Isolating the Drivers of Racial Inequities in Prostate Cancer Treatment

Black Americans are less likely to receive potentially lifesaving medical interventions compared with White Americans (1), even though they are more likely to die from a wide range of diseases and conditions than their peers (2). Prostate cancer is a primary example of this disconnect between disease burden and receipt of beneficial treatments. Black individuals experience a 2-fold higher risk of prostate cancer death than all other individuals in the United States (3), while they are also less likely to receive treatment (i.e., radical prostatectomy or radiation) for localized prostate cancer (4, 5), which has been shown to reduce metastases and death among individuals with clinically significant, higher-risk localized cancers (6–8). Previous studies demonstrate that similar treatment for localized prostate cancer leads to similar outcomes for Black individuals when matched with White individuals by disease severity (9, 10).

Black individuals with prostate cancer are less likely to receive guideline concordant and high-quality care, which likely contributes to a portion of the racial inequity in mortality in prostate cancer (11, 12). The management of localized prostate cancer includes observation (i.e., active surveillance or watchful waiting), surgery, radiotherapy, and in select cases, ablative therapy [i.e., high-intensity focused ultrasound (HIFU), irreversible electroporation (IRE), cryotherapy; refs. 13, 14]. The decision for management choice is dependent on the baseline health of the patient, their cancer aggressiveness, patient preferences and goals of care, and social and structural barriers to care (12). Guidelines recommend the utilization of watchful waiting—which is observation without frequent routine PSA testing or biopsy—in those with life expectancy less than 5–10 years. In healthy patients with prostate cancer with low-risk cancers (i.e., grade group 1), active surveillance with routine PSA testing and prostate biopsy is the preferred management strategy. For healthy patient with prostate cancer with intermediate (grade group 2–3 and/or PSA 10–20) and high (grade group 4–5 and/or PSA >20), definitive therapy is recommended in the form of surgery or radiation (15). If guidelines recommend the increased use of early detection for Black individuals at younger ages, there is a potential for treatment inequities to have a larger impact on mortality inequities.

While existing research demonstrates that prostate cancer treatment delivery is not equitable, the drivers of this inequity remain unclear. Identifying the drivers of inequitable treatment delivery is necessary to design interventions that support equitable care delivery and improve prostate cancer outcomes for Black individuals (12, 16). Some have posited that differences in individual patient health characteristics (e.g., cancer characteristics or comorbid conditions) drive differences in delivery of treatment. If so, interventions should address the underlying inequities that lead to these differences in patient health. Others have suggested that differences in delivery of treatment between Black individuals and White individuals are attributable to differences in physician and health system factors that influence access and delivery of care. In this case, interventions should support treatment delivery and access such that receipt of treatment among Black individuals is more equitable. Even though there is a long history of attributing racial differences in health outcomes to inherent patient health factors (17), it is unclear which of these two factors—patient health characteristics or treatment delivery—drive the receipt of lower treatment rates among Black individuals with localized prostate cancer.

There is a need to progress from describing prostate cancer treatment inequities to understanding why the inequity exists by explicating the drivers. To do so, we undertook an analysis that aimed to evaluate the role that individual patient health factors and health system level factors play in inequitable utilization of localized prostate cancer therapies. We hypothesize that differences in treatment between Black and White individuals are predominantly driven by differential care delivery, independent of baseline disease characteristics and comorbidities. Rejection of this hypothesis would instead imply that inequities in cancer characteristics or comorbid conditions play an important role in driving lower rates of treatment in localized prostate cancer.

This study adhered to the STROBE Statement guidelines for reporting observational studies (18). We analyzed a retrospective cohort from the Surveillance, Epidemiology, and End Result-Medicare claims data (SEER-Medicare) among patients with prostate cancer diagnosed between 2010 and 2015. We restricted our sample to men 66 years and older with a broad definition for clinically significant (grade group 2–5), non-metastatic cancer (M0), who were identified as either Black or non-Hispanic White by the SEER registry, which follows the definition of the U.S. Office of Management and Budget. The cohort was further limited to those who did not have managed care to allow for complete claims data coverage resulting in a cohort of 40,137 men.

We separate the drivers of treatment inequity between Black and White individuals with localized, clinically significant prostate cancer into two groups: (i) group-level differences in health predictors of treatment, which we define as differences in patient health, and (ii) group-level differences in how those health predictors associate with treatment, which we define as differences in health care delivery. Our approach uses the Kitagawa-Oaxaca-Blinder decomposition, a statistical approach grounded in economics originally used to investigate how wages of a disadvantaged group relate to those of an advantaged group (19–21). Utilizing this approach to assess drivers of inequities in prostate cancer gets us closer to understanding why current practices are inequitable, thus informing us how to achieve equitable care.

Patient health was defined by tumor characteristics, age, and comorbidities. Tumor characteristics included grade group and PSA level at diagnosis. PSA was modeled as a categorical variable (0–2, 2.1–4, 4.1–10,10.1–20, 20.1–50, 50.1+, missing) including a category for missing PSA. Age was modeled as four different categories to allow for non-linearities. We chose these variables as potential predictors due to their clinical significance. Demographic and tumor data came from the Patient Entitlement and Diagnosis Summary File. To model comorbidities, Medicare claims data were used to create binary indicators for the presence of each unique International Classification of Diseases, Ninth Revision, Clinical Modification (ICD‐9-CM) diagnosis code within 1 year of prostate cancer diagnoses. We kept the 1,000 most common codes and created an indicator for the presence of each of these diagnoses for each patient. Treatment was defined as the receipt of a radical prostatectomy or radiation as defined by the ICD‐9-Procedure Coding System (PCS) and Current Procedural Terminology (CPT) codes in Medicare claims within 12 months of diagnosis (Supplementary Data S1 for codes; ref. 22).

The Kitagawa-Oaxaca-Blinder method decomposes the difference in treatment rates between Black and White patients with prostate cancer into two portions: (i) patient health and (ii) health care delivery. Our application of this decomposition approach applies the least absolute shrinkage operator (lasso) prediction algorithm (23). Lasso simultaneously selected the important predictors of and modeled the association between baseline patient health and treatment to determine health care for White individuals. To estimate the expected percentage of Black individuals that would receive treatment had they been treated like the White individuals in this cohort, the health care model of White individuals was applied to the patient health predictors of Black individuals and then averaged.

Comparison of observed and expected average percentages for each group formed our application of the Kitagawa-Oaxaca-Blinder decomposition. The portion of the difference in treatment percentages between the groups driven by patient health was derived by comparing the observed percentage of White individuals currently treated with the expected percentage of treatment among Black individuals. This comparison holds health care constant but changes patient health. The portion driven by health care compares the expected percentage of treatment among Black individuals with the observed percentage of Black individuals currently treated because this comparison now holds patient health constant (both percentages use the patient health of Black individuals) but changes health care delivery. Any misspecifications in measuring patient health may be attributed to health care. Figure 1 demonstrates how difference scenarios of contributions of patient health and health care delivery to the difference in observed treatment of Black and White individuals.

We also followed the converse process to compare counterfactual and observed treatment percentages among White individuals had they been treated like Black individuals (i.e., if White individuals had the same association between baseline patient health predictors and treatment that Black individuals received) to check whether the choice of reference group impacts our conclusions. Statistical significance of the roles of patient health and health care delivery in treatment was estimated using a bootstrapped resampling approach (Supplementary Data S1; ref. 24).

The data used in this study are from the SEER-Medicare. These data were accessed and utilized under a data agreement with the SEER program. Because of the terms and conditions of this data agreement, we are unable to share the SEER data or any related datasets without explicit authorization from the SEER program. Researchers interested in accessing and using the SEER data for their own analyses are advised to contact the SEER program directly to inquire about data availability, terms of use, and the necessary permissions for data sharing.

Black individuals demonstrated a 6.5 percentage point lower probability of receiving treatment compared with White individuals in our sample, among whom the treatment rates were 72% and 78.5%, respectively. Baseline data are presented in Table 1 for the 40,137 individuals identified in our study sample. Black individuals comprised 13.5% of the cohort. Differences in clinical characteristics existed between White and Black individuals, where Black individuals had higher comorbidity and PSA scores while they were also younger on average.

Lasso selected important predictors of treatment specific to our context, which were age, grade group, PSA score, 31 comorbid conditions that predominantly decreased surgical appropriateness (the most influential being dementia, persistent mental disorders, paralysis agitans, cerebrovascular disease, and chronic kidney disease), and year of diagnosis. The predictive model for White individuals had an AUC operator curve of 0.78 on a 20% hold-out sample indicating a robust capture of the factors that predict treatment without major overfitting.

Figure 2 presents the average predicted and observed treatment rates over time. Predictions indicate that if Black individuals were treated like White individuals, they would have a higher average treatment percentage (77.5%) compared with their observed percentage (72.0%). White individuals treated like Black individuals would receive a lower percentage of treatment (71.0%) compared with the observed percentage (78.5%).

Figure 3 displays the observed inequity on the left, the three potential results in the middle, and the result of the Kitagawa-Oaxaca-Blinder decomposition on the right for the case where observed treatment percentages from Black and White individuals are compared with the predicted treatment percentage for Black individuals had they been treated like White individuals. Because Black individuals treated like White individuals would increase the percentage of treatment to 77.5%, which is much closer to the rate received by White individuals, it is the difference between being treated like a White individual and being treated like a Black individual that explains most of the difference (i.e., differences in health care delivery). Our results show that 85% of the treatment difference (5.5 of the 6.5 percentage point difference) by race is explained by differences in how prostate cancer health care is delivered, independent of patient health, even under the equal coverage of Medicare. The bootstrapped 95% confidence interval for health care portion is (95% CI: 4.6–6.6 or 74–94 of the total). The comparison of the treatment rate of Black individuals had they been treated like White individuals indicates that differences in patient health explain only 15% of the treatment difference (1.0 of the 6.5 percentage point difference). The bootstrapped 95% CI for health portion is (95% CI: 0.2–1.6 or 6–24 of the total). Similar results were obtained using the prediction model from the Black patient sample to form the counterfactual predictions.

We also investigated whether the finding of health care as the dominant driver of the inequity was driven by cancer aggressiveness, as measured by cancer grade group. We performed our main Kitagawa-Oaxaca-Blinder decomposition separately for samples of patients with grade group 2–3 and grade group 4–5 cancer. The definitive treatment rate differences between White and Black men were 6.65 percentage points for men with grade group 2–3 cancers, and 6.55 for men with grade group 4–5 cancers. For men with grade group 2–3 cancers, the health portion was 0.55 pp while health care was 6.1 pp (i.e., health care was 90% of the difference). For men with grade group 4–5 cancer, health was 0.94 pp while health care was 5.6 pp (health care was 85% of the difference). This additional analysis provides robustness to our main finding that health care is the dominant driver of the disparity.

Black individuals with prostate cancer are more likely to die but less likely to receive beneficial interventions (1, 2). Although inequities in treatment delivery have historically been attributed to differences in patient health (25), our approach using predictive modeling within the Kitagawa-Oaxaca-Blinder decomposition demonstrates the inadequacy of that explanation. Instead, our results show that 85% of the difference in treatment for prostate cancer in our Medicare cohort was driven by differences in how race and health factors translate into the delivery of health care. While patient health plays a statistically significant role, renewing the importance of addressing inequities in patient health (e.g., disparities in cardiovascular disease, diabetes, chronic kidney disease), inequity in health care delivery is the predominant driver of the difference in treatment between Black and White individuals with prostate cancer. The dominant role of inequities in health care delivery suggests a strong need to invest in patient-centered studies and interventions that support equitable care delivery. Although this difference in treatment does not explain current observations in mortality disparities, which are likely driven more by delayed diagnosis and higher cancer incidence, there must be a recognition that prostate cancer care is delivered inequitably based on race. Efforts to improve screening and early detection of prostate cancer in Black populations will potentially worsen prostate cancer disparities in mortality if they are not coupled with appropriate support and navigation to connect Black patients with prostate cancer to guideline-concordant therapies.

In this analysis, it is important to note that any contributions to the differences in treatment utilization between Black and White individuals that were not accounted for by the modeled health variables, including potential influences such as patient preferences, patient distrust/mistrust, and social determinant factors like education, were subsumed within the 85% of the Black/White difference attributed to health care delivery. While these factors may traditionally be considered as aspects of patient characteristics, categorizing them under health care delivery emphasizes their connection to the health care system. Factors such as patient preferences, patient mistrust, and education can be described within the term health care delivery because they intricately influence how health care services are provided and experienced (12). Patient preferences shape treatment decisions, patient mistrust affects health care engagement, and education levels impact health literacy and access to care. Recognizing these elements as part of health care delivery underscores the broader societal and individual dynamics that play a crucial role in shaping the overall effectiveness and equity of health care services (12).

The relationship between higher mortality rates and lower rates of potentially life-saving interventions reflects the deadly impact of inequities in both patient health and health care delivery for Black individuals in the United States (26–28). The objective of modeling within the Kitagawa-Oaxaca-Blinder decomposition is to accurately capture the relationship between patient health factors and health care delivery to assess the relationship between these two factors in driving disparate treatment rates. We utilize a novel predictive modeling-based decomposition approach that captures what it means for a Black individual to be treated like a White individual. Predictive modeling more accurately captures the relationships between patient health factors and health care delivery for a given group relative to standard statistical models, by selecting the important predictors that accurately capture how individuals are treated without overfitting the administrative data. Our approach can be used to quantify the roles of patient health and health care delivery in treatment inequities in other health domains.

Prostate cancer disparities have historically been attributed to differences in patient-level factors including mistrust, preferences, and competing comorbidities (12). However, equitable access and utilization of localized prostate cancer treatments lead to parity in prostate cancer outcomes for Black individuals (9, 10, 26, 27). Thus, prostate cancer treatment inequities reflect the intersection between structural determinants of equity, which include systemic racism and public policies, social determinants of health, and health care factors (12, 29). Our findings reinforce the importance of understanding the broader social and structural factors to understand the association between race and treatment inequities in prostate cancer. When combined with our previous work demonstrating Black Medicare beneficiaries are less likely to access high-quality prostate cancer treatments (30), it is imperative that we develop interventions and tools to support Black individuals in accessing and utilizing high-quality care.

This study allows us to move forward from describing the existence of prostate cancer treatment inequity to quantifying the magnitude by which patient health factors or health care delivery impact this inequity. Following our previously published conceptual model for studying prostate cancer inequities, these findings will be used to support a series of qualitative and community-partnered research endeavors that will develop sustainable solutions to help Black individuals with prostate cancer overcome health-system barriers to equitable treatment (12, 16). We aim to use a patient-centered and community-partnered approach to engage and empower Black communities and relevant stakeholders in the development of interventions for shared decision making and social support to help Black individuals with prostate cancer access equitable prostate cancer care. Community engagement strategies have shown success in reducing barriers and increasing equity in care, including early-stage cancers and prostate cancer specifically (2, 31–33). Our results demonstrate that there is a need for these and similar interventions in the delivery of prostate cancer care.

Our analysis is limited by the observational design. Administrative and claims data are also subject to limitations, which may include limitations in the capture of patient health factors from claims data. While we use prediction models and a wide array of potential health predictors to more accurately capture context-specific individual level health, unmeasured health factors can be mistakenly attributed to health care. We additionally recognize that our study is limited by a lack of granularity and availability of data to further identify which specific health care–related factors drive differences in treatment utilization, such as individual and structural racism, social determinants of health, access, health literacy, health care biases, and patient preferences. The use of a Medicare population limits our analyses to older patients (66 years of age or older) and an insured population. The presence of this inequity in this populations likely suggests that treatment inequities are likely worse in younger, uninsured prostate cancer populations. We aim to use qualitative research methodology to better understand which health care factors are the main drivers of the main observations of our study. We also recognize there is nuance in defining “clinically significant” prostate cancer beyond the terms we use; however, biopsy data are not available in SEER-Medicare. The strength of our study is in our methodologic approach, which has not been previously applied to understand the drivers of prostate cancer health disparities.

We describe an approach to quantify the roles of patient health factors and health care delivery as drivers of treatment disparities that can be broadly applied to health inequities. Our application of the Kitagawa-Oaxaca-Blinder approach to prostate cancer treatment inequities indicates that 85% of the treatment differences are driven by racial differences in health care delivery, independent of patient health factors. Action is needed to support Black individuals in accessing and utilizing high-quality prostate cancer care.

A. Basu reports personal fees from Salutis Consulting LLC outside the submitted work. J.L. Gore reports other support from Seagen Pharmaceuticals, Inc. and Immunity Bio outside the submitted work. Y.A. Nyame reports grants from Department of Defense, NIH, and Andy Hill CARE Fund during the conduct of the study; personal fees from OrthoClinical Diagnostics and ImmunityBio Inc outside the submitted work. No disclosures were reported by the other authors.

N. Hammarlund: Conceptualization, formal analysis, validation, visualization, methodology, writing–original draft. S.K. Holt: Data curation, formal analysis. A. Basu: Resources, writing–review and editing. R. Etzioni: Methodology, writing–review and editing. D. Morehead: Writing–review and editing. J.R. Lee: Writing–review and editing. E.M. Wolff: Writing–review and editing. J.L. Gore: Writing–review and editing. Y.A. Nyame: Resources, funding acquisition, writing–original draft, writing–review and editing.

This project was made possible in part by the Andy Hill Cancer Research Endowment (CARE) Fund (Y.A. Nyame, award number 2021-DR-01), the NCI SPORE (Y.A. Nyame, P50 CA097186), and the U.S. Department of Defense, Office of the Congressionally Directed Medical Research Programs (Y.A. Nyame) under Grant Number W81XWH2110531. This study used the linked SEER-Medicare database. The interpretation and reporting of these data are the sole responsibility of the authors. The authors acknowledge the efforts of the NCI; the Office of Research, Development and Information, Centers for Medicare & Medicaid Services Information Management Services; and the SEER Program tumor registries in the creation of the SEER-Medicare database. We would like to thank the Comparative Health Outcomes, Policy, and Economics (CHOICE) Institute at the University of Washington for obtaining the SEER data.