Annals of Oncology Advance Access originally published online on November 20, 2007
Annals of Oncology 2008 19(3):481-486; doi:10.1093/annonc/mdm486
breast cancer |
Correlation of changes between 2-year disease-free survival and 5-year overall survival in adjuvant breast cancer trials from 1966 to 2006
1 Department of Medical Oncology and Hematology, Princess Margaret Hospital, Faculty of Medicine, University of Toronto, Toronto
2 Biostatistics, Princess Margaret Hospital, Toronto, Canada
* Correspondence to: Dr E. Chen, Department of Medical Oncology and Hematology, Princess Margaret Hospital, Suite 5-719, 610 University Avenue, Toronto, Ontario, Canada M5G 2M9. Tel: +1 416-946-2263; Fax: +1 416-946-4467; E-mail: eric.chen{at}uhn.on.ca
 |
Abstract |
|---|
 Top Abstract introductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
Background: Although disease-free survival (DFS) is accepted as a valid end point in adjuvant breast cancer trials, improvement in 2-year DFS has never been formally established as an adequate correlate for 5-year overall survival (OS). We set out to ascertain if changes in 2-year DFS can be used to accurately predict 5-year OS changes.
Design: We conducted a systematic Medline search (1966–2006) for randomized adjuvant breast cancer trials of >100 patients per arm with 2-year DFS and 5-year OS data. A univariate regression model weighted by trial sample size was constructed to determine whether 2-year DFS differences between treatment arms within trials were predictive of 5-year OS differences.
Results: A total of 126 studies containing 149 treatment comparisons met the inclusion criteria. Difference in 2-year DFS was a significant predictor of difference in 5-year OS. For every 1% increase in 2-year DFS difference, the 5-year OS difference increased by 0.5%–0.55%. The proportion of variation explained ranged from 0.38 to 0.42, with a wide prediction interval.
Conclusion: There is a statistically significant correlation, of moderate strength, between difference in 2-year DFS between treatment comparisons and difference in 5-year OS but the correlation is not strong enough to be used as a predictor.
Key words: breast cancer, correlation, disease-free survival, overall survival
|
introduction |
|---|
|
TopAbstractintroductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
Breast cancer is the most prevalent cancer in the world and the leading cause of cancer mortality in women [1]. In the United States alone, an estimated 178 480 women will be diagnosed with breast cancer and 40 460 will die of the disease in 2007 [2]. The role of adjuvant chemotherapy was established in 1976 in improving both disease-free survival (DFS) and overall survival (OS) in women following resection of node-positive breast cancer [3]. Numerous adjuvant breast cancer trials have since continued to explore multiple treatment options in early stage breast cancer.
Improvements in OS, in particular 5-year OS, have traditionally been accepted as an efficacy end point for an intervention in adjuvant oncology trials. The definition is unambiguous and the outcome is one that most people can identify with. OS estimation in prospective trials, however, requires mature data with extended follow-up while 2- or 3-year DFS data are more readily available with shorter assessment period. Hence, an accurate predictor for OS, such as 2- or 3-year DFS, is a highly attractive alternative that allows the expeditious approval and delivery of effective interventions [4].
In 2004, Sargent et al. [5] presented an analysis of individual patient data pooled from 18 randomized adjuvant colon cancer trials in which a high correlation was demonstrated between DFS and OS, indicating that DFS after 3 years of median follow-up is an appropriate end point for adjuvant trials of fluorouracil-based regimens in colon cancer. Based on this finding and the results of the Multicenter International Study of Oxaliplatin/5-Fluorouracil/Leucovorin in the Adjuvant Treatment of Colon Cancer (MOSAIC) adjuvant colorectal cancer study which showed an improved 3-year DFS for participants on the oxaliplatin-containing arm [6], oxaliplatin-based chemotherapy was approved by the USA Food and Drug Administration (FDA) as an adjuvant treatment option in stage III colon cancer in 2004 [7].
Unlike colon cancer, DFS improvement has for some time now been regarded as an acceptable end point in adjuvant breast cancer trials despite the lack of data correlating its improvement with OS. This is illustrated by the acceptance of aromatase inhibitors as an adjuvant breast cancer treatment option based on trials with DFS benefit only [8–12]. Although overview data from the Early Breast Cancer Trialists’ Collaborative Group (EBCTCG) indicate that DFS improvements regularly predict improvements in OS [13], there has neither been a formal analysis undertaken to directly establish this correlation nor one to assess the degree of correlation. We therefore set out to ascertain whether changes in 2- or 3-year DFS can be used to accurately predict changes in 5- and 10-year OS in adjuvant breast cancer trials.
|
methods |
|---|
|
TopAbstractintroductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
We conducted a search of the Medline and PubMed databases for English language articles published in peer-reviewed journals between 1 January 1966 and 30 September 2006. Search terms were ‘breast neoplasms’, ‘adjuvant chemotherapy’, ‘antineoplastic agents, hormonal’, ‘antineoplastic agents’ and ‘hormone antagonists’. We included only randomized phase III trials with a minimum of 100 participants in each study arm, trials that compared systemic treatments (such as chemotherapy, hormonal therapy and immunotherapy) and those with 2- and 3-year DFS as well as 5- or 10-year OS data (or if they could be estimated from published Kaplan–Meier plot). We excluded studies investigating the effects of surgery, radiotherapy, and neo-adjuvant treatment and those that dealt with locally advanced or metastatic breast cancer. For studies with published updates, the latest report was used for acquiring DFS and OS statistics. Analysis was confined to the intention-to-treat population whenever possible.
multiple-arm trials
Due to the potential correlation of data in trials with multiple arms, only comparisons between the control and the other experimental arms were analyzed with no comparison made between the experimental arms. For example, for a study of A versus B versus C versus D, where A is the control arm, we only compared A versus B, A versus C and A versus D. Comparisons between B versus C, B versus D or C versus D were not carried out. Similarly, for studies with factorial design, only the factorial group comparisons were made and none were undertaken for the individual arms. We defined a control arm as the one which best reflects the standard of care for that time period. In the event in which a control arm is not clear, a consensus decision was reached after discussion among authors.
statistical methods
The method of ‘meta-analytic validation’ [14] was adopted which involves the derivation of a model from multiple trials to allow prediction of treatment effect on the true end point (i.e. difference in OS) from the treatment differences observed on the surrogate end point (i.e. differences in DFS). This method allows one to verify whether treatment effect on DFS is likely to predict effects on OS and that it does not require any of the treatment effects in the individual trials to be statistically significant. The association between difference in DFS and OS was calculated from regression calculations described previously in the literature [14, 15]. Prentice's operational criteria [16] for surrogacy were not used because they require a statistically significant treatment effect and/or individual patient data. As this is a literature-based analysis without individual patient data, our study goal was to explore whether 2- or 3-year DFS is an accurate predictor of, and not a surrogate for, the estimated 5- or 10-year OS.
Differences in 2- and 3-year DFS and the differences in 5- and 10-year OS were calculated as ‘experimental arm – control arm’. Study-specific covariates believed to potentially influence the relationship between DFS and OS differences were captured. These include menopausal status, intervention (e.g. chemotherapy or hormonal therapy), nodal status, hormonal status and the year accrual was completed. The base model is defined as the regression model including only one predictor (differences in 2-year DFS) and the full model is defined as the regression model which contained all study-specific covariates as described above. Generalized estimating equations (GEE) constructed with compound symmetry correlation structure were used to investigate which covariates were predictive of OS differences. The use of GEE was necessitated by the fact that some trials had more than one treatment comparison and there is likely some correlation between different treatment comparisons from the same trial. This correlation is accounted for by an estimated correlation matrix included in the GEE regression model which is not included in ordinary least-squares analyses. Regression statistics and 95% confidence intervals for these estimates were on the basis of the GEE analysis.
Estimates of the correlation coefficient (r), proportion of variation (R2), predicted estimates and 95% prediction intervals are ambiguous in the context of GEE models. As a result, two methods were used to obtain approximations of these statistics. First, an intraclass correlation coefficient was calculated to determine the amount of extra variability due to the added correlation from multiple treatment comparisons in the same trial. This extra variability was subsequently found to be minimal, therefore estimation using weighted linear regression methods, with weights proportional to sample size, would be suitable and allowed for direct calculation of these statistics. It is noted that the regression estimates provided by GEE analysis are the same as estimates calculated using weight linear regression, however, the amount of variability (or standard error) of the results is corrected. Second, the estimator as described in Chu et al. [17] was used, where R2 is defined as
 |
All analyses were carried out using SAS v9.0 for Windows (SAS Institute, Cary, NC) and plots were created using S-plus 2000 (Insightful Corp., Seattle, WA). All tests and confidence intervals were two-sided, and 
was set to 0.05.
primary analysis
The primary aim of the study was to assess how well one can predict 5- and 10-year OS differences between treatment arms using 2- or 3-year DFS differences. The primary outcome was defined a priori as 5-year OS difference between treatment arms, with 10-year OS difference being defined as a secondary outcome. The primary predictor of interest was defined as the 2-year DFS differences. It was noted during the analysis that there were insufficient numbers of trials which reported 10-year OS (n = 26); thus, 10-year OS prediction was not attempted. It was also found that results were similar whether 2- or 3-year DFS was used as the predictor, thus, only 2-year DFS and 5-year OS results are presented for brevity. Hazard ratios (HRs) were not considered for this analysis as majority of trials did not provide DFS HRs.
|
results |
|---|
|
TopAbstractintroductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
There are 126 studies included in this analysis. In all, 106 (84%) trials had two study arms (one comparison), 17 (13%) had three study arms (two comparisons) and three (2%) had four study arms (three comparisons). Study and treatment characteristics are summarized in Table 1. Median sample size per arm was 533 (range 216–3757) and median follow-up was 81 months (range 34–258 months). There was significant heterogeneity in the definition of DFS among trials. DFS was not defined in 36 (29%) studies. Of the 90 that did, all included distant relapses, 89 (99%) included local relapse, 38 (42%) included second malignancy, 62 (69%) included contralateral breast cancer and 63 (70%) included all deaths as part of the definition of DFS. Few studies reported breast cancer-specific mortality. Of the 149 treatment comparisons, 79 (53%) involved chemotherapy, 47 (32%) hormonal therapy, 16 (11%) both and 7(5%) others. The latter included studies investigating the effects of prednisone, immunotherapy, oophorectomy or sequencing of chemotherapy.
|
Two-year DFS was slightly better for the experimental arms compared with the control arms (85.0% versus 82.2%, median difference of 3.1%) as was 5-year OS (78.8% versus 78.0%, median difference of 2.0%, see Table 2).
|
Two-year DFS was statistically significantly predictive of 5-year OS univariately (base model) and after adjusting for all other predictors included in a multivariate model (full model, P < 0.001). The regression equations for the relationship between 2-year DFS and 5-year OS are shown in Figures 1 and 2. The only other statistically significant predictor of 5-year OS difference was nodal status. The relationship between 2-year DFS and 5-year OS within both subgroups was, however, similar (interaction term P = 0.50 for positive nodal status and P = 0.43 for negative nodal status).
|
|
The intratrial correlation coefficient was estimated to account for <1% of the total variability. This indicates that accounting for the intratrial correlation has a negligible effect on the outcome and using weighted linear regression models for estimation of r, R2, prediction estimates and prediction intervals is reasonable.
The estimated correlation coefficient (r) was 0.62 for the base model, 0.65 for the full model, 0.62 for node-positive trials, 0.63 for node-negative trials, 0.61 for hormonal therapy trials and 0.66 for trials with chemotherapy. The proportion of variation explained (R2) was 0.38 for the base model, 0.42 for the full model, 0.39 for node-positive trials, 0.39 for node-negative trials, 0.37 for hormonal trials and 0.43 for chemotherapy trials. As a result, regardless of the model used, less than half the variation of the differences in 5-year OS was accounted for, even in the full model. Similar results were observed when using the approximation of Chu et al. [17], with r (R2) estimates of 0.61 (0.37) for the base model, 0.63 (0.39) for the full model, 0.61 (0.37) and 0.58 (0.33) for node-positive and -negative trials and 0.62 (0.39) and 0.65 (0.42) for hormonal and chemotherapy trials. These results are shown in Figure 2.
Table 3 shows the predicted differences in 5-year OS and 95% prediction intervals for a range of differences in 2-year DFS and a range of potential trial sample sizes. The 95% prediction intervals remain relatively large, especially for trials with smaller sample sizes. In fact, if a future trial accrued a large sample size of 1000 patients and the 2-year DFS for the experimental arm was better than the control arm by 10%, the 95% prediction interval for 5-year OS difference would still range from –0.2% to 11%. Thus, one could not conclude that a statistically significant difference existed and one could argue that the experimental arm could still be inferior to the control arm.
|
|
discussion |
|---|
|
TopAbstractintroductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
The aim of this analysis was to ascertain whether differences in 2- or 3-year DFS could be used to accurately predict 5- or 10-year OS differences in adjuvant breast cancer trials. We have shown that although a statistically significant correlation of moderate strength exists between the two, the wide prediction intervals means that the correlation is not sufficiently strong for DFS to be used as a predictor for OS. This is perhaps not surprising when one considers that, with the exception of adjuvant trastuzumab (which did not fulfill the criteria of maturity for this analysis), the majority of adjuvant breast cancer trials published in the last several decades have only generated small, if any, incremental improvements in DFS and OS when compared with their respective control arms. Such improvements when present were generally <5%. The resultant limited variability in the predictor variable (i.e. 2-year DFS differences) ultimately diminishes the power to detect the correlation between changes in DFS and OS. Breast cancer is also a highly heterogeneous disease and there is a large number of factors, many of them unknown or not captured (e.g. HER-2 status), which may affect DFS and/or OS. Although we have shown nodal status to be a significant predictor of 5-year OS differences, the relationship between 2-year DFS differences and 5-year OS differences between node-positive and node-negative trials was similar in the analysis.
Data from the EBCTCG meta-analysis seem to indicate that the survival curves between control and tamoxifen or polychemotherapy-treated patients continue to diverge after 5 years and in fact, continue to do so beyond 10 years [13]. Thus, it is possible that the correlation between 2-year DFS difference may be stronger with difference in 10-year OS than with 5-year OS. Unfortunately, the paucity of published 10-year OS data means that this correlation could not be explored adequately. Similarly, breast cancer-specific mortality was not commonly reported, and thus, it could not be explored in detail.
Finally, one must consider the likelihood that DFS improvement may in fact have little or no impact on OS at all. It is conceivable that the same intervention employed at the point of disease relapse may very well delay the time to death to the same extent that it would have done if it had been administered in the adjuvant setting. Indeed, myriad subsequent-line therapies available in metastatic breast cancer may have influenced OS outcome while DFS remains unaffected, essentially disconnecting early effects of tumor control from OS.
This study was designed to ascertain whether one could accurately predict changes in OS from changes in DFS. It was not intended to explore the issue of whether DFS improvement is sufficient for the acknowledgment of efficacy for a particular intervention, which is a highly complex issue mired with budgetary, ethical and sociopolitical constraints. Indeed, intense debate arose recently as a result of the early termination of the MA-17 trial (which assessed the role of extended adjuvant hormonal therapy) on the basis of DFS, rather than OS benefit after a median follow-up period of only 2.4 years [8, 18–20]. One may argue that the use of DFS as a primary end point is not necessarily intended to infer a survival advantage and that a disease-free state alone may be meaningful to patients, for it is quality time spent without the symptoms of the relapsed disease and the accompanying treatment-related toxicity. In fact, DFS improvement alone has been sufficient for FDA approval of several drug-marketing applications in the past [12]. Conversely, this improvement requires intervention in the adjuvant setting which may potentially be more toxic and lead to deterioration in quality of life in otherwise healthy individuals, some of whom may never develop a recurrence.
Lastly, recent new oncology therapies have seen cancer treatment costs spiral beyond the means of many countries. For many public health care policy makers working in this current climate of economic constraints, the demonstration of DFS improvement alone may be insufficient for the approval of a new costly intervention when faced with other multiple competing health care needs.
There are several limitations to our study. First, the assessment was undertaken without individual patient data. It would have required access to individual records but is feasible nonetheless, as demonstrated by Sargent et al. [5] in the analysis of the adjuvant colorectal cancer studies and therefore exists as a potential for possible future exploration. Secondly, this analysis incorporated a very heterogeneous collection of studies, ranging from different types of chemotherapy, to hormonal treatment as well as to other interventions. The inclusion of such a large pool of varying interventions may have mitigated its external validity. We, however, do not believe this may have necessarily impacted on our study objective, i.e. to establish the link between changes in DFS and OS. Although changes in DFS and/or OS may be dependent on a particular intervention, there is no compelling reason to believe that the correlation between differences in DFS and OS could be influenced by the choice of the intervention alone. Certainly, the subgroup analyses indicated that the type of interventions studied (i.e. chemotherapy or hormonal therapy) or the nodal status had no effect on the ability to predict differences in 5-year OS from differences in 2-year DFS.
Finally, it is noteworthy that of the 126 studies included in this analysis, 21% did not state the definition of DFS while significant heterogeneity existed among the remaining 90 studies that did. It is evident that a standardized definition for breast cancer trial end points is required. Hopefully, the recent proposal for standardized definitions for efficacy end points in breast cancer trials [21] will be adopted to ensure consistent measurement of equivalent end points between all future studies.
In conclusion, there is a statistically significant correlation, of moderate strength, between differences in 2-year DFS between treatment arms and differences in 5-year OS, but due to the wide prediction intervals, the correlation is not sufficiently strong enough to be used as a predictor.
|
Acknowledgements |
|---|
|
TopAbstractintroductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
This study was presented in part at the 2007 American Society of Clinical Oncology Meeting. Potential conflicts of interests: Nil.
Received for publication July 18, 2007. Revision received September 13, 2007. Accepted for publication September 17, 2007.
|
References |
|---|
|
TopAbstractintroductionmethodsresultsdiscussionAcknowledgementsReferences |
|---|
1. Parkin DM. Global cancer statistics in the year 2000. Lancet Oncol (2000) 2:533–543.[CrossRef][Web of Science]
2. American Cancer Society. Cancer Facts and Figures. 2007 (2007) Atlanta, GA: American Cancer Society.
3. Bonadonna G, Moliterni A, Zambetti M, et al. 30 years’ follow up of randomised studies of adjuvant CMF in operable breast cancer: cohort study. BMJ (2005) 330:217–222.
4. Gill S, Sargent D. End points for adjuvant therapy trials: has the time come to accept disease-free survival as a surrogate end point for overall survival? Oncologist (2006) 11:624–629.
5. Sargent DJ, Wieand HS, Haller DG, et al. Disease-free survival versus overall survival as a primary end point for adjuvant colon cancer studies: individual patient data from 20,898 patients on 18 randomized trials. J Clin Oncol (2005) 23:8664–8870.
6. Andre T, Boni C, Mounedji-Boudiaf L, et al. Oxaliplatin, fluorouracil, and leucovorin as adjuvant treatment for colon cancer. N Engl J Med (2004) 350:2343–2351.
7. Food and Drug Administration. Eloxatin: New or Modified Indication (2004) Washington, DC: US Food and Drug Administration.
8. Goss PE, Ingle JN, Martino S, et al. A randomized trial of letrozole in postmenopausal women after five years of tamoxifen therapy for early-stage breast cancer. N Engl J Med (2003) 349:1793–1802.
9. Coombes RC, Kilburn LS, Snowdon CF, et al. Survival and safety of exemestane versus tamoxifen after 2–3 years’ tamoxifen treatment (Intergroup Exemestane Study): a randomised controlled trial. Lancet (2007) 369:559–570.[CrossRef][Web of Science][Medline]
10. Thurlimann B, Keshaviah A, Coates AS, et al. A comparison of letrozole and tamoxifen in postmenopausal women with early breast cancer. N Engl J Med (2005) 353:2747–2757.
11. Howell A, Cuzick J, Baum M, et al. Results of the ATAC (Arimidex, Tamoxifen, Alone or in Combination) trial after completion of 5 years’ adjuvant treatment for breast cancer. Lancet (2005) 365:60–62.[CrossRef][Web of Science][Medline]
12. Johnson JR, Williams G, Pazdur R. End points and United States Food and Drug Administration approval of oncology drugs. J Clin Oncol (2003) 21:1404–1411.
13. Early Breast Cancer Trialists’ Collaborative Group. Effects of chemotherapy and hormonal therapy for early breast cancer on recurrence and 15-year survival: an overview of the randomised trials. Lancet (2005) 365:1687–1717.[CrossRef][Web of Science][Medline]
14. Buyse M, Molenberghs G, Burzykowski T, et al. The validation of surrogate endpoints in meta-analyses of randomized experiments. Biostatistics (2000) 1:49–67.[Abstract]
15. Petrylak DP, Ankerst DP, Jiang CS, et al. Evaluation of prostate-specific antigen declines for surrogacy in patients treated on SWOG 99-16. J Natl Cancer Inst (2006) 98:516–521.
16. Prentice RL. Surrogate endpoints in clinical trials: definition and operational criteria. Stat Med (1989) 8:431–440.[Web of Science][Medline]
17. Chu T-M, Weir B, Wolfinger R. A systematic statistical linear modeling approach to oligonucleotide array experiments. Math Biosci (2002) 176:35–51.[CrossRef][Web of Science][Medline]
18. Burstein HJ. Beyond tamoxifen—extending endocrine treatment for early-stage breast cancer. N Engl J Med (2003) 349:1857–1859.
19. Bryant J, Wolmark N. Letrozole after tamoxifen for breast cancer—what is the price of success? N Engl J Med (2003) 349:1855–1857.
20. Buzdar AU, Hietanen P, Makela M, et al. Correspondence: letrozole in breast cancer. N Engl J Med (2004) 350:727–730.
21. Hudis CA, Barlow WE, Costantino JP, et al. Proposal for standardized definitions for efficacy end points in adjuvant breast cancer trials: the STEEP system. J Clin Oncol (2007) 25:2127–2132.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  B. Seruga and I. F. Tannock Up-Front Use of Aromatase Inhibitors As Adjuvant Therapy for Breast Cancer: The Emperor Has No Clothes J. Clin. Oncol., February 20, 2009; 27(6): 840 - 842. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||