Data source

Data were obtained from the NCDB, a hospital-based cancer registry containing data from over 1500 facilities accredited by the American College of Surgeons’ Commission on Cancer (CoC). Approximately 70% of all malignant cancers diagnosed in the United States are included in this dataset. Although the NCDB collects a large proportion of incident cancer diagnoses in the United States, selection bias may exist because only hospitals approved by the CoC contribute data to the NCDB. Available data elements, including sociodemographic characteristics, tumor characteristics, attributes of the treatment facilities, treatment, and survival outcomes, are abstracted from patient medical records by Certified Tumor Registrars. For cases with missing data elements, registrars may contact the treating physicians to obtain the necessary data to complete the record. Data submitted to the NCDB undergo rigorous data quality checks in line with standards set by the CoC. Case records that do not meet requirements are identified and returned to the hospital. All data are deidentified, and the study was considered exempt by the Ohio State University Institutional Review Board.

Study population

We conducted a retrospective cohort study of women diagnosed with EC (International Classification of Diseases for Oncology, Third Edition [ICD-O-3] primary site codes: C54.0–C54.3, C54.8–C54.9, C55.9] between 2004 and 2016 using data from the NCDB. ^, We identified 423,657 women ≥18 years of age at diagnosis who self-reported non-Hispanic White (hereafter, White) or non-Hispanic Black (hereafter, Black) race. We excluded women from this analysis for the following reasons: no surgical procedure (n=34,116); subtotal hysterectomy (n=4346); surgery not otherwise specified (n=866); missing stage (n=68,380); or histology types not classifiable as endometrioid/adenocarcinoma (ICD-O-3 morphology codes: 8140, 8380–8383, 8210, 8211, 8260–8263, 8560, 8570), serous (ICD-O-3: 8441, 8460, 8461), carcinosarcoma (ICD-O-3: 8950, 8951, 8980, 8981), mixed epithelial (ICD-O-3: 8323, 8255), or clear-cell (ICD-O-3: 8310) (n=9303) or ungraded endometrioid (n=32,713). We further excluded 5821 women with missing information on facility location, 21,384 women with missing zip code-level income, and 78 women who did not have information on follow-up time (or 0 months), resulting in a sample size of 246,650 ( Figure 1 ). Further exclusions for the analyses of radiation and chemotherapy refusal are described below.

Study flowchart demonstrating cohort selection and distribution of treatment status in the overall study population

Barrington et al. Treatment refusal and endometrial cancer survival disparities. Am J Obstet Gynecol 2022 .

Treatment refusal

Radiation and chemotherapy treatment were categorized as follows: (1) none, not recommended as part of the planned first course of therapy, (2) received as part of planned first course of therapy, (3) not recommended because of contraindications, (4) not administered because the patient died before planned therapy, (5) recommended but not administered because of unknown reasons, (6) recommended but refused by patient, (7) recommended but unknown whether administered, and (8) unknown whether recommended or administered. Figure 1 shows the distribution of these treatment categories in the sample of 246,650 women.

Because our goal was to understand how treatment refusal impacts survival, the radiation and chemotherapy analyses were limited to women who were recommended to have each specific treatment and either received (category 2) or refused (category 6) the treatment. Our definition of treatment refusal did not include patients who did not receive recommended treatment because of contraindications nor cases where treatment was recommended but not received for unknown reasons. In analyses of radiation refusal, we included 75,447 women who were recommended to receive radiation, whereas in analyses of chemotherapy refusal, we included 60,187 women who were recommended to receive chemotherapy. In Supplemental Table 1 , we compared demographic characteristics between women who were included in the analyses (recommended to receive treatment and either received or refused, with nonmissing data on exclusion criteria) and women who were recommended to receive treatment and either received or refused but were excluded because of missing data on exclusion criteria (eg, missing surgery, missing stage, etc.). In both analyses (refusal of radiation and refusal of chemotherapy), excluded women were more likely to be Black, younger at diagnosis, have comorbidities, be uninsured, and be less likely to have private insurance.

Covariates

Information on the following covariates was included: age at diagnosis (<50, 50–69, ≥70), Charlson-Deyo comorbidity score (0, 1, or ≥2), type of health insurance (none, private, Medicaid, Medicare), median income in zip code of residence in 2 categories (<$48,000 or ≥$48,000), percentage of residents by zip code who did not graduate from high school (≥21%, 13%–20.9%, 7%–12.9%, <7%), facility location (Northeast, Midwest, Mountain, Pacific Coast, South), facility type (community cancer, comprehensive community cancer, academic/research, integrated network cancer), 2009 American Joint Committee on Cancer pathologic stage (I, II, III, IV), and grade (1, 2, 3, applicable for endometrioid EC only). We combined grade and histology to create a histologic subtype variable with the following categories: low-grade endometrioid, high-grade endometrioid, serous, clear-cell, and carcinosarcoma. Median household income for each patient’s area of residence was estimated by matching the zip code of the patient recorded at the time of diagnosis against files derived from the 2012 American Community Survey data, spanning years from 2008 to 2012 and adjusted for 2012 inflation. Household income was categorized as quartiles based on equally proportioned income ranges among all US zip codes; we grouped the lower 2 and upper 2 quartiles. OS was calculated as the time from the date of diagnosis to the date of death; among women alive at the end of follow-up, the date of last contact was used as the censoring time.

Statistical analysis

We used multivariable-adjusted logistic regression to estimate adjusted odds ratios (aORs) and 95% confidence intervals (CIs) for associations between refusal of adjuvant treatment and epidemiologic, tumor, and facility characteristics in the overall study population and among Black and White women separately. We also examined the association between race and refusal of adjuvant treatment stratified by histology, on the basis of the hypothesis that treatment refusal among women with certain histologic diagnoses would be particularly harmful for survival. All variables shown in Table 1 were significantly related to treatment refusal in univariable models ( P <.05, data not tabled) and were therefore included in multivariable models. We included the same adjustment factors in the race-specific and histology-stratified analyses to facilitate comparisons. To test whether associations differed by race, we included a multiplicative interaction term between each assessed predictor and race in the multivariable models.

We used Kaplan-Meier curves and log-rank tests to evaluate the overall association between race and OS and the joint effect of race and refusal of adjuvant radiation or chemotherapy on OS. To quantify how much of the difference in survival by race might operate through different patterns of treatment refusal, we performed a causal mediation analysis. We applied the simulation-based structural equation modeling approach implemented in the R mediation package (R Foundation for Statistical Computing, Vienna, Austria). Mediation analysis explicitly examines how a third intermediate variable, the mediator (ie, treatment refusal), is related to the observed exposure–outcome (race–OS) relationship. Because the hazard ratio in a Cox proportional hazards model does not necessarily have a causal interpretation when the outcome is not rare, ^, we fit accelerated failure time models assuming a Weibull error distribution, and our outcome of interest was mean survival time. Mediation analysis seeks to partition the total effect (here, the mean difference in survival time between Black and White women) into the average causal mediation effect (ACME) and the average direct effect (ADE). The ACME, which captures how much the effect of race operates through the mediator (treatment refusal), is of central interest; elsewhere the terminology “natural indirect effect” or “pure/total indirect effect” is used. ^, The ACME is the difference in mean survival time expected if the distribution of race in our population remained the same, but the value of the mediator differed from the expected for each woman if her race was changed from White to Black. We modeled each woman’s likelihood of treatment refusal via logistic regression as a function of race and other covariates and used this model to estimate how a woman’s probability of treatment refusal changes if all variables are fixed except race, which was changed from White to Black. Thus, the ACME captures how much of the difference in mean survival time between Black and White women is attributable to racial differences in patterns of treatment refusal, adjusting for potential confounders. The ADE is the remaining effect of race on survival, that is, the extent to which the mean difference between Black and White women’s survival operates through pathways other than treatment refusal. Elsewhere the terminology “natural direct effect” or “pure/total direct effect” is used. ^,

Separately for chemotherapy and radiation, we estimated mediation effects overall and within categories defined by tumor histology. The mediator and outcome models were adjusted for the confounders listed in Table 1 . All analyses were performed in SAS version 9.4 (SAS Institute, Cary, NC) or R version 4.0.2. All P values were 2-sided.