Socioeconomic disparities in survival after high-risk neuroblastoma treatment with modern therapy

Abstract

Background:

Modern therapeutic advances in high-risk neuroblastoma have improved overall survival (OS), but it is unclear whether these survival gains have been equitable. This study sought to examine the relationship between socioeconomic status (SES) and OS in children with high-risk neuroblastoma, and to investigate whether SES-associated disparities have changed over time.

Procedure:

In this population-based cohort study, children <18 years diagnosed with high-risk neuroblastoma (diagnosis at age ≥12 months with metastatic disease) from 1991–2015 were identified through the National Cancer Institute’s Surveillance, Epidemiology, and End Results database. Associations of county-level SES variables and OS were tested with univariate Cox proportional hazards regression. For a sub-cohort diagnosed after 2007, insurance status was examined as an individual-level SES variable. Multivariable regression analyses with treatment era and interaction terms were performed when SES variables reached near-significance (p≤0.1) in univariate and bivariate modeling with treatment era.

Results:

Among 1,217 children, 2-year OS improved from 53.0±3.4% in 1991–1998 to 76.9±2.9% in 2011–2015 (p<0.001). In univariate analyses, children with Medicaid (hazard ratio [HR]=1.40, 95% confidence interval [CI]=1.05–1.86, p=0.02) and those in high-poverty counties (HR=1.74, CI=1.17–2.60, p=0.007) experienced an increased hazard of death. No interactions between treatment era and SES variables were statistically significant in multivariable analyses, indicating that changes in OS over time did not differ between groups.

Conclusions:

Low SES is associated with inferior survival in children with high-risk neuroblastoma. Survival disparities have not widened over time, suggesting equitable access to and benefit from therapeutic advances. Interventions to narrow existing disparities are paramount.

Introduction

Neuroblastoma is the most common solid extracranial tumor in childhood, with over 600 cases diagnosed per year in the United States.¹ High-risk neuroblastoma is associated with significant risk of relapse and death. However, advances in treatment for children with high-risk disease have led to impressive increases in survival over recent decades. Patients who receive the full complement of standard of care therapy (chemotherapy, radiation, surgery, autologous stem cell transplant (ASCT), and cytokine/immunotherapy) now experience a two-year overall survival (OS) as high as 86%, a striking survival gain over two decades.^2–5

While these therapeutic advances hold incredible potential for improving patient outcomes, differences in access along the cancer care continuum may simultaneously exacerbate disparities particularly in the setting of such complex and high-intensity treatments. Specifically, high-risk neuroblastoma treatment typically is limited to large referral centers that have the capacity to deliver ASCT, cytokine/immunotherapy, and advanced supportive care measures. This treatment lasts for around 18 months and consists of multiple prolonged inpatient admissions. Such a shift may disproportionately affect underserved patients through a variety of mechanisms including delays in diagnosis, geographic distance from referral centers, and differential or biased clinical trial enrollment on clinical trials that may cumulatively result in unequal survival gains.^6–8

A recent analysis of a heterogeneous population of patients with childhood cancer identified through the National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) database demonstrated that SES significantly mediated racial and ethnic survival disparities for a number of cancer diagnoses including neuroblastoma.⁹ This high-level analysis identified the applicability of SES to survival outcome disparities in pediatric cancer but did not examine disease-specific risk groups to differentiate outcome disparities due to stage versus SES-related access disparities. Other specific analyses of neuroblastoma cohorts have identified poverty as an independent risk factor for relapse and death and have found that minority patients have higher prevalence of high-risk disease.^10,11 However, these analyses were restricted to patients enrolled on clinical trials which poses limitations given the possibility that families of lower SES and without private insurance may have more difficulty accessing and participating in trials and experimental therapies. The SEER database, as a national population-based registry, mitigates this limitation. Our specific aims for this study were to: (1) describe differences in OS among patients with high-risk neuroblastoma by individual- and county-level SES characteristics; and (2) investigate whether changes in OS over time in patients with high-risk neuroblastoma differ by these characteristics.

Methods:

Selection of Study Population

The analytic cohort was selected using individual- and population-based cancer registry data from the National Cancer Institute’s SEER database using SEER*STAT 8.3.4 (Washington, D.C.). Pediatric patients (age <18 years) diagnosed with neuroblastoma from January 1, 1991 through December 31, 2015 were selected to allow for a minimum of 2-year follow-up at the time of SEER 8.3.4 release.¹² To approximate characteristics of high-risk (stage M) neuroblastoma according to the International Neuroblastoma Risk Group staging system used in most neuroblastoma clinical trials,¹³ we restricted analyses to patients with distant metastases and age ≥12 months at time of diagnosis. Based on data availability, for patients diagnosed from 1991–1992, we used the SEER-9 Registries; for patients diagnosed from 1992–2000, we used the SEER-13 Registries; for patients diagnosed from 2000 onwards, we used the SEER-18 Registries.¹⁴ Geographic distribution varies depending on era, with the most recent/expansive registries (SEER-18) covering approximately 28% of the U.S. population and including 18 geographic registries (Native Alaska; Connecticut; Detroit, Michigan; Atlanta, Georgia; rural Georgia; greater Georgia excluding the two previous regions; San Francisco-Oakland, California; San Jose-Monterey, California; Los Angeles, California; California excluding the previous metropolitan areas; Hawaii; Iowa; Kentucky; Louisiana; New Mexico; New Jersey; Seattle-Puget Sound, Washington; Utah).¹⁵ The study was deemed exempt from review by the Dana-Farber Cancer Institute’s Institutional Review Board (protocol 18-409).

Measures of Socioeconomic Status

We examined measures of SES at the county level for the entire cohort, and at the individual level for a sub-cohort of patients. SEER includes county-level variables from the American Community Survey (ACS) County Attributes data.¹⁶ County-level variables were determined based on the patient’s residency county code at diagnosis. SEER utilizes ACS data based on the cancer case/year of diagnosis. County-level measures of SES were chosen a priori based on previous SEER analyses and disparities literature.^17–19 We included county-based poverty (proportion of households living below the federal poverty level (FPL)), educational attainment (proportion of individuals in county >25 years of age with less than high school education), unemployment (proportion of individuals in county >16 years of age unemployed), language isolation (proportion of households in county with no household member age ≥14 years who speaks English), and urban-rural status (population >1 million vs. population of 250,000 to 1 million vs. population <250,000). These variables were defined per SEER and ACS.²⁰ To maximally highlight disparities should they exist, we dichotomized each county-level variable at the 90^th percentile cut-point (language isolation, education, unemployment) to define low-SES and high-SES cohorts. High-poverty counties were defined as those with greater than or equal to 20% of households living below 100% FPL.²¹ As a sensitivity analysis, we also analyzed county-based poverty using a cut-point at the 90^th percentile.

We examined individual-level insurance data for the sub-cohort of patients for whom it was available (diagnosed from 2007 onward) in addition to county-level. Insurance status was dichotomized as any public insurance coverage (i.e. Medicaid) versus non-Medicaid insurance (those with Medicaid as a second insurer were coded as Medicaid, per SEER database convention). Given its rarity in pediatrics, patients without documented insurance (N=14) were excluded. Patients with Medicaid insurance were a priori considered low-SES. Supplementary Table 1 details exact SEER variable names and descriptions.

Outcome

The primary outcome was OS, derived from SEER’s “Survival Months” attribute, defined as months from date of cancer diagnosis to date of death from any cause, censored at date of last contact. We used OS rather than cancer survival given its lack of ambiguity and frequency of use in oncology clinical trials,²² and the rarity of death from non-cancer causes among children with cancer. We reported 2-year OS given limitations of available follow-up data for our cohort at the time of analysis. Three-year OS, given its consistency with recent publications,^3,23 was included as supplementary data for the patients with sufficient follow-up data.

Covariates

Covariates included sex, race (white, black, or other), ethnicity (Spanish/Hispanic/Latino vs. non-Spanish/non-Hispanic/non-Latino) and diagnostic treatment era. We examined race and ethnicity as distinct constructs, given prior reports of differential health outcomes according to race and/or ethnicity.²⁴ Treatment eras were defined based on major advances in the standard-of-care for children with high-risk neuroblastoma: 1991–1998 (early treatment era) vs. 1999–2004 (multimodal treatment including ASCT)⁵ vs. 2005–2010 (improved supportive care) vs. 2011–2015 (immunotherapy)² to allow for exploration of the potential interaction between treatment era and SES.

Statistical analysis

Descriptive statistics were used to summarize baseline cohort characteristics. Kaplan-Meier curves of OS were generated for the overall cohort and stratified by county- and individual-level variables, as well as insurance for the post-2007 sub-cohort. OS was compared between groups using the log-rank test. Univariate Cox proportional hazards (Cox-PH) regression was used to test the association of each SES variable with OS. The proportional hazards assumption was tested by visually examining log-log plots and by testing the interaction of selected covariates with time. Our results indicated no violation of the proportional hazards assumption.

Bi-variable Cox-PH regression was used to jointly test the effect of each SES variable with (continuous) treatment era on OS, including SES variables with p≤0.1 in univariate analyses. Variables with p≤0.1 in bi-variable analyses were included in multi-variable models with treatment era as an interaction term to test for effect modification (i.e. any difference in change in OS over time between the SES groups). Multi-variable Cox-PH regression tested the SES variable, treatment era, and the interaction of treatment era with the SES variable. If the interaction term was statistically significant, this would provide evidence that there was a significant difference in the change in OS over time between examined SES groups.

Subjects with missing data (<2% in all measured variables) were excluded from analysis. Notably, insurance status was only available in SEER for those diagnosed after 2007. Thus, analyses of insurance were performed solely in the sub-cohort of patients diagnosed after 2007.

We performed multivariable regression analyses with county-level poverty, race, ethnicity, and treatment era to explore the relative contributions of these variables to survival. For the sub-cohort of patients with insurance data, we built a second multivariable model including insurance.

Statistical analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC). Two-sided p-values ≤ 0.05 were considered statistically significant. The data that support the findings of this study are publicly available through SEER.²⁵

Results:

Study population

The analytic cohort included 1,217 patients (Figure 1). 550 patients were diagnosed after 2007 and form the sub-cohort available for analyses of insurance status. Twenty-six percent of patients were non-White race and 20% were of Spanish/Hispanic/Latino ethnicity (Table 1). In the post-2007 sub-cohort (those for whom insurance data were available), 61% (338/550) were insured by Medicaid.

Figure 1.

Study Cohort

Table 1.

Individual- and County-level Patient Characteristics at Diagnosis (N = 1,217)

Individual-level variables	No. / N (%)
Sex
Female	505/1217 (41)
Male	712/1217 (59)
Race1
White	899/1211 (74)
Black	189/1211 (16)
Other minorities2	123/1211 (10)
Ethnicity
Non-Spanish/Hispanic/Latino	977/1217 (80)
Spanish/Hispanic/Latino	240/1217 (20)
Treatment era
1991 – 1998	215/1217 (18)
1999 – 2004	310/1217 (25)
2005 – 2010	391/1217 (32)
2011 – 2015	301/1217 (25)
County-level SES variables	Median (Range) or No. / N (%)
% of households in county below 100% FPL	9.17 (1.42–29.17)
% of individuals in county over the age of 25 years with a less than high school degree	16.57 (3.27–50.32)
% of households linguistically isolated	5.06 (0.00–19.17)
% of unemployment	7.37 (1.54–17.38)
Urban/rural status3
>1 million population	795/1204 (66)
250,000–1 million population	226/1204 (19)
<250,000 population	183/1204 (15)
Post-2007 sub-cohort	No. / N (%)
Insurance status
Other	212/550 (39)
Medicaid	338/550 (61)

¹Race unavailable for 6 individuals in cohort

²Other minorities include Asian or Pacific Islanders (N= 114) and American Indian/Alaska Natives (N=9)

³Urban/rural status unavailable for 13 individuals in cohort

Univariate analysis of SES factors and treatment era with OS

Median follow-up was 6.08 years [range=0–24.92 years]. OS improved by treatment era (p<0.001) with 2-year OS (±standard error) increasing from 53.0±3.4% to 76.9±2.9% between 1991–1998 and 2011–2015 (Figure 2). Similar trends were observed for 3-year OS for the patients with sufficient follow-up data (N=1032, Supplementary Table 2). In univariate analysis of OS (Table 2), increased hazard of death was seen in patients in high-poverty counties (≥20% of households below 100% FPL) compared to those in low-poverty counties (hazard ratio [HR]=1.74, 95% confidence interval [CI]=1.17–2.60, p=0.007). No other county-level SES factors were found to be statistically significant. In the post-2007 sub-cohort, individuals with any Medicaid experienced increased hazard of death compared to those with other insurance (HR=1.40, 95% CI=1.05–1.86, p=0.02).

Figure 2.

Kaplan Meier curves of Overall Survival (A. Overall Cohort (N=1217); B. Stratified by High- and Low-Poverty County (N=1217); C. Stratified by Insurance (post-2007 Sub-Cohort, N=550); D. Stratified by treatment era)

Table 2.

Univariate Analysis for Individual- and County-level Factors on Overall Survival (OS)

Prognostic Factors	No.	2-Year OS ± SE (%)	HR (95% CI)	P value
Sex
Female	505	68.6 ± 2.1	Reference	.66
Male	712	69.1 ± 1.8	1.04 (0.88, 1.22)
Race
White	899	69.3 ± 1.6	Reference	.46
Black	189	70.6 ± 3.4	0.94 (0.75, 1.18)
Other minorities	123	62.7 ± 4.6	1.15 (0.88, 1.49)
Ethnicity
Non-Spanish/Hispanic/Latino	977	69.0 ± 1.5	Reference	.10
Spanish/Hispanic/Latino	240	68.4 ± 3.2	1.18 (0.97, 1.43)
Treatment Era (trend)1
1991 – 1998	215	53.0 ± 3.4	0.79 (0.73, 0.86)	<.001
1999 – 2004	310	69.0 ± 2.6
2005 – 2010	391	72.9 ± 2.3
2011 – 2015	301	76.9 ± 2.9
% of households in county below 100% FPL
Low poverty (<20%)	1181	69.3 ± 1.4	Reference	.007
High poverty (≥20%)	36	55.4 ± 8.6	1.74 (1.17, 2.60)
% of individuals in county over the age of 25 years with a less than high school degree
<90th percentile (30%)	1115	69.0 ± 1.4	Reference	.48
≥90th percentile	102	68.6 ± 4.6	1.10 (0.85, 1.42)
% of households linguistically isolated
<90th percentile (13%)	1103	68.5 ± 1.5	Reference	.85
≥90th percentile	114	73.4 ± 4.4	0.97 (0.74, 1.29)
% of unemployment
<90th percentile (12%)	1102	68.3 ± 1.5	Reference	.54
≥90th percentile	115	74.9 ± 4.4	0.91 (0.68, 1.22)
Urban/rural status by population
>1 million	795	69.5 ± 1.7	Reference	.37
250,000–1 million	226	73.2 ± 3.1	0.94 (0.76, 1.16)
<250,000 population	183	63.6 ± 3.7	1.14 (0.91, 1.41)
Post-2007 sub-cohort
Insurance status
Other	338	79.6 ± 2.3	Reference	.02
Medicaid	212	69.9 ± 3.5	1.40 (1.05, 1.86)

¹Continuous treatment year was used in univariate analyses. HR indicates survival gain over time. For example, there is a 21% lower hazard of survival for patients diagnosed in 1999–2004 compared to those diagnosed in 1991–1998.

Multivariable analysis of change in OS by SES factors over time

In multivariable analysis, we included the SES variables that were near significant in bivariable analysis (p≤0.1), treatment era, and SES*treatment era interaction. None of these interactions were statistically significant (interaction p≥0.80), indicating that the changes in OS over time did not differ significantly between these SES groups. Hazard ratios with 95% confidence intervals are displayed in Figure 3.

Figure 3.

Hazard ratio plot with 95% CI of (A) Spanish/Hispanic/Latino and (B) High-poverty county in multivariate analyses controlling for treatment era and SES*treatment era interaction on overall survival. CI, confidence interval

Sensitivity analyses

A sensitivity analysis using the 90^th percentile as a cut-point to define a high-poverty county, while not statistically significant, did not differ greatly from the primary analysis (HR=1.15, 95% CI=0.89–1.49, p=0.27). In multivariable modeling, the interaction between treatment era and county-level poverty using this cut-point was similarly not statistically significant (p = 0.45).

Multivariable analyses including race and ethnicity in the model

In a multivariable regression model for the entire cohort considering county-based poverty, race, ethnicity, and treatment era, there was an increased hazard of death associated with higher county-based poverty (HR=2.08, p<0.001, Table 3) and a lower hazard of death associated with later treatment era (HR=0.78, p<0.001). Race and ethnicity were not significantly associated with survival.

Table 3.

Post-hoc Multivariable Models Including Race, Ethnicity, and Treatment Era

Individual and county level variables	Hazard Ratio	Adjusted P-value
Entire cohort (N = 1211) 1
County poverty dichotomized at ≥20% of households below poverty	2.08	<0.001
Race	Global p: 0.430
White	Reference	Reference
Black	0.98	0.85
Other	1.18	0.22
Spanish/Hispanic/Latino ethnicity	1.22	0.05
Treatment era	0.78	<.001
Post-2007 sub-cohort (N = 548) 2
County poverty dichotomized at ≥20% of households below poverty	2.38	0.001
Race	Global p: 0.542
White	Reference	Reference
Black	0.81	0.30
Other	1.07	0.81
Spanish/Hispanic/Latino ethnicity	1.07	0.72
Medicaid insurance	1.38	0.04
Treatment era	0.99	0.95

¹Race unavailable for 6 individuals in cohort

²Race unavailable for 2 individuals in sub-cohort

In the post-2007 sub-cohort, higher county-based poverty (HR=2.38, p=0.001) and any Medicaid insurance (HR=1.38, p=0.04) were statistically significantly associated with increased hazard of death. Race, ethnicity, and treatment era were not statistically significant. Treatment era was included in this second model for consistency although this sub-cohort notably only includes patients diagnosed after 2007, limiting power of this analysis.

Discussion:

In a representative population of U.S. children with high-risk neuroblastoma, children living in high-poverty counties experienced a 74% increased hazard of death compared to those living in low-poverty counties. In a sub-cohort of children with available insurance data, those with any Medicaid insurance experienced a 40% increased hazard of death compared to those with other insurance. As reported elsewhere⁴, OS for the entire cohort improved steadily over time (from 53% to 77%). Notably, while SES-related survival disparities persisted over time, they did not widen.

These findings build on prior work identifying SES-associated survival disparities in children with high-risk neuroblastoma treated on clinical trials¹¹ by investigating whether outcome disparities have increased with advances in care. Given the resource-intense and highly centralized nature of modern high-risk neuroblastoma treatment, this exemplifies the dual potential of precision medicine to improve outcomes while simultaneously worsening health disparities if these advances are not delivered equitably.²⁶ Although analyses of this cohort are limited by absence of patient-level treatment data in the SEER database, the equitable survival gains across all groups over time are encouraging and stand in contrast to widening disparities observed in other populations (e.g. asthma, adult cancers).^27,28 These data suggest that access to treatment advances with known survival benefits has been generally equitable among children with high-risk neuroblastoma, perhaps due to the high reliance on clinical trial enrollment and delivery of protocolized care in pediatric oncology.

That U.S. children of lower SES with high-risk neuroblastoma continue to die at higher rates than their higher SES counterparts must also be highlighted.¹¹ Despite focused policy statements and advocacy efforts identifying poverty as a major determinant and predictor of adverse health outcomes in children, we observed persistent and clinically meaningful survival disparities associated with SES. Recent studies in other cancer populations have similarly found insurance and neighborhood (i.e. county-based) poverty to be predictors of inferior outcomes.^29,30 Of note, children with high-risk neuroblastoma receive intensive (largely inpatient) treatment for 18-months at specialized, tertiary care centers. Therefore, this population is in many ways optimally positioned to minimize disparities in care. Our finding that survival disparities persist even in this population suggests that there may be other fundamental mechanisms driving SES-related gaps, warranting exploration of mechanisms beyond access to care.

Non-adherence to chemotherapy is a mechanism leading to disparate outcomes that has been well-described in pediatric acute lymphoblastic leukemia (ALL).^31,32 Future work could investigate if there are similar findings in high-risk neuroblastoma with outpatient administered GM-CSF injections and oral cis-RA. However, it is important to note that ALL treatment (which includes two years of outpatient-based therapy) contrasts sharply with the predominantly inpatient treatment of high-risk neuroblastoma. The intensity of high-risk neuroblastoma treatment including multiple prolonged admissions does raise the question of whether these admissions place excess and disproportionate burden on lower SES families who may have lower resources for childcare, missed work/compensation, and household material needs. A better understanding of the underlying mechanisms leading to the observed disparities will be critical to inform subsequent interventions. Some promising potential SES-based interventions across various healthcare settings include systematic provision of grocery and transportation services, clinic-based screening and referrals for community-based resources for a range of basic material needs, and free clinic-embedded tax services to assist families in obtaining earned income tax credits.^33–35

Our study benefits from a robust, population-based sample across more than two decades of high-risk neuroblastoma treatment, allowing for analysis of changes in survival over time. The SEER database is uniquely positioned as a population-based registry to provide disparities data as it is intentionally biased to oversample minority populations and those that have traditionally been underrepresented in clinical trials.³⁶ Interestingly, we did not identify racial and ethnic disparities observed in other neuroblastoma cohorts.^9,10 Importantly, our cohort differs from these previous studies in that we restricted analyses to those with high-risk disease, which is more prevalent in minority populations. As such, our data may thus reflect prior findings that inferior survival observed for Black patients compared to their white counterparts was attributable to their higher prevalence of high-risk disease at diagnosis.¹⁰ Our data build on recent publications demonstrating that SES mediates racial and ethnic survival disparities across pediatric cancer,⁹ and that low-SES is associated with inferior survival in the context of modern-era clinical trials.¹¹ These disparities disproportionately impact children of racial and ethnic minority status who disproportionately live in low-SES households due to structural disadvantages and biases. We importantly find that these disparities are not, however, worsening in the modern era of complex treatment delivery. Future attention to characterizing the relationships between SES and outcomes is essential to begin to narrow the survival gap.

There are important limitations to our data. Inherent to any large registry, SEER data are limited by missing data/unrecorded variables, coding reliability, and selection bias.³⁷ We utilized a proxy for high-risk (stage M) neuroblastoma in the absence of histological, genetic, and staging variables in SEER. While this proxy definition approximates elements of modern staging criteria, we may have misclassified children with lower risk disease. We similarly lacked access to patient-level treatment data, though our findings are consistent with SES-associated disparities in the clinical trial setting. Finally, we had access to individual-level (insurance) SES data for only the sub-cohort of patients diagnosed after 2007, limiting our ability to consider the impact of insurance across all treatment eras in this analysis. Given the magnitude of the effect of insurance status on survival, however, similar findings would be expected prior to 2007. SEER also codes patients with Medicaid-only (e.g. based on income eligibility) and Medicaid as a second-insurer identically. Consequently, some patients with Medicaid as a second insurer may have been misclassified as low-SES, an error which would bias toward the null, lending additional weight to our finding of survival disparities according to insurance status.

While significant SES-related survival disparities exist for children with high-risk neuroblastoma in the United States, these disparities have not widened over time. These findings suggest that the highly centralized and structured care delivery model of pediatric oncology allows for equitable integration of novel therapies into the standard of care. They suggest a model of care that could be applied to other patient populations—in oncology and more generally—for whom novel therapies and other resource-intensive treatment modalities are entering the clinical space. At the same time, low-SES patients continue to die at higher rates than their higher-SES counterparts, and efforts to reduce this inequity are essential. The complex interplay between poverty and other social factors, race/ethnicity, and cancer begs for clearer understanding so that we might better address differences in healthcare access and outcomes. Future investigations should focus on both elucidating mechanisms underlying SES-related disparities among oncology patients accessing complex but standardized treatments, as well as designing targeted interventions to improve OS in vulnerable groups.

Abbreviations Key:

OS

Overall survival

SES

Socioeconomic status

HR

Hazard ratio

ASCT

Autologous stem cell transplant

SEER

Surveillance, Epidemiology, and End Results

ACS

American Community Survey

FPL

Federal poverty level

Cox-PH

Cox proportional hazards