The influence of 21-gene recurrence score assay on chemotherapy use in a population-based sample of breast cancer patients
To quantify the influence of RS assay on changing chemotherapy plans in a general practice setting using causal inference methods.
We surveyed 3880 newly diagnosed breast cancer patients in Los Angeles and Georgia in 2013–14. We used inverse propensity weighting and multiple imputations to derive complete information for each patient about treatment status with and without testing.
A half of the 1545 women eligible for testing (ER+ or PR+, HER2−, and stage I–II) received RS. We estimate that 30% (95% confidence interval (CI) 10–49%) of patients would have changed their treatment selections after RS assay, with 10% (CI 0–20%) being encouraged to undergo chemotherapy and 20% (CI 10–30%) being discouraged from chemotherapy. The subgroups whose treatment selections would be changed the most by RS were patients with positive nodes (44%; CI 24–64%), larger tumor (43% for tumor size >2 cm; CI 23–62%), or younger age (41% for <50 years, CI 23–58%). The assay was associated with a net reduction in chemotherapy use by 10% (CI 4–16%). The reduction was much greater for women with positive nodes (31%; CI 21–41%), larger tumor (30% for tumor size >2 cm; CI 22–38%), or younger age (22% for <50 years; CI 9–35%).
RS substantially changed chemotherapy treatment selections with the largest influence among patients with less favorable pre-test prognosis. Whether this is optimal awaits the results of clinical trials addressing the utility of RS testing in selected subgroups.
KeywordsBreast cancer 21-gene recurrence score assay Chemotherapy Population studies
Breast cancer is an important paradigm for how advances in precision medicine may reduce overtreatment [1, 2]. Results from four genomic tests (ER/PR, HER2, 21-gene assay) and pathology largely determine clinician recommendations regarding adjuvant chemotherapy for most patients newly diagnosed with curable invasive breast cancer. The 21-gene recurrence score (RS) assay has rapidly diffused into clinical practice and is markedly influencing treatment decisions. Studies have reported correlates of the use of the RS assay, as well as the association of testing and test results with the use of chemotherapy [3, 4, 5, 6, 7, 8]. Although the test results clearly influence clinician recommendations, how testing itself influences the receipt of chemotherapy in different clinical subgroups in a community population has not been quantified. This question is particularly important as gene expression testing diffuses more broadly into the population [9, 10].
Selection effects confound the estimates of how testing influences receipt of chemotherapy because testing is not offered randomly. Patients are largely selected for testing by their physicians based on the pre-test likelihood of getting chemotherapy, which is influenced by clinical factors and patient preferences. Causal inference methods such as inverse propensity weighting can estimate the average causal effect of testing on chemotherapy use by attempting to create comparable groups of tested and un-tested people as in a randomized trial. However, these methods only estimate the net effect of testing on chemotherapy use, which can obscure the degree to which treatment plans are changed if, for a given population, there are changes in both directions in response to testing. It is also not possible to observe every patient’s treatment plan both with and without testing in a general practice setting in order to directly estimate the influence of testing.
We developed a method that enables more granular estimates of how the treatment plan would differ depending on whether a patient is tested or not, and for which patient subgroups RS testing most strongly influences the receipt of chemotherapy. We conducted our analysis using a large, diverse, contemporary population-based sample of patients newly diagnosed with early-stage breast cancer in 2013–14. We quantified (1) the overall change in chemotherapy use if the entire sample was tested with the RS assay and (2) the proportion of patients whose treatment plan would be changed in each direction by RS testing, both overall and in subgroups of interest.
The iCanCare study identified women with early breast cancer who were aged 20–79 years, diagnosed with ductal carcinoma in situ or invasive breast cancer, and reported to the Georgia or Los Angeles County surveillance, epidemiology, and end results (SEER) registry. Patients were sent surveys approximately 2 months after surgical treatment between July 2013 and September 2014. Patients were excluded if they had prior cancer, stage III or IV cancer, tumors >5 cm, or more than three positive lymph nodes. In Los Angeles County (only), Non-Hispanic Whites and African Americans under 50 were also excluded because of enrollment in other studies. African Americans and Latinas were oversampled in Los Angeles to ensure sufficient representation of minorities.
The registries identified eligible patients and administered the survey. Patients were sent a packet with a letter, survey materials, postage paid return envelope, and a $20 cash gift. A modified Dillman survey method was used to encourage response, and telephone interviews were used when requested (median time from diagnosis to survey completion was 6 months, SD 2.8 months) .
A total of 3880 patients were identified and sent a survey: 249 women were ineligible because they had exclusions noted above or were deceased, too ill to complete, or unable to complete a survey in Spanish or English; 2587 of eligible patients (71%) completed the survey. Genomic Health, Inc. provided a database with the RS assay results, which was linked with patients in the iCanCare study and SEER data. The analytical sample was limited to 1545 patients whose breast cancer was ER+ or PR+, HER2−, and stage I or II. The study protocol was approved by the University of Michigan, the University of Southern California, and Emory University.
The dependent variable was a binary variable that indicated the receipt of adjuvant chemotherapy via patient report. The primary independent variable was a binary variable indicating whether or not a patient received the RS assay. Covariates included patient demographics, clinical factors, and SEER site. Age at diagnosis, education, race/ethnicity, comorbidities, family income, insurance status, partner status, and employment status at diagnosis were obtained from surveys. Additionally, we asked patients how important it was for them to have as extensive treatment as possible and how important it was to have the newest and most advanced treatments. (5-point response categories from “not at all important” to “very important”). The variables were dichotomized to be 1 if patients rated “quite important” or “very important” and 0 otherwise. Cancer stage, tumor grade, tumor size, lymph node status, and progesterone receptor status were obtained from SEER clinical data.
We first used inverse propensity weighting (IPW) to estimate the overall causal effect of the RS assay on the use of chemotherapy in the sample population . The RS assay recipients were weighted by the inverse of the propensity score and the non-recipients were weighted by the inverse of one minus the propensity score. The propensity score was the probability of receiving the RS assay and calculated based on a logistic regression model. We included in the model the missing data patterns in addition to the covariates that affected either the RS assay or chemotherapy usage . The distributions of covariates between the RS assay recipients and non-recipients before and after IPW were compared using the Mantel–Haenszel test.
We then used a potential outcome multiple imputation approach (POMI) developed by us to estimate the causal effect (See online Appendix), now including details about whether or not testing would change treatment plans at an individual level and in key clinical subgroups. For this, we need to know the chemotherapy treatment status for each patient if tested with the RS assay, and if not tested. However, in a general practice setting, for each patient, the treatment status in only one of these two scenarios is available because she may not get tested; even if tested, her pre-test chemotherapy treatment plan is not usually available. Our approach uses sequential regression multiple imputation  to impute the unobserved (counterfactual) treatment status for each patient. Thus, for patients who received the RS assay, we imputed their chemotherapy treatment status without the assay, and we did the reverse for those who were not tested. The multiple imputation method accounts for uncertainty associated with the prediction. Additionally, the method simultaneously handles missing data that occur in other measures. The imputation models included the same comprehensive set of variables used in the IPW method.
We classified patients into four groups according to how the assay would have influenced their chemotherapy plan: (1) those who would never receive chemotherapy whether tested or not (never chemotherapy); (2) those who would have been treated without the test but would not if tested (chemotherapy discouraged); (3) those who would not have chemotherapy without the test but who would if tested (chemotherapy encouraged); and (4) those who would receive chemotherapy with or without the test (always chemotherapy).
POMI estimates the average causal effect of testing on chemotherapy use and the influence of testing by the direction and by the presence (vs. absence) of the causal effect of testing on treatment decisions. Both IPW and POMI methods minimize selection bias and control for observed confounders.
Patient characteristics: overall, before, and after inverse propensity weighting (IPW)
Overall (N = 1545)
No (N = 781)
Yes (N = 764)
No (N = 781)
Yes (N = 764)
Age at diagnosis (years)
Tumor size (cm)
0 main disease
1 main disease
>1 main diseases
Want newest and most advanced treatment
Want extensive treatment
High school/GED or less
Some college/technical school
College graduate or higher
Annual family income
Employment at diagnosis
Retired or not working
Los Angeles County
The percent of patients, with 95% confidence intervals, in each of the four groups according to how the RS assay influenced chemotherapy decisions: overall, by age, tumor grade, tumor size, and node status
Age at diagnosis (years)
Tumor size (cm)
The rapid adoption of RS testing into clinical practice underscores the commitment of medical oncologists to adopt the most precise evaluative testing algorithms to direct treatment decisions for patients with breast cancer. However, no study has quantified the causal effect of the dissemination of RS testing on altering chemotherapy plans and usage both in a community population as a whole and in key clinical subgroups.
We show that if the entire population represented by our diverse, contemporary sample of breast cancer patients with favorable prognosis (ER+ or PR+, HER2−, and stage I or II) was tested, RS testing would change treatment decisions in almost a third of patients. The change comprises 20% who were “chemotherapy discouraged” by RS (a change in decision against chemotherapy) and 10% “chemotherapy encouraged” (a change toward chemotherapy). Importantly, the subgroup whose decisions would be most likely to be changed by testing was patients with less favorable prognosis (Fig. 2). For example, for node-positive women, testing would change chemotherapy treatment plans of 44%, with a net reduction of 31%. For patients aged <50 years, the test changed treatment plans in 41%, with a net reduction of 22%; and these numbers were 43% and 30% in patients with tumor size >2 cm.
This finding is consistent with the observational evidence that the majority of patients tested receive low recurrence scores, even among those with less favorable pre-test prognosis . However, it provides a much better estimate less contaminated by selection effects related to who gets tested. Thus, we see patients with less favorable prognostic factors due to demography (e.g., young age) or clinical factors (e.g., larger tumor size or positive lymph nodes) are more likely to receive chemotherapy in the absence of RS, and thus are more likely to be “chemotherapy discouraged” by favorable RS results.
A number of prior studies have looked at the observed relationship of RS test results to chemotherapy use, but have reported only on the observable, RS-tested population. Since RS test is not offered randomly, such estimates are subject to selection bias and cannot be used to measure the population-wide effect of rapidly disseminating RS use both within and outside groups recommended for testing. It is also not feasible to obtain patients’ treatment plans before and after testing in a general practice setting. The two largest observational studies sampled populations from SEER-Medicare (N = 44,000) or selected cancer centers (N = 7400). Dinan et al. found no effect of testing on chemotherapy in women over 65 , and Hassett et al. found an odds ratio of 0.7 for the association of testing with chemotherapy . These large prior studies, however, are not population-based, restricted by age, location or both, and other than the SEER-Medicare study, excluded node-positive women for whom an increasing amount of testing and RS assay-based decision making is being done.
A number of studies have assessed the effect of RS testing on physician recommendations. Estimates for the proportion of cases in which testing would change the physician recommendation ranged from 33 to 43% of cases [17, 18]. These studies help clarify pieces of the decision-making process but cannot estimate the overall effect of testing on chemotherapy use across entire eligible population. Prior results are somewhat higher than the estimates we found for overall rates of change in chemotherapy decisions (30%). These prior studies provide evidence only about the stated preferences of the oncologists rather than the revealed choices, which are imperfectly represented by the chemotherapy received as the end result of the entire decision-making process in all its complexity.
Our study has several advantages: it is large, based on sampling from population-based SEER registry data across all ages, 20–79, and supplemented by comprehensive survey information about the subjects. Our analysis used causal methods to optimally balance our rich set of covariates for comparisons between tested and un-tested subjects. This is done in an effort to create estimates as close as possible to those that would be obtained from a randomized clinical trial, with virtually no selection bias affecting RS use. Our analyses also obtained more granular information about the direction and the presence of the influence of testing on chemotherapy decisions. However, there are some limitations to our study. The results of our study are limited to two large geographic area of the United States. Furthermore, some patient groups were not available for our sample frame because of commitments to other studies (Non-Hispanic Whites and African Americans under age 50 in Los Angeles County). Our study is observational and does not account for unmeasured factors. However, a particular strength of our analysis is the adjustment for a rich set of observed covariates including detailed information about socioeconomic status, patient attitudes, preferences and disease status collected through patient surveys.
Implications for clinical practice
Our findings reinforce that major advances in precision evaluate testing in curable breast cancer have markedly improved targeting of treatment to need. In our study, RS testing influenced the targeting of treatment to need in nearly one-third of patients. There is growing recognition of the need to address overtreatment in patients with favorable disease. A vital solution to this problem is to improve evaluative test algorithms in order to more accurately identify patients for whom treatment would have no benefit. Our results suggest that potential overtreatment in patients with the most favorable disease is already low and that the addition of RS testing in this clinical subgroup would yield a very small additional reduction in chemotherapy use. For example, the net effect of testing is about nil for patients over 65 with tumor size <1 cm and negative nodes. That this effect is not bigger, in part reflects the good news that relatively few of these patients would get chemotherapy, given current practice patterns, even in the absence of the more precise targeting that RS testing offers. However, clinicians may favor testing in these patients to identify the very few patients with high RS scores who would benefit by treatment or to assure patients who might be inclined toward chemotherapy that benefit would not outweigh risks.
Our results suggest that the biggest decrease in chemotherapy use as a result of RS testing would occur in patients with less favorable prognosis (e.g., those with demographic or anatomic factors suggesting worse outcome, such as young age or positive lymph nodes). There is mounting evidence that this may represent appropriate re-classification from higher to very low marginal benefit of adjuvant chemotherapy, as clinical trial-based evidence grows of the utility of genomic expression testing in clinical subgroups with less favorable pre-test prognosis [9, 10]. However, further refinement of the clinical utility of RS in these subgroups awaits the results of the RxPONDER trial , and it is even possible that some patients with node-positive disease who currently undergo testing may be ultimately undertreated as a result of testing. The advantage of the methods used in this paper is that we can start to anticipate how the spread of testing may change treatment patterns across the population, and thus can consider the appropriateness of those changes earlier in the course of dissemination of the test.
This work was supported by the National Cancer Institute (P01CA163233 to the University of Michigan). The collection of cancer incidence data used in this study was supported by the California Department of Public Health pursuant to California Health and Safety Code section 103885; Centers for Disease Control and Prevention’s (CDC) National Program of Cancer Registries, under cooperative agreement 5NU58DP003862-04/DP003862; the NCI’s Surveillance, Epidemiology and End Results Program under contract HHSN261201000140C awarded to the Cancer Prevention Institute of California, contract HHSN261201000035C awarded to the University of Southern California (USC), and contract HHSN261201000034C awarded to the Public Health Institute. The collection of cancer incidence data in Georgia was supported by contract HHSN261201300015I, Task Order HHSN26100006 from the NCI and cooperative agreement 5NU58DP003875-04-00 from the CDC. The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. The ideas and opinions expressed herein are those of the author(s) and endorsement by the State of California, Department of Public Health, the NCI, and the CDC or their Contractors and Subcontractors is not intended nor should be inferred. We thank Steve Shak MD and Genomic Health Inc. for collaboration on RS assay test linkage to the iCanCare data. We acknowledge the outstanding work of our project staff (Mackenzie Crawford, MPH and Kiyana Perrino, MPH from the Georgia Cancer Registry; Jennifer Zelaya, Pamela Lee, Maria Gaeta, Virginia Parker, BA, and Renee Bickerstaff-Magee from USC; Rebecca Morrison, MPH, Rachel Tocco, MA, Alexandra Jeanpierre, MPH, Stefanie Goodell, BS, and Rose Juhasz, PhD, from the University of Michigan). These people were compensated for their contributions to the work. We acknowledge with gratitude our survey respondents.
Compliance with ethical standards
Conflicts of interest
Allison Kurian has received research funding for work performed outside of the current study from Myriad Genetics, Invitae, Ambry Genetics, GenDx, and Genomic Health. The other authors have no conflicts of interest to disclose.
All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.
Informed consent was obtained from all individual participants included in the study. Consent was assumed when a completed survey was returned.
- 3.Partin JF, Mamounas EP (2011) Impact of the 21-gene recurrence score assay compared with standard clinicopathologic guidelines in adjuvant therapy selection for node-negative, estrogen receptor-positive breast cancer. Ann Surg Oncol 18(12):3399–3406. doi:10.1245/s10434-011-1698-z CrossRefPubMedGoogle Scholar
- 5.Dinan MA, Mi X, Reed SD, Lyman GH, Curtis LH (2015) Association between use of the 21-gene recurrence score assay and receipt of chemotherapy among medicare beneficiaries with early-stage breast cancer, 2005–2009. JAMA Oncol 1(8):1098–1109. doi:10.1001/jamaoncol.2015.2722 CrossRefPubMedGoogle Scholar
- 6.Hassett MJ, Silver SM, Hughes ME, Blayney DW, Edge SB, Herman JG, Hudis CA, Marcom PK, Pettinga JE, Share D, Theriault R, Wong YN, Vandergrift JL, Niland JC, Weeks JC (2012) Adoption of gene expression profile testing and association with use of chemotherapy among women with breast cancer. J Clin Oncol 30(18):2218–2226. doi:10.1200/JCO.2011.38.5740 CrossRefPubMedPubMedCentralGoogle Scholar
- 7.Jasem J, Amini A, Rabinovitch R, Borges VF, Elias A, Fisher CM, Kabos P (2016) 21-gene recurrence score assay as a predictor of adjuvant chemotherapy administration for early-stage breast cancer: an analysis of use, therapeutic implications, and disparity profile. J Clin Oncol. doi:10.1200/jco.2015.65.0887 PubMedGoogle Scholar
- 10.Cardoso F, van’t Veer LJ, Bogaerts J, Slaets L, Viale G, Delaloge S, Pierga J-Y, Brain E, Causeret S, DeLorenzi M, Glas AM, Golfinopoulos V, Goulioti T, Knox S, Matos E, Meulemans B, Neijenhuis PA, Nitz U, Passalacqua R, Ravdin P, Rubio IT, Saghatchian M, Smilde TJ, Sotiriou C, Stork L, Straehle C, Thomas G, Thompson AM, van der Hoeven JM, Vuylsteke P, Bernards R, Tryfonidis K, Rutgers E, Piccart M (2016) 70-gene signature as an aid to treatment decisions in early-stage breast cancer. N Engl J Med 375(8):717–729. doi:10.1056/NEJMoa1602253 CrossRefPubMedGoogle Scholar
- 11.Dillman D, Smyth J, Christian L (2009) Internet, mail, and mixed-mode surveys: the tailored design method, 3rd edn. Wiley, HobokenGoogle Scholar
- 12.Kurth T, Walker AM, Glynn RJ, Chan KA, Gaziano JM, Berger K, Robins JM (2006) Results of multivariable logistic regression, propensity matching, propensity adjustment, and propensity-based weighting under conditions of nonuniform effect. Am J Epidemiol 163(3):262–270. doi:10.1093/aje/kwj047 CrossRefPubMedGoogle Scholar
- 14.Raghunathan TE, Lepkowski JM, Van Hoewyk J, Solenberger P (2001) A multivariate technique for multiply imputing missing values using a sequence of regression models. Survey Methodol 27(1):85–96Google Scholar
- 15.National Comprehensive Cancer Network (2014) Breast cancer. https://www.nccn.org/professionals/physician_gls/f_guidelines.asp
- 16.Potosky AL, O’Neill SC, Isaacs C, Tsai HT, Chao C, Liu C, Ekezue BF, Selvam N, Kessler LG, Zhou Y, Schwartz MD (2015) Population-based study of the effect of gene expression profiling on adjuvant chemotherapy use in breast cancer patients under the age of 65 years. Cancer 121(22):4062–4070. doi:10.1002/cncr.29621 CrossRefPubMedPubMedCentralGoogle Scholar
- 19.SWOG (2011) SWOG RxPONDER trial (S1007): patient information. http://swog.org/Visitors/S1007/patients.asp