second, because EHRs are primarily used for administrative purposes, their structure can complicate data extraction and coding for population health studies. For our analysis, we used data from a highly curated EHR-derived datamart in which the EHRs are transformed, cleaned, and structured in a way that significantly simplifies extraction and analysis (Hurst et al., 2020). Large health systems should consider using similarly organized datamarts to lower the upfront costs associated with accessing EHRs, as well as partnering with local clinics to store and share patients’ EHRs from all local facilities into a single secure platform. Furthermore, while developing the coding scheme for health conditions can be complex for EHRs, this information can be shared across research teams to reduce burdens and ensure consistency of health condition definitions across studies.

Finally, even once individuals are captured in a health system, differences in the frequency and type of care they received may bias EHR-based analyses. For example, information biases are introduced when patients seek specialized care outside of a health system or discontinue follow-up appointments, so that one system’s EHRs do not provide a full picture of individuals’ health profiles (Bower et al. 2017; Phelan et al. 2017). These potential biases should be considered when operationalizing complex health outcomes that require appointments with specialists or multiple appointments or measurements that are taken over time.

Conclusion

EHRs are increasingly used by large health systems to store information on patient encounters and offer large amounts of information for local child population health and disparities. However, selection into health systems on individual and neighborhood characteristics may bias EHR-based analyses. The present study finds that most of a county’s child population is captured in the EHRs of a large health system and that the characteristics of pediatric patients in the EHRs are similar to those of the county’s child population, highlighting the suitability of using EHRs to evaluate local children’s health. This study also highlights the need to consider how physical and structural factors—such as proximity to other health systems, concentrations of poverty, and residential segregation—may introduce bias into EHR-based census tract analyses and provides recommendations on how to correct for potential biases related to census tract characteristics. Importantly, social scientists, health practitioners, and policymakers can combine EHRs with contextual data and health records from local health clinics to provide complete descriptions of the multiple dimensions of local children’s health and improve health equity across communities.

Highlights.

• EHRs from a large health system are suitable for studying local children’s health

• Local children in EHRs are demographically similar to county population estimates

• Rates of children captured in health system differ by census tract characteristics

• EHR-based analyses should adjust for local differences in capture rates