What Predicts an Advanced-Stage Diagnosis of Breast Cancer? Sorting Out the Influence of Method of Detection, Access to Care, and Biologic Factors

An advanced-stage diagnosis of breast cancer has long been associated with significantly poorer survival outcomes (1). Recent data show that women diagnosed at American Joint Committee on Cancer (AJCC) stage I have an overall 5-year relative survival rate near 100%, whereas the rate for those diagnosed at stage IV is 24% (2). Over two decades of investigations into the determinants of an advanced-stage diagnosis have yielded important findings.

Screening mammography has been consistently associated with earlier-stage detection of breast cancer, both in clinical trials (3–7) and in day-to-day practice (8–10). This is notwithstanding important complicating factors, including disagreement about appropriate screening strategies (11, 12) and variability in mammographic test sensitivity driven by certain biologic factors including mammographic breast density (13–15).

There are race/ethnicity differences in breast cancer stage at diagnosis (16–30), with African American women significantly more likely than white women to have a late-stage diagnosis. Insurance status is a significant independent predictor of stage, with women who are uninsured or enrolled in Medicaid less likely to access screening mammography (31, 32) and more likely to be diagnosed at later stage (20, 21, 33–35). There is a complex interplay involving the presence of comorbidities, access to care, detection by mammography, and stage (36, 37).

Taken together, these studies have significantly contributed to our understanding of factors associated with an advanced stage diagnosis of breast cancer. However, there are certain methodological considerations, not explored to date, with potentially important implications for the specification of models and interpretation of findings.

First, in virtually all studies, the likelihood of an advanced-stage diagnosis has been analyzed through a single-equation (typically logistic) regression model in which explanatory variables, including method of detection, were all regarded as independent, exogenous predictors of stage. In reality, the detection method is not a fixed, predetermined variable in the same sense as the individual's age or race/ethnicity. Rather, it can be regarded, as indeed it has been in the screening mammography trials, as an “exposure” influencing the “outcome” of stage at diagnosis.

Second, many analyses have relied heavily on cancer registry sources that do not routinely include several potentially important predictors of both stage and method of detection. Such unobserved variables may include breast density, which influences both mammography sensitivity (14, 15) and tumor aggressiveness (38); whether the woman is taking hormone replacement therapy, which may affect tumor development (39) and the perceived importance of regular screening; whether the woman has a family history of breast cancer; the nature of the woman's health care system (e.g., managed care vs. fee-for-service), which may influence both screening rates and the effectiveness of follow-up care (40); and certain health behaviors, e.g., excess alcohol consumption (41).

Third, to the extent such unobserved variables are important predictors of both stage and detection method, the statistical problem of endogeneity arises: the error structures for the regression models predicting stage and predicting method are then correlated (because they contain common unobserved variables). Without corrections for such potential endogeneity, estimates of the impact of predictors—such as method, race/ethnicity, and insurance—on stage are subject to bias (42).

In this article, we bring a new conceptual framework and accompanying statistical modeling strategy—built primarily around the method of instrumental variables (IV; refs. 42–47)—to a much-analyzed question: What predicts an advanced-stage diagnosis of breast cancer? Of particular interest is whether findings to date regarding the impact of method of detection, race/ethnicity, and insurance status on stage are sustained within this expanded framework.

Our maintained hypothesis regarding the causal factors leading to the breast cancer stage recorded at diagnosis (stage) is depicted in Fig. 1. Method of detection (method) plays a central role, and we specify two sets of variables that may influence both stage and method. One set contains variables associated with the woman's access to and utilization of health care (e.g., insurance status). The second set consists of variables associated with the aggressiveness and speed of the tumor's biologic development (e.g., grade), the sensitivity of detection methods (e.g., histology), or both [e.g., body mass index (BMI)]. Each set includes variables that, depending on the available data sources, may be observable (e.g., marital status) or unobservable (e.g., breast density, menopausal hormone therapy) to the investigator.

The principal source of data is the Breast and Prostate Cancer Data Quality and Patterns of Care Study (POC-BP), funded by the National Program of Cancer Registries (NPCR) of the Centers for Disease Control and Prevention (CDC) and involving investigators affiliated with population-based registries in seven states (California, Georgia, Kentucky, Louisiana, North Carolina, Minnesota, and Wisconsin) and the CDC. Institutional Review Board approval was obtained from all participating states, academic institutions, and government agencies.

Over 2007 to 2009, the POC-BP sampled NPCR patients diagnosed in 2004, with intensive re-abstraction of medical records from hospitals and outpatient facilities (including pathology laboratories, radiation facilities, surgical centers, and physician offices).

Our analyses included women ≥ 20 years of age diagnosed in 2004 with microscopically confirmed in situ or invasive primary breast cancer (International Classification of Disease-Oncology, 3rd Edition, site codes C50.0-C50.9) with no previous cancer diagnosis and meeting other standard exclusion criteria. Cases diagnosed at Veterans Affairs hospitals were excluded because of data availability limitations.

Cases were selected from the NPCR registries through single-stage random sampling stratified by race/ethnicity in all states and by other factors that varied by state (e.g., by urban/rural status in Georgia). A detailed account of data collection and quality assessment for POC-BP has been reported (48).

Patients were assigned an AJCC (Sixth Edition TNM) stage based on the collaborative stage algorithm in effect for 2004 diagnoses. There is wide variability in how previous studies have defined advanced (or late) stage of breast cancer: III or IV (24, 27, 30); II, III, or IV (19, 21, 22, 29); IIB, III, or IV (36). In response, we defined advanced stage on the basis of the pattern of decline by AJCC stage in 5-year overall survival rates. SEER*Stat analyses (2) on cases diagnosed in 2004–2010 and followed through 2011 yielded these 5-year survival percentages by AJCC stage: 0 (95.4), I (91.9), IIA (86.8), IIB (81.9), IIIA (76.9), IIIB (54.0), IIIC (59.6), and IV (20.8). Given the sharp drop-off between IIIA and IIIB, we designated “advanced” stage as diagnosis at IIIB, IIIC, or IV. All others were diagnosed at an “earlier” stage.

We defined a two-level variable, mammography and other, where “other” included detection by clinical breast examination (CBE), breast self-exam (BSE), or signs/symptoms. Each patient was assigned a detection method based on a detailed review of medical records at the site(s) where she received breast cancer care. The intent was to capture the initial detection-related event that triggered steps toward a definitive diagnosis. Thus, if an initial BSE led to a mammogram, which led eventually to a breast cancer diagnosis, the coded method of detection would be BSE; see Table 1 for more detail.

These included race/ethnicity, insurance status, comorbidity status [based on Piccirillo's Adult Comorbidity Evaluation (ACE) instrument (49)], marital status, age, and several area-level variables constructed from 2000 US Census data: urban/rural status, poverty status, and education status. These categorical variables are operationally defined in Table 1. 

In addition to race/ethnicity and age, we included BMI, tumor grade, and a constructed variable “mole-subtype” based on the patient's combined estrogen receptor (ER), progesterone receptor (PR), and HER2 status and intended to approximate the molecular subtype of the breast tumor (50). Specifically, a patient here may be “Luminal A” (ER⁺ and/or PR⁺, HER2⁻), “Luminal B” (ER⁺ and/PR⁺, HER2⁺), “triple negative” (ER⁻, PR⁻, HER2⁻), or “HER2 Type” (ER⁻, PR⁻, HER2⁺).

Our overall strategy is to compare conventional single-equation regression models with formulations designed to detect and correct for selection bias, under a variety of assumptions. Prototypically, the single-equation models will be binary logistic regressions with the log-odds of an advanced-stage diagnosis being a function of method of detection plus some combination of patient-level access/utilization factors and biologic factors. From the standpoint of Fig. 1, such single-equation models correspond to a conceptual framework that omits both arrows directed at method of detection.

To test and correct for any selection bias, our primary approach is the method of instrumental variables using the two-stage residual inclusion (2SRI) model (45–47, 51), which is especially well suited for nonlinear estimation, as here. Figure 2 is a transformation of Fig. 1 that depicts key aspects of our 2SRI econometric model, including the main observable and unobservable variables thought to be at play. To execute the 2SRI model, one estimates a first-stage regression in which the likelihood the patient receives the “intervention” (here, mammography) is a function of all observable predictors posited to influence the patient's “outcome” (here, advanced vs. earlier stage) plus instrumental variable(s), hypothesized to influence the patient's selection into intervention but not her outcome (except through their impact on choice of intervention). In the second-stage regression, the likelihood of advanced stage becomes a function of method of detection, the hypothesized access/utilization factors and biologic factors, and a variable consisting of the residuals computed from the first-stage regression and intended to both indicate the degree of selection bias and correct for it (42, 46).

Because the residual is a computed variable, thereby reflecting sampling error, we employed bootstrapping (100 iterations) to upwardly adjust the standard errors of coefficient estimates in the second-stage regression.

The IVs deployed here are seen in Fig. 2 and discussed further in Table 2: the patient's state of residence [mammography rates vary across the seven states, reflecting an underlying geographic variability in screening uptake (52)]; histology [because mammography is less sensitive for lobular tumors (52), while histology itself is posited not to be an important predictor of tumor aggressiveness after adjusting for other biologic variables]; and a constructed variable, mammography-capacity, indexing a woman's physical access to mammography in her county of residence. Because of state-imposed confidentiality requirements, mammography-capacity could not be computed for Minnesota.

The strength of the IVs is indexed by the magnitude of the F-statistic for the null they are jointly 0 in the first-stage regression; a frequently used, if informal, benchmark is that F ≥ 10 (42, 53).

For three predictor variables with substantial missing observations among the cases potentially available for analysis—mole-subtype (29.8%), BMI (22.0%), and grade (8.6%)—we used multiple imputation (MI) to assign values (54, 55). (No other variable was missing more than 3%, and most were missing under 1%.) When MI was applied to the 2SRI models, standard errors were constructed to reflect the sampling variability arising from both the computed residuals and the imputation process; see Supplementary Materials (Section A).

In regressions using data from single-stage sampling, where the weights are a function of predictor variables (here, for example, race–ethnicity), using sample weights can, at a minimum, reduce statistical precision (56); consequently, we did not weight the data in the base-case.

With binary logistic regression used throughout, our complementary measures of model performance are the coefficient of concordance (c-statistic) and the Hosmer–Lemeshow (H-L) goodness-of-fit test statistic (57). H-L indexes how well predicted and observed event rates (here advanced-stage) match up in subgroups (typically deciles) of the sample; the closer the match, the higher the P value.

Regression results are expressed as adjusted ORs with corresponding 95% confidence intervals; P values are two-sided, with P ≤ 0.05 as a benchmark for appraising statistical importance. Analyses used Stata, version 13.0 (Stata Corporation), and SAS, version 9.2 (SAS Institute, Cary, NC).

In summary, our base-case regression model for stage at diagnosis was a 2SRI specification estimated (1) without sample weights; (2) with missing values imputed for BMI, grade, and mole-subtype (but not other predictors); and (3) with the IV mammography-capacity included. As detailed below, we conducted extensive sensitivity analyses, including propensity score weighting as an alternative to IV for selection bias reduction (58–62).

From a total 11,643 qualifying breast cancer cases, re-abstraction was successfully completed on 9,142. Among these, 212 were not assigned an AJCC stage, and 230 additional patients were missing information on method of detection. Hence, 8,700 cases were potentially available for analysis (because, following standard practice, we did not apply MI to our two dependent variables, stage and method). Under base-case modeling assumptions (which exclude Minnesota), the corresponding sample has 7,503 patients. Among these, 762 (10.2%) were diagnosed at advanced stage (Table 1), and 3,718 (49.6%) cases overall were detected by mammography (Table 2).

For most predictor variables in Table 1, there were notable differences in the distribution of patients between advanced versus earlier stage at each level of the variable, and the corresponding unadjusted ORs from the binary logistic regression of the variable on stage were significant at P < 0.05 in most cases. For example, among those detected by mammography, only 2.6% were at advanced stage, whereas for those detected by some other method, 17.6% were advanced stage; the unadjusted OR for detection by mammography (vs. other method) being associated with advanced stage was 0.13 (P < 0.001).

Table 2 presents a parallel summary of information for the IVs. Of prime interest is the association between each IV and the likelihood of detection by mammography. For histology and mammography-capacity, the unadjusted ORs were significant, in the expected directions; although the state variable was not as strongly associated with method, we elected to retain it as an IV, given a priori expectations about geographic variations in screening practices.

The motivating purpose of the first-stage regression is to derive the “Method of Detection Bias Correction Factor” (Fig. 2)—that is, a variable consisting of that model's residuals which then enters the second-stage regression for stage. Regarding the statistical strength of the IVs, the F statistic (8 df) for the null that state, histology, and mammography-capacity are jointly noninfluential was 105.7 (P < 0.001), well above the benchmark of F ≥ 10. This estimated first-stage model is discussed in Supplementary Materials (Section B).

The right-hand portion of Table 3 displays the estimated second-stage regression for the base-case 2SRI model. The likelihood of an advanced-stage diagnosis is strongly negatively related to detection by mammography (OR = 0.04; P < 0.001). Of note, the first-stage residual is significantly positive (OR = 3.89; P = 0.02), consistent with selection bias in the “allocation” of women to mammography versus other.

Advanced stage was positively associated with being black (OR = 1.16; P = 0.07); not being privately insured, with Medicaid (OR = 1.48; P = 0.01) and Medicare (OR = 1.29; P = 0.04), and being uninsured (OR = 1.60; P = 0.07); being single (OR = 1.28; P = 0.06); having BMI ≥ 40 (OR = 1.62; P = 0.001); having a tumor of HER2 Type (OR = 1.40; P < 0.001); and having tumor grade that is moderately differentiated (OR = 2.36; P < 0.001) or else poorly or undifferentiated (OR = 3.91; P < 0.001). Advanced stage was negatively related to being diagnosed at age < 40 (OR = 0.48; P = 0.001) or between ages 40 and 49 (OR = 0.57; P < 0.001).

Results from the corresponding single-equation multivariable regression are in the left-hand portion of Table 3. Notwithstanding the significant bias correction term in the 2SRI model, there was general concordance in findings from the two models. The models had comparable within-sample predictive ability (c = 0.797 and 0.796), though the 2SRI model had a notably better HL statistic (P = 0.88 vs. 0.53).

Multiple model variants were analyzed where, in each case, we altered one key base-case provision while retaining the others; see Supplementary Tables S1–S6 in Supplementary Materials (Section D). These variants included models that (i) excluded the biologic variables associated with tumor progression (Supplementary Table S1); (ii) did not impute missing values (Supplementary Table S2); (iii) did employ the sample weights (Supplementary Table S3); (iv) excluded the IV mammography-capacity and thus included the Minnesota observations (Supplementary Table S4); and (v) used propensity score weighting as an alternative bias-reduction technique (Supplementary Table S5). For further appraisal of these estimated propensity score models, see Supplementary Table S6 and Supplementary Materials (Section C).

Overall, the findings from Supplementary Tables S1–S5 are broadly in tune with Table 3, but there are some notable differences, as discussed below.

Guided by a new conceptual framework and statistical modeling strategy, this article re-examines a much-investigated question: What predicts an advanced-stage diagnosis of breast cancer?

We found that detection by mammography is significantly negatively related to an advanced-stage diagnosis. Across all multivariable single-equation and propensity score–adjusted models, the OR (mammography vs. other) for advanced stage ranged narrowly from 0.13 to 0.15. Across the IV models, there was a consistent pattern: this adjusted OR was in the (much lower) 0.03–0.05 range, whereas the OR for the residual predictor variable—indicating the magnitude of selection bias—was in the 2.73–6.92 range. An OR > 1 implies a positive relationship between the algebraic sign of the residuals and the likelihood of advanced stage. We posit this reflects the fact (Table 1) that over half of all tumors were detected by some other method, thus generating negative residuals from the first-stage regression, and that over 82% of these were earlier stage. In effect, the estimated 2SRI model appropriately down-weighted the credit given to other methods for “detecting” earlier stage tumors.

In interpreting these results, it is noteworthy that relatively few studies have examined the impact of method of detection on stage at diagnosis (as opposed to the more common scenario of examining the relationship between observed or reported screening behavior on stage). Analyzing data from three screening trials, Shen and colleagues (63) found a “clear shift toward earlier stage” in cancers detected by mammography. This is in line with results reported by Malmgren and colleagues (9) from a prospective cohort study of women aged 40 to 49 diagnosed across 1990–2008, from a Wisconsin study of cases diagnosed across 1987–1990 (64), and from a 2001–2003 study of cases diagnosed in Detroit and Los Angeles (23).

Because we cannot observe in the POC-BP data the actual frequency and timing of a woman's screening mammography (or her CBE or BSE), the method of detection variable is not a direct measure of the effectiveness of mammography (or CBE or BSE) in averting an advanced-stage breast cancer. Thus, we cannot know for sure that a tumor diagnosed at advanced stage by other methods would have been found at an earlier stage if the woman had been getting regular mammograms. It is possible that she had been receiving mammography (at some rate), and a small but aggressive tumor was missed and subsequently emerged as an “interval” cancer of advanced stage. One role of the method of detection variable here is to account for the net influence of these unobserved (in our data) screening behaviors on stage at diagnosis. Further discussion about the role of the method of detection variable in these analyses, and its interpretation, can be found in Supplementary Materials (Section E).

The associations of race/ethnicity with stage within 12 alternative predictive models (including the base-case and Supplementary Tables S1–S5) are explored further in Table 4, with several implications. First, while Hispanics and Asian/Pacific Islanders (API), but not blacks, were significantly less likely to be detected by mammography than whites (Supplementary Materials, Section B), the only significant race/ethnicity difference in stage was between black and white women. Second, the odds of blacks being diagnosed at a later stage than whites varied across models 1 to 12 in the following general way: the richer the set of included covariates, the less influential was the race/ethnicity variable.

This pattern of findings underscores that the estimated magnitude of a race/ethnicity effect depends on the overall maintained hypothesis embodied in the chosen statistical model for stage. That said, a significant black–white difference in breast cancer stage at diagnosis has been reported almost without exception in US studies to date (16–29). Some studies have found that black–white differences are significantly reduced after accounting for screening history (22–24), while others have not (25, 29). We conclude that a black–white difference in stage is a robust finding, but one whose magnitude and interpretation can vary across data sources, study designs, time frames, and geographic settings.

Across the models in Table 4, the trend is clear: women without private insurance were significantly more likely to be diagnosed at advanced stage. Being uninsured had the greatest adverse impact, followed by having Medicaid, and then Medicare.

These findings generally align with earlier estimates (20, 21, 33–35). However, among the privately insured, we could not distinguish fee-for-service and managed care, and there may be a differential impact of coverage regime on method of detection and stage (65, 66). Overall, the likely route by which insurance influences stage is to increase the likelihood of detection by mammography (32); in parallel, insurance may increase the odds of timely diagnosis and treatment following a positive screen (8, 40, 67). Not directly accounted for here is whether the woman had a regular source of health care or received a provider recommendation for screening (8, 40).

Patients with severe comorbidity were much less likely to be detected by mammography (Supplementary Materials, Section B) and significantly more likely to be diagnosed at advanced stage in our single-equation models; however, while OR > 1 for severe comorbidity in all 2SRI models, it was generally not significant. As Fleming and colleagues (36) note, the consolidation of multiple comorbid conditions into a single metric, such as the ACE-27, may hide antagonistic effects of individual comorbidities on either screening or advanced stage (36). Yasmeen and colleagues (37) found that comorbidities were positively associated with mammography use and also with an advanced-stage diagnosis among women who were screened most frequently.

While lower SES has been associated with late-stage breast cancer diagnosis in several studies (68–70), the only significant effect here was that single women were generally at higher risk to advanced stage than married women. The area-level variables indexing education, poverty, and urban-rural status were not significant in any model variant.

Variables hypothesized to be associated with the aggressiveness and speed of tumor development generally performed as expected. Across models (see Table 3 and Supplementary Tables S1–S5), advanced stage was positively related to whether the tumor was HER2 Type, the tumor grade was not well differentiated, and the woman was morbidly obese (BMI ≥ 40); advanced stage was negatively associated with a diagnosis under age 50. While women with triple-negative disease were significantly less likely to be detected by mammography (Supplementary Materials, Section B), triple-negative status was not independently associated with an advanced-stage diagnosis in any model. As indicated in Fig. 2, a potentially important variable here not available in the POC-BP data set was breast density.

Although the joint impact of these factors on stage-at-diagnosis has not been previously evaluated, earlier studies support portions of our findings. For example, Kerlikowske and colleagues (71) found that screen-detected cancers were higher among overweight/obese women, whereas rates of advanced-stage diagnosis increased across BMI groups, controlling for mammography use; however, an earlier study did not find a connection between BMI and stage for screen-detected cancers (64). The complex interplay involving hormonal status, postmenopausal hormone use, age, menopause, BMI, mammography use and sensitivity, and stage remains a ripe topic for investigation (15, 39, 71–77).

In recent years, observational studies have examined the impact of various factors on breast cancer stage at diagnosis: method of detection (9); race, ethnicity, and socioeconomic variables (23); and biomedical variables such as BMI (72). This article adopts the perspective that the most conceptually and statistically defensible approach to understanding the influence of each such factor is to study them in concert (Figs. 1 and 2).

Overall, our findings about the determinants of advanced-stage align with those reported variously over the past two decades. What this study does provide, through its comprehensive modeling strategy and multiple sensitivity analyses, are new—and we think better grounded—estimates of the magnitude and statistical robustness of these posited predictors for stage.

What is needed going forward are continuing efforts to expand the empirical base so that the influence of potentially important unobservables (e.g., behavioral risk factors, health system effects, clinical variables not routinely collected in population-based studies) can be gauged in the context of an ever-more-richly specified model.

No potential conflicts of interest were disclosed.

The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the CDC.

Conception and design: J. Lipscomb, S.T. Fleming, A. Trentham-Dietz, G. Kimmick, X.-C. Wu, R.T. Anderson

Development of methodology: J. Lipscomb, A. Trentham-Dietz, X.-C. Wu, K. Zhang

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Lipscomb, A. Trentham-Dietz, X.-C. Wu

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J. Lipscomb, S.T. Fleming, G. Kimmick, X.-C. Wu, C.R. Morris, K. Zhang, R.A. Smith

Writing, review, and/or revision of the manuscript: J. Lipscomb, S.T. Fleming, A. Trentham-Dietz, G. Kimmick, X.-C. Wu, C.R. Morris, R.A. Smith, R.T. Anderson, S.A. Sabatino

Study supervision: J. Lipscomb, X.-C. Wu

The authors are indebted to the editors and two anonymous reviewers for multiple constructive comments and recommendations that have sharpened the interpretation of study findings and their implications. They are also grateful to Lyn Almon, Judy Andrews, and Kevin Ward at the Georgia Center for Cancer Statistics at Emory University for analytical file development and interpretation; to Trevor Thompson in the CDC Division of Cancer Prevention and Control for database development; and to Jennifer Wike at the DB Consulting Group (a contractor to the CDC) for ongoing logistical and management support of this CDC-supported study.

From the Breast and Prostate Cancer Data Quality and Patterns of Care Study, supported by the CDC through cooperative agreements with the California Cancer Registry (Public Health Institute; 1-U01-DP000260; to R. Cress), Emory University (1-U01-DP000258; principal investigator, J. Lipscomb), Louisiana State University Health Sciences Center (1-U01-DP000253; principal investigator, X.-C. Wu), Minnesota Cancer Surveillance System (1-U01-DP000259; principal investigator, C.I. Perkins), Medical College of Wisconsin (1-U01-DP000261; principal investigator, J.F. Wilson), University of Kentucky (1-U01-DP000251; principal investigator, S.T. Fleming), and Wake Forest University (1-U01-DP000264; principal investigator, R.T. Anderson). J. Lipscomb was supported also by Cancer Center Support Grant P30CA138292 to the Winship Cancer Institute of Emory University from the NCI.

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.