ejc supplements 10, no. 1 (2012) 13–19

Statistical methodology for personalized medicine:New developments at EORTC Headquarters since the turn ofthe 21st Century

L. Collette*, J. Bogaerts, S. Suciu, C. Fortpied, T. Gorlia, C. Coens, M. Mauer, B. Hasan,S. Collette, M. Ouali, S. Litiere, J. Rapion, R. SylvesterEORTC Headquarter, Statistics Department, Brussels, Belgium

article info

Keywords:BiostatisticsTranslational researchBiomarkersClinical trialsProgression-free survivalOncology

abstract

Since the creation of the EORTC Headquarters in 1974, major advances have beenmade in the methodology used in the design and analysis of cancer clinical trials.However, the speed of these developments in the fields of biology, medicine, andstatistics has greatly increased since the turn of the 21st century. These changeshave all become possible because of the increased computer power now available.The landscape of therapeutic anti-cancer development changed completely with

the advent of the so-called “targeted agents” that treat the underlying molecularbasis of the disease rather than the symptoms of tumor proliferation. The newchallenges posed by the clinical development of these numerous new agents thatare expected to work in often yet-to-be-identified subgroups of patients induced anew wave of methodological developments. Some of these were readily embracedby the EORTC Headquarters statistics department, while others met with opposition.Our assessment is still in progress in many areas. Trials tend to become increasinglycomplex so that their planning relies on an increasing number of unknownparameters that need to be monitored during the trial itself, one development being“adaptive” design methodology.

Because the knowledge required to design these new and more complex clinicaltrials and associated research program spans more disciplines (biology, genomics,radiology) and involves specialized knowledge within those disciplines, continuedsuccess requires further developing our partnerships with specialized departments(imaging, bio-informatics, biology, etc.) within EORTC Headquarters and in EORTCaffiliated centers as well as collaborations with specialized statisticians fromacademia.

© 2012 European Organisation for Research and Treatment of Cancer.

1. Introduction

Since the creation of the EORTC Data Center in 1974,major advances have been made in the methodologyused in the design and analysis of cancer clinical trials.

* Corresponding author. Laurence Collette, EORTC Head-quarters, Statistics Department, Avenue E. Mounier 83/11,1200 Brussels, Belgium. Tel.: +32 2 7741669.E-mail address: laurence.collette@eortc.be (L. Collette).

However, the speed of these developments in the fieldsof biology, medicine and statistics has greatly increasedsince the turn of the 21st century. These changes haveall become possible because of the increasing computerpower that is now available to these fields.

Until the late 1990’s, anti-cancer drugs undergoingclinical development were mostly non-targeted cytotoxicagents. Their clinical development was carried out inthree major steps: phase I studies to find the maximum

1359-6349 © 2012 European Organisation for Research and Treatment of Cancer. Open access under CC BY-NC-ND license.

Open access under CC BY-NC-ND license.

14 ejc supplements 10, no. 1 (2012) 13–19

tolerated dose of a drug whose activity (and toxicity)was assumed to increase with increasing dose; thenphase II trials to identify the cancer types in which thedrug shows some degree of biological anti-cancer activity(typically assessed in terms of tumor response rate), andfor the most promising drugs, one or several phase IIIcomparative studies to assess the relative benefit of thedrug in comparison to established standard therapy interms of a clinically relevant endpoint (most often overallsurvival).At that time, the methodology for both the design and

the statistical analysis of trials was well established, andstatisticians could do their job with a limited numberof tools. The most commonly used designs for phase IIwere developed by Gehan, 1 Fleming 2 and Simon 3 in,respectively, 1961, 1982 and 1989. In 1974, George andDesu 4 published formulae for the calculation of samplesize for phase III trials that could be done with a pocketcalculator and tables were published by Freedman 5 in1982. The Kaplan–Meier estimation of survival curveshas been known since 1958, 6 and Cox 7 published hisregression model in 1982. Randomizing treatments byminimization was published in 1976, 8 but was notadopted until the advent of the personal computer. Moreadvanced developments, such as equivalence studiesand survival estimation in the presence of competingrisks, were approached respectively by Blackwelder 9 andGray, 10 both in 1982, whereas the interim monitoringplans for phase III trials using alpha-spending functionswere published in 1993. 11

Up to 1995, the statistical analysis of studies at theEORTC Data Center was done using a home-madeVAX-VMS program, “SMART”, that directly accessed thecentral EORTC database. It could take several hours forthe analysis program of a phase III trial with a fewhundred patients to run. The results were then manuallytransferred from a line-listing into a document thatwould be faxed or mailed to the study coordinators.

Shortly thereafter, microcomputers became availablefor most EORTC collaborators, and the speed ofinformation exchange increased a lot thanks to email.Computer power also increased, and this enabledstatisticians worldwide to put into practice a numberof methodological developments whose implementationhad awaited the computing power that had beenheretofore unavailable.

In parallel, biology met informatics, giving rise to bioin-formatics and the sophisticated statistical algorithmsupon which DNA sequencing, microarray expressionprofiling and genomic sequence analysis rely. 12 Thegenomic revolution started the dream of personalizedmedicine whereby genetic information would helpdoctors select the most appropriate drugs to treat agiven patient’s disease. The landscape of therapeuticanti-cancer development changed completely with theadvent of the so-called “targeted agents” that treat the

underlying molecular basis of the disease rather than thesymptoms of tumor proliferation. The new challengesposed by the clinical development of these numerousnew agents that are expected to work in often yet-to-be-identified subgroups of patients induced a new waveof methodological developments. Some of these werereadily embraced by the EORTC Headquarters StatisticsDepartment, while others met with opposition. Ourassessment is still in progress in many areas. Trials tendto become increasingly complex so that their planningrelies on an increasing number of unknown parametersthat need to be monitored during the trial itself,one development being “adaptive” design methodology.Below we will discuss some of these challenges.

2. Early development of targeted agents

2.1. Phase I

Molecularly targeted agents are not expected to have thesame toxicity profile as the classical cytotoxic agents,and their degree of activity does not necessarily increasewith increasing dose. There is now evidence that theconventional methodology of phase I oncology clinicaltrials may not be appropriate for newer agents such asmolecularly targeted agents (MTAs). 13

The EORTC has launched a new project developing“Guidelines for the definition of Dose Limiting Toxicity(DLT) and Recommended Phase II dose (RP2D) for phase Iclinical trials testing MTAs as single agents”. This projectwill be conducted with partners in both academiaand industry and plans to build a data warehouseat the EORTC Headquarters of closed phase I clinicaltrials of single-agent MTAs. Exploration of the datawarehouse will provide information supporting a revisedDLT definition and approach to the RP2D. The proposalwill be published in a peer-reviewed journal. A parallelinitiative is being conducted in the US and Canada forphase I trials of combinations of agents.

2.2. Phase II

Although many new compounds enter clinical develop-ment in oncology, up to 70% of those entering Phase II failto make it to Phase III. 14 This can be partially explainedby the fact that although traditional designs developedin the 1980s for phase II trials (such as the Flemingdesign) were efficient at identifying inactive agents, theyhave low predictive value for positive phase III results. 15

They were also generally conducted with relatively highfalse-positive error rates to keep the false-negative errorrate low, because few compounds were available fordevelopment.Today, phase II trials need to be designed to have

an improved prediction of successful phase III studies.For certain cancers in which many new agents are

ejc supplements 10, no. 1 (2012) 13–19 15

in development, it is important to keep the risk of afalse-positive finding at the end of the phase II to aminimum. Furthermore, because the newer agents areexpected to slow down or stop tumor proliferation butare not expected to induce tumor response, progression-free survival is often the endpoint of today’s phase IIstudies. Because patients may remain progressive evenin absence of treatment, because new agents are oftentested upfront in combination with other agents knownto have some activity, and because targeted agents aresometimes developed in combination with diagnostictests and studied in subgroups of patients whose tumorbears the target of interest, phase II studies at the EORTCare nowadays most often randomized and include a ref-erence control treatment group. 16 EORTC Headquartershas adopted comparative screening phase II designs, 17,18

pick-the-winner designs, 19,20 and marker-based de-signs. 21 The EORTC also regularly conducts multi-tumorbasket phase II studies through its NOCI network.

In the EORTC Statistics Department, we are constantlyreassessing the methodology that is appropriate for ourphase II trials. Thus, when the Consensus Recommen-dations from the Clinical Trial Design Task Force of theNational Cancer Institute Investigational Drug SteeringCommittee 22 were released, we were pleased to see thatthis document perfectly reflected our current updatedprocedures.

3. Progression-free survival as endpoint

As more effective salvage treatments become available,many trials in the adjuvant or even metastatic settingnowadays have progression-free survival (PFS) as primaryendpoint. However, this endpoint is much more difficultto assess than overall survival: PFS is subject tomeasurement errors and imprecision and at risk ofinterpretation bias, since the imaging diagnostics involvea certain degree of human (subjective) interpretation.Because progression is assessed at scheduled follow-up visits, a failure that is diagnosed at a given follow-up visit has in fact occurred in the time interval fromthe preceding assessment to the present visit. Thisleads to an overestimation of the time to failure. 23,24

The frequency of assessments also directly impacts onthe estimated treatment differences, 23−25 even whenthe visits are scheduled symmetrically in the tworandomized treatment groups. Caroll 24 has shown thatthe hazard ratio is increasingly biased toward the nullhypothesis of no difference as the interval betweenvisits lengthens and the frequency of visits declines.As a consequence, statistical power decreases and thenumber of events needed to maintain the specifiedstatistical power increases.

From the above remarks, it is clear that artificialtreatment effect differences may be caused by an

asymmetric schedule of visits between the two groups orby systematically prolonged delays in observation timesin one arm compared to the other. Simulation studieshave demonstrated that differences in the timing ofdisease evaluations can significantly bias PFS analysesto the point of causing an apparent improvementin outcome when none exists. 23,26 Dancey et al. 27

recommend blinding in trials that use progression-free survival as endpoint. This is to prevent thatthe knowledge of the treatment group influences theinvestigator in the assessments that involve a greaterdegree of subjectivity (for example, review of images),or in the decision to delay treatment and/or visits onthe basis of toxicity or inconvenience for the patient.Physicians or patients may be biased towards earlierclaims of progression in the control arm that is generallythought to be the less effective treatment.

In the FDA Guidance, 28 a whole section is devotedto the statistical methods of analyzing progression-free survival endpoints. The guidance is often inter-preted as recommending that patients who stop takingrandomized therapy prior to documented progressionshould be censored at the time when the treatmentis stopped. 24 However, this causes obvious problemsin the analysis since such censoring is informative byconstruction. Indeed, patients who stop treatment inabsence of progression generally do so either becauseof toxicity or because of a general deterioration in theirstatus that is likely indicative of treatment failure. Insuch circumstances, if the prevalence of censoring differsbetween arms, naive censoring could lead to extremelybiased results: taken to the extreme, a treatment thatwould be so toxic that all patients would stop treatmentdue to toxicity would have an estimated progression-freesurvival rate of 100% when using the method describedabove. Importantly, central review of progression mayinduce the same type of problem if it is not conductedin real-time. If treatment is stopped on the basis ofprogression diagnosed by the treating physician andis later not confirmed by the central review, censoringthe data at the time of change of treatment will alsoinduce bias in the analysis. The EORTC strongly opposesthis type of analysis being the primary analysis of PFSin its trials. The EORTC statistics department is alsoinvestigating methods of interval censoring for analyzingthis type of endpoint and pays special attention toselect appropriate assessment schedules when designingstudies that use PFS as endpoint.

4. Predictive markers for response to therapy −Translational research

Many cancer treatments benefit only a small proportionof the patients. For such treatments, classical phase IIItrials with broad eligibility criteria are inefficient.With an

16 ejc supplements 10, no. 1 (2012) 13–19

Threshold � high PPV needed !

Biomarker expression

Relative(survival)

benefit over

standard treatmentMarker +Marker -

unselectedgroup

Prevalence M+: �“sensitive”

Prevalence M-: 1-�“insensitive”

Standard Experimental

R

Threshold � high PPV needed !

Biomarker expression

Relative(survival)

benefit over

standard treatmentMarker +Marker -

unselectedgroup

Prevalence M+: �“sensitive”

Prevalence M-: 1-�“insensitive”

Standard Experimental

R

Threshold � high PPV needed !

Biomarker expression

Relative(survival)

benefit over

standard treatmentMarker +

Relative(survival)

benefit over

standard treatmentMarker +Marker -Marker -

unselectedgroup

unselectedgroup

Prevalence M+: �“sensitive”

Prevalence M-: 1-�“insensitive”

Standard Experimental

R

Fig. 1 – Predictive markers for benefit from a new experimental treatment. Patients are classified based on a biomarkerexpression level threshold into two groups of differing sensitivity to the experimental treatment (M, Marker expression,PPV, Positive predictive value).

overall negative result, such studies fail to demonstratethe benefit of the new treatment when the proportion ofpatients who actually benefit from the new treatment issmall, and will result in the over-treatment of patientswho do not benefit from the new treatment. Today, thereis strong biological rationale explaining why some drugswork in only a selected subpopulation. The identificationof the subpopulation of interest is challenging; it requiresthe identification of predictive classifiers to assignpatients to subgroups that benefit (or not) from theinvestigated therapy. The predictive classifier may bebased on a single marker or on a group of markers,such as a molecular signature. The classifier requires inaddition the identification of one (or several) thresholdsto classify patients into homogeneous groups that sharea similar likelihood of benefiting from the new therapy(Fig. 1).

Before development of a predictive marker for clinicaluse by its formal testing in phase III marker-baseddesigns described below can begin, a preliminaryexploration is needed to identify the putative predictive(combination of) markers and the relationship betweenmarker expression and outcome. This is done at EORTCthrough the Translational Research Program.

Such exploration can be done by a retrospective studyof the marker in available tissue of patients with a knownoutcome who have been similarly treated in formerstudies, or by the prospective planning of marker studiesalongside or as part of ongoing trials.Together, the EORTCHeadquarters Statistics and Translational Research andImaging Departments have developed and acquired thestatistical methodology needed for the planning andhigh-level analysis of such translational research studies.In collaboration with the EORTC Imaging Group, the

Statistics Department is also developing methodology forthe exploration and validation of imaging markers.

5. Marker-based phase III trials

Before launching a phase III trial to confirm thatan investigational treatment works preferentially in asubgroup of patients defined by the expression of amarker, earlier research must have developed a validatedassay to measure with sufficient accuracy the biomarkerof interest and show that it has good reproducibility(if this is not attained in the multicenter setting, acentral laboratory should be used). The threshold todefine the subgroups thought to be more or less sensitiveto the treatment of interest must have been identified.This classifier must not only show a strong associationwith the outcome of interest (large hazard ratio orodds ratio) but must also show high specificity andpositive predictive value for the outcome according tothe treatment. There must be preliminary evidence thatthe experimental treatment works best in one of themarker groups (sensitive subgroup, say Marker +), andthe prevalence of the sensitive subgroup must be known.Notably, the marker needs not to be a biomarker butmay also be a diagnostic imaging marker or an earlyassessment of sensitivity to the treatment (such as withearly PET response). The above elements, while needed,are often not all fulfilled, thereby further complicatingthe design of the Phase III.With reasonable assumptions regarding these ele-

ments, a phase III trial may be envisaged. Such a trialmay take several forms 29 (Figure 2): either the new andthe standard treatment are tested in only the sensitive

ejc supplements 10, no. 1 (2012) 13–19 17

a) Enrichment design

b) Marker-by-treatment “all comers” design

c) Randomized marker-based strategy design

Treatment allocation based on

marker

Marker-: Std Trt

Marker+: New Trt

Standard Trt

New Trt

Randomize

Assess marker

R

Assess marker

Randomize

Standard Trt

New Trt

Marker +

Marker - Randomize

Standard Trt

New Trt

Standard trt

Assess marker Randomize

New trt

Off Study

Marker +

Marker -

Fig. 2 – Marker-based phase III designs.

subgroup (enrichment design, such as EORTC 40071testing lapatinib in HER2 positive gastro-esophagealjunction cancer 30), or all patients enter the trial and arerandomized, but a testing strategy is defined upfrontto test if the relative treatment benefit is present inonly one or in both subgroups (marker-by-treatment“all comers” design, such as the EORTC “p53” trial10994 that assessed taxane-based versus non-taxanebased chemotherapy for breast cancer 31). More rarely,the patients are randomized to one of two treatmentstrategies: (1) one that is based on knowledge of themarker or (2) a strategy that does not use the marker(strategy designs). Trials may also use hybrid designswhereby only a certain marker-defined subgroup ofpatients is randomized but the other patients are keptin the study and are assigned a specific treatment,such as in the EORTC MINDACT trial 32 in which apatient’s risk of metastases is assessed by two risk scores,one using genomic information, and the other usingstandard clinical and pathological information. Patientswith discordant risk assessments (i.e., the two scoresdisagree) are randomized to treatment according toclinical or genetic risk whereas the patients for whomthe two risk scores agree receive (high risk) or do notreceive (low risk) chemotherapy.

Over the past 15 years, EORTC Headquarters hasacquired expertise in the design and conduct of suchcomplex trials. We recommend the “all-comers” markerbased design when the new marker is not yet fullyestablished or validated. Because this design inevitablyleads to conducting a number of tests for treatmenteffect in both the whole group and in subgroups,the analysis plan must be fully pre-specified usingappropriate measures to control the overall type I errorrate (either through closed-testing procedures or usingsplit-alpha procedures 33). We also strongly advocate topre-specify stopping rules for futility or even inferiorityof the experimental arm in order to stop the study inthe “insensitive subgroup” and convert the trial to anenrichment design should there be strong evidence thatthe experimental treatment does not work or harmspatients not bearing the marker of interest. Wheneverfeasible, marker testing prior to entry in the study is alsoadvocated in order to allow a prospective stratification ofthe randomization and ensure balanced treatment armswithin subgroups and facilitate design adaptations. Anenrichment design is recommended when the markeris already well established or when it is unethical torandomize the Marker-negative patients. For this design,pretesting of the marker and a rapid assay turnover

18 ejc supplements 10, no. 1 (2012) 13–19

is mandatory. This design is more efficient when theproportion of sensitive patients is relatively low, so thattesting the new drug in the subgroup thought to beinsensitive is prohibitive. Experience has shown thatdue to the dilution of the treatment effect, randomizedstrategy designs in general require huge sample sizes andare thus rarely feasible.

Because of the numerous assumptions made in theirdesign, marker-based trials require careful monitoringof: (1) the assay success rate (and homogeneity betweenlaboratories when several are used), (2) the prevalenceof the marker subgroups, (3) the values of the endpointfor the reference arm in all subgroups. All of thesefactors influence the power of the tests in the variouspopulations and the timing of the analyses. Revisionof the sample size calculations must be consideredwhen necessary. Further trial adaptation may also resultfrom planned stopping rules or from new informationemerging from other trials. The pan-European PETACC-8trial, for example, started with an “all comers” designtesting the addition of cetuximab to FOLFOX in com-pletely resected stage III colon cancer, but ended, afterseveral amendments, with an enrichment design in thesubgroup of patients presenting with wild-type k-Ras!A well-informed and experienced IDMC is essential forperforming such a trial.

6. Prognostic factors and nomograms

Using the EORTC clinical trial databases, sometimescombined with data from other research organizations,the EORTC Statistics Department has also developedprognostic models and nomograms for a range ofconditions and endpoints over the years. 34−37 Thedevelopment of a prognostic model is not as simpleas using the data at hand to run an automatic modelselection procedure to produce one model that more orless fits the data. In order to be trustworthy and useful,prediction models need to:(1) use variables that are clinically relevant and are

routinely measured in patients;(2) be robust to small variations in the development

dataset;(3) show a good discrimination (separates groups with

differing probabilities of the outcome of interest);(4) show good calibration (predictions match observed

event rates) 38;(5) the model performance must be validated on an

independent dataset and it must be shown tooutperform all existing models for the intendedclinical use;

(6) model comparisons must be performed using thesame data for all models! 39;

(7) but also, and maybe most importantly, a predictionmodel is useful only if it can be shown that the

prediction tools impacts on clinical decisions and thatthey lead to improved treatment decisions.

This new challenge is similar to the assessment ofthe clinical utility of a new marker to guide treatmentdecisions and will require decision-analytic methods toforecast the impact of the tool. 40

7. Conclusions

Cancer clinical trials are becoming increasingly complexand challenging for the statistician. Meeting thesechallenges increasingly requires EORTC clinical trialstatisticians to acquire more knowledge in other fields(biology, genomics, radiology) to be able to developnew designs and new statistical methods. Becausethe knowledge required spans more disciplines andinvolves specialized knowledge within those disciplines,continued success requires further developing ourpartnerships with specialized departments (imaging, bio-informatics, biology, etc.) within EORTC Headquartersand in EORTC affiliated centers as well as collaborationswith specialized statisticians from academia.

8. Conflict of interest statement

The authors declare no conflicts of interest.

References

1. Gehan E. The determination of the number of patientsrequired in a preliminary and a follow-up trial of a newchemotherapeutic agent. J Chron Dis 1991;13:346−53.

2. Fleming T. One sample multiple testing procedure forPhase II clinical trials. Biometrics 1982;38:143−51.

3. Simon R. Optimal two-stage designs for Phase IIclinical trials. Control Clin Trials 1989;10:1−10.

4. Georges SL, Desu MM. Planning the size and durationof a clinical trial studying the time to some criticalevent. J Chron Dis 1974;24:15−24.

5. Freedman LS. Tables of the number of patients requiredin clinical trials using the logrank test. Stat Med 1982;1:121−9.

6. Kaplan EL, Meier P. Nonparametric estimationfrom incomplete observations. J Am Stat Assoc1958;53:457−81.

7. Cox DR. Regression models and life tables. J R Stat Soc B1972;34:187–220.

8. Freedman LS, White SJ. On the use of Pocock andSimon’s method for balancing treatment numbersover prognostic factors in the controlled clinical trial.Biometrics 1976;32(3):691−4.

9. Blackwelder WC. “Proving the Null Hypothesis” inclinical trials. Control Clin Trials 1982;3:345−53.

10. Gray RJ. A class of k-sample tests for comparing thecumulative incidence of a competing risk. Ann Stat1988;16:1141−54.

11. Fleming T, DeMets D. Monitoring of clinical trials:issues and recommendations. Control Clin Trials1993;14:183−97.

ejc supplements 10, no. 1 (2012) 13–19 19

12. Simon R. Bioinformatics in cancer therapeutics—hypeor hope? Nat Clin Pract Oncol 2005;2(5):223.

13. LoRusso PM, Boerner SA, Seymour L. An overview ofthe optimal planning, design, and conduct of phase Istudies of new therapeutics. Clin Cancer Res 2010;16(6):1710−8.

14. Kola I, Landis J. Can the pharmaceutical industryreduce attrition rates? Nat Rev Drug Discov 2004;3:711−5.

15. Ratain M, Sargent DJ. Optimizing the design of phase IIoncology trials: the importance of randomization. Eur JCancer 2009;45:275−80.

16. Van Glabbeke M, Stewart W, Armand JP. Nonrandomized phase II trials of drug combinations:often meaningless, sometimes misleading. Are therealternative strategies? Eur J Cancer 2002;38:635−8.

17. Rubinstein L, Crowley J, Ivy P, LeBlanc M, Sargent D.Randomized phase II designs. Clin Cancer Res2009;15(6):1883−90.

18. Korn EL, Arbuck SG, Pluda JM, Simon R, Kaplan RS,Christian MC. Clinical trial designs for cytostaticagents: are new approaches needed? J Clin Oncol2001;19:265−72.

19. Sargent D, Goldberg R. A flexible design for multiplearmed screening trials. Stat Med 2001;20:1051−60.

20. Simon R, Wittes R, Ellenberg S. Randomized phase IIclinical trials. Cancer Treat Rep 1985;69:1375−81.

21. McShane LM, Hunsberger S, Adjei AA. Effectiveincorporation of biomarkers into phase II trials. ClinCancer Res 2009;15(6):1898–905.

22. Seymour L, Ivy SP, Sargent D, et al. The design ofphase II clinical trials testing cancer therapeutics:consensus recommendations from the clinical trialdesign task force of the National Cancer InstituteInvestigational Drug Steering Committee. Clin CancerRes 2010;16(6):1764−9.

23. Panageas KS, Ben-Porat L, Dickler MN, Chapman PB,Schrag D. When you look matters: the effect ofassessment schedule on progression-free survival.J Natl Cancer Inst 2007;99(6):428−32.

24. Carroll KJ. Analysis of progression-free survival inoncology trials: Some common statistical issues. PharmStat 2007;6:99–113.

25. Gignac GA, Morris MJ, Heller G, Schwartz LH, Scher HI.Assessing outcomes in prostate cancer clinical trials: atwenty-first century tower of Babel. Cancer 2008;113(5):966−74.

26. Freidlin B, Korn EL, Hunsberger S, Gray R, Saxman S,Zujewski JA. Proposal for the use of progression-freesurvival in unblinded randomized trials. J Clin Oncol2007;25:2122−6.

27. Dancey JE, Dodd LE, Ford R, et al. Recommendations forthe assessment of progression in randomized cancertreatment trials. Eur J Cancer 2009;45:281−9.

28. U.S. Department of Health and Human Services Foodand Drug Administration. Guidance for Industry ClinicalTrial Endpoints for the Approval of Cancer Drugs and Biolog-

ics. 2007. Available at http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/ucm071590.pdf

29. Sargent DJ, Conley BA, Allegra C, Collette L. Clinicaltrial designs for predictive marker validation in cancertreatment trials. J Clin Oncol 2005;23:2020−7.

30. Roth A, Moehler M, Mauer M, et al. Lapatinibin combination with ECF/X in EGFR1 or HER2overexpressing first-line metastatic gastric cancer (GC):a Phase II randomized placebo controlled trial (EORTC40071). J Clin Oncol 2008;28(15):32s.

31. Bonnefoi H, Piccart M, Bogaerts J, et al. TP53 status forprediction of sensitivity to taxane versus non-taxaneneoadjuvant chemotherapy in breast cancer (EORTC10994/BIG 1-00): a randomised phase 3 trial. LancetOncol 2011;12(6):527−39.

32. Bogaerts J, Cardoso F, Buyse M, et al. Gene signatureevaluation as a prognostic tool: Challenges in thedesign of the MINDACT trial. Nat Cl Pract Oncol2006;3(10):540−50.

33. Hoering A, LeBlanc M, Crowley JJ. Randomized phase IIIclinical trial designs for targeted agents. Clin Cancer Res2008;14:4358−67.

34. Gorlia T, Van den Bent MJ, Hegi ME, et al. Nomogramsfor predicting survival of patients with newlydiagnosed glioblastoma: Prognostic factor analysis ofEORTC and NCIC trial 26981–22981/CE.3. Lancet Oncol2008;9(1):29−38.

35. Collette S, Collette L, Budiharto T, et al. Predictors ofthe risk of fibrosis at 10 years after breast conservingtherapy for early breast cancer − A study based on theEORTC trial 22881–10882 “Boost versus no Boost”. Eur JCancer 2008;44(17):2587−99.

36. Sylvester RJ, van der Meijden APM, Oosterlinck W, et al.Predicting recurrence and progression in individualpatients with Stage TaT1 bladder cancer using EORTCRisk Tables: a combined analysis of 2596 patients fromseven EORTC trials. Eur Urol 2006;49:466−77.

37. Van Glabbeke M, Verweij J, Casali PG, et al. Predictingtoxicities for patients with advanced gastro-intestinalstromal tumors (GIST) treated with imatinib: a studyof the European Organisation for Research andTreatment of Cancer, the Italian Sarcoma Group andthe Australian Gastro-Intestinal Trials Group. Eur JCancer 2006;42:2277−85.

38. Steyerberg EW, Vickers AJ, Cook NR, et al. Assessingthe performance of prediction models: A frameworkfor traditional and novel measures. Epidemiology2010;21:128−38.

39. Kattan MW. Factors affecting the accuracy of predictionmodels limit the comparison of rival prediction modelswhen applied to separate data sets. Eur Urol 2011;59:566−7.

40. Vickers AJ. Prediction models: revolutionary inprinciple, but do they do more good than harm? J ClinOncol 2011;29(22):2951−8.