Study Population
We obtained information from the CCR for female California residents ages 18 years and older at diagnosis, who were diagnosed with a first, primary invasive breast cancer [International Classification of Disease for Oncology, 3rd Edition, (ICD-O-3) site codes C50.0-50.9] during January 1, 2005 through December 31, 2015 (n=219,266). Patients were excluded from the analysis hierarchically as follows: diagnosis by death certificate or autopsy only (n=889) or diagnosis not microscopically confirmed (n=1,698); ICD-O-3 histologic type other than: 8000, 8001, 8010, 8020, 8022, 8050, 8140, 8201, 8211, 8230, 8255, 8260, 8401, 8453, 8480, 8481, 8500-8525, or 8575 (n=3,654); tumor size missing because unknown (n=8,347), no tumor noted (n=510), microscopic (n=2,250), diffuse (n=608), or mammographic diagnosis only (n=59); age <60 insured by Medicare (n=513); no follow-up (n=269); residential address that was uncertain or not geocodable (n=6,292). The study included 192,932 patients, of whom 94,076 were younger (age 18-59) and 98,856 were older (age 60 and above, up to age 109) patients.
Data Acquisition
Data from the CCR, mostly derived from the patient’s medical record, were used to obtain age at diagnosis, marital status, residential address at diagnosis, stage at diagnosis, tumor size (in centimeters [cm]), lymph node involvement, histology, grade (I, II, III/IV, or unknown), and hormone receptor [estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2)] status. The CCR followed patients for vital status, from linkage with vital records, to December 31, 2015 for this study.
For the variables of interest in the present report, we used data from the medical record to classify race/ethnicity as non-Hispanic white (NHW), non-Hispanic black, Hispanic, Asian/Pacific Islander (API), or other/unknown]; and primary and secondary source of payment were used to classify insurance status, as private only, Medicare only/Medicare + private, any Medicaid/military/other public, no insurance, and unknown); in the CCR, payer status is coded based on the most extensive insurance type across the diagnosis to treatment continuum. We used a multi-component measure of neighborhood socioeconomic (nSES), based on patients’ residential census block group at diagnosis. This measure incorporated the 2000 U.S. Census (for cases diagnosed in 2005) and the 2006-2010 American Community Survey data (for cases diagnosed in 2006 and forward) on education, occupation, unemployment, household income, poverty, rent, and house values (17, 18). Each patient was assigned a nSES quintile, based on the distribution of socioeconomic status across census block groups in California.
Statistical Analysis
Given the lack of standard for categorizing younger and older breast cancer patients, we used the median age of the study population as a cut-off (60 years); younger women included those age 18-59 and older included 60+ years. Differences in mortality for older and younger women were examined by two methods. First, comparisons were made between older and younger patients stratified by race/ethnicity, insurance status, and nSES, with younger women as the reference group. Next, models were stratified by age and comparisons were made between race/ethnicity (non-Hispanic white as the reference group), insurance status (private insurance as the reference group), and nSES quintiles (5th quintile as the reference group).
Descriptive statistics were calculated and reported as percentages for categorical data and means with standard deviation for continuous variables. Covariates were shown overall and for younger and older women. Covariates examined included: age (continuous and categorical), race/ethnicity, marital status, insurance status, nSES, whether the patient was seen at one or more of the NCI-designated Cancer Centers in California (NCICC) for her breast cancer, AJCC stage at diagnosis, tumor subtype, lymph node involvement, tumor size, tumor grade, and tumor histology.
Follow-up time was calculated as the number of days between the date of diagnosis and date of death from breast cancer (ICD 9/10 = 174/C50), the date of death from another cause, the date of last follow-up (i.e., last known contact), or the study end date (12/31/2015). There were 599 deceased patients with an unknown cause of death which were excluded from all models. Cox proportional hazards regression was used to estimate breast cancer specific hazard rate ratios (HR) and corresponding 95% confidence intervals (CI). Adjusted models were stratified by AJCC stage and adjusted for age at diagnosis (continuous), year of diagnosis (continuous), race/ethnicity, marital status, insurance status, nSES, whether the patient was seen at one or more of the NCICC in California for her breast cancer, tumor subtype, lymph node involvement, tumor size, tumor grade, and tumor histology. Fully adjusted models were additionally adjusted for clustering by block group, using a sandwich estimator of the covariance structure that accounts for intracluster dependence. The proportional hazards assumption was tested by examining the correlation between time and scaled Schoenfeld residuals for all covariates. The proportional hazards assumption was violated for AJCC stage at diagnosis, tumor subtype, and tumor grade. Stage was included as an underlying stratifying variable in the fully adjusted Cox regression models reported here, which allowed the baseline hazards to vary by stage. Additionally, stratifying the Cox model by tumor subtype and tumor grade did not meaningfully change the HR for the main effect of age, so these factors were simply adjusted for in fully adjusted models. Wald Type 3 tests for interaction between age group (18-59, 60+) and race/ethnicity, insurance status, and nSES and were computed using cross-product terms, in models adjusted for all statistically significant (p<0.05) interactions with age group (race/ethnicity, marital status, insurance status, NCICC, tumor subtype, and lymph node involvement). Wald tests for trend across nSES quintiles were computed using quintile number as an ordinal variable. All statistical tests were carried out using SAS software version 9.3 (SAS Institute).