Causal Mediation of Neighborhood-Level Pediatric Hospitalization Inequities Free

We conducted this work as part of the Responding to Identified Sociomedical risks with Effective Unified Purpose Study, which was approved by the Cincinnati Children’s Institutional Review Board. Our sample for this cross-sectional, population-wide, neighborhood-level study included all youth ≤18 years from Hamilton County, Ohio, who were hospitalized between July 1, 2016 and June 30, 2022. Cincinnati and Cincinnati Children’s are in Hamilton County, which includes 222 census tracts and ∼190 000 children and adolescents. Cincinnati Children’s captures ∼95% of all in-county hospitalizations. We extracted pertinent data from the electronic health record. We geocoded and merged residential addresses with extant place-based data using a standalone geomarker assessment tool.^24,25 

We defined the annual neighborhood-level pediatric hospitalization rate as the number of total hospital admissions per census tract per year divided by the total number of children residing within each census tract from the United States Census. We scaled the neighborhood-level pediatric hospitalization rate to hospitalizations per 1000 children per year for ease of interpretation. Our primary analysis looked at the all-cause hospitalization rate, but we repeated analyses for select conditions.

Selected a priori, we created condition-specific subsets of all hospitalizations based on the primary International Classification of Diseases, Tenth Revision, Clinical Modification diagnosis code and the Clinical Classifications Software Refined (CCSR) ontology from the Healthcare Cost and Utilization Project. We considered admissions related to asthma (CCSR category: “Asthma”), type 1 diabetes (CCSR category: “Diabetes mellitus, Type 1”), sickle cell anemia (CCSR category: “Sickle cell trait/anemia”), and psychiatric disorders (a subset of the CCSR category: “Mental Illness”). As with previous epidemiologic groupings of psychiatric disorders, we included only the CCSR subcategories of depressive disorders, bipolar and related disorders, other specified and unspecified mood disorders, anxiety and fear-related disorders, suicide attempt or intentional self-harm, obsessive-compulsive and related disorders, trauma- and stressor-related disorders, disruptive impulse-control and conduct disorders, personality disorders, feeding and eating disorders, somatic disorders, miscellaneous mental and behavioral disorders or conditions, neurodevelopmental disorders, and schizophrenia spectrum and other psychotic disorders.²⁶ 

We created a DAG to define the causes of neighborhood-level pediatric hospitalization rates related to racial or ethnic identity (captured as census tract-level racial or ethnic composition using data from the 2019 US American Community Survey) and to examine how racial and ethnic inequities in hospitalization rates are mediated. Specifically, we extracted estimates of the fraction of each census tract’s population identifying as 1 of 6 unique combinations of ethnicity (Hispanic/Latino or non-Hispanic/Latino) and race (white, Black, other). Given limited variation of the prevalence of individual groups across tracts (Supplemental Fig 2; eg, mean Hispanic/Latino population 2.1%, maximum 20%), we reduced these 6 categories into 1 estimate of the fraction of population that identified with any marginalized racial (Black, other) or ethnic (Hispanic/Latino) group for statistical modeling purposes. We conducted sensitivity analyses, using the measure with 6 instead of 2 levels.

To define the causal DAG, we used our previous causal inference studies, existing scientific literature, and input from clinical and population health experts.^22,27 –29  We adopted the assumption that “complex, historical processes” were at the root of neighborhood-level distributions of race, ethnicity, socioeconomic status, and a range of neighborhood-based exposures and characteristics.³⁰  This approach enabled us to quantify estimates of the direct-effect inequity operating through pathways not related to socioeconomic status, access, or environment (eg, discrimination and interpersonal racism). Alternatively, the direct-effect racial inequity could be interpreted as the difference that would remain if blocked pathways were set the same across racial and ethnic groups.

Various measures and indices are commonly used to capture multifaceted neighborhood-level exposures. Each has their own unique set of motivations, intended objectives, and caveats.^31,32  Below, we detail which census tract-level measures we employed to characterize factors identified within our DAG.^16,21,23 

To capture family socioeconomic status (SES) at birth at a tract-level, we used a previously-validated material deprivation index calculated from 6 Census American Community Survey variables.^29,33,34  This nationwide index is scaled and normalized to a 0 to 1 range (higher indicates more deprived) and has been associated with multiple pediatric health outcomes.^29,35,36 

To capture social context more fully, we used additional tract-level measures from the 2019 5-year US American Community Survey, including: fraction of households speaking a language other than English, fraction of housing units occupied by renters, and fraction of households headed by a single parent. Furthermore, we estimated the enforced housing infraction rates per census tract by geocoding records available on Open Data Cincinnati³⁷  to the census tract and dividing by the total number of households, estimated using 2020 US Census data. We estimated crime risk using a census block-level 7-year (2014–2020) average of total crime risk from Applied Geographic Solutions.³⁸  We used population weighting to interpolate census block to census tract crime risk.

For hospital access, we obtained drive time isochrones specific to the hospital as concentric polygons in 6-minute intervals from Openroute service.³⁹  We calculated tract-level drive time by taking the area-weighted average of drive times corresponding to isochrones that overlapped each census tract.

To capture environmental pollutants and amenities, we calculated the fraction of total land classified as greenspace. The National Land Cover Database classifies 30 × 30 m grid cells into 20 land-use categories, each then classified as “green” or “nongreen.”⁴⁰  We calculated tract-level percent greenspace by dividing the number of grid cells classified as green by the total number of grid cells overlapping each census tract. Furthermore, we employed a measure of traffic-related air pollution by calculating the fraction of households living within 400 m of an interstate or highway.

We used negative binomial regression models with an offset term equal to the number of children living in each tract divided by 1000 to model the pediatric hospitalization rate. We fit 1 regression model with marginalized composition as the only variable to identify the total racial and ethnic inequity in neighborhood, or census tract, pediatric hospitalization rates. We omitted tracts if the 2019 American Community Survey estimated 0 residents aged under 18. Model coefficients were scaled to a hypothetical 10% increase in the percentage of a census tract’s population that identified with a marginalized racial or ethnic group. Exponentiated model coefficients represented the estimated percent change in the hospitalization rate with the corresponding change in racial and ethnic group composition. We extended these models to estimate the “direct inequity” not mediated through DAG variables we could measure, adjusting for the area-level variables introduced above and listed in the Supplemental Information. The percentage mediated through structural and neighborhood characteristics was calculated as the direct hospitalization rate inequity divided by the total hospitalization rate inequity. We repeated analyses for the predefined condition subsets.

Of all Cincinnati Children’s hospitalizations between 2016 and 2022 for patients with Hamilton County addresses, we were able to successfully geocode 98.8% to a census tract. This enabled evaluation of 50 719 hospitalizations experienced by 28 390 patients: 12 684 hospitalizations for psychiatric conditions (25% of all hospitalizations), 2683 for asthma (5.3%), 597 for sickle cell anemia (1.2%), and 565 for type 1 diabetes (1.1%). The all-cause census tract-level hospitalization rate ranged from 10.9 to 143.0 per 1000 children (median 45.1; interquartile range: 34.5 to 60.1). The percentage of a census tract’s population that identified with a marginalized racial or ethnic group ranged from 1% to 95% (median 30%; interquartile range: 10 to 60). The histograms in Supplemental Fig 4 display the distribution of the racial/ethnic compositions across all tracts. Supplemental Figs 5 and 6 show all-cause and condition-specific hospitalization rates plotted as functions of each census tract’s percentage identified with groups that have been marginalized. Two tracts were excluded from the analysis for low numbers of children, leaving 220 tracts for statistical modeling.

The unadjusted model estimated that a 10% increase in the percentage of a census tract’s population that identified with a marginalized racial or ethnic group was associated with a 6.2% (95% confidence interval [CI]: 4.5 to 8.0) increase in the all-cause tract-level hospitalization rate (ie, total hospitalization rate inequity, Fig 2 and Table 1). This increase was greater for each condition-specific subgroup, 6.8% for psychiatric disorders, 9.7% for type 1 diabetes, 25% for asthma, and 46% for sickle cell anemia. Thus, neighborhoods with increased concentrations of marginalized racial or ethnic groups experienced higher hospitalizations rates for each assessed condition.

The community material deprivation index was the variable most strongly associated with both racial composition and hospitalization rates (0.45; 95% CI: 0.34 to 0.55; Fig 3).

Our causal DAG indicated that to estimate the effect of membership in marginalized racial or ethnic groups on pediatric hospitalization, we needed to block confounding by social context, hospital access, and environmental pollutants and amenities (Fig 1). After adjustment for community material deprivation, crime risk, primary language other than English, housing tenure, family composition, hospital access, greenspace, traffic-related air pollution, and housing conditions, the all-cause hospitalization rate inequity was reduced to 0.2% (95% CI: −2.2 to 2.7), suggesting that 96.6% of the total inequity is mediated via structural and neighborhood-level factors (Fig 2, Table 1). Among the condition-specific groups, the percentage of total inequity mediated via structural and neighborhood characteristics varied from 15.3% in sickle cell anemia to 98.4% for psychiatric hospitalizations. Asthma (52.7%) and type 1 diabetes (66.6%) fell in the middle. Overall, neighborhood characteristics and structural factors were responsible for nearly all the inequity in all-cause hospitalizations and most of the inequity for specific health conditions.

A sensitivity analysis using 6, instead of 2, racial and ethnic categories (Supplemental Fig 2), resulted in similar findings, with a total hospitalization rate inequity (for fraction “Black, not Hispanic/Latino” compared with fraction “white, not Hispanic/Latino”) of 6.0% (95% CI: 4.3 to 1.1) and a direct hospitalization rate inequity of 0.3% (95% CI: −2.0 to 2.7). A sensitivity analysis of all admissions, excluding 738 admissions related to sickle cell anemia, resulted in similar findings, with a total hospitalization rate inequity of 6.0% (95% CI: 4.3 to 7.7) and a direct hospitalization rate inequity of −0.1% (95% CI: −2.5 to 2.3).

We found significant associations between neighborhood measures of racial and ethnic composition and hospitalizations. Notably, neighborhood characteristics like socioeconomic deprivation, hospital access, and environmental pollutants accounted for nearly all the observed inequities in all-cause hospitalizations. When examining condition-specific effects, the magnitude of inequities explained by these neighborhood characteristics varied drastically. A comprehensive approach to evaluating links between such characteristics benefits from methods facilitating causal inference. DAGs can be used to elucidate interactions between the many conditions that influence childhood racial health inequities, accelerating analyses that could guide effective interventions.

Health inequities characterize a range of pediatric conditions. Multiple studies have investigated causal pathways linking racial inequities to social, economic, and environmental exposures. For example, racial differences in SES, community violence, and air pollution exposure have all been shown to contribute to pediatric health outcome inequities.^21,41 –45  The cumulative impact of these exposures was further illustrated in a study that found higher rates of hypertension in Black Americans when compared with rates of hypertension among Black people in Africa and the Caribbean, suggesting strong social and environment drivers.⁴⁶  Since race is a social construct, and racism is perpetuated by societal structures, no single factor can explain all health-related racial inequities for a given disease.^47,48 

Thus, deeply embedded, multiplicative factors such as structural discrimination, institutional racism, and residential segregation likely drive persistent associations. Many factors at the root of health inequities originate in place and in a history that influences why certain places promote health, whereas others impede health.⁴⁹  A recent study used a causal model to determine how much of segregation’s effect was mediated through SES, using the Getis-Ord statistic to capture segregation.⁵⁰  Here, we more directly evaluated racial composition and found further evidence that differential exposures to geomarkers, defined as “objective, contextual, or geographic measures that inform additional assessments and interventions,” generate inequities at the population level.⁵¹  Indeed, we found that geomarkers explained nearly all the identified racial inequities. A move toward health equity requires us to consider these exposures and address them in ways that overcome structural, systemic barriers continuing to impede meaningful progress.

Our findings provide additional evidence supporting structural racism as a root cause of inequities in pediatric population health outcomes.^{23,47,48,52 –54}  Racism affects health across structural, institutional, interpersonal, and internalized levels.²³  Although all levels of racism are important and may affect an individual’s experience of racism in health care settings, our results suggest that structural racism is responsible for the vast majority of racial inequities in hospitalization rates. To begin to address structural racism, we must critically assess and address the unjust, unequal distribution of power and resources driving differential geomarker exposure and differential health outcomes.⁴⁹ 

The amount of the total inequity explained by adjustment variables ranged from ∼15% for sickle cell anemia to nearly 100% for psychiatric disorders. Sickle cell anemia disproportionately affects those with African and Mediterranean ancestry.⁵⁵  In Greater Cincinnati, nearly all children treated for sickle cell anemia identify as Black or African American. Many live in segregated neighborhoods. As such, there was limited variability in assessed geomarkers. This led to poor model fit, limiting the conclusions we can draw from this finding. It may also be the case, for sickle cell anemia and asthma, which saw a decrease but not elimination of the racial inequity after adjustment, that unmeasured factors remain important. Our findings (the reduction or elimination of racial inequities upon adjusting for geomarkers situated within our DAG) generate hypotheses and should prompt ideas for action. For instance, interventions leveraging clinical-community partnerships may spring from assessed geomarkers. Access could be eased through altered bus routes, service delivery programs, or expanded telehealth capabilities. Concentration of deprivation may be altered through policies aimed at expanding opportunity, easing connection to public benefits, and informed placement of community resource centers. Environmental pollutants may be mitigated through proactive housing inspection, litigation against problem landlords, or connections to legal aid attorneys expert in housing rights.³¹  Our methods, and findings, do not allow us to determine which of these courses of action will be most appropriate or impactful. Rather, they help generate questions and ideas that, through meaningful engagement with patients, families, and partners, could yield testable interventions and policy changes.⁵⁶ 

Further, as pediatricians and pediatric researchers, we have an opportunity and a responsibility to act on existing evidence by partnering with our professional organizations and public officials to dismantle the effects of structural racism and promote health equity. The American Academy of Pediatrics recently published a policy statement⁵⁷  with recommendations on how to optimize public insurance, including expansion of access and coverage for interventions addressing health-related social needs. Relatedly, in November 2023, the White House released “The U.S. Playbook to Address Social Determinants of Health,” which expresses a vision where every American can “lead a full and healthy life within their community.” The playbook underscores the importance of data gathering and sharing, flexible funding to address social needs, and support for “backbone organizations.” Our findings illuminate the undeniable intersection between place and health, place and inequity, and the impediments to leading full, healthy lives within home communities. Bridging health and community data could, indeed, generate novel insights, inform impactful interventions, and build and grow multisector partnerships.

One limitation of our study is that we pursued this work in 1 county, potentially limiting generalizability; however, we believe our methodology is generalizable and could be applied elsewhere. We cannot rule out the possibility of residual confounding on our results. The interplay between race, ethnicity, socioeconomic factors, and health care outcomes is highly complex and dynamic, possibly involving feedback loops and interactions that even a well-constructed DAG might not capture. Our cross-sectional design did not allow us to enforce temporal ordering of exposures and outcomes. Additionally, we did not include patient-, family-, or household-level variables of potential relevance, including direct measures of racism as experienced at either population or individual levels.

We found that social, economic, and environmental exposures mediate greater than 96% of population-level racial and ethnic inequities in child hospitalizations. Interventions and policies that address such exposures and dismantle racist structures could promote better, more equitable child health outcomes.

The Responding to Identified Sociomedical risks with Effective Unified Purpose study research team includes (in alphabetical order): Chidiogo Anyigbo, Elizabeth Bishop, Lori Crosby, Magdely Diaz de Leon, John Egbo, Ben Foley, David Hartley, Adrienne Henize, Nana-Hawa Yayah Jones, Robert Kahn, Lauren Lipps, Alex Power-Hayes, Charles Quinn, Carley Riley, Laura Sandoval, Lisa Shook, and Jeffrey Steller.

CCSR

Clinical Classifications Software Refined

CI

confidence interval

DAG

directed acyclic graph

SES

socioeconomic status