Part XVI · Disparities, access, and health systems · Chapter 89

Disparities across the breast cancer continuum

Race, ethnicity and ancestry are three variables. Almost every claim in this literature depends on which one was actually measured.

1 · Race, ethnicity, and ancestry as three different variables

Host and tumor interactions ended with allostatic load. Among 4459 patients with stage I to III breast cancer, high allostatic load carried a hazard ratio of 1.46 for all-cause mortality against low allostatic load¹. That finding arrives in this part as biology, not as an addendum. It also arrives with a warning, because the same study found mean allostatic load higher in Black patients. A measured biological difference between groups is not evidence about the cause of the difference.

Three variables get used interchangeably in this literature, and they are not interchangeable.

Race is a social classification. Its categories are administrative, they are assigned or self-reported on a form, and they have been redefined repeatedly by the bodies that collect them. Ethnicity is a separate self-identified category of cultural or national origin. In United States datasets it is collected on a separate question, so a person has a race and an ethnicity rather than one combined label. Genetic ancestry is the estimated proportion of an individual genome derived from specified reference populations. It is continuous, it is estimated from markers, and it does not come in categories at all.

The measurement distinction is the whole chapter. Race is recorded. Ancestry is estimated from a sample. The two correlate and they do not coincide, and a study that classified participants by self-report has not measured ancestry however its discussion section is worded.

Structural racism is a fourth thing again. It refers to the mutually reinforcing systems of housing, education, employment, earnings, credit, health care and criminal justice through which societies distribute disadvantage by race². It is an exposure. Race is a label that predicts who receives that exposure.

2 · Structural racism and social determinants as measurable exposures

The claim that structural racism affects breast cancer outcomes is only useful if it can be measured. Measurement means choosing an indicator, and each available indicator stands for a different mechanism.

Historical mortgage lending discrimination is measured through archival lending maps. Contemporary residential segregation is measured through indices of dissimilarity and isolation computed from census tracts. Neighbourhood deprivation is measured through composite indices of income, education and employment. Individual experience of discrimination is measured through validated questionnaires. Insurance status is measured administratively. These are not four ways of measuring one exposure. They are four exposures.

The distinction matters because they enter the causal chain at different points and at different times. A lending map drawn in the 1930s acts through everything that happened to a neighbourhood since. An insurance category acts at the moment of diagnosis. An intervention aimed at one has no mechanism through which it could act on the other.

Social determinants of health are the upstream conditions. Health-related social needs are the individual-level consequences that present in clinic, and they are treated separately in Health-related social needs, definitions and data collection because they are collected differently and mean something different.

What makes these exposures rather than descriptions is that they have measurable biological readouts. Allostatic load indexes the accumulated physiological cost of adaptation across neuroendocrine, metabolic, inflammatory and cardiovascular systems³. In breast cancer it predicts mortality after adjustment for sociodemographic, clinical and treatment factors, with a dose-dependent gradient across quartiles¹. The pathway from exposure to outcome is partly traced. What is not quantified is how much of any particular gap each pathway carries.

3 · Redlining, residential segregation, and place-based exposure

The Home Owners' Loan Corporation graded American neighbourhoods in the 1930s on a four-point scale from A, best, to D, hazardous. The D grade was assigned substantially on the racial composition of the residents. Those maps governed mortgage access for decades. They are now used as an archival measure of a historical policy exposure attached to a place.

Two population-based registry studies illustrate both the strength and the trap of this design.

In New Jersey, 14 964 women with invasive breast cancer diagnosed between 2008 and 2017 were assigned the grade of their residential area. Compared with residence in a D-graded area, residence in an A-graded area carried an odds ratio of 0.34 for late-stage diagnosis, 0.67 for triple-negative subtype, and a hazard ratio of 0.48 for breast cancer death. Every one of those associations was present only among non-Latina White women. There was no evidence of association among non-Latina Black or Latina women⁴.

In New York State, 60 773 breast cancer cases diagnosed between 2008 and 2018 showed a graded relationship. Compared with an A grade, residence in B, C and D graded tracts carried 29%, 37% and 64% higher five-year all-cause mortality respectively. The associations persisted after adjustment for health insurance and treatment received. The elevated risk in D-graded areas was again observed specifically among non-Hispanic White cases⁵.

4 · Multilevel frameworks for disparities research

A multilevel framework organises determinants by the level at which they act. The distal level covers policy, economics and the social environment. The intermediate level covers neighbourhood, institution and social network. The proximal level covers individual biology, behaviour and the clinical encounter⁶. The framework's value is procedural. It forces a study to say which level its variables sit at, and therefore which level an intervention would have to act on.

The analytic method that follows is sequential adjustment. Variables are added to a model one level at a time, and the change in the estimate at each step is reported.

A study of 46 185 abnormal screening mammograms in 45 186 women at 109 facilities across six states is the clearest worked example in breast cancer. Compared with White women, Black women had a relative risk of 1.28 for not reaching biopsy within 90 days. Adjustment for individual, neighbourhood and non-facility health care factors moved it to 1.27. Adding the screening facility itself moved it to 1.20⁷.

Read that sequence carefully, because the lesson is not the size of the change. Individual and neighbourhood characteristics explained almost none of the difference. Which facility a woman was screened at explained a modest part of it. Most of it remained unexplained after every measured level had been entered.

An unexplained residual is not a biological residual. It is the part of the association that the measured variables did not capture. Treating it as evidence of an intrinsic difference is the exact inferential move that How disparities research can itself cause harm is about.

5 · Health-related social needs, definitions and data collection

A health-related social need is an individual-level, immediate and in principle actionable circumstance affecting health. The standard domains are food insecurity, housing instability, transportation difficulty, utility insecurity and interpersonal safety. These are the clinic-facing counterparts of the population-level social determinants described in Structural racism and social determinants as measurable exposures.

The distinction is not pedantic. A social determinant is a property of a population and is addressed by policy. A social need is a property of a patient and is addressed by referral. Collecting the second and calling it the first is a common category error in quality programmes.

Collection is by self-administered or interviewer-administered instrument, usually at registration or at a survivorship visit. Three properties of the instrument change the answer.

The first is language. In a primary care practice serving a predominantly Asian American population, 236 patients completed a screening tool offered in English and in Chinese. Half reported at least one need. Transportation needs were highest among those responding in traditional Chinese text, while food insecurity and housing instability were higher among those responding in simplified Chinese text⁸. Two groups given the same instrument in two scripts reported different needs. Language of administration is not a neutral formatting choice.

The second is setting. Needs reported at a registration desk and needs reported to a known clinician differ in disclosure.

The third is whether a referral pathway exists. Screening without a pathway measures need and does not meet it, and it imposes disclosure cost on the patient with nothing returned.

6 · Decomposing an outcome gap into biology, access, and delivery

This section carries the argument of the part. An outcome gap between populations has three possible sources. The tumours may differ. Access to diagnosis and treatment may differ. What the health system does to patients once they are inside it may differ. These are separable, and separating them requires study designs that hold one of the three fixed.

Start with the gap. Black women in the United States have 38% higher breast cancer mortality than White women despite 5% lower incidence. They are diagnosed with localised disease less often, 58% against 68%. They have the lowest survival of any racial or ethnic group for every breast cancer subtype and every stage except localised disease⁹.

The last clause is the one that matters. A gap explained by subtype distribution would close within subtype. This one does not.

Now hold access partly fixed. Among 177 075 women aged 40 to 64 diagnosed with stage I to III disease, 20% of those uninsured or covered by Medicaid presented at stage III, against 11% of the insured. Non-Hispanic Black women had an odds ratio of 1.46 for stage III presentation against non-Hispanic White women. Adjustment for insurance and other socioeconomic factors reduced it to 1.29. Insurance status mediated 45% to 47% of the racial difference in odds of locally advanced presentation¹⁰. Roughly half of the stage gap is coverage. Roughly half is not.

Now hold delivery fixed. TAILORx enrolled 9719 women with hormone receptor-positive, HER2-negative, node-negative disease, with adjuvant treatment assigned by protocol according to the 21-gene recurrence score. There was no substantial difference in 21-gene recurrence score, or in ESR1, PGR or HER2 expression, by race or ethnicity. Black women still had a hazard ratio of 1.60 for distant recurrence against White women, with limits of 1.07 to 2.41. For overall survival in the whole population it was 1.41, with limits of 1.05 to 1.90¹¹.

Now hold coverage fixed. The United States Military Health System covers care universally. Among 998 non-Hispanic Black and 3899 non-Hispanic White women with stage I to III disease, median times to surgery were 22 and 21 days. Among women who had breast-conserving surgery, Black women had a hazard ratio of 1.45 for death, with limits of 1.06 to 2.01. Adding time to surgery to the model did not change that estimate¹².

Four datasets, four different things held fixed, and a residual gap in each. That pattern is what rules out any single-factor explanation. It does not identify what remains, and no honest reading of these four studies claims that it does.

7 · Subtype distribution and young-onset disease across populations

Subtype distribution differs between populations, and the size and direction of the difference depend on age.

In the Carolina Breast Cancer Study, immunohistochemical surrogates were applied to 496 incident invasive cancers in a population-based sample that oversampled premenopausal and African American women. The basal-like subtype accounted for 39% of tumours in premenopausal African American women, against 14% in postmenopausal African American women and 16% in non-African American women of any age. Luminal A accounted for 36%, against 59% and 54%¹³. The HER2-positive, oestrogen-receptor-negative subtype did not vary by race or menopausal status.

Note what varies. It is not race alone. It is the interaction of race with menopausal status, which is why studies that do not stratify by age report smaller and inconsistent differences.

Incidence in young women is rising. Breast cancer incidence rose 1.4% annually in women younger than 50 between 2012 and 2021, against 0.7% annually in women aged 50 and over. That steeper rise in the younger group was statistically significant only among White women, which is the sort of qualification this chapter drops at its peril. Asian American and Pacific Islander women had the fastest increase in both age groups. Among women younger than 50, their rate rose from 57.4 per 100 000 in 2000 to 86.3 per 100 000 in 2021. That places them alongside White women at 86.4 and above Black women at 81.5⁹.

Two conclusions follow and they point in opposite directions, which is why both belong here.

Subtype distribution is a real contributor to the mortality gap between Black and White women in the United States. It is also insufficient, because Black women have the lowest survival within every subtype⁹. A composition effect that vanishes on stratification has been explained. This one does not vanish.

8 · Disaggregating Asian American, Pacific Islander, American Indian, and Hispanic populations

Aggregated categories are data-collection conventions. They are not populations, and the aggregate statistic can be the opposite of every statistic inside it.

Among 910 415 women with breast cancer in Surveillance, Epidemiology and End Results registries, 63 405 were recorded as Asian. As a single category, Asian women had the highest ten-year overall and cancer-specific survival of any racial or ethnic group. Disaggregated by origin, Southeast Asian women, meaning Cambodian, Laotian, Hmong or Thai, had the highest proportion presenting at stage IV at 5.9%, against 2.7% among Japanese women. Ten-year unadjusted cancer-specific survival was 89.4% among Japanese women, with limits of 88.7% to 90.1%, and 78.0% among Southeast Asian women, with limits of 74.1% to 81.3%. After adjustment, Southeast Asian women were the only Asian subgroup without better overall survival than White women¹⁴.

The aggregate reported a favoured population. One constituent group had outcomes as poor as any group in the dataset. The aggregate did not merely obscure that. It reversed it.

The same problem operates elsewhere with different arithmetic. Breast cancer mortality among American Indian and Alaska Native women has not changed since 1990, in a period when the overall breast cancer death rate fell by 44%⁹. That group is small in national datasets and is frequently combined or suppressed for cell-size reasons, which means the flat trend is often absent from the summary a reader sees.

Hispanic ethnicity is collected separately from race in United States datasets, so a Hispanic population contains people of several recorded races and several national origins. Reporting it as one stratum is the same operation performed on a different axis.

9 · Gene and environment interaction, social genomics, and biological embedding

Social adversity changes gene expression, and the change has a characterised signature.

Across urbanity, low socioeconomic status, social isolation, social threat and low or unstable social status, leukocyte transcriptomes show a common pattern. Proinflammatory gene expression is increased. Expression of genes involved in innate antiviral response and antibody synthesis is decreased. The pattern has been named the conserved transcriptional response to adversity, and the neural signal transduction pathways that produce it have been partly mapped¹⁵. The same programme has been detected in diseased tissue, including metastatic cancers.

This is what biological embedding means. It is not a metaphor. It is a measured transcriptional state produced by a social exposure.

The consequence for this chapter is precise, and it is the reason the section exists. A measured biological difference between two populations is entirely compatible with a purely social cause. An inflammatory signature, an allostatic load score, a cortisol rhythm and a leukocyte transcriptome are all downstream of exposure. None of them is evidence of a germline difference, and none of them licenses a claim about ancestry.

The candidate mediators connecting this to tumour behaviour are set out in Host and tumor interactions rather than here. Chronic adrenergic signalling, glucocorticoid rhythm and metabolic state each have demonstrated tumour-level effects. What remains unquantified is how much of any observed outcome gap each mediator carries.

The reverse inference deserves naming because it is common. Finding that a biological measure differs between racial groups is often reported as evidence that the gap is biological rather than social. The social genomics literature shows that the premise does not support the conclusion. A biological measurement tells you where in the body the exposure landed. It does not tell you what the exposure was.

10 · Ancestry-associated tumor biology, what replicates and what is confounded

This is the section where the discipline of the part is tested, because the findings are real and the most common description of them is wrong.

Begin with what was measured. In nearly all of the studies below, participants were classified by self-reported race. Ancestry was not estimated. The results are therefore race-associated. Calling them ancestry-associated is a substitution the data do not support.

One analysis combined institutional cohorts with The Cancer Genome Atlas, covering 663 to 711 White and 105 to 159 African American women. TP53 mutations were found in 42.9% of tumours from African American women against 27.6% from White women. PIK3CA mutations were found in 20.0% against 33.9%. Intratumour genetic heterogeneity was greater in tumours from African American women by 5.1 units, with limits of 2.4 to 7.7. PAM50 basal tumours accounted for 39.0% against 18.6%¹⁶.

A separate Cancer Genome Atlas analysis of 1104 patients reported the same directions. TP53 was mutated in 46% against 27% and PIK3CA in 23% against 34%. The same study then restricted to triple-negative disease, and within that subtype found no racial difference in mutation counts or in any high-prevalence gene¹⁷.

Whole-exome and RNA sequencing of tumours from 462 African American women with triple-negative breast cancer went further. The mutational portrait resembled that described in Asian and non-Hispanic White women. The study tested African ancestry directly and found no evidence of association between mutational features and ancestry proportion¹⁸.

A fourth study targeted the confounding rather than the association. It compared 357 participants in the Carolina Breast Cancer Study, 52% of them Black, with 948 Cancer Genome Atlas cases, 18% of them Black. Black participants were more likely to carry TP53 and FAT1 mutations and less likely to carry PIK3CA, CDH1, DDR2 and GATA3 mutations. Total aneuploidy was higher. Socioeconomic status was associated with different patterns of alteration in Black against non-Black women¹⁹.

Put together, the honest summary is narrow. The differences replicate across unselected cohorts. They attenuate or disappear within triple-negative disease. They vary with socioeconomic status. The one study that measured ancestry rather than race found no ancestry association. Much of what is described as ancestral tumour biology is subtype composition, reported at the level of the whole cohort.

Now the consequence, which is the point of putting this section in this chapter.

PIK3CA mutation is not only a biological observation. It is an eligibility criterion. Entry to INAVO120 required a PIK3CA alteration. SOLAR-1 enrolled two cohorts by mutation status and found benefit only in the mutated one. Approval of the drugs tested in both is restricted to tumours carrying the alteration, and they are prescribed on that basis^20,21. A population with a lower frequency of PIK3CA mutation has, by arithmetic alone, a lower rate of eligibility for that drug class. The treatment landscape it governs is developed in Endocrine resistance, and the presentations the sequence does not fit and Expression and alteration-defined treatment across subtypes.

That is where biology and access stop being separable. A difference in tumour genotype becomes a difference in what can legally be prescribed. No unfairness in the health system is required to produce it, and no amount of navigation, insurance or proximity to a cancer centre corrects it. It is also the strongest available argument for enrolling representative populations in the trials that define these criteria, since a biomarker-defined indication is only as generalisable as the population that established it.

11 · Screening uptake, diagnostic delay, and time to treatment

Three intervals get collapsed into the phrase diagnostic delay, and they have different causes. Whether the screening test happened. How fast an abnormal result was resolved. How fast treatment started after diagnosis.

The second interval has the cleanest data. Among 46 185 abnormal screening mammograms recommended for biopsy, 34.6% were not resolved within 30 days, 16.2% within 60 days and 12.2% within 90 days. Compared with White women, the relative risk of no biopsy within 30 days was 1.52 for Black women, 1.66 for Asian women and 1.50 for Hispanic women. At 90 days, elevated risk persisted significantly only for Black women, at 1.28 with limits of 1.11 to 1.47⁷.

The pattern across time windows is informative. Delay at 30 days affected several groups. Delay still present at 90 days was specific to one. Those are probably different mechanisms, and a single summary interval would have hidden both.

The third interval behaves similarly. In a population-based cohort of 2998 patients with invasive breast cancer, half of them Black, 35% had treatment initiation more than 30 days after diagnosis and 48% had prolonged treatment duration. Long travel time to both diagnosis and surgery was associated with delayed treatment only among Black women²².

The comparison that constrains all of this comes from a universal-coverage system. In the Military Health System the median time to surgery differed by one day between non-Hispanic Black and non-Hispanic White women. At the 75th percentile the difference was 3.6 days and at the 90th percentile 8.9 days¹².

12 · Guideline-concordant treatment receipt and completion

Receipt and completion are different outcomes with different determinants, and quality programmes routinely measure only the first.

Receipt is an event at a clinic. It depends on referral, on insurance authorisation and on whether the indicated test or treatment is available at the site. Completion is a process lasting months or years. It depends on transport, employment, childcare, toxicity management and cost, all of which are developed in Access, affordability, and financial toxicity.

Molecular testing shows the receipt problem plainly. Among patients with early-stage, oestrogen receptor-positive, HER2-negative disease in a population-based cohort, 62% had no 21-gene assay performed²². The assay changes whether chemotherapy is recommended. A patient who is not tested is not receiving individualised risk assessment, whatever the guideline says.

Completion shows the other problem. In a cohort of 8769 women prescribed adjuvant hormonal therapy, early discontinuation and non-adherence were common, and both were associated with patient characteristics rather than with tumour characteristics²³. Five years of daily medication is a delivery problem, not a prescribing decision.

The constraint on all of this comes from TAILORx. Where treatment was protocol-assigned and therefore concordant by construction, the outcome gap persisted¹¹. Guideline-concordant receipt is necessary. It is demonstrably not sufficient.

13 · Clinical trial representation and the mechanisms of exclusion

Representation in the trials that define standard of care is measurable, and the mechanisms that produce it are measurable separately.

Across 230 trials supporting United States Food and Drug Administration oncology drug approvals between July 2008 and June 2018, 63.0% reported at least one race. Only 7.8% reported all four major United States categories. Among the 112 293 enrolled participants, Black patients were 3.1%, Hispanic 6.1%, White 76.3% and Asian 18.3%. Measured against each group's share of United States cancer incidence, Black participants were 22% of expected and Hispanic participants 44% of expected, against 98% for White participants²⁴.

The proportional comparison is the one to quote. A raw percentage of participants confounds representation with the demographic composition of the country, and the expected-share figure does not.

Breast cancer specifically is no better. TAILORx enrolled 693 Black women among 9719 eligible participants, which is 7.1%¹¹.

Measuring the problem inside a trial network is a separate undertaking from measuring it across approvals. DARE is that undertaking in breast cancer, recording demographic representation and health-related social needs across the sites of a cooperative network rather than in the published results those sites produce. The distinction matters because the mechanisms below are properties of sites, and an approval-level count cannot see them.

Now the mechanism, in sequence. A meta-analysis of 13 studies covering 8883 patients traced the decision pathway. A trial was not available at the patient's own institution for 55.6% of patients, with limits of 43.7% to 67.3%. A further 21.5% were ineligible for the trial that was available. Then 14.8% were offered a trial and declined, and 8.1% enrolled. Structural and clinical barriers made participation unachievable for more than three of every four patients²⁵.

Read that ordering. Patient decision is the fourth step, and more than three quarters of patients never reach it.

The clause that this section exists to carry is the one about the fourth step. A systematic review of 20 studies reporting consent rates by race, covering the enrolment decisions of more than 70 000 individuals, found no meaningful difference in willingness. Across ten clinical intervention studies, African American patients consented at 45.3% against 41.8% for non-Hispanic White patients, an odds ratio of 1.06 with limits of 0.78 to 1.45. Hispanic patients consented at 55.9% against 41.8%, an odds ratio of 1.33 with limits of 1.08 to 1.65²⁶.

When a trial is offered, willingness to enrol does not differ by race. Where the difference is generated is in whether a trial exists at the site, whether the eligibility criteria admit the patient, and whether it is offered. Those are properties of the research system rather than of patients, and each of them is modifiable. The consequences for the interpretation of trial results belong to Clinical trials and evidence interpretation.

14 · Intersectionality, rural, immigrant, and underserved populations

Disadvantage does not stack additively, and the studies in this chapter demonstrate it repeatedly rather than asserting it.

Historical redlining grade predicted stage, subtype and mortality among White women and not among Black or Latina women in the same registries^4,5. Long travel time to diagnosis and to surgery predicted delayed treatment among Black women and not in the cohort overall²². A screening instrument produced different need profiles depending on which Chinese script the respondent chose⁸. In each case the effect of one axis was conditional on another.

That is what intersectionality means operationally. It is not a claim that some patients have more problems. It is a claim that a coefficient estimated in a pooled model may not apply to any subgroup, and that stratified estimates can differ in sign.

Rural residence is the axis most often reduced to a single variable. It bundles several distinct exposures: distance to the nearest mammography unit, distance to the nearest surgical and radiation facility, local workforce availability, and the economics of small hospitals. Geographic access is developed in Access, affordability, and financial toxicity, and the same bundling problem recurs there.

Immigration status carries its own set. Language of clinical encounter, eligibility for public coverage, documentation concerns affecting willingness to enter systems, and prior country of care all act at different points. Country of birth is routinely collected and routinely discarded in analysis, which forces migration-related exposure into a race or ethnicity variable that cannot represent it.

The practical implication is a reporting one. A study that pools and adjusts has estimated an average that may describe nobody. A study that stratifies loses power and gains meaning. Which is right depends on whether the question is how large is the gap, or where does the gap come from.

15 · How disparities research can itself cause harm

Research into group differences can do damage, and the mechanisms are specific enough to name.

The first is attributional closure. A gap attributed to biology stops generating questions about access. This is the reason Decomposing an outcome gap into biology, access, and delivery insists on holding one factor fixed at a time, and the reason Ancestry-associated tumor biology, what replicates and what is confounded states what was measured before it states what was found.

The second is the conversion of an observation into an instruction. Race correction in clinical algorithms took an observed association between a recorded racial category and a measured quantity and embedded it as a coefficient. The output then assigns different predicted values to two otherwise identical patients, and clinicians act on the output. Corrections built this way have operated across nephrology, cardiology, obstetrics, urology and oncology, frequently in the direction of assigning fewer resources²⁷. The failure is not statistical. A coefficient describing a population becomes a rule applied to an individual.

The third is publication of unreplicated ancestral claims. The ancestry-associated tumour biology literature contains findings that attenuate within subtype and that do not reproduce when ancestry is measured rather than inferred^17,18. Claims that enter clinical folklore are not withdrawn when the replication fails.

The fourth is aggregation. Reporting Asian American women as one category concealed a constituent group whose adjusted survival was no better than that of White women, in a dataset where the aggregate looked favourable¹⁴. A population with good aggregate outcomes attracts no programme funding, and the subgroup inside it stays invisible.

The fifth is deficit framing. A literature that describes a population only through its worse outcomes teaches clinicians an expectation. Expectation affects what is offered, and what is offered is the single largest determinant of trial participation²⁵.

References

All 27 references

1. Obeng-Gyasi S, Elsaid MI, Lu Y, et al. Association of allostatic load with all-cause mortality in patients with breast cancer. JAMA Netw Open 2023 6:e2313989. PMID 37200034
2. Bailey ZD, Krieger N, Agenor M, Graves J, Linos N, Bassett MT. Structural racism and health inequities in the USA: evidence and interventions. Lancet 2017 389:1453-1463. PMID 28402827
3. McEwen BS. Protective and damaging effects of stress mediators. N Engl J Med 1998 338:171-179. PMID 9428819
4. Plascak JJ, Beyer K, Xu X, Stroup AM, Jacob G, Llanos AAM. Association between residence in historically redlined districts indicative of structural racism and racial and ethnic disparities in breast cancer outcomes. JAMA Netw Open 2022 5:e2220908. PMID 35802373
5. Lima SM, Palermo TM, Lee FF, et al. Historical redlining and all-cause survival after breast cancer diagnosis. Cancer Epidemiol Biomarkers Prev 2025 34:904-913. PMID 40178940
6. Warnecke RB, Oh A, Breen N, et al. Approaching health disparities from a population perspective: the National Institutes of Health Centers for Population Health and Health Disparities. Am J Public Health 2008 98:1608-1615. PMID 18633099
7. Lawson MB, Bissell MCS, Miglioretti DL, et al. Multilevel factors associated with time to biopsy after abnormal screening mammography results by race and ethnicity. JAMA Oncol 2022 8:1115-1126. PMID 35737381
8. Tsui J, Yang A, Anuforo B, et al. Health related social needs among Chinese American primary care patients during the COVID-19 pandemic: implications for cancer screening and primary care. Front Public Health 2021 9:674035. PMID 34123992
9. Giaquinto AN, Sung H, Newman LA, et al. Breast cancer statistics 2024. CA Cancer J Clin 2024 74:477-495. PMID 39352042
10. Ko NY, Hong S, Winn RA, Calip GS. Association of insurance status and racial disparities with the detection of early-stage breast cancer. JAMA Oncol 2020 6:385-392. PMID 31917398
11. Albain KS, Gray RJ, Makower DF, et al. Race, ethnicity, and clinical outcomes in hormone receptor-positive, HER2-negative, node-negative breast cancer in the randomized TAILORx trial. J Natl Cancer Inst 2021 113:390-399. PMID 32986828
12. Eaglehouse YL, Georg MW, Shriver CD, Zhu K. Racial differences in time to breast cancer surgery and overall survival in the US Military Health System. JAMA Surg 2019 154:e185113. PMID 30673075
13. Carey LA, Perou CM, Livasy CA, et al. Race, breast cancer subtypes, and survival in the Carolina Breast Cancer Study. JAMA 2006 295:2492-2502. PMID 16757721
14. Yu AYL, Thomas SM, DiLalla GD, et al. Disease characteristics and mortality among Asian women with breast cancer. Cancer 2022 128:1024-1037. PMID 34792814
15. Cole SW. Human social genomics. PLoS Genet 2014 10:e1004601. PMID 25166010
16. Keenan T, Moy B, Mroz EA, et al. Comparison of the genomic landscape between primary breast cancer in African American versus white women and the association of racial differences with tumor recurrence. J Clin Oncol 2015 33:3621-3627. PMID 26371147
17. Ademuyiwa FO, Tao Y, Luo J, Weilbaecher K, Ma CX. Differences in the mutational landscape of triple-negative breast cancer in African Americans and Caucasians. Breast Cancer Res Treat 2016 161:491-499. PMID 27915434
18. Yao S, Wei L, Hu Q, et al. Mutational landscape of triple-negative breast cancer in African American women. Nat Genet 2025 57:2166-2176. PMID 40858906
19. Van Alsten SC, Love MI, Calhoun BC, et al. Genomic analysis reveals racial and age-related differences in the somatic landscape of breast cancer and the association with socioeconomic factors. Cancer Res 2025 85:1327-1340. PMID 39879109
20. Andre F, Ciruelos E, Rubovszky G, et al. Alpelisib for PIK3CA-mutated, hormone receptor-positive advanced breast cancer. N Engl J Med 2019 380:1929-1940. PMID 31091374
21. Turner NC, Im SA, Saura C, et al. Inavolisib-based therapy in PIK3CA-mutated advanced breast cancer. N Engl J Med 2024 391:1584-1596. PMID 39476340
22. Dunn MR, Li D, Emerson MA, et al. A latent class assessment of healthcare access factors and disparities in breast cancer care timeliness. PLoS Med 2024 21:e1004500. PMID 39621782
23. Hershman DL, Kushi LH, Shao T, et al. Early discontinuation and nonadherence to adjuvant hormonal therapy in a cohort of 8,769 early-stage breast cancer patients. J Clin Oncol 2010 28:4120-4128. PMID 20585090
24. Loree JM, Anand S, Dasari A, et al. Disparity of race reporting and representation in clinical trials leading to cancer drug approvals from 2008 to 2018. JAMA Oncol 2019 5:e191870. PMID 31415071
25. Unger JM, Vaidya R, Hershman DL, Minasian LM, Fleury ME. Systematic review and meta-analysis of the magnitude of structural, clinical, and physician and patient barriers to cancer clinical trial participation. J Natl Cancer Inst 2019 111:245-255. PMID 30856272
26. Wendler D, Kington R, Madans J, et al. Are racial and ethnic minorities less willing to participate in health research?. PLoS Med 2006 3:e19. PMID 16318411
27. Vyas DA, Eisenstein LG, Jones DS. Hidden in plain sight - reconsidering the use of race correction in clinical algorithms. N Engl J Med 2020 383:874-882. PMID 32853499