Population-Based Impact of Rurality and Neighborhood-Level Socioeconomic Disadvantage on Pediatric Cancer Mortality in Washington State

Despite advances in care that have resulted in an approximate 85% 5-year survival, cancer remains the leading non-injury cause of childhood and adolescent mortality in the United States (1, 2). Disparities in cancer-related outcomes persist (3), and mortality differences related to race and ethnicity (4–6), health insurance coverage (7, 8), and poverty/socioeconomic status (5, 9–13) are well-established. Geographic factors such as travel time to treatment center and rural/urban status have been studied to a lesser extent, but emerging evidence provides rationale for concern. Rural or geographically remote residence at cancer diagnosis is associated with unique challenges during treatment, lower school attendance, and increased caregiver stress (14, 15).

Limited investigation of pediatric cancer survival related to geographic factors has yielded mixed results to date, for both children with hematologic and solid organ malignancies (16–21). A national U.S. study using cancer registry data from the National Cancer Institute's Surveillance, Epidemiology, and End Results (SEER) Program found no difference in childhood cancer–related survival based on county-level rural–urban continuum code designation with adjustment for county-level median income and other factors (22). However, county-level measurement does not account for the considerable geographic variation present in many U.S. counties, and may potentially mask differences through misclassification (23, 24). For this reason, more precise geographic designations, when available, are favored over county-level measures, particularly when comparing individuals based on rurality (25).

We used population-based databases from Washington State (WA) to evaluate whether rural residence—in combination with socioeconomic factors—affects pediatric cancer–related mortality. Given the heterogeneity of disease and treatment characteristics by cancer type, we also examined whether any associations between these factors and mortality differed by cancer type.

This retrospective cohort study used a database linkage between WA cancer registries (1992–2013) and state birth (1974–2013) and death (1992–2013) records as a part of an existing study to identify individuals born in WA who were diagnosed with cancer at <20 years of age from 1992 to 2013 (26). We obtained cancer data from two population-based cancer registries in WA that together capture all cases of pediatric cancer in the state: (i) the SEER program-funded Cancer Surveillance System (CSS) and (ii) the Washington State Cancer Registry (WSCR), funded by the Centers for Disease Control and Prevention's National Program of Central Cancer Registries. Both cancer registries provided diagnosis date, International Classification of Childhood Cancer (ICCC) third edition cancer type, International Classification of Diseases for Oncology (ICD-O) code, demographic information at diagnosis (age, sex, race, and ethnicity), basic initial treatment information (chemotherapy, radiotherapy, and/or surgery), and census block group-level residence at diagnosis. Sex assigned at birth was identified from birth certificate records.

Diagnoses representing non-malignant disease—typically non-malignant central nervous system (CNS) tumors, non-melanomatous skin cancers, or cervical carcinoma in situ—were excluded from analysis (n = 641). Cases missing data for race and ethnicity (n = 57) or census designation (n = 15), and those living in locations missing key indicators for calculation of area-level predictor variables (n = 39), were also excluded, 111 (2.5%) total cases. We determined vital status via CSS/WSCR's quarterly follow-up and the state death registry, the latter of which maintains agreements with the National Center for Health Statistics and the WA Department of Health to identify deaths to residents occurring out of state. We evaluated all-cause mortality, as 94% of deaths listed cancer as the cause of death.

This study was approved by the Institutional Review Boards of the Fred Hutchinson Cancer Center and the WA Department of Health. The requirement for written informed consent was not required for this study.

Census tracts are relatively permanent statistical subdivisions of a county that each contain approximately 4,000 residents (1,500 housing units) and are designed to be relatively homogeneous with regard to demographic characteristics and living conditions (27). Census block groups are more refined geographic subdivisions of census tracts, containing approximately 1,500 people each and closely resembling neighborhoods (27). All study cases were geocoded to 2010 U.S. Census block groups.

Using 2010 Census Rural–Urban Commuting Area (RUCA) codes corresponding to census tract at diagnosis, residences were classified as rural (primary RUCA code 4–10) or non-rural (primary RUCA code 1–3), consistent with recommendations (25, 28).

Area deprivation index (ADI) is a composite measure consisting of 17 items concerning income, employment, education, and housing, developed from U.S. Census data and the American Community Survey to proxy neighborhood-level socioeconomic disadvantage (29). ADI has been used to assess the impact of neighborhood socioeconomic disadvantage on healthcare access and outcomes, including pediatric cancer (11, 12). Scores by the census block group are reported as national percentiles, with 1 indicating the lowest level of socioeconomic disadvantage and 100 corresponding to the greatest socioeconomic disadvantage. State-level norms have been established, including for WA. A priori, we categorized all WA census block groups as either having high ADI or low ADI, using a cutoff value of the highest WA quintile to define high ADI neighborhoods. The 80th percentile ADI cutoff value that we used for WA corresponded to approximately the 66th percentile of social deprivation on the national scale.

We estimated 5- and 15-year overall survival (OS) by rural residence and ADI classification with Kaplan–Meier survival analysis. Differences in OS were compared between neighborhood groups using pairwise log rank tests. We then used Cox proportional hazard models to estimate mortality hazard ratios (HR) with 95% confidence intervals (CI) according to rural and ADI status as our main predictors of interest. Cox models were adjusted for race and ethnicity, sex assigned at birth, age at diagnosis (categorically: <1, 1–4.9, 5–9.9, 10–14.9, and ≥15 years), birth year (ordinally: by decade), and ICCC category. Multivariate models notably did not include insurance information, due to incomplete data in the cancer registry (n = 2,629 missing). Receipt of chemotherapy, surgery, and/or radiotherapy was not included in multivariate models, as these were likely mediators, rather than confounders, of the exposures and outcomes of interest. We did, however, perform descriptive statistics to compare rates of these treatments by neighborhood classification. Proportionality assumptions for each model were verified by plotting and testing Schoenfeld residuals (30). To address potential challenges to proportionality assumptions, we performed stratified Cox regression for some variables, allowing for separate baseline hazards by age at diagnosis, birth year, and ICCC. To test for effect modification between rurality and ADI classification, we also included an interaction term between rural status and high ADI. Because this term was significant, we performed Kaplan–Meier estimates and Cox proportional hazards models with rurality and ADI status categorized into four mutually exclusive groupings: non-rural with low ADI (referent), non-rural with high ADI, rural with low ADI, and rural with high ADI. Results with cell sizes <5 were suppressed in all tables.

We performed two sensitivity analyses to examine the robustness of our main Cox regression models and ADI definitions. First, we created an identical hazards model to the main analysis, but instead defined high ADI using the highest WA quartile rather than the highest quintile. This cutoff value corresponded to the 44th percentile on the national scale. We tested a statistical interaction term between rural residence and high ADI, followed by analysis by neighborhood grouping. Next, we tested an interaction term between rural residence and linear national ADI percentile, rather than a binary variable, in an otherwise identical multivariate model.

We also performed separate Cox regression sub-analyses by cancer type: Acute lymphoblastic and myelogenous leukemias (ALL and AML), lymphoma, CNS malignancies, and other solid tumors (neuroblastoma, renal tumors, hepatoblastoma, bone tumors, soft tissue sarcomas, and malignant germ cell tumors). Categories were determined using ICCC group types and ICD-O-3 morphology codes. Specifically, ALL was defined using the ICCC classification of leukemia, in addition to one of the following ICD-O-3 codes: 9811/3–9818/3 or 9835/3–9837/3. AML was defined by having the ICCC leukemia classification plus one of the following ICD-O-3 codes: 9840/3, 9860/3, 9861/3, 9865/3–9874/3, 9876/3–9912/3, or 9930/3. Multivariate HRs of these hazards models adjusted for the same variables as the main model, except rural residence and high ADI were adjusted for individually, as opposed to a multi-level combined variable, because there was no significant statistical interaction between these predictors in cancer type-specific sub-analyses. All analyses were conducted with R 4.1.3 (R Foundation) and used the “survival” package (31).

The data generated in this study are not publicly available to protect patient privacy and confidentiality, and are available only with separate approval from the WA Institutional Review Board.

The study sample included 4,306 cases after exclusions (51.7% male; 77.7% non-Hispanic White), with 20% of the study population residing in a neighborhood classified as high ADI and 13% residing in a rural neighborhood at the time of diagnosis (Table 1). Proportions of cases identified as having Hispanic ethnicity were greater in rural and/or high ADI neighborhoods compared with the total cohort. Distribution of sex and diagnosis type did not seem to differ in rural and/or high ADI cases. Cancer survivors were followed-up for a median of 5.0 years (inter-quartile range, 1.0–11.5). Over the study period, 18.9% of the cohort died, with 94% of deaths reported as cancer-related.

Estimated 5-year OS for the referent group (children from non-rural neighborhoods with low ADI) was 80.9% (Fig. 1). For children from neighborhoods that were non-rural with high ADI, rural with low ADI, and rural with high ADI, 5-year OS was 66.4%, 69.4%, and 66.9%, respectively. Estimated 15-year OS in the referent group was 75.8%. For comparison neighborhoods, 15-year OS was 57.0%, 66.9%, and 56.2%. Compared with the referent category, OS was significantly lower for all comparison groups by pairwise log-rank testing at 15 years (Fig. 1). Survival did not differ significantly between the three comparison categories themselves. OS at 5 years differed similarly via log-rank testing.

In adjusted Cox proportional hazards models, residence in neighborhoods categorized as non-rural with high ADI, rural with low ADI, or rural with high ADI was associated with greater mortality relative to living in a non-rural neighborhood with low ADI, with HRs of 1.64 (95% CI, 1.32–2.04), 1.50 (95% CI, 1.12–2.02), and 1.53 (95% CI, 1.16–2.01), respectively (Table 2). In adjusted analyses, Hispanic ethnicity was also associated with increased mortality compared with those identifying as non-Hispanic White.

In our proportional hazards model, the interaction term between rural status and high ADI was significant (P = 0.03), suggestive of effect modification. Therefore, we also plotted Kaplan–Meier survival curves stratified by rurality and ADI status (Fig. 2). Among children from non-rural neighborhoods, we found a survival difference based on ADI classification (log-rank test, P < 0.001). However, for children from rural neighborhoods, ADI did not have an appreciable survival effect (log-rank test, P = 0.58). Similarly, rural status had a strong effect on survival among children from low ADI neighborhoods (log-rank test, P < 0.001), but not among children from high ADI neighborhoods (log-rank test, P = 0.60). We observed similar results in our stratified Cox proportional hazards models. In the non-rural subset of children, the adjusted mortality Cox HR for high ADI versus low ADI was 1.76 (95% CI, 1.43–2.17; P < 0.0001). In the rural subset, no association with ADI was identified (HR, 1.07; 95% CI, 0.72–1.58; P = 0.75). Similarly, when stratified by ADI category, the association between rurality and survival varied. Rural residence was associated with mortality in the low ADI subset (HR, 1.56; 95% CI, 1.17–2.07; P = 0.002), but not in the high ADI subset (HR, 0.94; 95% CI, 0.68–1.29; P = 0.68). There was no statistical interaction identified between rurality and ADI category in cancer type-specific sub-analyses (Table 3); therefore, we did not calculate HRs for neighborhood groupings outside of the main analysis.

When individual associations of rural residence and high ADI for different cancer types were examined (Table 3), rural status (HR, 1.82; 95% CI, 1.01–3.27), but not high ADI (HR, 1.16; 95% CI, 0.69–1.97) was associated with greater mortality among children with ALL. For children with solid tumors, high ADI was associated with mortality (HR, 1.62; 95% CI, 1.16–2.26), but not rural status (HR, 1.06; 95% CI, 0.72–1.56). Similar results were observed among children with CNS tumors, for whom high ADI (HR, 1.47; 95% CI, 1.01–2.13) but not rural status (HR, 1.08; 95% CI, 0.70–1.68) was associated with mortality. No associations for ADI or rurality were detected in children with AML or lymphoma, the two smallest cancer-type categories.

The Cox hazards model with ADI corresponding to the highest WA quartile, rather than quintile, was similar to the main analysis. The interaction between rural status and ADI remained significant (P = 0.036). HRs by neighborhood grouping were also similar. Residence in neighborhoods categorized as non-rural with high ADI, rural with low ADI, or rural with high ADI was associated with HRs of 1.75 (95% CI, 1.41–2.08), 1.49 (95% CI, 1.06–2.07), and 1.60 (95% CI, 1.23–2.07), respectively. In our sensitivity analysis using national ADI percentile, the statistical interaction between rural status and ADI in the multivariate model also remained significant (P = 0.005).

In this population-based study of more than 4,000 children with cancer in WA, we observed mortality differences associated with rural residence and neighborhood-level socioeconomic disadvantage defined by geocoded address information. Compared with children living in non-rural neighborhoods with lower socioeconomic disadvantage at diagnosis and after adjusting for important patient- and disease-related factors, children in neighborhoods with rurality and/or high poverty experienced greater mortality. Socioeconomic status also modified the effect of rurality on mortality. Rural residence was associated with greater mortality among children living in neighborhoods with low socioeconomic disadvantage, but not for children in neighborhoods with greater socioeconomic disadvantage. We also observed variability in the relative impact of rurality and neighborhood poverty by cancer type. Most notably, among children with ALL, rural residence was associated with mortality, whereas among children with CNS and solid tumors, neighborhood socioeconomic status had a stronger association.

Our finding of an association between mortality and rural residence in children with cancer in WA is novel. Although associations between rurality and cancer-related mortality have been demonstrated in adults (32–34), our findings contrast with prior pediatric studies in the U.S., which were conducted with SEER national cancer registry data over similar time periods to our study and did not detect an association between county-level rurality and mortality (22, 35). However, county-level designations of rurality are subject to greater misclassification compared with the more granular census block group and tract-level definitions we used (23–25). Another potential explanation for this discrepancy is that WA cancer cases may differ from other U.S. regions with respect to sociodemographic factors, geography, distribution of cancer centers, and/or legislative policy. In the 2020 Census, WA's rural distribution (16.0%) is similar to the U.S. aggregate (19.3%), ranking 35th of 50 states (36). WA's rates of poverty (18.2%) and health insurance coverage (93.6%) in children <18 years rank similarly (19th and 25th, respectively; ref. 36). Conversely, WA has comparatively greater educational attainment (31.1% of adults >25 years with a bachelor's degree or higher, ranked 11th; ref. 36). However, given prior work associating geographic distance or rurality with cancer survival outcomes in specific populations (18, 20), these findings may be applicable to other states, and possibly much of the U.S.

The demonstrated association of neighborhood-level poverty with mortality corroborates previously published survival disparities in multiple pediatric cancer types (8–12). This association may be reflective of the impact of household socioeconomic status on survival, although effects of the neighborhood environment itself—such as food deserts, healthcare resources, or community support—likely exert additional influence beyond individual household financial resources. However, our study design cannot evaluate the impact of household poverty itself. In addition, historical race–based discriminatory policies, such as redlining, have been known to shape neighborhood socioeconomic factors that are associated with poorer pediatric health outcomes (37). Although not the main focus of this study, survival disparities related to race and ethnicity were also identified in our multivariate analysis, consistent with existing literature (4–6). The impact of systemic racism on pediatric cancer mortality deserves continued attention.

The mechanisms by which rural/non-rural survival differences may occur are not fully understood. Travel time to cancer care could inherently contribute to poorer outcomes. Geographically distant families, in the U.S. and elsewhere, are known to spend significant time and resources on travel, which may increase stress and difficulty with treatment adherence (14, 38). The stronger associations with rurality observed among patients with ALL—for which treatment is largely outpatient, is longer in duration, and involves frequent travel—may support this, as other authors have also hypothesized (18). Barriers to accessing emergency care are also documented, and patients in communities with fewer pediatric healthcare resources could be at greater risk for complications such as severe infection (15), or therapy-related comorbidities that may contribute to later mortality differences. It is also possible that children in rural locations have less access to diagnostic resources, leading to delayed diagnosis and more advanced disease at presentation, as observed in adults (39). However, pediatric research to date has not necessarily supported this (21). Finally, rural-dwelling patients may experience greater difficulty accessing clinical trials or specialized treatment for relapsed or refractory cancer, access to which may mediate important survival disparities in some marginalized populations (40).

In adults with cancer, whereas clear disparities exist for rural patients, receiving treatment at an academic cancer center versus a community clinic is associated with improved survival and mitigates much of this risk (41). Such results are likely confounded by socioeconomic status or other factors, as rural adults traveling to central locations may be more affluent or have different functional status compared with those receiving local care. Given the centralized nature of pediatric cancer care (42), it is also likely that proportionately fewer rural children receive cancer treatment at non-tertiary centers compared with adults. However, our study did not explicitly confirm this, as the cancer registry data did not precisely define patients’ treating institutions. Further investigation will help elucidate mechanisms by which geographic survival disparities may develop.

Mechanisms underlying the observed effect modification between ADI and rurality are also unclear. Our observation of ADI-related survival disparities being greater in non-rural versus rural areas could indicate non-rural locations to be more sensitive to neighborhood poverty, whether due to the built environment, systemic discrimination or racism, public policy, or other factors. Conversely, survival disparities related to rurality—apparent in low ADI neighborhoods only—may suggest that rurality affects certain protective factors more commonly present in wealthier versus less-resourced areas. Finally, rural residence, itself associated with higher ADI (43), may be associated with other unmeasured socioeconomic disadvantage incompletely accounted for by ADI in our analyses. Ultimately, understanding these mechanisms will be critical to tailor interventions and policymaking to address these factors independently.

Interventions to address potential gaps are lacking in pediatrics, but initiatives to address other disparities may provide insight. For rural adults, telehealth and other strategies to decentralize specialty oncology care have been successful (44–46). In pediatric oncology, a program providing remote hospice support to rural populations was feasible (47). Similarly, increased participation from community centers for supportive care and oncologic emergencies could ameliorate difficulties that geographically remote families report during treatment (15). Rigorous financial support for housing and transportation is another promising strategy, the feasibility and effectiveness of which are currently being studied in pediatric populations with low household income (48). If effective, similar approaches could be considered for families traveling long distances for care.

This study has several limitations. First, socioeconomic status was measured at the census block group level but not at the household level, potentially understating the effect of household socioeconomic factors. Nevertheless, census block group data represent a more focused measurement of area-level socioeconomic status than used in many prior studies, reducing misclassification of neighborhood deprivation. Although our analysis adjusted for several important patient and disease factors—including treatment era, proxied by age at diagnosis and birth year—the extent of missing insurance data precluded its inclusion as a covariate. The relatively small samples sizes of some disease types should also prompt caution in the interpretation of our sub-analyses. In addition, our study time period consists largely of years preceding the Affordable Care Act, which may have benefitted healthcare access disproportionately for patients identified as rural and/or impoverished (49, 50), potentially limiting generalizability to the current era. Finally, measurements of census tract/block group data were also taken at a single point in time and may introduce misclassification. However, RUCA classifications in WA appear to be relatively stable over the study period (25).

In summary, we observed novel pediatric cancer survival disparities associated with rural residence in combination with the well-known impact of neighborhood socioeconomic disadvantage over a 20-year period in WA. We also observed effect modification between rurality and neighborhood socioeconomic disadvantage with regard to mortality. Future investigation is necessary to confirm these results in other U.S. populations, and to examine how rurality and socioeconomic factors impact other health-related outcomes such as hospital utilization, treatment complications/morbidity, health-related quality of life, financial toxicity, and clinical trial enrollment. Additional analyses incorporating geographic distance to care as a possible mechanism underlying these disparities are also warranted. Such strategies may elucidate mechanisms and mediators of survival disparities. Investigation into the variability of effects by cancer type is also merited. Ultimately, interventions to address these disparities will be necessary.

E.J. Chow reports grants from Alex's Lemonade Stand Foundation during the conduct of the study. No disclosures were reported by the other authors.

T.J.D. Ohlsen: Conceptualization, data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. D.R. Doody: Data curation, software, formal analysis, validation, methodology, project administration, writing–review and editing. B.A. Mueller: Data curation, funding acquisition, methodology, writing–review and editing. A.D. Desai: Conceptualization, supervision, methodology, writing–review and editing. E.J. Chow: Conceptualization, supervision, funding acquisition, methodology, writing–review and editing.

This work was conducted with grant support from the Alex's Lemonade Stand Foundation for Childhood Cancer awarded (to B.A. Mueller and E.J. Chow), as well as cancer registry support from #HHSN261201300012I, N01-CN-005230, N01-CN-67009, N01-PC-35142, and HHSN261201000029C from the NCI SEER program, with additional support from the Fred Hutchinson Cancer Center, and CDC #DP12–1205 DP003899–02.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.