|Home | About | Journals | Submit | Contact Us | Français|
Recent studies suggest trends toward more mastectomies for primary breast cancer treatment. We assessed survival after mastectomy and breast-conserving surgery (BCS) with radiation for early-stage breast cancer among non-selected populations of women and among women similar to those in clinical trials. Using population-based data from Surveillance Epidemiology, and End Results cancer registries linked with Medicare administrative data from 1992 to 2005, we conducted propensity score analysis of survival following primary therapy for early-stage breast cancer, including BCS with radiation, BCS without radiation, mastectomy with radiation, and mastectomy without radiation. Adjusted survival was greatest among women who had BCS with radiation (median survival = 10.98 years). Compared with this group, mortality was higher among women who had mastectomy without radiation (median survival 10.04 years, adjusted hazard ratio (HR) = 1.19, 95% confidence interval (CI) = 1.14–1.23), mastectomy with radiation (median survival 10.02 years, HR = 1.20, 95% CI = 1.14–1.27), and BCS without radiation (median survival 7.63 years, HR = 1.81, 95% CI = 1.70–1.92). Among women representative of those eligible for clinical trials (age ≤70 years, Charlson comorbidity score = 0/1, and stage 1 tumors), there were no differences in survival for women who underwent BCS with radiation or mastectomy. In conclusion, after careful adjustment for differences in patient, physician, and hospital characteristics, we found better survival for BCS with radiation versus mastectomy among older early-stage breast cancer patients, with no difference in survival for BCS with radiation versus mastectomy among women representative of those in clinical trials. These findings are reassuring in light of recent trends towards more aggressive primary breast cancer therapy.
Results from randomized controlled trials demonstrating equivalent outcomes for mastectomy and breast-conserving surgery (BCS) with radiation for women with early-stage breast cancer have been available since the early 1980s [1–13]. In 1990, the National Institutes of Health Consensus Development Conference recommended BCS as the preferred therapy for most women with early-stage breast cancer because it preserves the breast .
Nevertheless, substantial regional variation in the use of mastectomy versus BCS remains. In 2005, rates of mastectomy per 1,000 Medicare beneficiaries living in South Dakota were more than five times those for women in Vermont (1.93 vs. 0.38 per 1,000 female Medicare enrol-lees) . The reasons behind these variations are not clear, but women report that their surgeons’ recommendations are very influential in their decisions, and those who choose mastectomy over BCS most frequently report doing so because they fear that removal of the lump only “will not get it all” [16–19]. Recent research also suggests a trend towards more aggressive surgical treatment for breast cancer, with increases in mastectomy rates  as well as bilateral mastectomies among women with unilateral breast cancer [21–23].
These findings suggest that some women and their physicians perceive a benefit to mastectomy over BCS with radiation despite clinical trials demonstrating equivalent outcomes. Although clinical trials are the gold standard for examining treatment efficacy, only 2.5% of cancer patients enroll in clinical trials . Those that enroll in trials tend to be younger and healthier than other patients and they are typically cared for by providers who practice at or are affiliated with cancer centers [25–28]. Moreover, women in trials of primary breast cancer therapies predominantly had stage I tumors.
Through careful design and analysis, observational data may address questions regarding treatment effectiveness that are not feasible to answer within the context of randomized trials. We used cancer registry data linked with Medicare claims to assess survival up to 12 years after primary breast cancer treatment. Specifically, we questioned whether survival after mastectomy and BCS with radiation for early-stage breast cancer were equivalent among non-selected populations of women. In addition, we assessed whether outcomes varied by patient age, level of comorbid illness, or tumor size and specifically assessed outcomes among relatively younger, healthier women with smaller tumors—women represented in clinical trials.
We used the Surveillance, Epidemiology, and End Results (SEER)-Medicare data for this analysis . The SEER program of the National Cancer Institute collects uniformly reported data from population-based cancer registries; data from 12 registries representing 14% of the United States population have been available since 1992; in 2000, the SEER program expanded to include greater California, New Jersey, Louisiana, and Kentucky and currently represents 26% of the population [30, 31]. For each incident cancer, registrars collect information on patient demographics, month and year of diagnosis, and tumor characteristics.
Since 1991, the SEER data have been merged with Medicare administrative data (successfully linking data for more than 94% of SEER patients aged ≥65 years at diagnosis ). The Medicare claims data for this study included the Hospital Outpatient Standard Analytic file (outpatient facility services), the 100% Physician/Supplier file (physicians’ services and other medical services), and the Medicare Provider Analysis and Review (MEDPAR) file (inpatient services). The study protocol was approved by the institutional review boards at Harvard Medical School and the University of Iowa.
We selected women with a first diagnosis of breast cancer during 1992–2002 who were ≥66 years old and enrolled in Parts A and B of fee-for-service Medicare as of 1 year before their breast cancer diagnosis (N = 104,532). We excluded 1,224 women whose diagnosis was reported only by autopsy or death certificate and 2,332 patients with histologies suggesting a cancer other than primary breast cancer. We restricted the cohort to the 72,723 women with stage I or II breast cancer and then excluded 3,176 women who disenrolled from Part A or B of Medicare or joined a HMO or died within 9 months of diagnosis (to maximize ascertainment of primary treatment), or had no claims during the period from 45 days before diagnosis through 195 days after diagnosis (because of concern for an inaccurate match). We excluded 407 women who did not undergo primary surgery for their stage I or II cancer, resulting in 69,140 women with early-stage invasive breast cancer who underwent primary surgical therapy.
We classified primary therapy as mastectomy with radiation, mastectomy without radiation, BCS with radiation, or BCS without radiation using registry data and claims [32–35]. We examined four groups rather than focusing only on BCS with radiation and mastectomy because many elderly women undergo BCS without radiation and because radiation following mastectomy is increasingly being used to treat selected patients with stage II breast cancer. Moreover, we did not want to define eligibility criteria based on treatment choices that may have occurred after the surgery decision (e.g., decisions about radiation). For surgery, we coded the most definitive procedure in either source. For radiation, we considered a patient to have radiation if reported by the registry or if at least one code for radiation was in the claims. Codes for identifying treatments are included in Appendix Table 1 (see electronic supplementary material).
We assessed overall mortality through 2005 and breast-cancer-specific mortality through 2004. Mortality data were obtained from Medicare enrollment data and SEER registry data based on a National Death Index match.
SEER registrars document each patient’s age at diagnosis, race, Hispanic ethnicity, marital status, history of other cancer, year of diagnosis, residence in a metropolitan county, and SEER area. We assessed comorbid illness using Diagnostic Cost Groups (DCGs) [36, 37], a risk-adjustment tool used by the Centers for Medicare & Medicaid Services to predict future costs and disease burden for Medicare beneficiaries based on clinical diagnostic information from inpatient and ambulatory claims. DCGs summarize 182 conditions (compared to 18 captured by the Charlson score ). We used 2000 Census data to obtain information on education and income by census tract of residence. We classified income and education into quartiles within registry and then combined across registries to avoid misclassification due to regional variations in income and education levels. Tumor characteristics included American Joint Committee on Cancer (AJCC) stage, 3rd edition  (staging for breast cancer did not change appreciably during the study period), grade, tumor size, histology, estrogen, and progesterone receptor status. We assessed receipt of chemotherapy based on procedure codes through 1 year following diagnosis (Appendix Table 1 (see electronic supplementary material)).
For each surgeon, we used the physician Unique Provider Identification Number (UPIN) to link with American Medical Association data to obtain information on physician sex, specialty, year of medical school graduation, and foreign graduate status .
We used data from the Hospital File for information on hospital characteristics , including bed size, hospital ownership, urban/rural location, teaching status, medical school affiliation, National Cancer Institute cancer center, participation in clinical trials, presence of an American College of Surgeons approved cancer program, and Radiation Therapy Oncology Group participation. We documented availability of radiation therapy if the hospital had submitted at least 10 claims for radiation therapy.
We calculated each physician’s and each hospital’s total surgical volume by summing the number of mastectomies and BCS procedures performed by the surgeon or at the hospital, respectively, each year. Each patient was given a value for physician volume and hospital volume based on the surgical volume during the year the patient had surgery, categorized into quartiles. Determinations of volume were based on women aged ≥66; we had no data on procedures for patients in Medicare managed care or patients aged <66; however, others have demonstrated that hospital volume for fee-for-service Medicare beneficiaries correlates well with total volume [42–44]. Hospital volume and radiation availability were assessed only for hospitals located in SEER areas because we could not reliably capture care delivered outside of SEER areas (2427 (3%) of women were treated in 329 hospitals located outside of the SEER areas).
We were unable to match physician data for 6,707 (9%) patients and hospital data for 6,172 (9%) patients. We used simple imputation to randomly assign missing data points to categories with probabilities based on the observed distribution among women with non-missing data. Results were similar when we categorized missing data as a separate variable; we present results using the imputed data.
We used χ2 tests to examine bivariate associations between patient, tumor, physician, and hospital characteristics and primary treatment. We then used propensity score methods to assess the association between type of primary treatment and mortality. We used propensity score methods instead of multiple regression to control for observed differences between treatment groups because regression-based approaches have been shown to be sensitive to model misspecification, particularly when models include many potential confounders and there are large observed differences between treatment groups [45, 46]. First, we used a generalized logistic regression model to predict treatment, categorized into four groups (mastectomy with radiation, mastectomy without radiation, BCS with radiation, BCS without radiation) . Covariates in the propensity score models included all variables in Appendix Table 2 (see electronic supplementary material). We included a large number of covariates, including covariates that were likely to be collinear, to control for as many observable characteristics as possible . We then used the estimated regression coefficients and observed covariates to estimate a propensity score for each women defined as her predicted probability of receiving her observed covariate (pobs). We developed a propensity score weight for each women as a function of her observed covariates equal to the inverse of her propensity score (i.e., 1/pobs; note each women received a single weight based on her observed treatment) . Weighting by this inverse propensity score weight gives additional weight to women less likely to receive their observed treatments while also down-weighting observations with characteristics most strongly associated with the observed treatment producing balance in observed characteristics across the four treatment groups. We tested for balance in the groups after propensity weighting by fitting weighted regression models to each covariate separately, using tests that account for unequal weighting of observations. We used weighted Kaplan–Meier curves to compare overall and breast-cancer-specific adjusted mortality across the four treatment groups. We then used weighted Cox proportional hazards models to assess the association of each treatment with overall and breast-cancer-specific survival, specifically assessing (1) time to death of any cause and (2) time to death from breast cancer. To control for small residual imbalance in key covariates even after propensity score weighting, we included age, stage, comorbidity, SEER registry, year of diagnosis, receipt of chemotherapy, and treatment in a teaching hospital in the Cox regression models. Results of two-group comparisons looking at any two of the four groups were similar to the four-group analyses.
We repeated all analyses for a cohort of women who were well represented in clinical trials: women aged ≤70 years with stage I tumors and Charlson score of 0 or 1 (based on the Klabunde modification  in the year before diagnosis). We also repeated analyses among all women aged 70 and older and in elderly women of all ages with stage I disease. In each case, we refit the propensity score model using the same variables described above.
Finally, because propensity score analyses can only control for observed characteristics, we examined the robustness of estimated treatment effects to unobserved confounders [50, 51]. To do this, we assumed there exists an unobserved variable (or collection of variables), such as receipt of tamoxifen, associated with both treatment and survival. We then updated our estimates of survival differences after controlling for observed covariates and this new unobserved covariate using the following formula : RRadj = RRPropensity Score/A, where , RRPropensity Score is the relative risk of death associated with the treatment estimated using propensity score methods, Γ is the relative risk of death associated with the confounding variable of interest, and P1 and P0 are the prevalence of the confounder in the two treatment groups. We obtained estimates of Γ, P1, and P0 through analyses of supplementary data or from the literature. We considered four example unmeasured confounders: receipt of tamoxifen (which is not captured in SEER-Medicare data), performance status (which independently predicts survival following cancer therapy), smoking status (a health behavior), and college education (a better measure of socioecomonic status than our area-level education variable). An overview of randomized controlled trials found that tamoxifen led to a 25% relative reduction in the risk of death . We estimated likelihood of receiving tamoxifen by choice of primary treatment using data from the National Cancer Institute’s Patterns of Care study . Among clinical trial populations, poor performance status can increase the risk of mortality by approximately twofold . Because we did not have data on differences in performance status associated with primary treatment choice, we considered a range of effects based on observed differences in comorbitidy across the treatment groups. The impact of smoking status on survival in cancer patients has been evaluated in clinical trial populations [54, 55] and observational cohorts [56, 57]; a 20 pack-year smoking history was associated with a 25% increase risk of death compared to never smokers in stage III colon cancer patients. The impact of education on survival in cancer patients has been associated with up to a 30% decrease in mortality in clinical trial  and population-based cohorts . We considered a range of effects for the relationship between smoking status and education and treatment choice. Finally, we considered a hypothetical unobserved confounder whose effect is equal to the combination of age, stage and comorbidity (all strongly associated with treatment and survival). Specifically, we estimated the fraction of women age 66–69, diagnosed with stage I disease, with Charlson comorbidity score equal to 0 in each treatment group. We used hazard ratios (HRs) from the Cox regression model to estimate the combined effect of age, stage, and comorbidity on survival.
The mean (SD) age of the cohort was 75.9 (6.6) years. Most women were white and living in a major metropolitan area or metropolitan county. Additional characteristics are included in Table 1 and Appendix Table 2 (see electronic supplementary material). Type of primary breast cancer therapy varied by patient and tumor characteristics as well as physician and hospital characteristics (Table 1 and Appendix Table 2 (see electronic supplementary material)). For example, younger and healthier women and married women more often had BCS with radiation than mastectomy, while the oldest and sickest women more often had BCS without radiation. Women with more advanced tumors and less-favorable histologies more often had mastectomy than BCS with radiation. Patients of female physicians more often had BCS than mastectomy, as did patients at larger hospitals and hospitals with medical school affiliations, comprehensive cancer center designation, American College of Surgeon-approved cancer programs, onsite radiation, and higher breast cancer surgery volume. Differences across treatment groups were dramatically reduced after propensity score weighting (Appendix Table 3 (see electronic supplementary material)), although a few small differences remained; we account for these residual differences by including control variables in the Cox regression models as described above.
Adjusted survival was greatest among women who had BCS with radiation (median survival 10.98 years) (Fig 1a). Compared to these women, all-cause mortality was higher among women who had mastectomy without radiation (median survival 10.04 years, adjusted HR 1.19, 95% confidence interval (CI) 1.14–1.23), mastectomy with radiation (median survival 10.02 years, HR 1.20, 95% CI 1.14–1.27), and BCS without radiation (median survival 7.63 years, HR 1.81, 95% CI 1.70–1.92) (Table 2). These differences were similar when examining mortality from breast cancer (Fig 1b and Table 2).
Among women representative of those eligible for clinical trials (age ≤70 years with Charlson comorbidity score of 0 or 1 and stage 1 tumors), there were no differences in all-cause or breast-cancer-specific survival for women who underwent BCS with radiation or mastectomy with or without radiation (Fig. 2a, b). Women who had BCS without radiation had lower breast-cancer-specific and all-cause survival. For other groups (women older than age 70 and women with stage I disease), results were similar to our primary analyses (data not shown), although the slight increased adjusted HRs for breast-cancer-specific mortality among stage I women undergoing mastectomy without radiation (1.08 [0.95–1.23]) or with radiation (1.19 [0.95–1.49]) versus BCS with radiation were not statistically significant.
We performed sensitivity analyses to assess whether unmeasured confounders might explain observed survival differences (Table 3). Adjusting for differences in the receipt of tamoxifen had little impact on estimates. Adjusting for potential differences in performance status had larger effects. For example, if 15% of women who underwent mastectomy without radiation had poor performance status compared to only 5% of women receiving BCS with radiation (second row, Table 3), then adjusting for this difference would have decreased the observed HR from 1.19 to 1.08 (although this ratio would remain statistically significant). College education, smoking status, and the combined effect of age, comorbidity, and stage could each explain a small but important portion of the observed effects. However, adjusted HRs remained statistically significant after adjustment for each of these factors individually. The combined effect of all of these factors could completely explain observed survival differences between women who received BCS with radiation and those who received mastectomy with radiation and could reverse the survival advantage of BCS with radiation versus mastectomy without radiation.
We examined survival up to 12 years after primary breast cancer therapy among women in population-based registries from across the United States using rigorous statistical methods and found that women undergoing BCS with radiation had a greater likelihood of survival than women who underwent mastectomy. Among a subset of women who represent those enrolled in clinical trials, we found no differences in survival, consistent with the evidence from clinical trials.
Our findings are consistent with a recently reported study demonstrating that distant disease-free survival and overall survival for women with one to three positive lymph nodes was better for women who had segmental mastectomy with radiation than for those treated with mastectomy without radiation . Thus, radiation may have an important role in breast cancer control. The current understanding of breast cancer emphasizes both the importance of local control as well as the possibility of systemic spread . A 2005 meta-analysis by the Early Breast Cancer Trialists’ Collaborative Group examined 78 randomized clinical trials evaluating the extent of surgery and the use of radiation therapy and demonstrated benefits to radiation therapy beyond reductions in local recurrence . For every four local recurrences that were prevented at 5 years, there was one fewer breast cancer death at 15 years. These benefits were present regardless of whether the decrease in local recurrence was via more extensive surgery or radiation. Radiation also improved 15-year survival for women undergoing mastectomy, a finding also noted in observational studies of women with high-risk cancers undergoing mastectomy [62, 63].
Because results of observational analyses may be influenced by the presence of unobserved confounders, we rigorously tested the robustness of our estimates to a series of potential unobserved confounders. These analyses suggest that there would have to be very large effects of multiple unobserved confounders to explain our findings or demonstrate a benefit of mastectomy over BCS. Thus, although our effect sizes may be increased due to unobserved confounders, we can be comfortable rejecting the possibility that mastectomy is better than BCS with radiation. This finding is important because recent data suggest trends towards more aggressive surgery [20–23], and rates of mastectomy remain quite high in some areas in the United States .
The survival benefit to BCS with radiation over mastectomy that we and others  have observed has not been observed in clinical trials. One possible explanation for the different findings is that although clinical trials are the gold standard for examining treatment efficacy, only 2.5% of cancer patients enroll in clinical trials , and participants tend to be younger and healthier than other patients and are typically cared for by providers who practice at or are affiliated with cancer centers [25–28, 64]. Women in trials of primary breast cancer therapies were predominantly younger than age 70, had few comorbid illnesses, and had stage I tumors. In our analyses, we identified women who were most like those in the trials and found similar long-term survival following BCS with radiation and mastectomy, consistent with results of clinical trials. It may be that the benefits of radiation in addition to BCS are greatest in women with larger tumors or those cared for outside of major cancer centers.
A second possible explanation is that unobserved confounders contributed to our findings . The patients in our cohorts differed in many ways that could explain longer-term survival (patients undergoing BCS without radiation were older and sicker than other women, patients undergoing mastectomy with or without radiation had more advanced tumors). We used careful propensity score adjustment and tested the sensitivity of our analyses to unobserved confounders. As noted above, although unmeasured confounders could possibly explain the survival advantage we observed, it is unlikely that the ability to adjust for unobserved confounders would lead to a conclusion that mastectomy is better than BCS with radiation.
Given the importance of local tumor control, it is important to note that 10% of women over the age of 65 underwent BCS without radiation. These women tended to be older and sicker than other women, but we observed worse survival even among younger and healthier women, who may be most likely to benefit from the addition to radiation.
Several previous studies have examined breast cancer treatment effectiveness using observational data. Two studies assessed outcomes associated with mastectomy or BCS using instrumental variable analysis [66, 67]. The first, which examined women diagnosed with stage II breast cancer in Iowa during 1989–1994, suggested that, for patients whose care was influenced by distance to a radiation facility, mastectomy was associated with better survival than BCS with radiation . The second examined a national sample of Medicare beneficiaries diagnosed with stage I or II disease during 1992–1994. The instrumental variables analysis found no difference in 3-year survival by treatment type, although the estimates were large and unstable, and an ordinary least-squares approach found a survival advantage of BCS with radiation over mastectomy , similar to our findings. These inconsistent findings of these two articles may be because the first study examined only stage II patients in one state, while the second study included all patients with breast cancer undergoing mastectomy or BCS.
Our findings should be viewed in light of several limitations. First, we studied only older women with breast cancer residing in SEER areas, so the generalizability of our findings to other patients requires further study. Second, we had no information about recurrence and limited information about the breast cancer care that women received following their primary therapy. Third, although our propensity score methods allowed adjustment for observed confounders, our measures of comorbidity were based on administrative data only, and we could not account for possible unobserved confounders, such as hormonal therapy. Still, our sensitivity analyses suggest that the effects of one or more confounders would need to be very large to explain our findings. Finally, we used death certificate data to assess cause of death, and misclassification is possible .
In summary, we found that the type of primary therapy for early-stage breast cancer is related to many patient, tumor, surgeon, and hospital characteristics, and even after careful adjustment for such factors, there is no evidence to suggest benefits to mastectomy over BCS with radiation; rather, women who underwent BCS with radiation had better survival than women who underwent mastectomy. These findings are reassuring in light of recent trends towards more aggressive primary breast cancer therapy.
This study was funded by the National Cancer Institute, grant R01 CA104118 to Dr. Keating. The authors would like to thank Joshua Angrist, Ph.D., Amitabh Chandra, Ph.D., Richard Gelber, Ph.D., Douglas Staiger, Ph.D., and Alan Zaslavsky, Ph.D., for their helpful advice on analyses during an Advisory Meeting as well as Rachel Freedman, M.D., for helpful comments on an earlier draft of the manuscript and Garrett Kirk for administrative assistance. The authors also thank Linda C. Harlan, Ph.D. for sharing data from the National Cancer Institute’s Patterns of Care study that allowed us to estimate rates of tamoxifen use by treatment group.
Electronic supplementary material The online version of this article (doi:10.1007/s10549-010-0865-4) contains supplementary material, which is available to authorized users.
Nancy L. Keating, Division of General Internal Medicine, Department of Medicine, Brigham and Women’s Hospital, Boston, MA, USA; Department Health Care Policy, Harvard Medical School, 180 Longwood Avenue, Boston, MA 02115, USA.
John M. Brooks, College of Pharmacy, The University of Iowa, Iowa City, IA, USA; Department Health Care Policy, Harvard Medical School, 180 Longwood Avenue, Boston, MA 02115, USA.
Elizabeth A. Chrischilles, College of Public Health, The University of Iowa, Iowa City, IA, USA.
Eric P. Winer, The Dana-Farber Cancer Institute, Boston, MA, USA.
Kara Wright, College of Public Health, The University of Iowa, Iowa City, IA, USA.
Rita Volya, Department Health Care Policy, Harvard Medical School, 180 Longwood Avenue, Boston, MA 02115, USA.