Differences in Breast Cancer Survival by Molecular Subtypes in the United States

This article is featured in Highlights of This Issue, p. 617

Breast cancer molecular subtypes have been defined based on gene expression profiling (1). The major subtypes of breast cancer are approximated by the joint expression of three tumor markers: estrogen receptor (ER), progesterone receptor [PR, where ER and PR status are jointly assessed as hormone receptor (HR) status], and HER2 status. The four main molecular subtypes approximated by joint HR/HER2 status are: HR⁺/HER2⁻ (i.e., approximating Luminal A subtype), HR⁺/HER2⁺ (Luminal B), HR⁻/HER2⁺ (HER2-enriched), and HR⁻/HER2⁻ (triple-negative; refs. 2–4). Recently, these three tumor markers have become part of routine data collection by the population-based Surveillance, Epidemiology, and End Results (SEER) cancer registries covering about 30% of the U.S. population (3, 5, 6).

Important differences in demographic and clinical characteristics in occurrence of molecular subtypes among women have been described using U.S. population–based cancer registry data (3, 4, 7). In contrast, breast cancer–specific survival for each subtype is poorly documented at the national level. The use of Herceptin and other targeted therapies for HER2-positive breast cancer has been clearly shown to improve survival (8–10). The clinical benefits of these and other treatment advances underscore the need for national population-based data that are specific to subtype. Although some studies have assessed outcomes by molecular subtypes, they were based on relatively small observational studies or confined to a specific geographic region (11–13). Recently, a few studies utilized data from the population-based California Cancer Registry (14–18). However, most of the prior studies examining breast cancer prognosis by subtypes suffered from a large number of women with missing information on receptor status, where missing receptor statuses ranged from 12% to 33% of overall breast cancer cases (3, 11, 17–19). Furthermore, women with missing receptor status tended to be minorities, older, of lower socioeconomic status, had no insurance, or had advanced disease (3). With registry data, we are unable to assume that receptor information is missing completely at random; therefore, failing to account for missing information could bias breast cancer prognosis in the general population.

Our primary aim was to present the first report of nationally representative breast cancer–specific survival estimates by four main molecular subtypes. We provide an assessment of demographic (age, race, ethnicity, socioeconomic status, marital, and insurance status) and clinical (tumor stage, size, nodal status, and Bloom–Richardson grade) differences in breast cancer survival by subtypes using SEER data covering approximately 30% of the U.S. general population. Our second aim was to develop an imputation algorithm to fill in missing receptor status, enabling us to obtain accurate estimates of breast cancer survival in the population.

We identified females with invasive breast cancer diagnosed during 2010–2013 (ICD-O-3 site codes C500-509 excluding histology codes 9050–9055; 9140; 9590–9992). Cases were followed until December 31, 2014 (allowing a maximum follow-up of 59 months). We used data from 17 population-based cancer registries that participate in the SEER program. Data from the Alaska Native Registry were excluded (n = 273), since this registry does not provide any additional information on ethnic origin of these cases, i.e., whether they are of Hispanic versus non-Hispanic origin, a distinction necessary for analyses by mutually exclusive race-ethnic subgroups (i.e., non-Hispanic white, non-Hispanic black, non-Hispanic Asian Pacific Islander (API), non-Hispanic American Indian/Alaskan Native, and Hispanic). We began the analysis in 2010 when information on HER2 status was first captured uniformly across all SEER registries (information on ER and PR status in the registries has been collected since 1990). In constructing the study cohort, we applied several exclusion criteria typically used with SEER data for survival analyses: (i) cases diagnosed by autopsy or death certificate (n = 937), (ii) alive with no survival time (n = 466), (iii) missing cause of death (COD, n = 584). We further restricted analysis to cases that were a woman's first or only breast cancer to create a more homogenous group, because a prior cancer diagnosis may affect patient prognosis. The final analytic set consisted of 196,094 females diagnosed with invasive breast cancer.

Study variables such as ER, PR, and HER2 status, demographic characteristics (e.g., age, race, ethnicity, year of diagnosis, marital, and insurance status), and tumor stage, size, nodal status, and Bloom–Richardson grade were ascertained across SEER registries using standardized coding rules based on hospital medical records and pathology reports. The four main breast cancer molecular subtypes were derived based on the joint expressions of the three tumor markers explained in detail elsewhere (3). Area-based indicators of patient poverty were derived at the county level from 2010 U.S. Census data and included as tertiles of the observed distribution in the U.S. general population. The proportion of the population in a patient's county of residence who lived below the poverty level was stratified as <10.0% (high socioeconomic status, SES), 10%–19.9% (medium SES), ≥20% (low SES), or unknown. The poverty indicator was used as a surrogate for low, medium, or high SES. Similarly, area-based indicators of urbanicity were derived at the county level from 2010 U.S. Census data and included as the median of the observed distribution in the U.S. general population. The proportion of the population in a patient's county of residence living in urban areas was stratified as <50% (less urban), ≥50% (more urban), or unknown.

Underlying CODs were ascertained by cancer registries from death certificate codes obtained from the National Center for Health Statistics. To correct for known errors with COD attribution, the SEER program recently developed a special COD variable that maps underlying CODs to the primary cancer diagnosis (20). We used this variable to assign a broad set of CODs to capture deaths due to breast cancer among women with an incident breast cancer diagnosis.

The three receptor statuses (ER, PR, and HER2) had a considerable amount of missing information. Overall, 16,563 (8.4%) women had a missing HER2 status, 7,965 (4.0%) were missing ER status, and 8,763 (4.5%) were missing PR status. As a result, molecular subtypes could not be derived for 16,996 (8.7%) female breast cancer cases (Table 1). Of these 16,996 cases with missing molecular subtypes, 8,996 (52.9%) women had a known HR status but missing HER2 status, 7,567 (44.5%) women were missing both HR and HER2 statuses, and 433 (2.5%) women had a known HER2 status but missing HR status. In addition, some degree of missing data was present in clinical and demographic variables, which has been documented before (2, 3, 21, 22). For this reason, we used a sequential regression multiple imputation technique (also referred to as fully conditional specification) to impute HER2 status, ER status, PR status, and clinical/demographic variables with sporadic missingness (23). The idea behind this imputation technique is intuitive: to model each variable with missing observations conditional on all other variables (including those with missing values) and stochastically impute from these the conditional distributions. With a sporadic missingness pattern as is present in the SEER data, the imputation procedure cycles through these conditional models to produce a specified number of imputed datasets. Thus, the final imputed dataset would contain not only imputed HER2, ER, and PR statuses but also other variables that initially had missing information (e.g., stage, grade, node, etc.). Molecular subtype was not directly imputed, but rather derived for all cases based on observed and imputed HER2, ER, and PR status.

Supplementary Table S1 lists variables used for multiple imputation of missing HER2 status, along with the percentage of missing data for each variable. These variables were chosen based on demographic and clinical importance associated with HER2 status (2–4), and included year of diagnosis, age at diagnosis, poverty status (county level), urban indicator (county level), race, ethnicity, registry, reporting source, marital status, insurance, ER status, PR status, American Joint Committee on Cancer (AJCC) 7th stage, tumor size, nodal status, Bloom–Richardson grade, histology, survival time, vital status, and surgery. In addition, we used vital status and the Nelson–Aalen estimator of the cumulative baseline hazard in the imputation model for imputing missing HER2 status for the survival analyses (24). Poverty, urban, race, and surgery covariates had very little missing information (<1% missing). Therefore, we used a single imputation model controlling for age, Hispanic ethnicity, registry, and year of diagnosis to fill in missing information for these four covariates before performing multiple imputation on the remaining covariates in the breast cancer dataset. Because grade had the largest amount of missing information (i.e., 13.8% women with missing grade), we assessed the imputation procedure's sensitivity to inclusion or exclusion of this variable by repeating the imputation, excluding grade from the imputation model. We found no major influence on the imputation results and therefore decided to keep grade in our final model.

We imputed missing HER2 status under the missing at random (MAR) assumption, as the standard sequential regression multiple imputation technique assumes the MAR mechanism. The MAR assumption is not testable; however, we evaluated known and unknown HER2 status by demographic and clinical variables used in the imputation model (Supplementary Table S2). Missing HER2 status appeared to be associated with several of these variables, thus providing plausibility for the MAR assumption after conditioning on these associated variables with a missing pattern of HER2 status. We used proc mi with fcs in SAS 9.3 (SAS Institute) and generated 25 imputed datasets for analyses. We generated 25 imputations based on the rule that the number of imputations should be at least equal to the percentage of incomplete cases (23). Finally, we analyzed each imputed dataset separately, and then estimates were combined using Rubin's rules (25).

We present estimates of 4-year breast cancer–specific survival according to molecular subtypes and clinical and demographic factors. Breast cancer–specific survival was calculated using the actuarial method with the SEER special COD variable (20). Deaths due to breast cancer were treated as the event and other causes of death as the censoring indicator. Survival times were censored at loss to follow-up, death from causes other than breast cancer, or December 31, 2014, whichever occurred first. Finally, a Cox proportional hazards model was used to evaluate the association between molecular subtypes and breast cancer–specific survival after controlling for demographic and clinical factors. Although a previous study showed survival curves by HR⁺ or HR⁻ subtypes cross over when examining over a longer follow-up time (26), however, over the short follow-up period considered in this analysis, the proportionality assumption by molecular subtypes for the Cox model was tested [i.e., by looking at log-log (survival probability) over log (survival time) by molecular subtypes] and appeared to be valid.

A total of 196,094 invasive breast cancer cases were diagnosed and reported to the SEER-18 excluding Alaska registries during 2010–2013. The proportions of women with each molecular subtype were 130,543 (66.6%) for HR⁺/HER2⁻, 21,136 (10.8%) for triple-negative (i.e., HR⁻/HER2⁻), 19,016 (9.7%) for HR⁺/HER2⁺, 8,403 (4.3%) for HR⁻/HER2⁺, and 16,996 (8.7%) unknown (Table 1). Subtype distributions varied by age, race, ethnicity, county-level poverty and urbanicity, insurance, marital status, stage, grade, tumor size, and nodal status. Compared with women with HR⁺/HER2⁻ subtype (the most common subtype), those diagnosed with the other three subtypes were somewhat more likely to be younger, belong to minority groups, living in counties with higher poverty levels, and had later stage, larger tumors, positive nodal status, and higher Bloom–Richardson grade.

Women with missing molecular subtypes tended to be older (12.5% missing among age 75+ vs. 7.5% age <50), minorities, poor, living in less urban areas, uninsured, and diagnosed with more advanced stage disease (17.2% stage IV vs. 7.1% stage I; Table 1). Supplementary Fig. S1 shows observed and imputed distributions for HER2 status by age at diagnosis based on one of the 25 imputed datasets (similar results were found when we repeated the analysis with a second imputed dataset). After imputation, the frequency distribution was shifted to the older age group from the younger age group (Supplementary Fig. S1A; Fig. 1B). This is expected because a larger percentage of older women were missing HER2 status than younger women. The percentages of HER2⁺ and HER2⁻ tumors were similar across ages before and after imputation (Supplementary Fig. S1C; Fig. 1D), even though older women had more missing information.

We present 4-year estimates of breast cancer–specific survival by molecular subtypes before and after imputation (Fig. 1). Before imputation, the survival curves were overestimated, whereas after imputation, the survival curves were shifted downward. This downward shift in survival after imputation is consistent with our finding that women with missing receptor status had worse prognostic features (Table 1; Supplementary Table S2; Supplementary Fig. S2). Based on imputed datasets, the best survival was observed among women with the HR⁺/HER2⁻ subtype (survival rate 92.5% at 4 years), followed by HR⁺/HER2⁺ (survival rate 90.3%), HR⁻/HER2⁺ (survival rate 82.7%), and finally worst survival for the triple-negative subtype (survival rate 77.0%).

Although molecular subtypes affected survival, stage at diagnosis appeared to be the most powerful factor (Fig. 2; Supplementary Fig. S2). For example, based on imputed datasets, the survival rate for the triple-negative subtype (known for poor prognosis) was 95.1% among stage I disease, yet dropped to 11.2% among those with stage IV disease (Fig. 2; Supplementary Table S3). Survival rates were similar between the HR⁺/HER2⁻ (the most common subtype known to have the best prognosis) and HR⁺/HER2⁺ subtypes among women with stage I and II disease but not stage IV disease, where survival rate was significantly higher for HR⁺/HER2⁺ than HR⁺/HER2⁻ subtypes (45.5% HR⁺/HER2⁺ vs. 35.9% HR⁺/HER2⁻). In Supplementary Table S3, we summarized survival estimates by molecular subtypes and other demographic and clinical characteristics.

After controlling for demographic and clinical factors, breast cancer–specific survival was significantly worse among the triple-negative subtype [HR, 2.5; 95% confidence interval (CI), 2.4–2.6], HR⁻/HER2⁺ (1.2; 95% CI, 1.1–1.3), but not for the HR⁺/HER2⁺ subtype (0.7; 95% CI, 0.7–0.8), compared with HR⁺/HER2⁻ as the referent subtype over this entire observation period (Table 2). Other factors associated with worse breast cancer–specific survival included advanced disease stage (32.7; 95% CI, 29.9–35.6), high grade (2.1; 95% CI, 1.9–2.2), non-Hispanic black race (1.1; 95% CI, 1.0–1.1), no surgery (2.8; 95% CI, 2.7–3.0), and poverty (low SES: 1.2; 95% CI, 1.1–1.3). Being married and having any type of insurance had a protective effect on survival.

This is the first study to assess breast cancer survival according to molecular subtypes and important clinical and demographic features using the largest population coverage to date at the national level in the modern treatment era and correcting for missing information with unknown receptor status to obtain a more accurate estimate of breast cancer prognosis in the population. We found that breast cancer cases with HR⁺ subtypes are associated with the best prognosis, also shown in previous studies (17, 18, 27). By contrast, women with HR⁻ subtypes, especially those with triple-negative disease, suffer the worst prognosis, likely because of the lack of a receptor target (e.g., ER, PR, HER2) for therapy (18, 19, 28). In addition, stage is one of the most powerful factors determining survival outcomes. For example, among those with stage I disease, survival rate was greater than 95% regardless of subtypes after 4 years of follow-up. Among those with stage IV disease, women with the HR⁺/HER2⁺ subtype appeared to have better survival than women with the HR⁺/HER2⁻ subtype, even after controlling for other clinical and demographic factors. A similar pattern in terms of favorable survival with the HR⁺/HER2⁺ subtype was also shown using California Cancer Registry data (17, 18).

The triple-positive subtype has the advantage of expressing all therapeutic targets (ER, PR, HER2), which likely accounts for the superior survival we observed. Recently, a profusion of HER2-targeted therapies has entered clinical practice, including the combination of trastuzumab, pertuzumab, and chemotherapy which has yielded a median survival of nearly 5 years (56.5 months), substantially exceeding earlier median survival estimates in range of 2 years (8). Additional HER2-targeted options include trastuzumab emtansine and lapatinib (9, 10). These agents may be combined with endocrine therapy, chemotherapy, or each other, and are given sequentially in various permutations for many lines of treatment. Thus, treatment of the HR⁺/HER2⁺ subtype has advanced beyond what is available for other types of breast cancer. Our observation of a clearly superior survival with HR⁺/HER2⁺ disease in stage IV, and less so in earlier stages, might be due to the fact that stage I–III patients receive early HER2-targeted therapy, which may select for disease that is more resistant to HER2-directed therapies at the time of distant metastatic recurrence. Therefore, stage IV patients have the opportunity for maximal benefit from HER2-targeted therapy, because they have not already developed resistance to it.

This is the first time tumor receptor status was imputed for survival analyses using population-based registries. Use of multiple imputation to fill in missing receptor status is important. If we do not impute cases with missing receptor status, we overestimate survival because those with missing receptor status tend to have worse prognostic features. We used a rich set of covariates in the imputation model to make the MAR assumptions viable. Moreover, when we evaluated missing HER2 status by clinical, demographic, and breast cancer survival, missingness was explained by these factors, making MAR assumption more plausible. One limitation of our imputation model was that we did not have information on some other potential predictors of missing HER2 status such as breast cancer risk factors, treatment, and comorbid conditions. The SEER program recommends using imputed datasets when assessing breast cancer survival in the U.S. general population. These imputed datasets can be made available through SEER*Stat software to researchers in the future (29).

There are several strengths of our study. Our results are population-based and incorporate high-quality U.S. cancer registry data. SEER registries reliably capture breast cancer cases in their catchment areas, and they have complete follow-up information for greater than 95% of cases, so reporting of survival is reliable (30). Our results are more generalizable than those from single centers, and clinical trials are unlikely to include representative samples of older, sicker, and low-income patients (31, 32). Therefore, our study best reflects outcomes among unselected U.S. breast cancer cases experiencing typical patterns of care.

A few limitations of our study should be noted. Because we relied on data collected by cancer registries, we did not have detailed information on use of treatments including endocrine, HER2-directed therapy and chemotherapy, individual-level SES, breast cancer risk factors, or comorbid conditions. Controlling for these factors, if possible, would likely reduce confounding and improve our understanding of survival differences by subtype. The maximum follow-up time (59 months) was limited by the year in which relatively complete receptor data became available in SEER. In particular, over a decade, the use of Herceptin and other targeted therapy dramatically changed prognosis for breast cancer, especially for HER2-positive tumor. However, central cancer registries collected HER2 status beginning with cases diagnosed only recently (from 2010+), limiting our ability to assess long-term survival trend from breast cancer in the early part of Herceptin era.

In conclusion, we found that breast cancer prognosis in the U.S. population varies significantly based on molecular subtypes and associated clinical and demographic features. We also found that for de novo metastatic disease, women with the HR⁺/HER2⁺ subtype (once considered a poor prognostic feature) have better survival than those with the HR⁺/HER2⁻ subtype (often considered the best prognostic feature). This remarkable divergence of survival curves is likely attributable to major advances in HER2-targeted treatment (8–10, 33). As follow-up time increases, SEER data can be used to monitor and better understand the impact of how targeted therapies are contributing to reduce breast cancer mortality in the U.S. population.

R. Andridge is a consultant/advisory board member for NCI. No potential conflicts of interest were disclosed by the other authors.

Conception and design: N. Howlader, K.A. Cronin

Development of methodology: N. Howlader, K.A. Cronin

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): N. Howlader

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): N. Howlader, R. Andridge

Writing, review, and/or revision of the manuscript: N. Howlader, K.A. Cronin, A.W. Kurian, R. Andridge

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): N. Howlader

Study supervision: N. Howlader

This study was supported by the Surveillance Research Program in the Division of Cancer Control and Population Sciences at the NCI.