Duke-IndustryStatisticsSymposium (DISS2016 ......DISS2016,Durham,NC SHORTCOURSES–SEPTEMBER14,2016...
Transcript of Duke-IndustryStatisticsSymposium (DISS2016 ......DISS2016,Durham,NC SHORTCOURSES–SEPTEMBER14,2016...
..
Duke-Industry Statistics Symposium
(DISS 2016)
PrecisionMedicine in Cancer Research
https://sites.duke.edu/diss2016
DukeUniversity Department of Biostatistics and Bioinformatics
September 14-16, 2016
MillenniumHotel, 2800 CampusWalk Ave, Durham, NC
DISS2016, Durham, NC
KEY EVENTSAND SPONSORS
DukeUniversity Department of Biostatistics and Bioinformatics 2
DISS2016, Durham, NC
FOREWORD
Dear Colleagues,
The annual symposium is organized by the Department of Biostatistics and Bioinformatics,
Duke University School ofMedicine, and co-sponsored by Amgen, Boehringer-Ingelheim,
ICSA, NC-ASAChapter, PAREXEL, Quintiles and SAS. It was established 4 years ago to
discuss challenging issues and recent advances related to the clinical development of drugs
and devices and to promote research and collaboration among statisticians from industry,
academia, and regulatory agencies.
The theme of the DISS 2016 is “PrecisionMedicine in Cancer Research.” The first day is
devoted to six short courses. The 2nd daywill start with opening remarks and keynote
speech. Dr Eric Peterson, the Director of Duke Clinical Research Institute, and the
representatives from the co-sponsors will give the opening remarks. Dr Lisa LaVange, the
Director of theOffice of Biostatistics from FDA/CDER, will present the keynote speech on
“PrecisionMedicine Initiatives at FDA.” The rest of the 2nd day and the 3rd daymorning
consist of 17 parallel scientific sessions. A poster session will also be held.
We hope you find the symposium informative and useful.
Sincerely yours,
TheOrganizing Committee of DISS2016
Duke University Department of Biostatistics and Bioinformatics 3
DISS2016, Durham, NC
EXECUTIVE COMMITTEEANDORGANIZERS
EXECUTIVE COMMITTEE
Elizabeth DeLong, Chair (Duke) Stephen George (Duke)
René Kubiak (Boehringer-Ingelheim) Kerry Lee (Duke)
Steve Snapinn (Amgen) Fred Snikeris (PAREXEL)
Terry Sosa (Quintiles) Maura Stokes (SAS)
Yi Tsong (FDA)
ORGANIZINGCOMMITTEE
Cliburn Chan (Duke) Shein-Chung Chow, Co-Chair (Duke)
Terry Hyslop (Duke) Qi Jiang (Amgen)
René Kubiak (Boehringer-Ingelheim) DebbieMedlin (Duke)
Marlina Nasution (PAREXEL) Kouros Owzar (Duke)
Frank Rockhold (Duke) Terry Sosa (Quintiles)
Yi Tsong (FDA) Sharon Updike (Duke)
XiaofeiWang, Co-Chair (Duke) YuanWu (Duke)
...
Special thanks to John Bauman fromQuintiles and Craig Ou, Fan Li, Jingyi Lin, Ke Song,
NancyQi, YanlinMa and Yiling Liu fromDuke B&B.
Duke University Department of Biostatistics and Bioinformatics 4
DISS2016, Durham, NC
LETTER FROMDEANANDREWS
DukeUniversity Department of Biostatistics and Bioinformatics 5
DISS2016, Durham, NC
KEYNOTE SPEECH
...
Title: PrecisionMedicine Initiatives at FDA
Time: September 15 9:30-10:30
Speaker: Lisa LaVange, PhD
Director Office of Biostatistics
Office of Translational Sciences
Center for Drug Evaluation and Research (CDER)
US FDA
Lisa LaVange, PhD, is Director of theOffice of Biostatistics in the Center for Drug Evaluation
and Research (CDER), US Food andDrug Administration (FDA). As Director, she oversees
approximately 195 statistical reviewers and staff members involved in the development
and application of statistical methodology for drug regulation. Prior to joining the FDA, Dr
LaVangewas Professor andDirector of the Collaborative Studies Coordinating Center (CSCC)
in the Department of Biostatistics, Gillings School of Global Public Health at the University
of North Carolina at Chapel Hill (UNC), where she served as Principal Investigator (PI) of
the coordinating centers for several large-scale multi-center clinical trials, epidemiology
studies, and patient registries. Before joining academia, Dr LaVange spent 10 years in the
pharmaceutical industry and 16 years in non-profit research. She is a Fellow of the American
Statistical Association, served as President of the Eastern North American Region of the
International Biometric Society (IBS; 2007), and served on the IBS Executive Board (2013-2014).
She was formerly co-editor of the Journal of Pharmaceutical Statistics and editor-in-chief of
the ASA-SIAMbook series
Duke University Department of Biostatistics and Bioinformatics 6
DISS2016, Durham, NC
SHORTCOURSES – SEPTEMBER 14, 2016
Short Courses Time Room
Registration & Light Breakfast 8:00-9:00 Foyer (3rd)
C1: Adaptive Clinical Trial Design - Case Studies 9:00-12:00 Brightleaf E (3rd)
Shiowjen Lee (FDA), Annie Lin (FDA)
C2: Statistical Procedures for Interim 9:00-12:00 Brightleaf F (3rd)
Analysis in Clinical Trials
Gordon Lan (JnJ)
C3: Biomarker Utilities in Adaptive Trials 9:00-12:00 Brightleaf G (3rd)
Robin Bliss (Veristat), JingWang (Gilead)
Boxed Lunches 12:00-1:00 Foyer (3rd)
C4: Adaptive Designs for Dose-Finding Studies 1:00-4:00 Brightleaf E (3rd)
SandeepMenon (Pfizer), Inna Perevozskaya (Pfizer)
C5: Patient-ReportedOutcomes: 1:00-4:00 Brightleaf F (3rd)
Measurement, Implementation and Interpretation
Joseph Cappelleri (Pfizer)
C6: Analytical Similarity Assessment 1:00-4:00 Brightleaf G (3rd)
Shein-Chung Chow (Duke), Yi Tsong (FDA)
DukeUniversity Department of Biostatistics and Bioinformatics 7
DISS2016, Durham, NC
C1: ADAPTIVE CLINICAL TRIALDESIGN – CASE STUDIES
...
Title: Adaptive Clinical Trial Design – Case Studies
Time: September 14 9:00-12:00
Instructors: Shiowjen Lee (FDA) and Annie Lin (FDA)
There has been considerable interest among pharmaceutical and other medical product developers
in adaptive clinical trials, in which knowledge learned during the course of a trial affects ongoing
conduct or analysis of the trial. Following the release of the FDA draft Guidance document on
adaptive design clinical trials in early 2010, expectations of an increase in regulatory submissions
involving adaptive design features, particularly for confirmatory trials, were high. There are
indeed concerns regarding the statistical issues and operational challenges in conducting
adaptive design clinical trials. Wewill share our experiences in the reviews of adaptive design
proposals, including surveys performed regarding regulatory submissions of adaptive design
proposals as well as case studies which have been reviewed. Wewill also provide general
recommendations for developing proposals for such trials. Ourmotivation in instructing
this short course is to encourage the best study design proposals to be submitted to FDA.
Sometimes these can be adaptive and sometimes a simpler design is most efficient.
C2: STATISTICAL PROCEDURES FOR INTERIMANALYSIS IN CLINICAL TRIALS
...
Title: Statistical Procedures for Interim Analysis in Clinical Trials
Time: September 14 9:00-12:00
Instructors: Gordon Lan (Johnson & Johnson)
This short course provides an introduction to the design and interim analyses of clinical trials
with the following topics: Sample size and information, Some fundamental statistical tools for
interim data analyses (The trend of the data, Conditional power, Group sequential methods),
Use of theWisconsin software to design sequential trials, Adaptive designs,Survival data
analysis, and Design of multiregional clinical trials (Time permitting).
Duke University Department of Biostatistics and Bioinformatics 8
DISS2016, Durham, NC
C3: BIOMARKERUTILITIES INADAPTIVE TRIALS
...
Title: Biomarker Utilities in Adaptive Trials
Time: September 14 9:00-12:00
Instructors: Robin Bliss (Veristat) and JingWang (Gilead)
In this short course, wewill discuss the opportunities and challenges in biomarker utilization
and personalizedmedicine, covering both classical and adaptive designs with biomarkers.
Wewill discuss the design options for biomarkers with very strong credentials, strong credentials
andweak credentials. Related statistical theories and analysis strategies will be coveredwith
case studies. By attending this session, participants will learn some recent developments
in biomarker study from statistical perspective and share their experiences and practical
problems concerning the biomarker utility in drug development.
C4: ADAPTIVEDESIGNS FORDOSE-FINDING STUDIES
...
Title: Adaptive Designs for Dose-Finding Studies
Time: September 14 1:00-4:00
Instructors: SandeepMenon (Pfizer) and Inna Perevozskaya (Pfizer)
Adaptive designs have been increasing in popularity over the past decade. FDA has released
its draft guidance on adaptive designs in 2010, in which it particularly encouraged the use of
“well understood” designs in exploratory space (i.e. Phase I and Phase II studies). Such studies
are often referred to as adaptive dose-escalation and adaptive dose-response designs, respectively.
When carefully planned and used appropriately, theymay bemore efficient than traditional
designs in determining the target dose given limited budget.
C5: PATIENT-REPORTEDOUTCOMES
...
Title: Patient-ReportedOutcomes: Measurement, Implementation and
Interpretation
Time: September 14 1:00-4:00
Instructors: Joseph C. Cappelleri (Pfizer)
Duke University Department of Biostatistics and Bioinformatics 9
DISS2016, Durham, NC
This short course will provide an exposition on healthmeasurement scales – specifically, on
patient-reported outcomes. Some key elements in the development of a patient-reported
outcome (PRO)measure will be noted. The core topics of validity and reliability of a PRO
measure will be discussed. Exploratory factor analysis and confirmatory factor analysis, mediation
modeling, item response theory, longitudinal analysis, andmissing data will among the topics
considered. Approaches to interpret PRO results will be elucidated in order tomake results
useful andmeaningful. Illustrations will be provided through real-life examples and also through
simulated examples using SAS.
C6: ANALYTICAL SIMILARITYASSESSMENT
...
Title: Analytical Similarity Assessment
Time: September 14 1:00-4:00
Instructors: Shein-Chung Chow (Duke) and Yi Tsong (FDA)
For assessment of biosimilarity of biosimilar products, the United States (US) Food andDrug
Administration (FDA) proposed a stepwise approach for providing totality-of-the-evidence
of similarity between a proposed biosimilar product and a US-licensed (reference) product.
The stepwise approach starts with assessment of critical quality attributes that are relevant
to clinical outcomes in structural and functional characterization inmanufacturing process of
the proposed biosimilar product. FDA suggests that these critical quality relevant attributes
be identified and classify into three tiers depending their criticality or risking ranking. To
assist the sponsors, FDA also suggests some statistical approaches for assessment of analytical
similarity for critical quality attributes (CQAs) from different tiers, namely equivalence test
for Tier 1, quality range approach for Tier 2, and descriptive raw data and graphical comparison
for Tier 3. In this short course, wewill give an overview of the equivalence tests in terms of
bioequivalence and biosimilarity and therapeutically equivalence and focus on analytical
similarity assessment for identified CQAs at various stages of manufacturing process of the
proposed biosimilar product. In addition, challenging issues such as (i) sample size determination
for reference and test lots required for a valid and reliable equivalence test, (ii) fixed criterion
versus random criterion approach, (iii) alternativemethods to the FDA’s recommended tiered
approaches (mainly for CQAs from Tier 1 and Tier 2) are discussed.
Duke University Department of Biostatistics and Bioinformatics 10
DISS2016, Durham, NC
SYMPOSIUM - SEPTEMBER 15, 2016
Time Room Events
8:00-9:00 Foyer (3rd) Registration & Light Breakfast
9:00-9:30 Greenbriar Ballroom (2nd) Welcome
Elizabeth DeLong (Duke B&B)
Opening Remarks
Eric Peterson (DCRI)
Terry Sosa (Quintiles)
James Love (Boehringer-Ingelheim)
9:30-10:30 Greenbriar Ballroom (2nd) Keynote Address:
PrecisionMedicine Initiatives at FDA
Lisa LaVange (FDA)
10:30-10:45 Foyer (3rd) Coffee Break
10:45-12:00 Brightleaf E (3rd) Parallel Session S1A
Basket Designs in Oncology Trials
Brightleaf F (3rd) Parallel Session S1B
Multi-Regional Clinical Trials (MRCTs)
Brightleaf G (3rd) Parallel Session S1C
Methods in Cancer Pharmacogenomics
12:00-1:30 Foyer (3rd) Lunch and Poster Session
1:30-2:45 Brightleaf E (3rd) Parallel Session S2A
Adaptive Designs for Clinical Trials
Brightleaf F (3rd) Parallel Session S2B
Collaboration to Accelerate Drug Development
Brightleaf G (3rd) Parallel Session S2C
Methods for Electronic Health Records
2:45-3:00 Foyer (3rd) Coffee Break
3:00-5:00 Brightleaf E (3rd) Parallel Session S3A
Application of Genetic Information in Trials
Brightleaf F (3rd) Parallel Session S3B
Current Issues in Cancer Phase II Trials
Brightleaf G (3rd) Parallel Session S3C
Discovery Science for Immunotherapy Trials
5:00-7:00 Greenbriar (2nd) Social Mixer and Poster Session
...
Poster Session will be on Sep. 15, 12:00-1:30 and 5:00-7:00 in Greenbriar (2nd) Foyer.
Social Mixer will be on Sep. 15, 5:00-7:00 in Greenbriar Ballroom (2nd).
Duke University Department of Biostatistics and Bioinformatics 11
DISS2016, Durham, NC
PARALLEL SESSIONS - SEPTEMBER 15, 10:45-12:00
...
S1A: Opportunities and Challenges with Basket Designs in Oncology Trials
• Organizers: Marlina Nasution (PAREXEL), XiaofeiWang (Duke)
• Speakers: Richard Simon (NCI), Amanda Redig (DFCI)
• Discussant: Daniel Sargent (Mayo)
S1B: Key Consideration onMulti-Regional Clinical Trials (MRCTs)
• Organizers: Qi Jiang (Amgen)
• Speakers: Steven Snapinn (Amgen), Gordon Lan (JnJ), Bruce Binkowitz (Merck)
S1C: Principles andMethods in Cancer Pharmacogenomics
• Organizers: Kouros Owzar (Duke)
• Speakers: Jichun Xie (Duke), Raluca Gordan (Duke), Federico Innocenti (UNC)
Duke University Department of Biostatistics and Bioinformatics 12
DISS2016, Durham, NC
PARALLEL SESSIONS - SEPTEMBER 15, 1:30-2:45
...
S2A: Recent Advances in Adaptive Designs for Clinical Trials
• Organizers: Qi Jiang (Amgen)
• Speakers: JingWang (Gilead), Qi Jiang (Amgen), Yeh-Fong Chen (FDA)
S2B: Collaboration to Accelerate Development of EffectiveOncologyMedications
• Organizers: Marlina Nasution (PAREXEL), XiaofeiWang (Duke)
• Speakers: Rajeshwari Sridhara (FDA), SharonMurray (PAREXEL), Daniel Sargent
(Mayo)
S2C:Methods and Applications for Electronic Health Records
• Organizers: Terry Sosa (Quintiles), John Bauman (Quintiles)
• Speakers: Walter Boyle (Sutherland Healthcare), Joseph Lucas (Duke), Ben
Goldstein (Duke)
Duke University Department of Biostatistics and Bioinformatics 13
DISS2016, Durham, NC
PARALLEL SESSIONS - SEPTEMBER 15, 3:00-5:00
...
S3A: Application of Genetic Information in Oncology Clinical Trial Design
• Organizers: Renè Kubiak (Boehringer-Ingelheim), Marlina Nasution (PAREXEL)
• Speakers: Suzanne Dahlberg (DFCI), SumithraMandrekar (Mayo), Yijing Shen
(Genentech)
S3B: Current Issues in Cancer Phase II Trials
• Organizers: Terry Sosa (Quintiles), XiaofeiWang (Duke)
• Speakers: Sin-Ho Jung (Duke), Ilya Lipkovich (Quintiles), Kevin Liu (JnJ)
• Discussant: Gary Koch (UNC)
S3C: Discovery Science for Immunotherapy Trials
• Organizers: Terry Sosa (Quintiles), Cliburn Chan (Duke)
• Speakers: KentWeinhold (Duke), Eric Groves (Quintiles), Lynn Lin (Penn State),
Radleigh Santos (TPIMS)
Duke University Department of Biostatistics and Bioinformatics 14
DISS2016, Durham, NC
SYMPOSIUM - SEPTEMBER 16, 2016
Time Room Events
8:00-9:00 Foyer (3rd) Registration & Light Breakfast
9:00-10:15 Brightleaf E (3rd) Parallel Session S4A
Addendum to Statistical Principles for Clinical Trials (ICH E9)
Brightleaf F (3rd) Parallel Session S4B
Causal Inference in Cancer Clinical Research
Brightleaf G (3rd) Parallel Session S4C
NewDevelopments in Survival Analysis for Cancer Research
Greenbriar A (2nd) Parallel Session S4D
Opportunities and Challenges in Design and Analysis
of Immunotherapies Trials
10:15-10:30 Coffee Break
10:30-12:30 Brightleaf E (3rd) Parallel Session S5A
Safety and Benefit Risk Analysis in Drug Development
Brightleaf F (3rd) Parallel Session S5B
Statistical Issues Related to Progression-free Survival
andOverall Survival
Brightleaf G (3rd) Parallel Session S5C
Current Issues in Biosimilar Studies
Greenbriar A (2nd) Parallel Session S5D
Exposure ResponseModeling in the Pharmaceutical Industry
DukeUniversity Department of Biostatistics and Bioinformatics 15
DISS2016, Durham, NC
PARALLEL SESSIONS - SEPTEMBER 16, 9:00-10:15
...
S4A: Addendum to Statistical Principles for Clinical Trials (ICH E9): Choosing
Appropriate Estimands andDefining Sensitivity Analyses in Clinical Trials
• Organizers: Terry Sosa (Quintiles), John Bauman (Quintiles)
• Speakers: CraigMallinckrodt (Eli Lilly), Bohdana Ratitch (Quintiles), Devan
Mehrotra (Merck)
S4B: Causal Inference in Cancer Clinical Research
• Organizers: XiaofeiWang (Duke), Terry Hyslop (Duke)
• Speakers: Donglin Zeng UNC), XiaofeiWang (Duke), Jeremy Taylor (UMICH)
S4C: NewDevelopments in Survival Analysis for Cancer Research
• Organizers: YuanWu (Duke)
• Speakers: Danyu Lin (UNC), Butch Tsiatis (NCSU), Jason Fine (UNC)
S4D: Opportunities and Challenges in the Design and Analysis of Immunotherapies
Trials
• Organizers: Susan Halabi (Duke)
• Speakers: Kay Tatsuoka (BMS), Susan Halabi (Duke), PralayMukhopadhyay
(AstraZeneca)
Duke University Department of Biostatistics and Bioinformatics 16
DISS2016, Durham, NC
PARALLEL SESSIONS - SEPTEMBER 16, 10:30-12:30
...
S5A: Safety and Benefit Risk Analysis in Drug Development
• Organizers: Qi Jiang (Amgen)
• Speakers: Frank Rockhold (Duke), Chunlei Ke (Amgen), OlgaMarchenko
(Quintiles)
S5B: Statistical Issues Related to Progression-free Survival andOverall Survival
• Organizers: Renè Kubiak (Boehringer-Ingelheim), XiaofeiWang (Duke)
• Speakers: Jim Love (Boehringer-Ingelheim), Richard Cook (Waterloo), Terry
Therneau (Mayo)
• Discussant: Stephen George (Duke)
S5C: Current Issues in Biosimilar Studies
• Organizers: Shein-Chung Chow (Duke),Yi Tsong (FDA)
• Speakers: XiaoyuDong (FDA), Meiyu Shen (FDA), Aili Cheng (Pifzer)
S5D: Exposure ResponseModeling in the Pharmaceutical Industry
• Organizers: Dalong Huang (FDA)
• Speakers: Yaming Hang (Biogen), BretMusser (Merck), Dalong Huang (FDA)
Duke University Department of Biostatistics and Bioinformatics 17
DISS2016, Durham, NC
PARALLEL SESSIONS - ABSTRACTS
S1A: OPPORTUNITIES ANDCHALLENGESWITHBASKETDESIGNS
INONCOLOGYTRIALS
Richard Simon (NCI) – A Bayesian Design for Basket Clinical Trials
Amajor focus of oncology drug development involves use of tumor genomics to guide the use
of molecularly targeted drugs. In some cases the anti-tumor effect of a drug is mediated by its
effect on a de-regulatedmolecular target whose role in the pathophysiology of the tumor is
well understood. In those cases the development of the drug and a companion diagnostic in
a histologic type of cancer is relatively straightforward. Activity of a drug against tumors of
a histologic type bearing a genomic alteration does not always, however, imply that the drug
will be active against tumors of other histologic types bearing the same alteration. Also, even
for a single histologic type, theremay bemultiple alterations in the same pathway (or gene)
of interest and performing a separate clinical trial for each alteration is generally not feasible.
These uncertainties must generally be resolved in earlier phase clinical trials. For this reason,
a new type of early phase clinical trial has arisen, the “basket trial”. The basket trial represents
an early phase II discovery trial in which patients with defined genomic alterations but multiple
histologic types of tumors are selected to discover in which histologic types of tumors the targeted
drug is active. If the selection includes a variety of types of genomic alterations or a variety of
mutated genes, the basket trial may also be designed to determine which alterations in which
genes sensitize the tumor to the drug. Basket trials are discovery trials rather than hypothesis
testing trials; promising results of drug activity for a subset should be confirmed in amore focused
follow-up trial. Here I will describe a design for planning, monitoring and analyzing basket trials.
A website for using the new design is available at https://brpnci.shinyapps.io/BasketTrials/ and
the software is available at GitHub in the “Basket Trials” repository of account brbnci.
Amanda Redig (DFCI) – Precisionmedicine and the evolution of clinical trial design
Scientific advances of themodern era have begun to challenge earlier views of oncology in
which patients were treated with an exclusive focus on a tumor’s tissue of origin. The translation
of next-generation sequencing (NGS) into oncology practice has begun to demonstrate that
while a tumor’s primary site of origin matters, so too does its genetic landscape. However, despite
the tremendous promise of this new era of oncology, several challenges have emerged in the
translation of these new developments to the clinical trials arena. First, a genetic classification
and treatment strategymay not always follow the traditional boundaries of histopathology.
Yet how are such patients to be identified and directed towards appropriate clinical trials in
a way that provides clinically meaningful endpoints? Second, despite increasing recognition
of the importance of genomic analysis in oncology practice, evaluating targeted therapies can
present a formidable challenge when themutation(s) in question are rare and found across
DukeUniversity Department of Biostatistics and Bioinformatics 18
DISS2016, Durham, NC
disease types. Some eminently targetable mutationsmay be so rare they are only discovered
in the context of a negative trial. As our ability to probe the genome of an individual tumor
continues to expand, so toomust strategies for clinical trial design. Basket trials are a new and
evolving form of clinical trial design and are predicated on the hypothesis that the presence of
amolecular marker predicts response to a targeted therapy independent of tumor histology.
In many cases, a basket trial may actually contain several independent and parallel phase II
trials. However, the success of a basket trial depends in large part upon the strength of the data
linking target and targeted therapy. For this trial design to work, two key conditionsmust be
met: the tumormust depend upon the target pathway and the targeted therapymust reliably
inhibit the target. Several ongoing and recently published basket trials illustrate boEth the
strengths andweakness of this approach and provide insight into ways to improve trial design
while also serving as an important reminder that applying precisionmedicine sometimesmeans
looking for patient benefit when the n = 1.
S1B: KEY CONSIDERATIONONMULTI-REGIONAL CLINICAL TRIALS (MRCTS)
Steve Snapinn (Amgen) – Some Thoughts on Subgroup Analyses,With Emphasis on
Regional Subgroups inMultiregional Trials
Subgroup analyses have always been an important part of the analysis of nearly every clinical
trial. However, as long as they have been done they have faced sharp criticism, particularly
due to their relatively small sample size and their great number, leading to high rates of type I
error as well as type II error. One common approachwhen evaluating subgroup analyses is to
assume that the treatment effect is consistent across subgroups unless there is strong evidence
to the contrary, typically based on a significant treatment-by-subgroup interaction. However,
increasingly, many stakeholders find this approach to be fundamentally flawed. Notably, when
presentedwith the results of a multi-regional trial, many regional regulatory authorities want
direct evidence that the treatment is safe and effective in their region, rather than lack of evidence
that its efficacy varies across regions. In this presentation I will provide some thoughts on this
issue, touching on topics such as the goal of a subgroup analysis, the definition of region, and
proper role of an assessment of consistency.
Gordon Lan (JnJ) –MRCT designmodels and drop-min data analysis
In recent years, developing pharmaceutical products via amultiregional clinical trial (MRCT)
has becomemore popular. Many studies with proposals on design and evaluation ofMRCTs
under the assumption of a common treatment effect across regions have been reported in the
literature. However, heterogeneity among regions causes concern that the fixed effects model
for combining informationmay not be appropriate forMRCT. In this presentation, wewill discuss:
The use of the fixed effect model, the continuous random effect model, and the discrete random
effect model for the design and data analysis ofMRCTs. Numerical examples will be provided to
illustrate the fundamental differences among these threemodels. Consistency and inconsistency:
DukeUniversity Department of Biostatistics and Bioinformatics 19
DISS2016, Durham, NC
Wewill provide examples of inconsistency, and discuss the use of drop-the-min data analysis
when the region withminimum treatment effect is excluded from theMRCT data analysis. We
provide a solution first formulated within the fixed effects framework, and then extend it to
discrete random effects model.
Bruce Binkowitz (Merck) – Regions in PrecisionMedicine Clinical Trials – Complementary
or Superfluous?
Precisionmedicine is an approach to disease treatment and prevention that seeks tomaximize
effectiveness by taking into account individual variability in genes, environment, and lifestyle.
Multiregional clinical trials are an efficient way to examine disease treatments and prevention
across a diverse population including ethnic factors, culture, and other intrinsic and extrinsic
factors with region standing in as a surrogate for some combination of these factors. This presentation
will discuss the relationship between the concepts, and discuss the concept of region in a future
setting of precisionmedicine research.
S1C: PRINCIPLES ANDMETHODS IN CANCER PHARMACOGENOMICS
Jichun Xie (Duke) –Multiple Testing of General Dependence byQuantile-Based Contingency
Tables with an Application in Identifying Gene Co-expression Network Change Associated
with Cancer Survival
Gene co-expression networks describe the interactions among a set of regulators and other
substances in the cell to govern the gene expression levels of mRNA and proteins. One popular
way to estimate them is to infer the network between gene expression levels, which can be
formulated as a high dimensional network estimation/testing problem. The existingmethods
focus on networksmeasuring linear dependence or rank association, which cannot always represent
gene regulatorymechanisms. Without parametric assumptions on themarginal distributions
of continous random variables and their dependence structure, we propose a test statistic to
test whether the expression levels of two genes are independent. Based on the test statistic,
we further propose amultiple testing procedure that can simultaneously test independence
between all pairs of variables conditioning on other covariates. The numerical experiments
show that the performance of our method significantly outperform other methods when complex
dependence structures exist. Evenwhen non-linear or non-rank-associated dependence exists,
the proposedmethod performsmuch better in both validity and efficiency. Theoretically, we
prove the proposedmethod can control false discovery rate (FDR) under the desired level.
We apply this method on a gastric cancer data to investigate the change in gene co-expression
networks of patients with early or late prognostic stages, and show that ourmethod can identify
more changes that relate to the survival of the patients.
Raluca Gordan (Duke) – Assessing the effect of somatic mutations in non-coding regions
Recent whole-genome sequencing studies of paired normal-tumor samples revealed that the
DukeUniversity Department of Biostatistics and Bioinformatics 20
DISS2016, Durham, NC
majority of somatic mutations fall within non-coding genomic regions, where they have the
potential to affect regulation of gene expression. Here, we present a newmethod for assessing
the impact of non-codingmutations on the DNA-binding level of regulatory proteins called
transcription factors (TFs). TFs bind short DNA sites, typically in the neighborhood of the regulated
genes, and promote or repress gene expression. To quantitatively predict the effect of mutations
on TF-DNA binding, we use high-throughput in vitro data from protein-bindingmicroarray
(PBM) experiments for hundreds of mammalian TFs. Each PBMdata set is specific to one TF,
and it contains quantitativemeasurements of the biding specificity of that TF for 40,000 60-bp
DNA sequences. We use PBMdata to develop k-mer-based linear regressionmodels using
ordinary least squares (OLS).We use the estimated regression coefficients, as well as the variance-
covariancematrix, to compute: 1) a quantitative prediction of the change in TF binding due to
eachmutation, and 2) our confidence that the changewe predict is significant, given themodel
characteristics. Our approach is novel because, by using OLS and the variance in the coefficient
estimates, our predictions of the effects of mutations on TF-DNA binding implicitly take into
account the quality of the training data andmodel. As a result, in the case of poor data that
leads to poor predictivemodels, we require a larger change in TF binding level for amutation
to be called significant. To validate the quality of our predictions, we leverage high-throughput
enhancer assays where all possible single base-pair mutations in specific regulatory regions
have been tested directly for their effect on gene expression. We find that our TF-DNA binding
models can explain about 50% of the effect on gene expression caused by nucleotidemutations
in regulatory regions. Thus, we are confident that our newmethodwill be instrumental in prioritizing
non-coding somatic mutations for further computational and experimental analyses. This is a
joint work with Jingkang Zhao, Dongshunyi Li and AndrewAllen fromDuke.
Federico Innocenti (UNC) – Cancer treatment and use of germline genetics for determining
precision in drug therapy
The field of pharmacogenomics is focused on the characterization of genetic factors contributing
to the response of patients to pharmacological interventions. Drug response and toxicity are
complex traits; therefore the effects are likely due tomultiple genes. The investigation of the
genetic basis of drug response has evolved from a focus on single genes to relevant pathways
to the entire genome. The scope of this talk is to provide the current status of germline genomic
studies in cancer patients treated with chemotherapy, themethods for discovery and implications
for patient care. Large clinical trials enrichedwith genome-wide association studies (GWAS)
provide an unprecedented opportunity for a comprehensive and unbiased assessment of the
heritable factors associated with drug response. In oncology, germline genomics is still relatively
unexplored, particularly in reference to biomarkers of patient survival. In this presentation,
results fromGWAS efforts in oncology and the use of somatic information from the cancer
genome are presented, within the large context of achieving precision in treatment. The focus
of the GWAS discoveries in relation to overall survival will be on VDR and gemcitabine in pancreatic
DukeUniversity Department of Biostatistics and Bioinformatics 21
DISS2016, Durham, NC
cancer, as well as AXIN1 and colorectal cancer patients treated with combination of chemotherapy
and targeted therapies. Moreover, prospective validation for dose assignment will be also presented
for irinotecan based upon UGT1A1 genetics, based upon results if phase I trials where dosing
has been individualized on the basis of the UGT1A1*28 allele. Challenges and opportunities in
translating heritable genomic discoveries to patients are also discussed.
S2A: RECENTADVANCES INADAPTIVEDESIGNS FORCLINICAL TRIALS
JingWang (Gilead) – Biomarker-driven adaptive designs for precisionmedicine
Precisionmedicine has paved the way for a new era of delivering tailored treatment options to
patients according to their biological profiles. In combination with innovative adaptive design,
this has presented drug developers unprecedented opportunities to engage novel thinking to
accelerate drug discovery. Adaptive design options such as: adaptive accrual design, adaptive
threshold design, adaptive signature design, and cross-validated adaptive signature design will
be discussed. Related statistical theories and analysis strategies will also be coveredwith case
studies.
Qi Jiang (Amgen) – Key Statistical Consideration on PlatformDesign
A draft adaptive design guidance was released by FDA in 2010, and utilization of adaptive designs
in drug development has been increasing. Still, even greater use of adaptive designs is well
needed, especially in terms of less well-understood andmore complicated adaptations. In this
presentation we’ll present key statistical issues and consideration on platform design. This is a
joint work with Dr Chunlei Ke (Amgen).
Yeh-Fong Chen (FDA) – Development of GeneralMethodology for NASH Studies
Utilizing the Seamless Adaptive Design
The prevalence of non-alcoholic fatty liver disease including non-alcoholic steatohepatitis (NASH)
is increasing worldwide. NASH is the secondmost common indication for liver transplantation
and is expected to be the leading indication by 2020. In light of the increasing prevalence and
burden of disease, it is imperative to develop therapeutic strategies for patients with NASH.
Although it is feasible to conduct clinical trials with liver transplantation or death as the clinical
outcome endpoint, it may take 10 to 20 years for NASH patients to develop cirrhosis or other
liver-relatedmorbidity andmortality. In addition, evenwith the accelerated regulatory approval
to use surrogate endpoints, a study needs to be at least one year long to detect clinically meaningful
effectiveness of a study drug. Seamless adaptive clinical designs by rolling over patients from
phase 3 to phase 4 can be adopted to shorten the duration of a new drug development program
for NASH. In this presentation, wewill focus on the statistical considerations and the application
of the seamless design for NASH and hopefully stimulate broader discussions about the advantages
and drawbacks of other innovative designs. This is a joint work with Dr Shein-chung Chow
(Duke).
DukeUniversity Department of Biostatistics and Bioinformatics 22
DISS2016, Durham, NC
S2B: COLLABORATIONTOACCELERATEDEVELOPMENTOF EFFECTIVE
ONCOLOGYMEDICATIONS
Rajeshwari Sridhara (FDA) – Regulatory collaborations with academia, industry and
international regulatory agencies
FDA collaborates with external stakeholders onmultiple issues to promote science based drug
development. In many of these collaborations FDA and in particular CDER statisticians play
an important role. This presentation will include examples of collaborations between CDER
statisticians and external stakeholders including academia and pharmaceutical/biotech industry
focusing on oncology/hematology drug development.
SharonMurray (PAREXEL) –Working Together ToOperationalize Complicated Phase
I/II Studies: An Innovative Collaboration between Sponsor and CRO
PAREXEL is working collaboratively with various pharmaceutical companies on a number of
Phase I/II studies with complex study designs andmultiple data reviews during the course of the
study. These range from first-time-in-human trials to first-time-in-combination trials or phase
IIb trials withmultiple interims. Studies may havemore than one decision gate or study part,
including for example a dose selection phase, a verification phase to confirm the selected dose,
and a third phase where subjects are randomized to the experimental therapy or a comparator.
Data reviews are required at each decision gate, prior tomoving to the next phase. This talk
will describe the responsibilities of the sponsor and the CRO (PAREXEL) prior to study start-up,
during study conduct, and at the analysis stage using one of the studies as an example. Suggestions
will be provided for recommendedways of working to ensure successful delivery.
Daniel Sargent (Mayo) – Research Collaboration in Oncology: Academic perspectives
Success in today’s clinical research environment requires multiple levels of collaboration, between
different disciplines, institutions, and funding sources. In this talk I will present experiences
and perspectives on oncology clinical trials collaboration from an academic perspective, which
must satisfy multiple stakeholders including patients, funders, the scientific community, and
regulatory agencies. Issues that will be highlighted are the ever-evolving academic/industry
relationship, the growing role of large, collaborative academic-based (but often industry funded)
groups, and the rapidly growing emphasis on data sharing of individual patient data from completed
clinical trials.
S2C:METHODSANDAPPLICATIONS FOR ELECTRONICHEALTHRECORDS
Walter Boyle (SutherlandHealthcare) - The Promise and Pitfalls of ElectronicMedical
Records
To researchers and analysts in the healthcare field, the ElectronicMedical Record (EMR) is one
of themost exciting sources of data available. At first glance an EMRwould seem to have the
DukeUniversity Department of Biostatistics and Bioinformatics 23
DISS2016, Durham, NC
potential to give an almost holistic representation of a patient’s health. However, that potential
comes hand in handwith a variety of pitfalls which can stop a research project before it even
gets started. This presentation will use a previous research project rooted in EMR data to discuss
some of these issues and help you to understandwhy they happen and to even potentially overcome
them.
Joseph Lucas (Duke) - Tracking the Effects of Healthcare Innovations through the use
of electronic health records
Changes in incentives and business models are leading health systems to rapidly innovatemany
aspects of care. However, the effects of those innovations are often not being tracked. In scientific
terms, this is like running an experiment without paying any attention to the outcome. Without
understanding the effect of healthcare innovation on both patient health and efficiency, we
lose the opportunity to learn from the changes that are implemented. Changes in healthcare
delivery should be treated as a full-fledged experiment; a hypothesis should be generated,
tools for measuring the outcome should be developed and the implementation of the changes
should be conducted in a way to allow testing of the hypothesis. In this talk wewill discuss some
tools we have developed for ongoing tracking of healthcare innovations. Our approach uses a
Bayesian framework to continuously update our understanding of the effects of the innovation
and presents those results in results in a context that is useful for supporting decisionmaking by
health system administrators.
BenGoldstein (Duke) - Informed Presence in the Analysis of Electronic Health Records
An increasingly popular data source for clinical analyses are Electronic Health Records (EHRs).
They are often readily available and contain detailed information on large amounts of patients.
Owing to this density of information problems of unmeasured covariates is often a secondary
concern. Instead, a challenge in the analysis EHRs is a form of selection bias: Informed Presence.
As others have noted patients in EHRs are sicker than the general population. Moreover, since
individuals are only observedwhen they have amedical encounter, i.e. sick, there is high potential
for selection effects and bias in association studies. In this talk wewill discuss different forms
of informed presence and illustrate how they can bias typical analyses. Wewill also discuss
different means of addressing these biases.
S3A: APPLICATIONOFGENETIC INFORMATION INONCOLOGYCLINICAL
TRIALDESIGN
SuzanneDahlberg (DFCI) - Incorporation of Genetic Information in Clinical Trials
The discovery of driver oncogenes has dramatically impacted clinical research, placing great
focus on the biological causes of dramatic responses to therapy among populations that do not
derive treatment benefit on average. The corresponding successful development of therapies
targeting those genetic abnormalities has prompted us to investigate tumor biology routinely as
DukeUniversity Department of Biostatistics and Bioinformatics 24
DISS2016, Durham, NC
integral or integrated biomarkers in the trial setting. How to incorporate the science statistically
is not a one-size-fits-all approach, particularly as the field rapidly transitions its focus to the
development of immune checkpoint inhibitors; study design depends in part on the phase of
the trial, the prevalence and strength of evidence of the biomarker, preferred endpoints, and
the assay or test used to detect it. Using examples from the oncology literature, I will discuss
how each of these considerations impacts the design and conduct of a clinical trial.
SumithraMandrekar (Mayo) - Clinical Trial Designs in the Era of PrecisionOncology
Clinical trial design strategies have evolved as ameans to accelerate the drug development
process so that the right therapies can be delivered to the right patients. With these changes
in the science of oncology have come changes to the waywe design and perform clinical trials.
Increasingly common are trials tailored to detect enhanced efficacy in a patient subpopulation,
e.g., patients with a known biomarker value or whose tumors harbor a specific genetic mutation.
Earlier biomarker-based designs typically assessed a single targeted therapy in a single disease
typewith 1 or 2molecular groups. These include enrichment, marker-stratified, andmarker
strategy designs. Newer biomarker-based designs expand on the earlier ones by includingmultiple
targeted therapies, multiple disease types, and/or multiple molecular groups. These include
modified strategy designs, umbrella trials, Bayesian biomarker-adaptive designs, and basket
trials. In this talk, I will discuss a number of trials that are examples of these biomarker-based
designs either in a proof of concept early phase setting or in amore definitive Phase III setting.
These include the National Cancer Institute’s precisionmedicine initiative trials such as the
ALCHEMIST, and LungMAP, as well as other trials such as SHIVA,Matrix, the FrenchNational
Cancer Institute’s AcSé (Secured Access to Innovative Therapies) program, and ASCO’s Targeted
Agent and Profiling Utilization Registry trial.
Yijing Shen (Genentech) –Operational Adaptive Biomarker Development Strategy
for the AtezolizumabNSCLC Program
Identifying accurate diagnostic cutoffs and suitable endpoints to use in pivotal oncology studies
is often challenging, especially in the immunotherapy setting where the early efficacy endpoints
such as ORR and PFSmay not be strongly associated with OS. In addition, delayed treatment
effect in this setting requires larger studies and longer follow up to estimate the treatment
benefit to inform pivotal study designs. This abstract describes the clinical development strategy
for the second/third line AtezolizumabNSCLC program conducted by Genentech/Roche. Atezolizumab
is a humanized anti-PDL1 antibody that inhibits the binding of PD-L1 to PD-1 and B7.1. A Phase
Ia trial showed single-agent activity in NSCLC patients with the objective response rate (ORR)
associated with PD-L1 expression on tumor-infiltrating immune cells (IC) and/or tumor cells
(TC). After atezolizumab demonstrated single agent activity with differential ORR by PD-L1
expression level, the atezolizumab clinical and biomarker development plan was designedwith
limited atezolizumab randomized data and from in-class agents whichmade evaluation of efficacy
endpoints and optimal PD-L1 expression cutoff challenging, particularly in the context of accelerated
DukeUniversity Department of Biostatistics and Bioinformatics 25
DISS2016, Durham, NC
drug development. Four parallel 2L+NSCLC studies were designed to provide dynamic information
regarding the optimal diagnostic tissue type, cutoffs, and endpoints. Two of these were phase 2
trials (one PD-L1 selected single arm and one all comer randomized) designed to better estimate
the treatment benefit and inform potential changes of the other two registrational study designs
(one PD-L1 selected single arm and one all comer randomized). This operational adaptive strategy
provided flexibilities and helped refine the final registrational trial designs. In the presentation,
strategic context, design details, andmodification scenarios of the 2L+NSCLCAtezolizumab
studies will be discussed. This is a joint work with Zhengrong Li, Pei He and Jing Yi.
S3B: CURRENT ISSUES IN CANCER PHASE II TRIALS
Sin-Ho Jung (Duke) – Statistical Issues for Design and Analysis of Phase II Cancer
Clinical Trials
Phase II trials have been very widely conducted and published every year for cancer clinical
research. In spite of the fast progress in design and analysis methods, single-arm two-stage
design is still themost popular for phase II cancer clinical trials. Because of their small sample
sizes, statistical methods based on large sample approximation are not appropriate for design
and analysis of phase II trials. As a prospective clinical research, the analysis method of a phase
II trial is predetermined at the design stage and it is analyzed during and at the end of the trial
as planned by the design. The analysis method of a trial should bematchedwith the design
method. For two-stage single arm phase II trials, Simon’s method has been the standards for
choosing an optimal design, but the resulting data have been analyzed and published ignoring
the two-stage design aspect with small sample sizes. In this talk, I review analysis methods
that exactly get along with the exact two-stage designmethod. I also discuss some statistical
methods to improve the existing design and analysis methods for single-arm two-stage phase II
trials.
Ilya Lipkovich (Quintiles) – Biomarker-driven seamless phase II/III trials for a rare
disease
In this presentation we consider a seamless Phase II/III design that was recently used in a clinical
trial in patients withmesothelioma, a rare cancer found in the lining surrounding the lungs and
other organs. The new experimental treatment was compared to placebo using an adaptive
biomarker-driven design. After the first stage of the seamless trial (end of Phase II) has been
completed, data are evaluated and themost promising biomarker that helps predict treatment
response is identified using the SIDESmethodology (Lipkovich et al, 2011 and Lipkovich and
Dmitrienko, 2014). Based on predictive power evaluated after the first stage, a decision is made
whether to terminate the trial for futility, continue the trial without any changes in the overall
patient population, adjust the sample size (target number of events) in the same population, or
focus on a subset of the overall population based on the selected biomarker (biomarker -positive
subpopulation), possibly in combination with testing the overall effect via an appropriate multiple
DukeUniversity Department of Biostatistics and Bioinformatics 26
DISS2016, Durham, NC
testing procedure. We present the results of a simulation study that evaluates key operating
characteristics of this design under different scenarios. A joint work with Alex Dmitrienko (Mediana).
Kevin Liu (JnJ) – Biomarker Enrichment Design for Early PhaseOncology Studies
Oncology drug development has been increasingly shaped bymolecularly targeted agents
(MTAs), which often demonstrate differential effectiveness driven by the expression profile
of their molecular targets in tumors. Innovative statistical designs have been proposed to tackle
this new challenge, e.g., Freidlin (2005), Jiang (2007) and Freidlin (2010). All of these are essentially
adaptive confirmatory Phase 3 designs that combine the testing of treatment effectiveness in
the overall population with a possible pathway to identify a potential sensitive subpopulation.
We argue that, in cases that there are strong biological rationale and preclinical evidence to
support that aMTAmay provide differential benefit a general patient population, it is imperative
that early phase POC studies be designed to specifically address the biomarker-related questions,
e.g., subgroup selection, biomarker threshold evaluation, in order to improve the efficiency of
development. In this presentation, wewill discuss statistical enrichment strategies for Phase
2 oncology designs, both in single-arm and in randomized settings. Also to be discussed is the
likely challenge of the lack of a reliable assay in the earlier development phases. To this extent,
wewill discuss the impact of measurement error on the operation characteristics of these designs,
which are evaluated through simulations. This is a joint work with Dr Hong Tian from JnJ.
S3C: DISCOVERY SCIENCE FOR IMMUNOTHERAPY TRIALS
KentWeinhold (Duke) – The Immunologic Basis for Current Cancer Immunotherapies
Despite concerted efforts over the past 30+ years, it was not until very recently that immunotherapeutic
approaches against cancer began to showmore than just rare anecdotal successes. Much of the
current success is the result of technological advances that have re-shaped our understanding
of the patient’s immune response to their actively growing tumor. This, coupled with the development
of novel biological treatmentmodalities, has brought about a revolution in cancer immunotherapy
that has, in certain instances, resulted in response rates exceeding 30-40%. The challenges
brought about by this ‘paradigm shift’ are now centered on: 1) the identification of baseline
‘pharmacodynamicmarkers’ that predict which patients would benefit from a specific immunotherapeutic
approach, and 2) the profiling of immunologic components within both the peripheral circulation
and the tumormicroenvironment in search of specific biomarkers that track with clinical efficacy.
This introductory overview presentation will focus on the process of T cell activation, maturation,
exhaustion, and regulation as well as effector T cell responses to highly conserved tumor associated
antigens (TAA) and neoantigens generated by somatic mutations. Finally, the concept of ‘hot’
versus ‘cold’ tumormicroenvironments and their impact on specific immunotherapeutic strategies
will be discussed.
DukeUniversity Department of Biostatistics and Bioinformatics 27
DISS2016, Durham, NC
Eric Groves (Quintiles) – Immunotherapies: Opportunities and Challenges
Until recently therapies that manipulate the immune system to achieve anti-cancer benefits
have had limited success. Coley’s toxin, IL-2, Interferon, various vaccines, all have produced
tantalizing but infrequent results. But with the advent of CTLA-4, PD-1 and PD-L1 inhibitor
therapies, dramatic results have been observedwith evidence of long term benefit. This discussion
will focus on a brief historic review, present some current data and then discuss the opportunities
and challenges that future development of these and additional novel immune therapies may
offer.
Lynn Lin (Penn State) – Characterizing Antigen-specific T-cell Functional Diversity in
Single-cell Expression Data
Rapid advances in flow cytometry and other single-cell technologies have enabled high-dimensional
measurement of individual cells in a high-throughput manner so that many new and long-standing
questions about cell population heterogeneity can now be addressed. One specific hypothesis
is that some characteristic or quality of a subset of antigen-specific T cells involved in immune
function (in particular, “olyfunctional T-cells” which are capable of simultaneously producing
multiple effector cytokines) is associated with protective immunity from infectious diseases.
During this talk, I will present a novel statistical framework for unbiased polyfunctionality analysis
of antigen-specific T-cell subsets defined from high-dimensional single-cell assays and demonstrate
how it can be used to comprehensively unravel rare signals associated with vaccine efficacy
frommultiple clinical datasets that would bemissed by traditional analyses.
Radleigh Santos (TPIMS) – Using Fusion of ResponseMetrics andMonte Carlo Simulation
to Determine Immune Response in Cancer Immunotherapy Patients
Determining immune responders in the post-treatment clinical context of cancer immunotherapy,
in which patients are treated with one ormore antigens for the purpose of eliciting an immune
response against the cancer can be challenging. In general, the effectiveness of threshold-based
criteria, such as spot count difference from control in ELISpot, can vary widely, depending on
patient population latent response and also on experimental choices that increase background
variation such as IVS testing. On the other hand, inferential statistical tests such asmDFR or the
binomial test can be impacted by varying numbers of samples per patient and also by general
differences in patient population distribution of response. Furthermore, measuring differences
between pre- and post-treatment response using either a direct statistical test, or a difference
of some kind between independently determined pre- and post-treatment response are options
when determining immune responders. The end result is that no single approach is applicable in
all cases; this, in turn, can lead to the data itself dictating the definition of immune responder, a
non-objective process that is difficult to apply broadly. In this presentation, a novel heuristic for
determining immune responders usingmultiple metrics combined in a fusion scoring approach
will be shown. Monte Carlo simulation is then used to put these fusion scores into a clear context
for the purpose of assigning responder status to individual patient samples, and hence to each
DukeUniversity Department of Biostatistics and Bioinformatics 28
DISS2016, Durham, NC
patient . A specific implementation of this approachwill be shown using data from a recent
phase 2 glioblastoma immunotherapy trial (ICT-107) in which HLA-A2 patients were treated
with six synthetic peptides. Patient samples were tested for immune response using both ELISpot
andMultimer, and it will be shown how themethodwas used for both types of data. Promising
associations between responders designated in this manner and survival endpoints suggest that
this method of designating patients as immune responders captures some of the underlying
mechanism of action of this treatment.
S4A: ADDENDUMTOSTATISTICAL PRINCIPLES FORCLINICAL TRIALS (ICH
E9): CHOOSINGAPPROPRIATE ESTIMANDSANDDEFINING SENSITIVITY
ANALYSES IN CLINICAL TRIALS
CraigMallinckrodt (Lilly) - Overview of Estimands, Estimators, and Sensitivity for
Longitudinal Clinical Trials
Recent research has fostered new guidance on preventing and treatingmissing data. Consensus
exists that clear objectives should be defined along with the causal estimands; trial design and
conduct shouldmaximize adherence to the protocol specified interventions; and, a sensible
primary analysis should be used along with plausible sensitivity analyses. Two general categories
of estimands are: effects of the drug as actually taken (de-facto, effectiveness) and effects of the
drug if taken as directed (de-jure, efficacy). Fundamental, design, and analysis considerations for
common estimands will be discussed. Examples are used to illustrate the benefit from assessing
multiple estimands in the same study. General approaches to sensitivity analyses will be introduced
and subsequent speakers in this session will elaborate on specific approaches.
Bohdana Ratitch (Quintiles) - Imputation-based Analysis Strategies in the Presence of
Treatment Non-Adherence
A definition of estimand for a clinical trial may take into account how non-adherence to randomized
treatment, e.g., early discontinuation or initiation of rescue therapy, would be accounted for in
the estimate of treatment effect, in alignment with study objectives. Ideally, this should lead to
a design of a clinical study where all data relevant for the defined estimand can be collected for
all subjects. In practice, however, some amount of missing data can be expected inmost clinical
trials andwithmost estimands, due to subjects missing planned study visits or withdrawing
from the study prematurely. In some cases, for ethical reasons, it may also be impossible to
obtain usable (non-confounded) data for a specific estimand for some subjects during the periods
of non-adherence. In case of missing data or data unusable for a given estimand, analysis methods
have to rely on some strategies of handling unusable or unavailable outcomes in amanner that
is consistent with the planned estimand. Wewill review several analysis strategies that are
based on subject-level imputation usingmultiple imputationmethodology. Wewill focus on
several variants of reference-based and delta-adjustment approaches and discuss the type of
DukeUniversity Department of Biostatistics and Bioinformatics 29
DISS2016, Durham, NC
assumptions about the outcomes of non-adhering subjects that can be implemented using these
strategies.
DevanMehrotra (Merck) - Clinical Trials with Dropouts: Proposed Estimand-Aligned
Primary and Sensitivity Analyses
In a typical randomized clinical trial comparing two treatments (test, control), the endpoint of
interest (e.g., change from baseline in HAMD-17 at 6 weeks) is not observed for dropouts. The
resultingmissing data problem is commonly tackled by invoking amissing at random (MAR)
assumption and proceeding with amixedmodel repeatedmeasures (MMRM) analysis. In some
settings, theMAR assumptionmay be reasonable for the control treatment (often placebo)
but not for the test treatment (experimental drug). In such cases, theMMRM-based estimated
between-treatment difference in endpoint means tends to be biased for the estimand of interest.
We propose a simple alternative approach in which the implicitly imputedmean for test-arm
dropouts in theMMRManalysis is explicitly replacedwith either the estimatedmean for all
control-arm patients (primary analysis) or the estimatedmean for control-arm dropouts only
(sensitivity analysis); patient-level imputation is not required. An additional sensitivity analysis
is proposed in which a common “bad” outcome is imputed for all the test and control arm dropouts
followed by a between-treatment comparison of trimmedmeans using quantile regression. All
analyses address the same estimand. A real dataset and simulations are used to support the key
messages.
S4B: CAUSAL INFERENCE IN CANCERCLINICAL RESEARCH
Donglin Zeng (UNC) – Estimating Treatment Effects with Treatment Switching via
Semi-Competing RisksModels: An Application to a Colorectal Cancer Study
Treatment switching is a frequent occurrence in clinical trials, where, during the course of the
trial, patients who fail on the control treatmentmay change to the experimental treatment.
Treatment switching creates statistical challenges for estimating the causal effect of the treatment.
Analyzing the data without accounting for switching yields highly biased and inefficient estimates
of the treatment effect. In this talk, in order to accurately assess the treatment effect, we propose
a novel class of semiparametric semi-competing risks transition survival models to accommodate
switch. Theoretical properties of the proposedmodel are examined and an efficient expectation-
maximization algorithm is derived for obtaining themaximum likelihood estimates. Simulation
studies are conducted to demonstrate the superiority of themodel compared to the intent-to-treat
analysis and other methods proposed in the literature. The proposedmethod is applied to analyze
data from the panitumumab study.
XiaofeiWang (Duke) – Bias-adjusted Kaplan-Meier Survival Curves forMarginal
Treatment Effect in Observational Studies
For time-to-event outcome of multiple treatment groups, the Kaplan-Meier estimator is often
DukeUniversity Department of Biostatistics and Bioinformatics 30
DISS2016, Durham, NC
used to estimate survival functions of treatment groups and computemarginal treatment effects,
such as difference of survival rates between treatments at a landmark time. The Kaplan-Meier
estimates and the derived estimates of marginal treatment effect are uniformly consistent
under general conditions when data are from randomized clinical trials. For data from observational
studies, these statistical quantities are often biased due to treatment-selection bias. Propensity
score basedmethods, such as the inverse probability of treatment weighting, estimate the
survival function by adjusting for the disparity of propensity scores between treatment groups.
Unfortunately, themisspecification of the regressionmodel will lead to biased estimates in
these existingmethods. Using an empirical likelihood (EL) method in which themoments of the
covariate distribution of treatment groups are constrained to equality, we obtain consistent
estimates of the survival functions and themarginal treatment effect through themodified
Kalpan-Meier estimator. Equatingmoments of the covariates distribution between treatment
simulates the covariate distribution if the patients had been randomized to these treatment
groups. We established the consistency and the asymptotic limiting distribution of the proposed
EL estimators. We demonstrated that unlike propensity scoremethods, the consistency of the
proposed estimator does not depend on a correct specification of amodel. Covariates subsets
on bias control are also discussed. Simulation was used to study the finite sample properties
of the proposed estimator and compare it with existingmethods. The proposed estimator is
illustrated with observational data from a lung cancer observational study to compare two
surgical procedures in treating early stage lung cancer patients.
Jeremy Taylor (UMICH) – Estimating the causal effect of salvage hormone therapy in
prostate cancer
Causal effects of interventions can be assessed by considering what the subject’s outcome
would have been if they had taken the intervention and compare that with what it would have
been had they not taken the intervention. After initial treatment for localized prostate cancer
patients monitor their PSA values, and a rise in PSA suggests that the cancer may be regrowing
and it would bewise to start a new therapy to prevent or delay the occurrence of clinical symptoms.
Androgen deprivation (hormone) therapy can delay the recurrence of prostate cancer, however
it has some side effects. In deciding whether to start hormone therapy some important considerations
are; what is the chance of recurrence in the next few years and how effective will hormone
therapy be at reducing that. A complication to estimating the effect of hormone therapy is that
it is given by indication, i.e. those with rising PSA levels and at greater risk of recurrence tend to
have received it. Using data from patients initially treated with radiation therapy, in this talk I
will present and compare three approaches to estimating the effect of hormone therapy. One
based on joint longitudinal and survival models, one based on sequential stratification and one
based onmarginal structural models. A second important question is to determine the optimal
treatment regime for when to start salvage hormone therapy for a population to follow. We
consider a class of regimes in which hormone therapy is first givenwhen PSA is rising and it first
DukeUniversity Department of Biostatistics and Bioinformatics 31
DISS2016, Durham, NC
crosses a threshold b. For each regime, from the observational data, wemodel the probability
of adherence to that regime, using random forests to estimate this probability and use inverse
probability weighting to estimate the expected survival curve under that regime, and define the
optimal regime as the onewith the largest mean restricted survival distribution.
S4C: NEWDEVELOPMENTS IN SURVIVALANALYSIS FORCANCERRESEARCH
Danyu Lin (UNC) –Maximum Likelihood Estimation for Semiparametric Regression
Models with Interval-Censored Data
Interval censoring arises frequently in clinical and epidemiological studies, where the event
or failure of interest is not observed at an exact time but is rather known to occur within an
interval induced by periodic monitoring. We formulate the effects of potentially time-dependent
covariates on the interval-censored failure time through semiparametric regressionmodels,
such as the Cox proportional hazardsmodel. We study nonparametric maximum likelihood
estimation with an arbitrary number of monitoring times for each subject. We devise an EM
algorithm that converges stably for arbitrary datasets. We then show that the estimators for
the regression parameters are consistent, asymptotically normal, and asymptotically efficient
with an easily estimated covariancematrix. In addition, we extend the EM algorithm and asymptotic
theory to competing risks andmultivariate failure time data. Finally, we demonstrate the desirable
performance of the proposed numerical and inferential procedures through extensive simulation
studies and applications to real medical studies.
Butch Tsiatis (NCSU) – Inference on treatment effects from a randomized clinical trial
in the presence of premature treatment discontinuation: The SYNERGY trial
The SYNERGY trial was a randomized, open-label, multi-center clinical trial designed to compare
two anti-coagulant drugs on the basis of various time-to-event endpoints. As usual, the protocol
dictated circumstances, such as occurrence of a serious adverse event, under which it wasmandatory
for a subject to discontinue his/her assigned treatment. In addition, as in the execution of many
trials, some subjects did not complete their assigned treatment regimens but rather discontinued
study drug prematurely for other, “optional” reasons not dictated by the protocol; e.g., switching
to the other study treatment or stopping treatment altogether at their or their provider’s discretion.
In this situation, as an adjunct to the usual intent-to-treat analysis, interest may focus on inference
on the “true” treatment effect; i.e., the difference in survival distributions were all subjects in
the population to follow the assigned regimens and, if to discontinue treatment, do so only for
mandatory, but not optional, reasons. Approaches to inference on this effect used commonly
in practice are ad hoc and hence are not generally valid. We use SYNERGY as amotivating
case study to propose generally-applicable methods for estimation and testing of this “true”
treatment effect by placing the problem in the context of causal inference on dynamic treatment
regimes. Analysis of data from SYNERGY and simulation studies demonstrate the utility of the
methods. This is a joint work with DrMarie Davidian andDrMin Zhang.
DukeUniversity Department of Biostatistics and Bioinformatics 32
DISS2016, Durham, NC
Jason Fine (UNC) - Dependent Censoring and Competing Risks: Confusion and Controversy
Historically, censoring has played a dominant role in survival analysis, with the typical development
in an introductory course placing heavy emphasis on right censoring and the assumption of
independence necessary for the validity of standard analyses for the event time. Competing
risks, in which there aremultiple causes of failure, are widespread in contemporary applications
in the biological and biomedical sciences. Such issues are generally treated superficially in the
framework of right censoring, where the independence assumptionmay be violated when a
competing event occurs. This has lead to confusion regarding the appropriate analysis of competing
risks data: the conventional wisdom is that if the competing risks are independent, then standard
methods for right censored data should be employed. However, the interpretation of such
analyses may be problematic, as the quantity being estimated corresponds to a setting where
competing events do not exist, whichmay not be practically relevant. On the other hand, if
dependence is a possibility, then the standard recommendation is that alternative approaches
may be needed. Again, care is needed, as issues of interpretation are closely tied to themethod
of analyses. When the competing event is treated as a censoring event, issues of interpretation
may arise, just as under independence. This talk will survey these issues, highlighting how the
common strategy for teaching survival analysis has lead to controversy regarding the handling
of competing events. I will suggest an alternative approachwhich places censoring and competing
risks on equal footing, providing a clear and explicit understanding of key interpretive issues
and assumptions underlying the available analyses. Real data examples will be used to illustrate
themain points.
S4D: OPPORTUNITIES ANDCHALLENGES IN THEDESIGNANDANALYSIS
OF IMMUNOTHERAPIES TRIALS
Kay Tatsuoka (BMS) - Alternatives to traditional endpoints and analysis methods in
immuno-oncology trials
Traditional endpoints, design and analysis methods in Oncologymay not be appropriate for
Immuno-Oncology. Novel endpoints based on datamining will be explored as alternatives.
Their relationship to long-term survival will be discussed. Additionally, the issues related to
event-based designs owing to non-proportional hazards will be illustrated using examples.
Some possible solutions to overcome these issues will be discussed.
SusanHalabi (Duke) - Challenges in the Design of Oncology Trials with Immunotherapies
Several oncology trials with immunotherapies have shown that the treatments have impact on
overall survival but not progression-free survival. Evenwith the survival, delayed treatment
effect of at least 3months has been observed suggesting that the proportional hazards assumption.
The violation of the proportional hazardsmodel will also have an impact onmonitoring the trial.
Late stage designs will be discussed and examples will be used to illustrate howwe addressed
DukeUniversity Department of Biostatistics and Bioinformatics 33
DISS2016, Durham, NC
some of the challenges in the design.
PralayMukhopadhyay (AstraZeneca) - Statistical Considerations in the development
of novel Immune-Oncology Agents
The development of immune-oncology (IO) agents poses some unique challenges that may not
be fully addressed by the use of traditional statistical approaches. Onewell known problem
is the anticipation of a delayed treatment effect and the need for sample size considerations
assuming non-proportional hazards (PH).We discuss the impact of a delayed treatment effect
and implications on sample size and study duration during the design stage. We also explore the
appropriateness of using a weighted logrank test versus the un-weighted test for comparing
two treatment arms in this situation. Another challenge is the overall characterization of clinical
benefit of these agents, where the treatmentmay be effective in only a subset of patients but
resulting in long term remissions that are a hallmark of IO therapy. We evaluate the strengths
and limitations of available measures such as the hazard ratio (HR), medians and restricted
mean survival time (RMST) in describing overall treatment benefit. An alternative approach
using the generalized pairwise comparisonmethod is evaluated in this setting.
S5A: SAFETYANDBENEFIT-RISKANALYSIS INDRUGDEVELOPMENT
Frank Rockhold (Duke) – Benefit to risk considerations in ongoingmonitoring of clinical
trials
The overall goal of the clinical trial is to assess a primary objective and endpoint (usually a benefit)
over the background of secondary endpoints including patient safety. The objective of trial
monitoring is to integrate the information efficacy and safety information in an integratedmanner
tomake ongoing decisions about whether to continue the trial as is, have the design altered,
or prematurely discontinue based on the benefits and harms they are observing in the trial. In
other words, the IDMC is taskedwith creating a “benefit to risk” picture for the trial patients
and future patients on an ongoing basis. The science of Benefit to risk for quantitatively summarizing
completed trials (one ormany) has evolved over the past decade. The purpose of this talk to is to
continue our exploration of how onemight apply these techniques in amoremore structured
and systematic way in an IDMC setting.
Chunlei Ke (Amgen) – Benefit-risk Assessment Using Number Needed to Treat and
Number Needed to Harm
Number needed to treat (NNT) is a useful measure to translate the therapeutic effect of a drug
to clinical practice. So is the number needed to harm (NNH) for the potential risk. NNT and
NNHprovide a simple approach to assisting the benefit-risk assessment (BRA) of a drug. Time-to-event
endpoints are commonly used in oncology clinical trials. NNT has been extended to time-to-event
endpoints but these extensions have some limitations. Wewill present a hazard-based NNT
using an additive hazardsmodel. Estimation and inference procedures will be provided. Graphical
DukeUniversity Department of Biostatistics and Bioinformatics 34
DISS2016, Durham, NC
methods and formal statistical tests will be proposed to evaluate the assumption of additivity. A
multivariate additive hazardsmodel will be used to estimate NNT andNNH jointly. Inference
on the NNH andNNT ratio for BRAwill be explored as well. The proposedmethods will be
illustrated with some real clinical trial dataset.
OlgaMarchenko (Quintiles) – Safetymonitoring in oncology clinical trials
TheNIH requires data and safetymonitoring, generally, in the form of Data and SafetyMonitoring
Boards (DSMBs) for phase III clinical trials. For earlier trials (phase I and phase II), a DSMB
may be appropriate if the studies havemultiple clinical sites, are blinded (masked), or employ
particularly high-risk interventions or vulnerable populations; otherwise, safety monitoring
should be performed by a study investigator and a study safety team. A formal stopping rule for
toxicity can serve as a useful reference for a DSMB or a safetymonitoring teamwhen reviewing
the totality of toxicity data in oncology trials. Phase I trials in oncology are conducted among
cancer patients and typically with an assumption that the benefit of the cancer treatment will
increase with dose. Severity of toxicity is also expected to increase with dose, so the challenge
is to increase the dose without causing an unacceptable toxicity to patients. The goal of Phase
I trials is to identify themaximum tolerated dose (MTD). Phase II studies are conducted at the
MTD estimated from phase I and they evaluate whether a new drug has sufficient efficacy to
warrant further development and refine the knowledge of its safety profile. In Phase II trials a
toxicity rule prescribes if the trial needs to be stopped early due to levels of toxicity higher than
expected. In Phase III trial we look at the benefit-risk trade-off to decide if the trial needs to
be stopped. In this presentation, the overview of safety monitoring strategies using Frequentist
and Bayesian approaches together with case studies will be presented and discussed. The discussion
will be focusedmainly on safetymonitoring in Phase II and Phase III trials.
S5B: STATISTICAL ISSUES RELATED TOPROGRESSION-FREE SURVIVALAND
OVERALL SURVIVAL
Jim Love (Boehringer Ingelheim) –Observations on the relationship between progression-free
survival and overall survival
This talk will use published results from trials in lung cancer to support the following assertions:
The PFS-OS relation is qualitative, as well as quantitative. Qualitative factors include: the type
of comparator; extent of post-progression anti-cancer treatment; and sub-groups defined by
the targetedmechanism of drug action. Despite substantial effects on PFS, the effect onOS can
be equivocal, and can differ within sub-groups. Clear effects onOS can sometimes be demonstrated
in sub-groups defined by themechanism of action despite extensive “cross-over”. On the other
hand, significant OS can be seen despite amoremodest effect on PFS. Sub-groups driving the
DukeUniversity Department of Biostatistics and Bioinformatics 35
DISS2016, Durham, NC
significant effect onOS in the overall results can sometimes be inferred, but not clearly identified.
Richard Cook (Waterloo) – The analysis of progression-free survival, overall survival
andmarkers in cancer clinical trials
Cancer clinical trials are routinely designed on the basis of event-free survival timewhere the
event of interest may represent a complication, metastasis, relapse, or progression. This talk
is concernedwith a number of statistical issues arising with use of such endpoints including the
interpretation of composite endpoints, analyses involving dual censoring schemes for component
endpoints, and the causal interpretation of effects. Remarks will also bemade on the analysis of
longitudinal and survival data.
Terry Therneau (Mayo Clinic) – UsingMulti-stateModels to Understand Cancer Data
This talk will show howwewere able to sort out and better understand puzzling results from a
cancer trial by usingmulti-state prevalence (Aalen-Johansen) curves to track patients’ progress
through the cycle of induction, response, stem cell transplant, and relapse. Tools for this are part
of the standard R survival package and are very easy to use.
S5C: CURRENT ISSUES IN BIOSIMILAR STUDIES
XiaoyuDong (FDA) - Exact Test Based Approach for Equivalence Test with Parameter
Margin
Equivalence test has a wide range of application in pharmaceutical statistics in which we need to
compare the similarity between two groups. In recent years, equivalence test has been used in
assessing analytical similarity between a proposed biosimilar product and a reference product.
More specifically, themean values of the two products for a given quality attribute are compared
against an equivalencemargin of±f × σR, which is a function of the reference variability.
In practice, this margin is unknown and is estimated from the sample as±f × SR. If we use
this estimatedmargin with the classic t-test statistics on the equivalence test on themeans,
both Type I and Type II errors may inflate. To resolve this issue, we develop an exact-based test
method and compare this methodwith other proposedmethods, such asWald test, constrained
Wald test and Generalized Pivotal Quantity in terms of Type I error and power. Application of
thosemethods on data analysis is also provided in this paper. This work focuses on the development
and discussion on the general statistical methodology and is not limited to the application of
analytical similarity.
Aili Cheng (Pifzer) – A Further Look at the Current Equivalence Test for Analytical
Similarity Assessment
Establishing analytical similarity is the foundation of the biosimilar product development. Although
there is no guidance on how to evaluate analytical data for similarity, the FDA recently suggested
the equivalence test for themean difference between innovator and the biosimilar product
DukeUniversity Department of Biostatistics and Bioinformatics 36
DISS2016, Durham, NC
as the statistical similarity assessment for Tier 1 quality attributes (QAs), which are defined
as theQAs that are directly related to themechanism of action. However, our mathematical
derivation and simulation work has shown that type I error is typically increased inmost realistic
settings when an estimate of sigma is used for the equivalencemargin, which cannot be corrected
by increasing sample size. The impact of the constant c on type I error and sample size adjustment
in the imbalanced situation will be discussed in this presentation as well. This is a joint work
with Dr Neal Thomas from Pifzer.
Meiyu Shen (FDA) – Statistical considerations regarding to correlated lots in analytical
biosimilar equivalence test
In the evaluation of the analytical similarity data, an equivalence testing approach for most
critical and quantitative quality attributes, that are assigned to Tier 1 in their proposed three-tier
approach, was proposed (Tsong, Dong, and Shen, 2016). Food andDrug Administration (FDA)
has recommended the proposed equivalence testing approach to sponsors throughmeeting
comments for Pre-Investigational NewDrug Applications (PINDs) and Investigational New
Drug Applications (INDs) since 2014. FDA has received some feedbacks on statistical issues of
potentially correlated reference lot values subjected to the equivalence testing since independent
and identical observations (lot values) from the proposed biosimilar product and the reference
product are assumed. In this article, we describe onemethod proposed by Yang et al (2016) in
biosimilar submissions for correcting the estimation bias of the reference variability so as to
increase the equivalencemargin and its modified versions for increasing the equivalencemargin
and correcting standard errors in the confidence intervals assuming that the lot values are
correlated under a few known correlationmatrices. Our comparisons between these correcting
methods and no correction for bias in the reference variability under several assumed correlation
structures indicate that all correctingmethods would increase type I error rate dramatically
but only improve the power slightly for most of the simulated scenarios. For some particular
simulated cases, type I error rate can be extremely large (e.g., 59%) if the guessed correlation is
larger than the assumed correlation. Since the source of a reference drug product lot is unknown
in nature, correlation between lots is a design issue. Hence, to obtain independent reference lot
values by purchasing reference lots at a wide timewindow often is a design remedy for correlated
reference lot values. This is a joint work with Dr TianhuaWang andDr Yi Tsong from FDA/CDER.
S5D: EXPOSURE RESPONSEMODELING IN THE PHARMACEUTICAL INDUSTRY
YamingHang (Biogen Idec) – Important Issues Relevant to Statisticians in Exposure-Response
Modeling – Illustrated by Case Studies
Over the years, pharmacometrics work including exposure-responsemodeling and simulation
have been appliedmore broadly to aid drug development as well as to influence the regulatory
decisionmaking. Depending on the purpose of themodeling work and nature of the data, same
data set can be analyzed in very different approaches. In turn, the requirement for statistical
DukeUniversity Department of Biostatistics and Bioinformatics 37
DISS2016, Durham, NC
rigor and robustness of themodel varies. For example, a proportional hazardmodel might be
sufficient to describe the relationship between exposure and risk of certain event, however,
it is most likely insufficient to predict the incidence rate under different complicated dosing
regimens. Through a few case studies, the presenter will discuss the important issues related
to exposure-response analysis that are fit-for-purpose in nature. These topics include but are
not limited to: underlying physiology/mechanism, experimental design, selection of exposure
metric, proper statistical methodology andmodel qualification. The audience will also be given
a taste of how exposure-responsemodeling can be used to answer questions during the drug
development, or impact the labeling language.
BretMusser (Merck) – Dose-Response and Exposure-Response Analyses in Dose
Selection
Dose selection is themost critical decision in a clinical development program; that is, choosing
the right dose in a given patient population for a certain indication. Traditionally, dose selection
decisions have been driven by dose-response analyses, but more recently exposure-response
analyses have grown in prominence to both supplement and replace dose-response analyses.
Exposure-response analyses have beenwidely used to support new target populations (such as
pediatric or geriatric populations), dosing regimens, dosing forms and routes of administration,
and are increasingly used to support even the primary dose selection decisions. In this talk,
dose-response and exposure-response analyses will be compared and contrasted, and the strengths
and limitations of each will be discussed. Examples will be provided to illustrate themajor points.
Joint work with Ronda K Rippley (Merck)
Dalong Patrick Huang (FDA/CDER) – Some Statistical Issues in Exposure-Response
Modeling using Concentration-QTcData
The ICH E14 clinical guidance has been revised (December 2015) and now enables the use
of exposure response (ER) analysis applied to early phase clinical data to provide definitive
evidence of the lack of a QT effect of a drug in development. The overview of the revised guidance
will be provided. Wewill present statistical reviewers’ perspectives in concentration-QTcmodeling
based on FDA’s QT-interdisciplinary reviews and simulations.
DukeUniversity Department of Biostatistics and Bioinformatics 38
DISS2016, Durham, NC
POSTER SESSION - ABSTRACTS
Design and statistical analysis of method transfer studies for biotechnology products
Meiyu Shen (FDA), Lixin Xu (FDA), Yi Tsong (FDA), Juhong Liu (FDA)
During the biotechnology product development, a new analytical procedure regarding the
choice of analytical instrumentation andmethodology is often carefully selected based on the
intended purpose and scope of the analytical method. After an analytical method is successfully
validated and implemented, this method including the standard operating procedure (SOP)
will be followed during the life cycle of the product. The life cycle management of analytical
methods also includes but not limited to trend analyses onmethod performance at regular
intervals to evaluate the need to optimize the analytical procedure or to qualify even revalidate
all or a part of the analytical procedure, development and validation of a new or alternative
analytical method for a new impurity, and transferring a well-developed analytical method from
an original laboratory to a new contract laboratory or a new proposed production site. Method
transfer is a common practice during the life cycle management of analytical methods. Since the
analytical method to be transferred has been already thoroughly evaluated and fully validated
for its intended purpose at the original laboratory, themain purpose of method transfer studies
is usually through a qualification process to determine if the two laboratories providing comparable
results across the range of interest, and to assure themethod after transfer is still suitable
for its intended use. Inconsistent advice is often seen regarding testingmaterials, statistical
methods, and acceptance criteria in the literature. Furthermore, there is no detail on advising
the design and analyses for method transfer studies in the regulatory guidance fromUS Food
andDrug Administration (FDA) and other Agencies or authorities. Wewill propose a design and
a statistical equivalence analysis with a sample size dependent margin for themean comparison
of the proposed two laboratories (the original laboratory and the receiving one) in method
transfer studies in this poster based on our review experience.
Assessment of prescribability based on two one-sided probabilities
HuzhangMao (UTexas), Xiaoyu (Cassie) Dong (FDA), Yi Tsong (FDA)
Drug prescribability is defined as the choice for prescribing an appropriate drug product for a
new patient between a reference drug product and a test drug product. Biological drug products
aremade via living systems and are complex and variable in nature. Unlike generic versions
of small molecule drug products that contain the exact same structure as the innovative drug,
biological products can only be similar to reference product. Hence, average bioequivalence
between test and reference drug products need to be confirmed before prescribability can
be claimed. Prescribability is usually assessed through comparison of themarginal response
distributions between test and reference products by including the total variability. Using the
two one-sided tests approach developed byDong et al (2014), parallel arm design and 2 × 2
crossover design were investigated to assess prescribability in terms of type I error rate, statistical
DukeUniversity Department of Biostatistics and Bioinformatics 39
DISS2016, Durham, NC
power and sample size requirement. Our results indicated that both parallel arm design and 2 ×
2 crossover design can be used to assess prescribability, In addition, compared to parallel arm
design, crossover design is more efficient and thus requires smaller sample size by removing
the variance due to subject-by-sequence interaction, which is attributed to the fact that each
subject serves as his/her control in either sequence.
Dilemma on Conditional Power Based Sample Size Re-estimation
Xiaoyu Cai (GWU), Yi Tsong (FDA),Meiyu Shen (FDA), Yu-TingWeng (FDA)
Conditional powermethodwas widely discussed in the literature on adaptive un-blinded sample
size re-estimation (SSR) designs. The common application of conditional powermethod can be
divided into three categories: (a) Futility assessment (b) Decision on sample size adjustment and
(c) Decision on sample size increment. The conditional power basedmethods have advantages
of controlling type I error rate andmaintaining conditional power of the final test at a desired
level. However, likemost of other designs, it is not a universal optimal design, even among different
approaches of adaptive designs. Moreover, the performance of conditional powermethod
highly relies on some operational factors and sample distribution parameters such as the information
fraction, the true treatment and so on. The problem is, except the basic statistical operating
characteristics (type I error rate, power and so on), we have plenty of different criteria to compare
different designs but hardly identify amost important criteria. For example, wemay get better
estimation of conditional power but lost efficiency under the same information fraction. Several
problemsmaymake it a dilemmawhether or not it is worthwhile to use conditional power based
adaptive design in the clinical trials. In this study, wewill point out some of these potential
problems.
Percentile Estimation and Applications
Qi Xia (Temple), Yi Tsong (FDA), Yu-TingWeng (FDA)
Percentile is ubiquitous in statistics and plays a significant role in the day-to-day statistical
application. Not only it can be applied to screening and confirmatory cut-point determination
in immunogenicity assays but also the general percentile formulation enriches the statistical
literature for mean comparison between reference group and test group in bioequivalence or
biosimilarity studies, with the analytical biosimilarity evaluation and scaled average bioequivalence
as special cases. Shen et al. (2015) proposed and compared exact based approachwith some
approximated approaches in one sample scenario for cut-point determination. However, exact
based approach has the issue of computational time complexity. In this poster, we explored
more approximated approaches for percentile estimation such asMethod of Variance Estimates
Recovery (MOVER) based approaches andModified Large Sample (MLS) approaches. All these
approximated approaches are comparedwith exact based approach in one or two sample scenarios.
The applications and performance comparison for each approach are displayedwith numerical
results.
DukeUniversity Department of Biostatistics and Bioinformatics 40
DISS2016, Durham, NC
Statistical considerations regarding to correlated lots in analytical biosimilar equivalence
test
TiamhuaWang (FDA),Meiyu Shen (FDA), Qi Xia (Temple), Yi Tsong (FDA)
In the evaluation of the analytical similarity data, FDA has received some feedbacks on statistical
issues of potentially correlated reference lot values subjected to the equivalence testing since
independent and identical observations (lot values) from the proposed biosimilar product and
the reference product are assumed. In this work, we describe onemethod proposed by Yang et
al (2016) in biosimilar submissions for correcting the estimation bias of the reference variability
so as to increase the equivalencemargin and its modified versions for increasing the equivalence
margin and correcting standard errors in the confidence intervals assuming that the lot values
are correlated under a few known correlationmatrices. Our comparisons between these correcting
methods and no correction for bias in the reference variability under several assumed correlation
structures indicate that all correctingmethods would increase type I error rate dramatically
but only improve the power slightly for most of the simulated scenarios. Since the source of
a reference drug product lot is unknown in nature, correlation between lots is a design issue.
Hence, to obtain independent reference lot values by purchasing reference lots at a wide time
window often is a design remedy for correlated reference lot values.
An application ofmodel-fitting formarginal structural modeling in the context of an
observational cohort studywithmissing exposures
Hyang Kim (PAREXEL)
Assessing treatment effectiveness in longitudinal, observational data can be complex, because
treatments are not randomly assigned and patients can change treatment at any time at the
discretion of the investigator depending on changes in confounder. Hence, there are some
confounding of the effect of treatment by a time-varying variable which is affected by previous
exposure and can also subsequently influence treatment changes. Marginal structural models
(MSM) estimation employed in which the goal is to obtain coefficients to create weights so that
treatment exposure is not temporally confounded. However, missing in covariates/confounders
measurements are unavoidable with longitudinal, observational data and it is directly violated
theMSMmodeling assumptions. MSMpermits us to create a pseudo-population in which treatment
exposure is no longer temporally confounded. The pseudo-population can then be used in a
straightforwardmanner with an appropriate regression estimator to derive treatment effects.
A simulation study is conducted to examine bias in estimation of the effect of treatment on
recurrent event rates when ignoringmissing in covariates and further themodel assumptions
ofMSM.We demonstrate implementation of themethod in a longitudinal cohort study setting
where the effect of treatment is estimated on binary outcomes with both time-fixed and time-varying
potential confounders. Puttingmodel misspecification problem aside, simulation study shows
that strength of confoundingmay lead to bias in estimation regardless missing in covariates.
Modeling for assessing treatment effective in presence of time varying confounders andmissing
DukeUniversity Department of Biostatistics and Bioinformatics 41
DISS2016, Durham, NC
exposures on the confounders is a real challenge. A customized approach is required to estimate
causal treatment effects where time-varying confounders are affected by previous exposure
andmissing.
DukeUniversity Department of Biostatistics and Bioinformatics 42
DISS2016, Durham, NC
ORGANIZERS, SPEAKERS, INSTRUCTORS, POSTER PRESENTERS
Name Affiliation Name Affiliation
Amanda Redig DFCI Annie Lin FDA
BenGoldstein Duke Bohdana Ratitch Quintiles
BretMusser Merck Bruce Binkowitz Merck
Butch Tsiatis NCSU Chunlei Ke Amgen
Cliburn Chan Duke CraigMallinckrodt Lilly
Dalong Huang FDA Dan Sargent Mayo
Danyu Lin UNC DavanMehrotra Merck
DebbieMedlin Duke Donglin Zeng UNC
Elizabeth DeLong Duke Eric Groves Quintiles
Eric Perterson Duke Federico Innocenti UNC
Frank Rockhold Duke Fred Snikeris PAREXEL
Gary Koch UNC Gordon Lan JnJ
Ilya Lipkovich Quintiles Inna Perevozskaya Pfizer
James Love Boehringer-Ingelheim Jason Fine UNC
Jeremy Taylor Michigan Jichun Xie Duke
JingWang Glead John Bauman Quintiles
Joseph Cappelleri Pfizer Joseph Lucas Duke
Kay Tatsuoka BMS Ke Song Duke
KentWeinhold Duke Kerry Lee Duke
Kevin Liu JnJ Kouros Owzar Duke
Lisa LaVange FDA Lynn Lin PSU
Mark Chang Veristat Marlina Nasution Paraxel
Maura Stokes SAS Meiyu Shen FDA
NancyQi Duke OlgaMarchenko Quintiles
PralayMukhopadhyay Astrazeneca Qi Jiang Amgen
Radleigh Santos TPIMS Rajeshwari Sridhara FDA
Raluca Gordan Duke Renè Kubiak Boehringer-Ingelheim
Richard Cook Walterloo Richard Simon NIH
Robin Bliss Veristat SandeepMenon Pfizer
SharonMurray Paraxel Sharon Updike Duke
Shein-Chung Chow Duke Shiowjen Lee FDA
Sin-Ho Jung Duke Stephen George Duke
Steve Snapinn Amgen SumithraMandrekar Mayo
Susan Halabi Duke Suzanne Dahlberg DFCI
Duke University Department of Biostatistics and Bioinformatics 43
DISS2016, Durham, NC
Terry Hyslop Duke Terry Sosa Quintiles
Terry Therneau Mayo Walter Boyle Sutherland Healthcare
XiaofeiWang Duke XiaoyuDong FDA
Yaming Hang Biogen Yeh-Fong Chen FDA
Yi Tsong FDA Yijing Shen Genentech
YuanWu Duke
Duke University Department of Biostatistics and Bioinformatics 44
DISS2016, Durham, NC
ATTENDEE LIST
Name Affiliation Name Affiliation
John Bauman Quintiles Anna Bellach Copenhagen
Gerald Belton NCSU Parul Bhargava Shire
Bruce Binkowitz Merck Robin Bliss Veristat
Walter Boyle Sutherland AILI CHENG Pfizer
Joseph Cappelleri Pfizer Cliburn Chan Duke
Stephen Chang Pharmacycl Ling Chen FDA
Mei Chen Medpace Yeh-Fong Chen FDA
Shein-Chun Chow Duke Richard Cook Waterloo
Mary Cooter Duke Doug Criger PPD
SuzanneDahlberg DFCI QianyuDang FDA
Sargent Daniel Mayo XiaoyuDong FDA
Dexiang Gao Colorado ConwayGee Chiltern
Stephen George Duke BenGoldstein Duke
Raluca Gordan Duke Lin Gu Duke
SusanHalabi Duke Yaming Hang Biogen
JimHerndon Duke ShuyenHo PAREXEL
ChuyunHuang PAREXEL Dalong Huang FDA
WayneHuggins RTI FEDERICO INNOCENTI UNC
James Imus Quintiles Qi Jiang Amgen
Teri Jimenez PAREXEL Sin-Ho Jung Duke
Natalia Kan-Dobrosky PPD Chunlei Ke Amgen
Wang Kehui PPD Brian Kilgallen UCB
Shiowjen Lee FDA Lynn Lin PSU
Min Lin FDA Ilya Lipkovich Quintiles
Kevin Liu Janssen Lan Liu INC Resear
Shubin Liu PPD James Love Boehringer
Joseph Lucas Duke CraigMallinckrodt Eli Lilly
SumithraMandrekar Mayo OlgaMarchenko Quintiles
MalickMbodj FDA CynthiaMcShea UCB
DebbieMedlin Duke DevanMehrotra Merck
SandeepMenon Pfizer ZhuangMiao FDA
PralayMukhopadhyay AstraZenec SharonMurray PAREXEL
BretMusser Merck Marlina Nasution PAREXEL
Tatiana Nevmyrych PPD Wei Pan Duke
Duke University Department of Biostatistics and Bioinformatics 45
DISS2016, Durham, NC
Inna Perevozskaya Pfizer Matthew Phelan Duke
XINYUEQI Duke Bohdana Ratitch Quintiles
Amanda Redig DFCI Frank Rockhold Duke
Kingshuk Roy Choudhur Duke Christel Rushing Duke
Radleigh Santos TPIMS Justin (Zo Shang PAREXEL
Lynn Shemanski CRAB Meiyu Shen FDA
Yijing Shen Genentech Qing Shi PAREXEL
Richard Simon NIH Steven Snapinn Amgen
Fred Snikeris PAREXEL Terry Sosa Quintiles
Rajeshwari Sridhara FDA Patricia Stephenson Rho
Thais Talarico Duke Jiali Tang PPD
Jeremy Taylor Umichigan Therneau Terry Mayo
Samantha Thomas Duke Tracy Truong Duke
Anastasios Tsiatis NCSU Yi Tsong FDA
Fabiana Vazquez Duke JINGWANG Gilead
YU-TINGWENG FDA GuanfangWang Yahoo
LiweiWang PPD XiaofeiWang Duke
KentWeinhold Duke YuanWu Duke
Jichun Xie Duke Qing Yang Duke
Lisa Ying BMS JIAYIN ZHENG Duke
Donglin Zeng UNC eric groves Quintiles
frances wang Duke tianhuawang FDA
DukeUniversity Department of Biostatistics and Bioinformatics 46
DISS2016, Durham, NC
FLOORPLAN
The symposiumwill be held atMillenniumHotel Durham, 2880 CampusWalk Avenue, Durham,
NC., 27705. MillenniumHotel Durham is minutes away fromDukeUniversity Hospital and
Duke University and an easy drive to downtownDurham.
OPENINGREMARKS, KEYNOTE SPEECH, PARALLEL SESSIONS,
POSTER SESSION, SOCIALMIXER
SHORTCOURSES, PARALLEL SESSIONS
Duke University Department of Biostatistics and Bioinformatics 47
DISS2016, Durham, NC
Department of Biostatistics and Bioinformatics
Duke University School ofMedicine
2424 Erwin Road, Suite 1102
Durham, NC 27710
U.S.A.
Duke University Department of Biostatistics and Bioinformatics 48