Method of Detection and Breast Cancer Survival Disparities in Hispanic Women

Differences in cancer survival between race or ethnic groups are often preceded by differences in stage or tumor characteristics at diagnosis (1-3). Hispanic and African American women are frequently diagnosed at a later stage and with negative prognostic features, such as greater tumor size, higher grade lesions, and a higher proportion of hormone receptor–negative tumors than white non-Hispanic women (1, 2, 4), and experience greater breast cancer–related mortality in comparison. Such tumor characteristics may be a product of several factors, but those most strongly suspected include access to or utilization of mammography screening, and more aggressive tumor biology, which results in faster growing tumors. In New Mexico, Hispanic women have had at least a 1.5-fold greater risk of breast cancer–specific mortality than white non-Hispanic women since 1995. A proportion of that difference seems to be attributable to the characteristics of the cancer at diagnosis. However, because we have limited understanding of how to directly prevent advanced tumors, earlier steps in the causal sequence may provide clues to mortality reduction. Breast cancer screening is believed to influence cancer survival by creating a “stage shift” to detection of earlier stage, less aggressive tumors (5, 6). However, screening outside of randomized trials is associated with two biases that provide a false appearance of survival benefit (7, 8): (a) lead-time bias, in which screening artificially seems to extend the cancer survival time by reducing the age and stage at diagnosis, although screened women do not live to an older age than they would have if unscreened; and (b) length bias, which is defined as the propensity of screening to pick up slower-growing cancers with a favorable prognosis, because such cancers have a longer preclinical phase in which they can be screen-detected. In contrast, more aggressive tumors quickly progress to symptoms, which lead to more frequent detection outside regular screening intervals. Because it is difficult to adjust for these biases, or to quantify them, nonrandomized studies are likely to show a survival advantage of screening, regardless of whether a true benefit of screening exists.

However, the method of breast cancer detection (symptomatic versus screen-detected) is a screening-related measure that has been independently related to prognosis after comparing women diagnosed at the same stage, and after adjusting for markers of tumor aggressiveness, approaches aimed at reducing lead-time and length biases, respectively (9-13). Thus, method of detection seems to provide independent information regarding breast cancer outcome beyond a stage shift. We sought to determine whether method of detection was independently related to survival under similar conditions in our data, and if so, whether and to what extent it influenced the disparate survival in New Mexico Hispanic women.

All white women residents diagnosed with an invasive breast cancer in the six-county central New Mexico (NM) region (Bernalillo, Sandoval, Santa Fe, Socorro, Torrance, and Valencia counties, which include 50% of NM residents) from January 1, 1995 through December 31, 2004 were eligible for the study. Population-based cancer reporting has been conducted in New Mexico since 1973 by the Surveillance Epidemiology End Results (SEER) program funded by the National Cancer Institute. Breast cancer cases diagnosed only through death certificate or autopsy studies were not included, nor were women who received a diagnosis of ductal carcinoma in situ. Eligible women were considered “Hispanic” if so identified by SEER through the use of the North American Association of Central Cancer Registries Hispanic Identification Algorithm. Race/ethnicities other than white were not included because the small number of cases precluded most analyses, and women with unknown ethnicity (n = 113) were excluded. We matched included women against breast cancer imaging records collected in 1992-2004 by the New Mexico Mammography Project, a participating site in the Breast Cancer Surveillance Consortium (14).

Data available through the New Mexico Mammography Project, specifically the indication for exam, as well as the assessment information collected according to the Breast Imaging Reporting and Data System (BI-RADS) established by the American College of Radiology, were used to ascertain the method of detection of breast cancer. Women who received a routine bilateral mammogram with an indication of “screening” within 270 days prior to the SEER date of breast cancer diagnosis were considered screen-detected as long as: (a) neither the woman, her radiologist, nor the radiology facility records listed breast cancer symptoms or an earlier nonscreening exam during that period, and (b) the BI-RADS assessment code for the exam result was not “1” (normal) or “2” (benign), but was instead noted a suspicious finding and/or recommended further follow-up. Women whose earliest exam within the 270 days prior to diagnosis had an indication suggesting symptoms, such as “evaluation of breast problem” (BI-RADS indication, 4) or who self-reported symptoms were considered “symptomatic at diagnosis.” Women whose only mammographic exam occurred within 30 days after the SEER date of breast cancer diagnosis were also included in this category, because SEER records the date of diagnosis as the first date of evidence of breast cancer, which sometimes precedes breast imaging. Women who reported breast augmentation were excluded(n = 69) due to the potential diminishment in screening accuracy (15).

Women were followed through May 2009 by the NM SEER registry, and vital status was determined by probabilistic matching to the NM State Vital Statistics Bureau files and the National Death Index of the National Center for Health Statistics using SEER*DMS (Data Management System) software. Vital status was verified by submission of files to the Centers for Medicare and Medicaid Services, and any indeterminate matches were resolved by individual review. Hispanic women have a lower age-adjusted all-cause mortality rate than non-Hispanic white women (4), which can mask any elevation in cause-specific mortality rates. Thus, only deaths attributed to breast cancer were considered events in the analysis. Women whose underlying cause of death was classified as “Breast Cancer” [International Classification of Diseases (ICD) version 9 code 174, or ICD-10 codes C50.1-C50.9] were considered “cases” in the analysis. Women with an unknown cause of death (n = 311) were excluded.

Cox proportional hazards models were constructed to evaluate potential contributors to breast cancer–specific survival. In these models, months of follow-up served as the time variable, and deaths attributable to breast cancer as the events of interest. Deaths due to other causes were censored. Hazard ratios (HR) and 95% confidence intervals (95% CI) were calculated. In initial analyses, we sought to determine whether method of detection was related to survival, independently of established prognostic variables. Cox models containing a term for “method of detection” were fit separately by tumor size group, lymph node status, and hormone receptor status, to ascertain whether method of detection was related to survival in each stratum. The relationship between method of detection and breast cancer survival was also examined separately by ethnicity, and differences in the hazard ratio according to ethnic group were assessed by fitting an interaction term.

We next evaluated whether method of detection and other prognostic factors contributed to differences in breast cancer survival between Hispanic and non-Hispanic white women. We developed a baseline model to estimate a HR for women of Hispanic ethnicity, adjusted for age and diagnosis year. To determine whether method of detection influenced the disparate survival in Hispanic women, we added the detection method variable to the baseline model, and noted any change in the Hispanic ethnicity HR. Any decline in the elevated Hispanic HR with the introduction of the detection method variable suggests the proportion of the HR that may be attributed to the included variable. This proportion was estimated by determining the absolute change in the HR from the unadjusted to the adjusted model as a fraction of the excess hazard ratio in the unadjusted model [(HR_U-HR_A)/(HR_U-1)] (16). For example, if an unadjusted HR of 1.6 (HR_u) after adjustment for an included variable changed to 1.4 (HR_A), the proportion of the HR of 1.6 that could be attributed to the included variable would be: (1.6-1.4)/(1.6-1.0), or 33%. In addition, because this measure can be asymmetrical when calculated using HRs rather than regression coefficients, we also calculated the proportion of change in the parameter estimate due to adjustment ([1-(log HR_A/log HR_u)]; ref. 17).

Potential confounding factors evaluated included age at cancer diagnosis (5-year age groups) and census tract income (proportion living below the poverty level), education (proportion with less than a high school education), and rural residence (proportion of census tract considered rural), calculated with ethnicity-specific U.S. census information (included as continuous variables to allow adequate cell sizes). Other factors assessed for their role in Hispanic survival differences included American Joint Committee on Cancer tumor size and histologic grade (indicator variables for T₂, T₃/T₄, and grade 2, grades 3/4), lymph node involvement (yes/no), and estrogen receptor (ER) and progesterone receptor (PR) status (both positive, versus not). Her2/neu was not assessed routinely during most of the study period. In addition, we examined SEER information on the first course of therapy (surgery, radiotherapy, chemotherapy, hormonal therapy).

All models were adjusted for age at diagnosis. The proportional hazards assumption was verified by examination of cumulative sums of martingale residuals (18) as implemented in SAS software, v9 (SAS Incorporated). As the combined estrogen and progesterone receptor status variable was nonproportional (19, 20), the receptor data were modeled using piecewise linear covariates for 0 to 24, 25 to 84, and ≥85 months of follow-up. All analyses were approved by the Human Research and Review Committee of the University of New Mexico Health Sciences Center.

A total of 5,067 eligible women diagnosed with invasive breast cancer from 1995 to 2004 were followed until May 2009 (median, 87 months; range, 1-172 months). At least one mammography report from the 270 days prior to or 30 days after diagnosis was available for 87.1% of women (n = 4,402). Missing reports were not more likely to be from residents of lower-income, lower-education, or rural census tracts, but were more likely to be from women diagnosed with stage III or stage IV disease. Of those with exam records, 88.4% of women (n = 3,891; 76.8% of all eligible) could be classified as symptomatic at diagnosis or screen-detected. The remaining women either had a missing or inconsistent BI-RADS code, or had experienced symptoms and signs that required a short-term follow-up for some time, with the indication for the initial exam unclear. Hispanic women were equally as likely as other breast cancer cases to have available a mammogram that led to diagnosis. However, a slightly higher proportion of Hispanic than other women (20% versus 15.1%) had only one mammography record in the NM Mammography Project files (most often, the one that led to diagnosis), suggesting that they may have received less previous screening. Among the final study population of 3,891 women, 490 deaths attributable to breast cancer occurred (12.6%; Table 1).

To determine whether the method of detection had an influence on breast cancer survival, we first used previously established methods to verify that this variable was independently related to survival among women with tumors of the same size and hormone receptor status (ER and PR positive, versus not). An increased risk of breast cancer–related death was apparent when examined within each tumor size, lymph node status, and hormone receptor group (data not shown). However, within strata defined by both a specific tumor size and specific receptor status, detection method was less strongly related to survival in the non-ER/PR-positive groups, probably due to small cell sizes. These data provide additional evidence that symptomatic status at diagnosis might be an independent prognostic indicator, as indicated by other studies (9-13).

Women with symptom-detected cancers were 2.3 times more likely to die of their breast cancer than those who were screen-detected (95% CI, 1.9-2.6; Table 2). After adjustment for tumor characteristics associated with lead-time and length bias, those detected with symptomatic cancer remained at 1.6-fold increased risk for breast cancer–related death. Hispanic women were more likely than non-Hispanic white women to be symptomatic at diagnosis (48.3% versus 39.4%; Fig. 1). However, the HR associated with method of detection did not differ between the two groups (P = 0.11 for difference; Table 2).

In a Cox proportional hazards model, Hispanic women had a 1.5-fold increased risk of breast cancer–related death, compared with non-Hispanic white women (95% CI, 1.3-1.9). However, when a variable indicating method of detection was included in the Cox proportional hazards model, the increased risk of death among Hispanic women declined only slightly (Table 3). These results suggest that at most, only a modest proportion (10%) of the 0.5-fold excess risk was due to being diagnosed when symptomatic, in comparison with being detected at screening.

To determine how other prognostic indicators affected the Hispanic survival disparity after inclusion of method of detection in the Cox model, we added additional terms for tumor size, histologic grade, and nodal status. The 1.5-fold increased risk of breast cancer death in Hispanic women was reduced to 1.2 (95% CI, 1.0-1.5). The HR did not decline further with the addition of variables indicating either hormone receptor status or first course of treatment (Table 3). Although the proportion of excess risk that is potentially attributable to each modeled variable can be calculated, because these variables are correlated, that proportion is dependent upon the sequence in which they enter the model. Therefore, the attributable proportion is best interpreted as the sum effect of several variables. For variables included in Table 3, it is approximately 50% (1.50-1.23/1.50-1). We also entered method of detection into a model after adjusting for tumor characteristics. As expected, it accounted for even less of the Hispanic survival difference (4%; data not shown). To the extent that method of detection remains correlated with lead-time and length biases and is also an indicator of some true benefit of screen detection, inclusion in the model prior to tumor characteristics allows some adjustment for all of those screening-related factors as potential contributors to Hispanic survival differences. However, the Hispanic HR changed very little, suggesting that although method of detection does influence survival, it does not account for most of the differential survival by ethnicity in this population.

Method of detection and tumor characteristics continued to be strongly and independently related to breast cancer–specific survival after mutual adjustment. These results point towards additional factors that may influence the probability of breast cancer survival, possibly prior to diagnosis and treatment.

Differences in breast cancer–specific survival by race/ethnic group have not proven to be easy to ameliorate. Although disparities in survival are often preceded by differences in stage at diagnosis, and thus adjustment for stage or related factors reduces the disparity HR, diagnosis stage or tumor characteristics in and of themselves are not preventable. The HR reduction due to these factors merely points to earlier events in the causal sequence for possible prevention efforts. One of the leading suspected contributors to survival disparities is lack of access to, or utilization of, mammographic screening. Another is tumor biology. Method of detection is associated with both in that women who receive regular screening but who have biologically aggressive tumors may well be symptomatic at diagnosis, as will nonscreeners.

Breast cancer detected when symptomatic versus by screening was related to a 1.6-fold increased risk of breast cancer–related death in our study, after adjustment for other prognostic indicators. However, a diagnosis when symptomatic rather than at screening is not a strong contributor to the disparate survival evident in Hispanic women in our data. Our results should be understood in light of the strengths and weaknesses of our study. Our findings may be limited by the misreporting of screen-detected versus symptomatic status. Such misclassification could bias the HR associated with method of detection towards the null (HR, 1.0) if unrelated to later survival (nondifferential misclassification). The adjusted HR for detection method has ranged from 1.3 in survival studies with 5 years or less of follow-up (11, 21), to 1.9 in studies of tumor recurrence with 9.5 median years of follow-up (9), and our results fall squarely in that range. However, an attenuated HR for detection method could diminish the proportion of the Hispanic survival difference that would be accounted for by this variable. In contrast, it is possible that detection method is not independently related to survival, or related only weakly, and that the 1.6-fold HR among women symptomatic at diagnosis is artificially elevated due to residual lead-time and length bias. In three studies that adjusted for markers of tumor aggressiveness, such as Her2/neu, tp53, and Ki-67, method of detection remained independently related to breast cancer outcomes (9, 11, 22). Method of detection has also been related to breast cancer outcomes among strata defined by stage, lymph node status, and other standard tumor characteristics that might be shifted by screening and thus associated with lead-time bias (9, 10). Thus, method of detection has seemed to be an independent prognostic factor when evaluated to the extent of current knowledge. The strengths of our study include its population-based approach, with inclusion of 77% of eligible invasive breast cancer cases in the defined diagnosis years and counties. We also used electronic data files collected by standardized methods, and classified women as symptomatic or screen-detected by applying a data analysis algorithm, blinded to outcome.

Our findings of only a moderate influence of method of detection on breast cancer survival disparities in Hispanic women, if replicated by other studies, suggest that we need to look elsewhere for sources of this difference. Although the gap in breast cancer survival between ethnicities could possibly be reduced if method of detection was equivalent between the two groups, our data suggest that a much larger independent contribution is made by negative prognostic features of the tumor at diagnosis. Differences in tumor biology and their underlying causes are a potentially fruitful area for exploration to further understand and move towards the goal of eradicating survival differences.

No potential conflicts of interest were disclosed.

Grant Support: Carl C. and Marie Jo Anderson Foundation, by National Cancer Institute (NCI) Cooperative agreement U01 CA69976 to the New Mexico Mammography Project as a member of the Breast Cancer Surveillance Consortium, NCI Contract NO1-PC-35138 to the New Mexico Tumor Registry, and by the University of New Mexico Cancer Center, a recipient of NCI Cancer Center Support Grant P30-CA118100.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.