BOOK OF ABSTRACTS - Τμήμα Στατιστικήςemribs/page/EMR2011/book of abstracts... ·...

of 122 /122

Embed Size (px)

Transcript of BOOK OF ABSTRACTS - Τμήμα Στατιστικήςemribs/page/EMR2011/book of abstracts... ·...

6TH CONFERENCE OF THE EASTERNMEDITERRANEAN REGION OF THE

INTERNATIONAL BIOMETRIC SOCIETY

BOOK OF ABSTRACTS

HERSONISSOS - CRETE8-12 MAY 2011

BOOK OF ABSTRACTS

SIXTH CONFERENCE OF THE EASTERN

MEDITERRANEAN REGION OF THE

INTERNATIONAL BIOMETRIC SOCIETY

(EMR-IBS)

HERSONISSOS, CRETE, 8-12 MAY 2011

Preface

We are very glad to welcome you at Hersonissos, Crete at the heart of the Mediter-ranean sea, for the 6th conference of the Eastern Mediterranean Region of the Inter-national Biometric Society (EMR-IBS).

The 6th EMR-IBS conference coincides with the 10-year anniversary of the forma-tion of the EMR region of IBS, which includes the following countries: Cyprus, Egypt,Greece, Israel, Jordan, Palestinian National Authority, Turkey, Saudi Arabia and re-cently, Bulgaria. This is the last of a very successful series of conferences started inAthens, Greece in 2001, followed by Antalya, Turkey in 2003, Corfu, Greece in 2005,Eilat, Israel in 2007 and nally Istanbul, Turkey in 2009.

The conference is dedicated to the memory of Steve Lagakos who, in his eorts ofpromoting Biostatistical Science in the Region, has been a great supporter of EMR,participating actively in all of the Regional conferences. In 2005, he gave a full-dayworkshop on Clinical Trials to medical professionals to provide the funding to sup-port student participation in the 3rd EMR conference. He had the vision of makingthe region an educational center for young Biostatisticians and Clinical Trialists. Hecollaborated with Postgraduate Studies Programs where he served as a Visiting In-structor for many years and he co-founded in 2007, Frontier Science Foundation-Hellas(FSF-H), in Athens, Greece where he served as a member of its Board of Directors.

Frontier Science & Technology Research Foundation (FSTRF) sponsors the La-gakos Memorial Lectures, with invited talks by world renowned statisticians to bepresented throughout the meeting. Frontier Science Foundation-Hellas (FSF- H) issupporting the participation of young researchers in his memory, providing 3 studentawards to students with the best submitted abstracts. These talks will be presentedat a special session on the last day of the conference.

A complete list of all the abstracts of the papers to be presented in the conferencecan be found in this book. A detailed index of all authors can be found at the end tofacilitate easy search.

We hope that you will enjoy the 6th EMR-IBS Conference honoring the memoryof Steve Lagakos and celebrating the 10 year anniversary of EMR.

Ranny DafniDimitris KarlisStergios Tzortzioson behalf of the LOC and SC.

3

Scientic Committee

ChairsMarvin Zelen Harvard University, USAUrania Dafni University of Athens, Greece

MembersChristel Faes Hasselt University, BelgiumCostas Fokianos University of CyprusLaurence Freedman Bar Ilan University, IsraelConstantine Gatsonis Brown University, USALupe Gomez Universitat Politecnica de Catalunya, SpainErgun Karaagaoglu Hacettepe University, TurkeyDimitris Karlis Athens University of Economics, GreeceKlea Katsouyanni Medical School, University of Athens, GreeceMarkos Koutras University of Pireaus, GreeceGeert Molenberghs University of Hasselt, BelgiumIoannis Ntzoufras Athens University of Economics, GreeceSharon-Lise Normand Harvard University, USABenjamin Reiser University of Haifa, IsraelStergios Tzortzios University of Thessaly, GreeceMaria-Grazia Valsecchi Prevenzione e Biotecnologie Sanitarie, ItalyConstantin Yiannoutsos Indiana University, USADavid Zucker University of Haifa, Israel

Local Organizing Committee

ChairStergios Tzortzios University of Thessaly

Vice ChairDimitris Karlis Athens University of Economics

MembersGregory Chlouverakis University of CreteUrania Dafni University of AthensNikolaos Demiris Agricultural University of AthensVissarion Gousios University of ThessalyJoanna Moschandreas University of CreteChristos Nakas University of ThessalyIoannis Ntzoufras Athens University of EconomicsDemosthenes Panagiotakos Harokopio University of AthensAris Perperoglou University of East AngliaGiota Touloumi University of Athens

4

Organizedby Eastern Mediterranean Region (EMR) International Biometric Society (IBS)

Coorganizers Frontier Science & Technology Research Foundation (FSTRF) Frontier Science Foundation-Hellas (FSF-H) Laboratory of Biostatistics, University of Athens

Athens University of Economics and Business University of Thessaly

ConferenceSponsors Eastern Mediterranean Region (EMR) International Biometric Society (IBS) Frontier Science & Technology Research Foundation (FSTRF) Frontier Science Foundation-Hellas (FSF-H)

Contents

One Man's Meat: Nutrition Biomarkers in Chronic DiseasePrevention ResearchLaurence Freedman . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1Clinical Trials and the Growth of RegulationsDavid L. DeMets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3Confronting the Challenges of Subgroup Analyses in Clinical TrialsRichard D. Gelber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4Building Global Capacity in StatisticsRonald E. LaPorte, Ph.D. for the Supercourse team . . . . . . . . . . . . . . . . . . . . . . . . 5An Overview of Screening Programs for the Early Diagnosis ofChronic Diseases: Issues and ProblemsSandra Lee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6Inference on Treatment Eects from a Randomized Clinical Trialin the Presence of Premature Treatment Discontinuation: TheSYNERGY TrialAnastasios A. Tsiatis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7Augmented Cross-Sectional Studies with Abbreviated Follow-up forEstimating HIV IncidenceBrian Claggett, Stephen W. Lagakos, Rui Wang . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9Nonparametric Estimation of Distribution Functions in MeasurementErrors ModelsItai Dattner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10Modelling Health Surveillance Data via a Bivariate INAR(1) ProcessXanthi Pedeli, Dimitris Karlis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11Dependence Calibration in Conditional Copulas: A NonparametricApproachElif F. Acar, Radu V. Craiu, Fang Yao . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Unbalanced and Partial Group Sequential Methods for NormalResponses in Clinical TrialsAtanu Biswas, Shirsendu Mukherjee, KyungMann Kim . . . . . . . . . . . . . . . . . . . . 14Improved Estimation of Survival Probabilitiesfrom Two-phase Stratied SamplesNorman E. Breslow, Thomas Lumley, Jon A. Wellner . . . . . . . . . . . . . . . . . . . . . . 15More Robust Doubly Robust EstimatorsMarie Davidian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16Order Restricted Inference for Multivariate Binary Data withApplicationsOri Davidov, Shyamal Peddada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17k-FWER Control without p-value Adjustment, with Application toDetection of Genetic Determinants of Multiple Sclerosis in ItalianTwinsLivio Finos, Alessio Farcomeni . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18Comparative Eectiveness Research: A Methodologic IntroductionConstantine Gatsonis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19Use of Composite Endpoints in Cardiovascular Device Trialsinvolving Coronary StentsGuadalupe Gmez, Urania Dafni, Moises Gmez . . . . . . . . . . . . . . . . . . . . . . . . . . . 20Treatment Noncompliance in Studies ofAdjuvant Chemotherapy for Breast CancerRobert J. Gray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21Estimation of Sensitivity of Chest X-ray and Cancer PreclinicalSojourn Time for Lung Cancer Screening TrialsPing Hu, Philip Prorok . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22Inference under Biased Sampling with Application to Infection DataMicha Mandel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23Issues in ROC Surface Analysis with an Application to ExternallyValidated Cognition in Parkinson Disease ScreeningC.T. Nakas, T.A. Alonzo, J.C. Dalrymple-Alford, T.J. Anderson . . . . . . . . . . . . . 24Formulation of Recommendation Domains for Sugarcane Varieties:Using Modied Stability Analysis and Best Linear UnbiasedPredictorNjuho, P.M., Sewpersad, C.N., Redshaw K A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25Massively Parallel Nonparametrics in NeuroimagingPhilip T. Reiss, Lei Huang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Statistics in a Translational Neuroscience ProgramAllan R. Sampson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

ii

Generation Times in Epidemic ModelsGianpaolo Scalia Tomba . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28Real Time Prediction for Dengue Epidemic in Havana (2001) UsingModel Averaging MethodsZiv Shkedy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29A Measure of Explained Variation for Event History DataJanez Stare, Maja Pohar Perme, Robin Henderson . . . . . . . . . . . . . . . . . . . . . . . . . 30Modeling Reductions in Breast Cancer MortalityEster Vilaprinyo, Montserrat Rue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31Accounting for Prediction Error in Environmental Exposure whenRelating Public Health to Environmental FactorsLinda J. Young, Kenneth K. Lopiano, Carol A. Gotway . . . . . . . . . . . . . . . . . . . . . 32A Regularization/Extrapolation Corrected Score Method forNonlinear Regression Models with Covariate ErrorDavid M. Zucker, Malka Gorne, Yi Li, Donna Spiegelman . . . . . . . . . . . . . . . . . 33Modelling the Mean and Covariance Structure for ContinuousBounded Longitudinal DataMouna Akacha, Jane L. Hutton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35Estimates of Clinically Useful Measures in Survival AnalysisFederico Ambrogi, Elia Biganzoli, Patrizia Boracchi . . . . . . . . . . . . . . . . . . . . . . . 36Penalized Likelihood Methodology and ApplicationsE. Androulakis, C. Koukouvinos, F. Vonta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37Performance of Markers for Censored Failure Time Outcome:Nonparametric Approach Based on ProportionsLaura Antolini . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38Forecasting Longitudinal Multivariate Binary DataOzgur Asar, Ozlem Ilk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39Eects of Covariate Omission when Fitting the Fine-Gray Model toData from Randomized Controlled TrialsGiorgos Bakoyannis, Giota Touloumi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40Modeling the Non-Inherited Maternal Antigens Eect in Multi-CaseFamiliesB. Balliu, R. Tsonaka, D. van der Woude, S. Bohringer, J.J. Houwing-Duistermaat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41Hierarchical Testing of Subsets of HypothesesYoav Benjamini, Marina Bogomolov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42Doubly Robust and Multiple Imputation Based GeneralizedEstimating EquationsTeshome Birhanu, Geert Molenberghs, Cristina Sotto, Michael G. Kenward . . . 43

iii

The Inuence of Rotation Type on the Repeatability of DietaryPatterns Derived through Principal Component AnalysisVassiliki Bountziouka, Demosthenes B. Panagiotakos . . . . . . . . . . . . . . . . . . . . . . . 44Analysis of Multirater Ordinal Data: An IRT ApplicationRek Burgut, Yasar Sertdemir, Ilker Unal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45Adaptive Policies for Sequential Sampling under IncompleteInformation and Side ConstraintsApostolos Burnetas, Odysseas Kanavetas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46Comparison of MB-MDR to BOOST and RAPID for DetectingEpistasis in UnrelatedsT. Cattaert, F. Van Lishout, J. Rial, K. Van Steen . . . . . . . . . . . . . . . . . . . . . . . . . 47Modelling Occupancy-Abundance Patterns in Supra-Specic Taxaof Soil Invertebrates from Zambia and IndiaLegesse Kassa Debusho, G. Sileshi, Mujeeb Rahman . . . . . . . . . . . . . . . . . . . . . . . . 48Herd-Prevalence based on Aggregate Testing of AnimalsChristel Faes, Marc Aerts, Saskia Litiere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49On a Poisson Mixture Model for Count Time SeriesKonstantinos Fokianos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50Trends in Mammographic Breast Density and Risk of Breast CancerCarles Forn, Marisa Bar, Nria Tor,, Montserrat Ru . . . . . . . . . . . . . . . . . . . 51The Presence of the Absence in a Geriatric Cohort: FunctionalDecline Curve Accounting for AttritionGeva, D, Shahar, D.R., Harris, T.B., Researchers of Health ABC, Friger, M. . . 52A Classication of the Main Biometrical Methodologies Applied inAgricultural Experimentation and ResearchV. Gousios, S. Tzortzios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53Measuring Follow-up Completeness inSurvival StudiesErika Graf, Elmar Abelein, Martin Schumacher . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54Statistical Methods for Evaluating Prognostic Features of SingleNucleotide Polymorphisms (SNPs) in Critical Signal Pathways forRenal Cell Carcinoma (RCC)P. Kathryn Gray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55Longitudinal CART and its Application in Neuroimaging StudiesJaroslaw Harezlak, Madan G. Kundu, Constantin T. Yiannoutsos . . . . . . . . . . . . 56Marginalized Models for BivariateLongitudinal Binary DataOzlem Ilk, Michael J. Daniels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

iv

An Improved Cusum Procedure for Detectionof Outbreaks in Poisson Distributed Health EventsRobert Jonsson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58Bayesian Semiparametric Modeling for Nonignorably MissingCovariatesZeynep I. Kalayloglu, Olcay ztrk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59Importance of Hazard Functions for Lifetime DataPinar Gunel Karadeniz, Nural Bekiroglu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60Quality Adjusted Survival Analysis of Postmenopausal BreastCancerSevilay Karahan, Osman Saracbasi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61A Graphical Approach for Adaptive Clinical Trials Testing MultipleHypothesesFlorian Klinglmueller, Martin Posch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62Assessing Non-Inferiority in a Gold Standard Design for Retentionof Eect Hypotheses - A Semiparametric ApproachKarola Kombrink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63Robust Estimation in Cox Proportional Hazards ModelHande Konsuk, Meral Cetin, Oniz Toktamis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64Variable Selection and Computation of the Prior Probability of aModelChristos Koukouvinos, Christina Parpoula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65Cross-Validation Prior Choice in Bayesian Probit Regression withMany CovariatesDemetris Lamnisos, Jim E. Grin, Mark F.J. Steel . . . . . . . . . . . . . . . . . . . . . . . . 66A Longitudinal Study of the Eect of Physical Activity on Risk ofHeart Failures: Methods and ApplicationsDaniela Mariosa, Rino Bellocco, Ylva Lagerros, Olof Nyrn, Weimin Ye,Johan Sundstrm, Hans-Olov Adami, Erik Ingelsson . . . . . . . . . . . . . . . . . . . . . . . . 67Handling Missing Data: A Strategy Based on Multiple Imputationand Bayesian AnalysisSebastien Marque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68Meta Analysis for Summarizing Results of Simulation Studies:One-Way ANOVA and its Some Alternatives CasesMehmet Mendes, Soner Yigit Erkut Akkartal, Amit Mirtagioglu . . . . . . . . . . . . . . 69Calibration Comparison of Instruments for Measuring Particle-MassConcentrations, in the Presence of Random EectsS. D. Oman, B. Maoz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70A Simulation Study Based on Nonlinear Mixed Eects ModelsMehmet N. Orman, Selcuk Korkmaz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

v

Utilizing the Clinical DataFax System from Randomization to theCompletion of a Clinical TrialZekai Otles, Xiang Li, Diane Pauk, KyungMann Kim . . . . . . . . . . . . . . . . . . . . . . 72Nonparametric Regression of Mean and Covariance Structures forLongitudinal DataJianxin Pan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73Censoring Biomarker Measurements due to Treatment Initiation:Ignorable or Not?Nikos Pantazis, Giota Touloumi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74Optimal Variable Selection for Regression ModelsSusana Prez-lvarez, Christian Brander, Guadalupe Gmez . . . . . . . . . . . . . . . . 75Cost-Eectiveness of HIV Tropism Testing to Inform AntiretroviralTreatment with MaravirocN. Prez-lvarez, G. Gmez, R. Paredes, B. Clotet . . . . . . . . . . . . . . . . . . . . . . . . . 76Semi-Competing Risks with Interval-Censored Intermediate EventNria Porta, M.Luz Calle, Guadalupe Gmez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77Reduced Rank Hazards Regression with Fixed and Time VaryingEects of the CovariatesAris Perperoglou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78A Nonparametric Approach for Estimating the Survival Functionsfrom Case-Control Family DataNadia Pogrebinsky, Malka Gorne and Li Hsu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79Using Scan Statistics on Multiple Processes with Dependent Signalsand Assessing its Distribution, with Application to Sequence SearchAlong the GenomeAnat Reiner-Benaim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80Building Maintenance: A Time-to-Event Approach and a SimulationStudyCarles Serrat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81The Use of Multi-State Models in the Analysis of Semi-CompetingRisks DataFotios Siannis, Jessica Barrett, Vern Farewell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82Statistical Properties of Heterogeneity Measures in Meta-AnalysisBahi Takkouche, Polyna Khudyakov, Julian Costa-Bouzas, Donna Spiegelman . 83Inference in Generalized Linear Regression Models with a CensoredCovariateJohn V. Tsimikas, Leonidas E. Bantis, Stelios D. Georgiou . . . . . . . . . . . . . . . . . . 84

vi

An Ecient Algorithm to Perform Multiple Testing in EpistasisScreeningFrancois Van Lishout, Tom Cattaert, Jestinah M. Mahachie John, LouisWehenkel, Kristel Van Steen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85Marginal Structural ModelsUnder Dierent Mechanisms of MissingnessGeorgia Vourli, Giota Touloumi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86Spatio-Temporal Analysis of Breast Cancer Mortality RisksUgarte, M.D., Goicoa, T., Etxeberria, J., Militino, A.F., Polln, M . . . . . . . . . . 87Marginal Distribution Estimation from Double-Sampled CompetingRisks DataConstantin T. Yiannoutsos, Menggang Yu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88Estimator for the Correlation of Recurrent Events in Comparison tothe Wei Lin and Weissfeld MethodZakiyah Zain, John Whitehead . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89Bayesian Sample Size Estimates for One Sample Test in ClinicalTrials with Dichotomous OutcomesBoris Zaslavsky . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90Analytic Approaches for Eye-Specic Outcomes: One Eye or Two?Anna Karakosta, Maria Vassilaki, Sotiris Plainis, Nazik Hag Elfadl, MiltiadisTsilimbaris, Joanna Moschandreas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92Confounding Techniques in Experimental Design: Results of anExperiment on Aniba Rosaeodora in the Central AmazonTeresa A.Oliveira, Roberval M. B. Lima, Amlcar Oliveira . . . . . . . . . . . . . . . . . . 93Visualization in Joint Regression Analysis Using RAmlcar Oliveira, Teresa A. Oliveira . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94A GEE Approach for Poisson Correlated DataM.C. Pardo, R. Alonso . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95Spatial-Mathematic Methods for Analysis of Indicators of MortalityGeorgia Pistolla, Poulikos Prastakos, Maria Vassilaki, Anastas Philalithis . . . . 96Adaptive Hypothesis Multi-stage Phase IIDesign for Time-to-Event EndpointStavroula Poulopoulou, Dimitris Karlis, Constantin T. Yiannoutsos, UraniaDafni . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97Estimation of Catalan Breast Cancer Survival Functions, Correctedfor Lead Time and Length BiasAlbert Roso, Carles Forne, Montserrat Rue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

vii

Molecular and Epidemiological Characterisation of HIV- 1 InfectionNetworks Involving Transmitted Drug Resistance Mutations inNorthern GreeceSkoura L, Haidich AB, Metallidis S, Buckton AJ, Mbisa JL, Pilalas D,Papadimitriou E, Papoutsi A, Valagouti D, Tsachouridou O, Antoniadou ZA,Kollaras P, Nikolaidis P, Arvanitidou M, Malisiovas N . . . . . . . . . . . . . . . . . . . . . 99Modelling Long-Term Trends in Cancer Mortality in Greece UsingJoinpoint RegressionSpyridonidou Christina, Samoli Evi, Stasinopoulos M., Touloumi Giota . . . . . . . 100Development and Evaluation of an Entropy Index as a GaitVariability Measure in Orthopaedic PatientsGN Tzagarakis, SD Tsivgoulis, NA Kampanis, PG Katonis, GI Chlouverakis . . 101Eects of Antibiotics on MRSA Carriage DynamicsEleni Verykouki, Philip D. O'Neill and Theodore Kypraios . . . . . . . . . . . . . . . . . . 102Testing for a Changepoint in the Cox Survival Regression ModelDavid M. Zucker, Sarit Agami, Donna Spiegelman . . . . . . . . . . . . . . . . . . . . . . . . . 103Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

viii

Keynote Lecture

ix

One Man's Meat: Nutrition Biomarkers in ChronicDisease Prevention Research

Laurence FreedmanGertner Institute for Epidemiology and Health Policy Research, Sheba Medical Center, TelHashomer 52621, Israel [email protected]

Abstract. Variation between individuals both in their dietary intake and in their biologicalresponses to such diet makes nutritional epidemiology a fascinating topic. However, dicul-ties in accurately quantifying dietary intake through self-reporting make it dicult to studyhypothesized associations between diet and disease. One response to these challenges hasbeen the emergence of biomarkers for dietary intake. I review this development and the usesof such biomarkers in chronic disease prevention research. I focus on several dierent rolesthat biomarkers can play in this endeavor, including assessing compliance and mediationanalysis in prevention trials, calibration of self-report instruments, adjustment of estimatedrelative risks for dietary measurement error in observational studies, and enhancing the sta-tistical power of such studies. Examples are provided from the Women's Health Initiativedietary intervention trial, the Observing Protein and Energy study and the Carotenoids inEye Disease Study.

Keywords

biomarkers, cohort studies, nutrition, measurement error, prevention trials

References

Freedman LS, Kipnis V, Schatzkin A, Tasevska N, Potischman N. Can we use biomarkersin combination with self-reports to strengthen the analysis of nutritional epidemiologicstudies? Epidemiologic Perspectives & Innovations 2010, 7:2.

Freedman LS, Tasevska N, Kipnis V, Schatzkin A, Mares J, Tinker L, Potsichman N. Gains instatistical power from using a dietary biomarker in combination with self-reported intaketo strengthen the analysis of a diet-disease relationship - an example from CAREDS.Am J Epidemiol 2010; 172:836-842.

Prentice RL, Shaw PA, Bingham SA, et al. Biomarker-calibrated Energy and Protein Con-sumption and Increased Cancer Risk Among Postmenopausal Women. Am J Epidemiol.2009;169(8):977-989.

1

Lagakos Memorial Lectures

2

Clinical Trials and the Growth of Regulations

David L. DeMetsDepartment of Biostatistics and Medical Informatics University of Wisconsin-Madison,USA

Abstract. The randomized clinical trial has been the primary research method for evaluat-ing new drugs, biologics, devices, procedures and behavioral modications. Industry conductsmany randomized clinical trials to evaluate their new products or novel application of ex-isting products. These products must be approved for clinical use by regulatory agencies ofcountries in which the product will be marketed. The regulatory bodies have produced reg-ulations that guide the design, conduct, analysis and presentation of these trials, especiallytrials which are going to be used for the registration or approval process. There is no singleset of guidelines but among them are those of the US Food and Drug Administration and theInternational Conference on Harmonization (ICH). Trials are increasingly being conductedworldwide and so requirements may dier across countries or regions. While the principles ofthese guidance documents are generally consistent with fundamentals of clinical trials, theirinterpretation and application have become more complex, time consuming and expensiveover the past decade. We will discuss some of these practices and whether they lead to bettertrials. Recent experience suggests that the current trends cannot be sustained.

Plenary Talk

3

Confronting the Challenges of Subgroup Analysesin Clinical Trials

Richard D. GelberDana-Farber Cancer Institute, Harvard School of Public Health, Harvard Medical Schooland Frontier Science and Technology Research Foundation, Boston, MA, USA

Abstract. Proper conduct and interpretation of subgroup analyses in clinical trials arechallenging.[1] Such analyses are subject to unacceptably high levels of false-positive andfalse-negative error rates. If improperly interpreted, they can lead to erroneous conclusionsthat can have substantial detrimental eects on patient care, either denying eective treat-ment to some patients, or encouraging use of ineective treatments for others. Steve Lagakospublicized these concerns, but also recognized the importance of subgroup analyses to pro-vide guidance for care of individual patients and to generate hypotheses to be tested infuture clinical investigations. He wrote in 2006, [2] "When subgroup analyses are properlyconducted, presentation of their results can be informative, . . . avoiding any presentation ofsubgroup analyses because of their history of being over interpreted is a steep price to payfor a problem that can be remedied by more responsible analysis and reporting."

In this presentation I will review challenges of subgroup analyses using examples fromclinical trials evaluating adjuvant therapies for breast cancer. Steve Lagakos served as amember of the Data and Safety Monitoring Committees for these trials and his wise counselcontributed substantially to their successful and ethical completion. Interpretation of sub-group analyses based on dichotomizing a continuous measure such as age or a biomarkerrepresents a specic challenge. The Subpopulation Treatment Eect Pattern Plot (STEPP)method was developed to explore treatment eect heterogeneity as a function of a covari-ate measured on a continuous scale.[3-6] STEPP has been applied in a variety of settingsto highlight patterns of treatment eect dierences based on the biological diversity of theunderlying disease process and therapeutic response.[7,8] Examples will be presented tostimulate discussion.

References[1] Wang R, Lagakos SW, Ware JH, Hunter DJ, Drazen JM. Statistics in medicine - reporting

of subgroup analyses in clinical trials. N Engl J Med 2007; 357:2189-94.[2] Lagakos SW. The challenge of subgroup analyses{reporting without distorting. N Engl J

Med 2006; 354:1667-9.[3] Bonetti M, Gelber RD. A graphical method to assess treatment-covariate interactions

using the Cox model on subsets of the data. Stat Med 2000; 19:2595-609.[4] Bonetti M, Gelber RD. Patterns of treatment eects in subsets of patients in clinical

trials. Biostatistics 2004; 5:465-81.[5] Bonetti M, Zahrieh D, Cole BF, Gelber RD. A small sample study of the STEPP approach

to assessing treatment-covariate interactions in survival data. Stat Med 2009; 28:1255-68.

[6] Lazar AA, Cole BF, Bonetti M, Gelber RD. Evaluation of treatment-eect heterogene-ity using biomarkers measured on a continuous scale: Subpopulation Treatment EectPattern Plot (STEPP). J Clin Oncol 2010; 28: 4539-44.

4

Building Global Capacity in Statistics

Ronald E. LaPorte, Ph.D. for the Supercourse teamDepartment of Epidemiology, University of Pittsburgh, USA [email protected]

Abstract. World wide there are over 300 times more clinicians than there are statistics.There are 20 times more morticians there are those trained in statistics. In many developingcountries there are no statistical training programs. This is at a time when there has beenan explosion of data. There is a critical need for more individuals to be trained in statisticsin developed and developing countries alike. Ideally we would like to build more masters andPh.D. programs in developing countries. However, this is not practical as there typically isnot the expertise for teaching, and in these days the costs are prohibitive. We have takena dierent approach for work for development in statistics and other areas. Our goal is notto produce Ph.D.s in statistics, but rather to build awareness and interest in statistics forstudents. We want to double the training in statistics world wide in the next 5 years.

In most medical, and nursing schools world wide in 4-6 year training students might have15 minutes training in statistics. The reason that there is a paucity is that few faculty canteach about statistics. We are changing this. Our approach is simple, we have a network ofover 50,000 faculty interested in global health and prevention from 174 countries. From thisnetwork we have collected 4800 top quality lectures, 75 from Nobel Prize winners. We makethe lectures available in a free open source library (www.pitt.edu/ super1/). We feed thelectures back to the faculty and the world, and they are then able to teach in areas thatmay not be their primary areas of expertise, as they have top quality, up to date lectures.We have already doubled the training of global health in the world. There are 31 dierentlanguages represented in the supercourse. In the past year our lectures have taught over 6million people. We have distributed these lectures to all medical, public health and nursingschools in the world.

We propose to build a statistical supercourse where we collect the top lectures of statisticsand make these available for free. We currently have about 30 statistical and research designlectures available. We give all the credit for the lectures to the authors, as we are like an artgallery, and the lectures are like the art. A statisical Supercourse would help build awarenessof statistics to many 1000s more students than are available now.

Panel Discussion

5

An Overview of Screening Programs for the EarlyDiagnosis of Chronic Diseases: Issues and Problems

Sandra LeeDana-Farber Cancer Institute and Harvard School of Public Health, Department ofBiostatistics and Computational Biology

Abstract. The screening of asymptomatic individuals for chronic disease is a public healthinitiative that is rapidly growing. This is especially true in cancer where there are expandingearly detection programs in breast, cervical, colorectal, lung, prostate and stomach cancers.The basic idea motivating the screening asymptomatic populations is that diagnosing thedisease early before it becomes symptomatic may result in better prognosis. We have de-veloped stochastic models to describe the early detection process. Our model can projectvarious outcomes of screening programs, including the mortality reduction associated withscreening programs. Our theoretical results have been applied to the problems arise in breastcancer screening. In this talk, the stochastic model and model assumptions will be presented.In addition, the following topics and examples will be discussed: i) periodic vs. risk-basedscreening, ii) mammogram screening in younger women (ages between 40-49), iii) overdiagno-sis from breast cancer screening, iv) contribution of mammogram screening in the US breastcancer mortality reduction between 1975-2000 and v) collaborations with the U.S. Preven-tive Services Task Force in updating the U.S. mammogram guideline in 2009. The last twoexamples are noteworthy as the modeling result had a signicant impact in understandingthe role of breast cancer screening in the U.S. population.

6

Inference on Treatment Eects from a RandomizedClinical Trial in the Presence of PrematureTreatment Discontinuation: The SYNERGY Trial

Anastasios A. TsiatisDepartment of Statistics, North Carolina State University, Raleigh, NC, [email protected]

Abstract. The Superior Yield of the New Strategy of Enoxaparin, Revascularization, andGlYcoprotein IIb/IIIa inhibitors (SYNERGY) trial was a randomized, open-label, multi-center clinical trial comparing two anticoagulant drugs (enoxaparin and unfractionated hep-arin, UFH) on the basis of various time-to-event endpoints. In contrast to those of otherstudies of these agents, the primary, intent-to-treat analysis did not nd sucient evidenceof a dierence, leading to speculation that premature discontinuation of the study agentsby some subjects might have attenuated the treatment eect. As is of the case in such tri-als, some subjects discontinued (stopped or switched) their assigned treatment prematurely,either because occurrence of an adverse event or other condition under which discontinu-ation was mandated by the protocol or due to other reasons, e.g., switching to the othertreatment at his/her provider's discretion (with more subjects switching from enoxaparin toUFH than vice versa). In this situation, interest often focuses on \the dierence in survivaldistributions had no subject discontinued his/her assigned treatment," inference on whichis often attempted via standard analyses where event/censoring times for subjects discon-tinuing assigned treatment are articially censored at the time of discontinuation. However,this and other common ad hoc approaches may not yield reliable information because theyare not based on a formal denition of the treatment eect of interest. We use SYNERGYas a context in which to describe how such an eect may be conceptualized properly and topresent a statistical framework in which it may be identied, which leads naturally to theuse of inverse probability weighted methods. This is joint work with Min Zhang (Universityof Michigan), Marie Davidian (North Carolina State University), and Karen Pieper and KenMahaey (Duke Clinical Research Institute)

Keywords

inverse probability weighting, potential outcomes, proportional hazards model

7

Student Awards

8

Augmented Cross-Sectional Studies withAbbreviated Follow-up for Estimating HIVIncidence

Brian Claggett1, Stephen W. Lagakos and Rui Wang21 Department of Biostatistics, Harvard University School of Public Health,655 Huntington Avenue, Boston, Massachusetts 02115, [email protected]

2 Department of Biostatistics, Harvard University School of Public Health,655 Huntington Avenue, Boston, Massachusetts 02115, U.S.A. [email protected]

Abstract. Cross-sectional HIV incidence estimation based on a sensitive and less-sensitivetest oers great advantages over the traditional cohort study. However, its use has beenlimited due to concerns about the false negative rate, reecting the phenomenon that somesubjects may remain negative permanently on the less-sensitive test. Wang and Lagakos(2010) propose an augmented cross-sectional design which provides one way to estimatethe false-negative rate and subsequently incorporate this information in the cross-sectionalincidence estimator. In an augmented cross-sectional study, subjects who test negative onthe less-sensitive test in the cross-sectional survey are followed forward for transition into thenonrecent state, at which time they would test positive on the less-sensitive test. However,considerable uncertainty exists regarding the appropriate length of follow-up and the falsenegative rate. In this paper, we assess the impact of varying follow-up time on the resultingincidence estimators from an augmented cross-sectional study, evaluate the robustness ofthese estimators to assumptions about the existence and the size of the subpopulation whowill remain negative permanently, and propose a new estimator based on abbreviated follow-up time (AF). Compared to the original estimator, the AF estimator allows shorter follow-uptime and does not require estimation of the mean window period, dened as the average timebetween detectability of HIV infection with the sensitive and less-sensitive tests. It is shownto perform well in a wide range of settings. We discuss when the AF estimator would beexpected to perform well and oer design considerations for an augmented cross-sectionalstudy with abbreviated follow-up.

Keywords

augmented, cross-sectional studies; false negative; incidence estimators

9

Nonparametric Estimation of DistributionFunctions in Measurement Errors Models

Itai DattnerDepartment of Statistics, University of Haifa, Haifa 31905, [email protected]

Abstract. Many practical problems are related to the estimation of distribution functionswhen data contains measurement errors. For example, consider the estimation of the preva-lence of a disease which is determined by some underlying biomarker, measured with error,having value greater than some known constant. Another example is the estimation of thearea under the receiver operating characteristic curve which is widely used in biostatisticalstudies. These two examples deal with some functionals of the distribution function.

In this work we consider the problem of nonparametric estimation of some smooth func-tionals in measurement errors models. We study minimax complexity of this problem whenthe unknown distribution has a density belonging to the Sobolev class, and the error densityis ordinary smooth or supersmooth. We develop rate optimal estimators based on directinversion of the empirical characteristic function.

For the problem of estimating a distribution function we also propose an adaptive versionof the estimator and illustrate its superiority with respect to other methods both theoreticallyand through simulations. A real example of estimating hypertension prevalence is discussed.Extensions to other important estimation problems are also studied.

Keywords

adaptive estimator, deconvolution, error in variables, prevalence

This work is based on the PhD thesis of the author done under the supervision of Prof.Alexander Goldenshluger and Prof. Benjamin Reiser.

10

Modelling Health Surveillance Data via a BivariateINAR(1) Process

Xanthi Pedeli and Dimitris KarlisDepartment of Statistics, Athens University of Economics and Business{xpedeli,karlis}@aueb.gr

Abstract. Non{negative integer{valued time series are often encountered in many dierentscientic elds, usually in the form of counts of events at consecutive time points. Manyrepresentative examples of such data can be found in epidemiology. Due to their frequentoccurrence, a wide variety of models appropriate for treating such data have been proposedin the literature (Grunwald et al., 2000). Among the most popular are the INteger{valuedAutoRegressive (INAR) models (McKenzie, 1985; Al{Osh and Alzaid, 1987). In this work,we extend the simple INAR(1) model to the 2{dimensional space. In this way we dene abivariate INAR(1) process (BINAR(1)), we examine its marginal properties and proposedthe method of maximum likelihood for the estimation of its unknown parameters. Exten-sions to incorporate covariate information are discussed while emphasis is placed on modelswith bivariate Poisson and bivariate negative binomial innovations. Other distributional as-sumptions allowing for overdispersion and negative correlation are also briey considered.To motivate the model we use the example of syndromic surveillance during Athens 2004Olympic Games and investigate the potential of lower eectiveness of the system duringweekends compared to weekdays.

Keywords

counts, BINAR, Poisson, negative binomial, syndromic surveillance

References

AL{OSH, M.A. and ALZAID, A.A. (1987): First|Order Integer|Valued AutoregressiveProcess. Journal of Time Series Analysis, 8, 261{275.

GRUNWALD, G.K., HYDMAN, R.J., TEDESCO, L. and TWEEDIE, R.L. (2000): Non|Gaussian Conditional Linear AR(1) Models. Australian & New Zealand Journal ofStatistics, 42, 479{495.

McKENZIE, E. (1985): Some Simple Models for Discrete Variate Time Series. Water Re-sources Bulletin, 21, 645{650.

11

Invited papers

12

Dependence Calibration in Conditional Copulas: ANonparametric Approach

Elif F. Acar1, Radu V. Craiu2, and Fang Yao21 McGill University, Department of Mathematics and Statistics2 University of Toronto, Department of Statistics

Abstract. The study of dependence between random variables is a mainstay of statistics.In many cases, the strength of dependence between two or more random variables varies ac-cording to the values of a measured covariate. We develop inference for this type of variationvia a conditional copula model in which the copula is parametric and its parameter variesas the covariate. We propose a nonparametric procedure based on local likelihood to esti-mate the functional relationship between the copula parameter and the covariate, derive theasymptotic properties of the proposed estimator and outline the construction of pointwisecondence intervals. We also contribute a novel conditional copula selection method basedon cross-validated prediction errors and a generalized likelihood ratio-type test to determineif the copula parameter varies signicantly. We derive the asymptotic null distribution ofthe formal test. Using a subset of the Matched Multiple Birth dataset, we demonstrate theperformance of these procedures via analysis of gestational age-specic twin birth weights.

Keywords

conditional copula, covariate adjustment, generalized likelihood ratio test, local like-lihood

13

Unbalanced and Partial Group Sequential Methodsfor Normal Responses in Clinical Trials

Atanu Biswas1, Shirsendu Mukherjee2, and KyungMann Kim31 Indian Statistical Institute, Kolkata, India [email protected] Behala College, Kolkata, India shirsendu [email protected] University of Wisconsin-Madison, 600 Highland Avenue, Madison, WI 53792-4675, [email protected]

Abstract. Group sequential methods for a two treatment clinical trial with normal re-sponses are discussed. First we consider the case where the sample sizes for two treatmentsare possibly unequal between the two groups due to an unequal randomization. Then wediscuss group sequential design in the context of a historical-control study, that is, underthe partial sequential sampling scheme, in which the samples on one treatment, say control,are available at the outset, and the samples on the other treatment, say experimental, areobtained in the group sequential way. We discuss the cases of known and unknown variancefor unbalanced and partial group sequential set up. All the procedures are discussed withsimulation studies.

Keywords

clinical trials, group sequential methods, multivariate t distribution, partial sequentialsampling, type I error spending function

References

SCHEFFE, H. (1943): On solution of the Behrens-Fisher problem based on the t-distribution.Annals of Mathematical Statistics, 14, 35{44.

LAN, K.K.G. and DEMETS, D.L. (1983): Discrete sequential boundaries for clinical trials.Biometrika, 70, 649{663.

14

Improved Estimation of Survival Probabilitiesfrom Two-phase Stratied Samples

Norman E. Breslow1, Thomas Lumley2, and Jon A. Wellner31 University of Washington, Seattle, USA [email protected] University of Auckland, NZ [email protected] University of Washington, Seattle, USA [email protected]

Abstract. A standard problem in clinical epidemiology is the estimation of 5 year survivalprobabilities for cancer patients using information on age and stage of disease at diagnosis andthe histology and molecular characteristics of the tumor. Cost considerations may limit thenumber of stored tissue samples that are utilized in expensive bioassays. Stratied samplingis then useful to identify the most informative tissue samples to send to bioassay, assumingthat demographic, clinical and outcome data are already available for all subjects. Predictionof survival or other outcomes at designated times after diagnosis is generally accomplishedusing parametric or semiparametric models for failure time data, in particular, the Coxproportional hazards model.

The standard method for estimating Euclidean parameters in (semi)parametric modelsfrom stratied samples is Horvitz-Thompson weighting of the estimating equations, using asweights the inverse sampling probabilities. In previous work we have shown how adjustmentof the weights, either by their calibration to known totals of auxiliary variables or theirestimation using these same variables, can markedly improve the eciency of estimationof Euclidean parameters, e.g., of hazard ratios in the Cox model. By separating likelihoodcalculations, based on the model, from those on weak convergence of the inverse probabilityweighted empirical process, based on the sampling design with or without adjustment, ourgeneral results apply to a variety of (semi)parametric models. Here we apply these results tojoint estimation of Euclidean and innite dimensional parameters, in particular, to predictionof survival outcomes for individual patients using the hazard ratios and baseline hazard ofthe Cox model.

References

BRESLOW, N.E. and WELLNER, J.A. (2007,2008) Scand J Stat, 34, 86{102; 35, 186{192BRESLOW, N.E. and LUMLEY, T. et al. (2009) Am J Epidemiol, 35, 1398{1405BRESLOW, N.E. and LUMLEY, T. (2009) et al. Stat Biosci, 1, 32{49LUMLEY, T. (2010): Complex Surveys, Wiley, New York.

15

More Robust Doubly Robust Estimators

Marie DavidianDepartment of Statistics, North Carolina State University, Raleigh, NC, [email protected]

Abstract. Considerable recent interest has focused on doubly robust estimators for a pop-ulation mean outcome in the presence of incomplete data, which involve models for boththe propensity score and the regression of outcome on covariates. These estimators havethe appealing property that they are consistent for the true population mean even if one ofthe outcome regression or propensity score models, but not both, is misspecied. . However,the usual doubly robust estimator may yield severely biased inferences if neither of thesemodels is correctly specied and can exhibit nonnegligible bias if the estimated propensityscore is close to zero for some observations. We propose alternative doubly robust estimatorsthat achieve comparable or improved performance relative to existing methods. This is jointwork with Weihua Cao (US Food and Drug Administration) and Anastasios Tsiatis NorthCarolina State University).

Keywords

causal inference, missing at random, no unmeasured confounders, outcome regression

16

Order Restricted Inference for Multivariate BinaryData with Applications

Ori Davidov1 and Shyamal Peddada21 Department of Statistics, University of Haifa, ISRAEL2 Biostatistics Branch, National Institute of Environmental Health Sciences, USA

Abstract. In many applications such as toxicology, epidemiology, genetics and the socialsciences, researchers collect multivariate binary response data under two or more, naturallyordered experimental conditions. In such situations one is often interested in using all binaryoutcomes simultaneously to detect order related trends among the experimental conditions.We show that the problem of determining dose related trends can be formulated as a mul-tivariate stochastic ordering problem. We develop a methodology for the analysis of K 2multivariate binary distributions which are subject to such order restrictions. Our simula-tion studies indicate that order restricted estimators of the population proportions are oftenmore ecient in terms of their mean squared error than the unconstrained estimators. Thereduction in total mean squared error can be as much as 40%. We propose a test that incor-porates the ordering of all binary outcomes simultaneously. This test is shown to be muchmore powerful than the unrestricted Hotelling T 2 type procedure. We also compared theproposed test to procedures which combine m univariate tests for order. In particular westudied several union intersection type tests and a Bonferroni based test. Our simulationssuggest that the proposed method competes well with these alternatives. The gain in poweris often substantial. The proposed methodology is illustrated by applying it to a two yearrodent cancer bioassay data obtained from the US National Toxicology Program (NTP).

Keywords

binary data, dose response studies, stochastic order relations, order restricted statis-tical inference

17

k-FWER Control without p-value Adjustment, withApplication to Detection of Genetic Determinantsof Multiple Sclerosis in Italian Twins

Livio Finos1 and Alessio Farcomeni21 University of Padua - Italy [email protected] Sapienza - University of Rome - Italy [email protected]

Abstract. We show a novel approach for k-FWER control which does not involve anycorrection, but only testing the hypotheses along a (possibly data-driven) order until asuitable number of p-values are found above the uncorrected level. p-values can arise fromany linear model in a parametric or non parametric setting. The approach is not only verysimple and computationally undemanding, but also the data-driven order enhances powerwhen the sample size is small (and also when k and/or the number of tests is large). Theprocedure retains the error control under independence and weak dependence of p-values(see Farcomeni, 2007 and Clarke and Hall, 2009). As an alternative, a simple correction isproposed in order to retain the control under any dependence. We illustrate the method onan original study about gene discovery in multiple sclerosis, in which were involved a smallnumber of couples of twins, discordant by disease.

The paper is now pubblished on Biometrics (Finos and Farcomeni 2010) and the methodsare implemented in an R package (someKfwer), freely available on CRAN.

Keywords

data driven order, gene discovery, multiple sclerosis, multiple testing

References

FARCOMENI, A. (2007). Some results on the control of the false discovery rate underdependence. Scandinavian Journal of Statistics, 34, 275 - 297.

CLARKE, S. and HALL, P. (2009). Robustness of multiple testing procedures against de-pendence. Annals of Statistics 37, 332-358.

FINOS, L. and FARCOMENI, A. (2010). k-FWER Control without p-value Adjustment,with Application to Detection of Genetic Determinants of Multiple Sclerosis in ItalianTwins. Biometrics,DOI: 10.1111/j.1541-0420.2010.01443.x.

18

Comparative Eectiveness Research: AMethodologic Introduction

Constantine GatsonisCenter for Statistical Sciences Brown University Providence RI 02912

Abstract. Comparative Eectiveness Research (CER) is an initiative with wide rangingimplications for biomedical research and health care policy. The term "CER" refers to abody of research that generates and synthesizes evidence on the comparison of benets andharms of alternative methods to prevent, diagnose, treat, and monitor clinical conditions,or to improve the delivery of care. The evidence from Comparative Eectiveness Researchis intended to support clinical and policy decision making at both the individual and thepopulation level. The mandate of CER places a premium on the study of outcomes thatare of primary relevance to patients and on the derivation of conclusions that can informindividual patient choices.

The broad scope of CER requires a wide array of methodological approaches. CERresearch may include both randomized and observational primary studies as well as researchsynthesis. In this presentation we will discuss the research questions addressed by CER andsome of the statistical challenges they present.

19

Use of Composite Endpoints in CardiovascularDevice Trials involving Coronary Stents

Guadalupe Gmez1, Urania Dafni2, and Moises Gmez11 Universitat Politecnica de Catalunya [email protected] University of Athens [email protected]

Abstract. Composite endpoints (CE) are nowadays used commonly as primary endpoint toassess the ecacy of a new treatment. It is sometimes unclear, and often controversial, whichendpoint should be used in randomized clinical trials. For instance, in cardiovascular devicetrials involving coronary stents the often rare events of cardiovascular death and myocardialinfarction are combined with more common events such as target lesion revascularization(TLR), target vessel revascularization (TVR) or non cardiovascular death. Although theseevents are combined with the aim to increase the statistical power of the study, addingless specic components might in fact lead to loss of power to detect the true treatmentdierences. In addition, improvement in the composite does not necessarily mean an improvedsurvival. In other occasions, the usefulness of a CE relies in combining outcomes to havea better description of the disease process. Very often CEs are chosen for their potentialfor statistical eciency and as a way of dealing with the issues of multiple testing andcompeting risks. In cardiovascular trials, the use of a CE is furthermore intricate due to therelative importance to patients of the dierent components as well as the magnitude of theeect of treatment across the component endpoints. Gmez and Lagakos (2011) developed astatistical methodology to derive guidelines for deciding whether to expand a study primaryendpoint from E1 (cardiovascular death and myocardial infarction, say) to the composite ofE1 and a secondary endpoint E2 (TLR, say). Their method considers the asymptotic relativeeciency (ARE) of a logrank test for comparing treatment groups with respect to E1 versusthe composite endpoint of E1 or E2. The ARE depends on the marginal distributions of thetimes until E1 and E2, the correlation between these times, the treatment group dierenceswith respect to E1 and E2, and the pattern and amount of censoring. The Gmez andLagakos method is illustrated on two case studies. A set of recommendations is provided toobtain more ecient results in the area of coronary stents.

Keywords

Asymptotic Relative Eciency; Composite outcome; Logrank test; Statistical PowerReferences

FERREIRA-GONZLEZ, N. and others (2007). Problems with use of composite end pointsin cardiovascular trials: systematic review of randomized controlled trials. British Med-ical Journal, doi: 10.1136/bmj.39136.682083.AE

GMEZ and LAGAKOS (2011). Statistical considerations when using a composite endpointfor comparing treatment groups (submitted)

20

Treatment Noncompliance in Studies ofAdjuvant Chemotherapy for Breast Cancer

Robert J. Gray1Dana-Farber Cancer Institute, Boston, MA [email protected]

Abstract. The Trial Assigning IndividuaLized Options for Treatment (TAILORx) uses Ge-nomic Health's Oncotype DXTM 21 gene Recurrence Score (RS) assay to select adjuvanttreatment for hormone receptor positive, node negative breast cancer. This assay was devel-oped to predict recurrence risk (Paik et al, 2003) and retrospective analyses suggest RS maypredict whether patients will benet from chemotherapy (Paik et al, 2006). In TAILORx,patients with an RS < 11 are assigned to receive hormonal therapy (HT) only, patients withan RS > 25 are assigned to receive HT plus chemotherapy and patients with RS in range11-25 are randomized to HT alone or HT plus chemotherapy. Accrual to TAILORx wascompleted in 2010, with 11,233 patients screened for RS evaluation and 6,907 patients withRS 11-25 randomized. Approximately 7% of the patients assigned to HT alone have receivedchemotherapy and 17% of the patients assigned to receive chemotherapy have refused it.There is substantial variation in treatment preference over the range of recurrence scores,and based on prior retrospective data, it is expected that the magnitude of the treatmenteect may also vary with RS. Methods for accounting for this noncompliance in the designand analysis will be discussed and examples of noncompliance in other breast cancer studieswill be presented.

Keywords

noncompliance, breast cancer, recurrence score

References

Paik, S., Shak, S., Tang, G., et al (2004): A Multigene Assay to Predict Recurrenceof Tamoxifen-Treated, Node-Negative Breast Cancer. The New England Journal ofMedicine, 351, 2817-2826.

Paik, S., Tang, G., Shak, S., et al (2006): Gene Expression and Benet of Chemotherapyin Women With Node-Negative, Estrogen Receptor-Positive Breast Cancer. Journal ofClinical Oncology, 24, 3726-3734.

21

Estimation of Sensitivity of Chest X-ray andCancer Preclinical Sojourn Time for Lung CancerScreening Trials

Ping Hu and Philip ProrokNational Cancer Institute/National Institute of Health, USA [email protected]

Abstract. The eectiveness of cancer screening depends crucially on two elements: thepreclinical sojourn time (that is, the duration of the preclinical screen-detectable period)and the sensitivity of the screening test. Chest x-ray has historically been employed mostfrequently as the major screening test for lung cancer. Little is known about the accuracy ofChest x-ray in community practice.

To investigate this issue, one possibility is to use the available methods in prior literature.However, these methods rst have largely concentrated on breast cancer screenings andsecond are assumed 100It is clear that the data from most cancer screening trials do notsupport a zero false positive rate. Therefore, it would be interesting to generalize thesecommonly used methods by considering specicity also as a parameter and to estimatemean sojourn time/mean lead time, sensitivity and specicity simultaneously. It would bealso interesting to demonstrate whether the existing methods used in breast cancer screeningcould be used in lung cancer screening. New method is applied to the data from the lungcomponent of PLCO cancer screening trial and Yunnan Tin Miners Lung Cancer study inChina.

Keywords

sojourn time, lead time, sensitivity, specicity, chest x-ray, lung cancer

References

Auvinen, A. et al. (2009): Test Sensitivity in the European Prostate Cancer Screening Trial:Results from Finland, Sweden, and the Netherlands. Cancer Epidemiol Biomarkers,18(7): 2000-2005.

Shen, Y. and Zelen, M. (1999): Parametric estimation producers for screening programmes:Stable and nonstable disease models for multimodality case nding. Biometrika, 86,503-515.

22

Inference under Biased Sampling with Applicationto Infection Data

Micha MandelThe Hebrew University of Jerusalem [email protected]

Abstract. Selection bias is a general term for observed data having a dierent (biased) lawthan that of the target population. Correction of selection bias requires understanding andappropriate modelling of the biasing mechanism. A common and very important exampleof selection bias arises in cross-sectional designs in which data are collected on individualswho are available at a given place and time window; estimation methods for the lifetimedistribution function are well developed under the assumption of steady state. In this talk,I review basic results in the analysis of data obtained in cross-sectional designs and discussdierent aspects of the steady state assumption. I then discuss inference under cross-sectionaldesigns for a population that can be joined in xed and known time points. I demonstratedierent ideas using several data sets collected by the Israeli Ministry of Health.

Part of the work is joint with Yosi Rinott, Rebecca Betensky, Ronen Fluss, Laurence Freed-man, and the Department of Health Services and Research, the Ministry of Health, Israel.

Keywords

cumulative incidence, prevalence, truncation, weighted analysis

23

Issues in ROC Surface Analysis with an Applicationto Externally Validated Cognition in ParkinsonDisease Screening

C.T. Nakas1, T.A. Alonzo2, J.C. Dalrymple-Alford3, and T.J. Anderson31 Laboratory of Biometry, University of Thessaly, Phytokou str, 38446 Volos, [email protected]

2 Division of Biostatistics, University of Southern California, Arcadia, CA 91066, [email protected]

3 Van Der Veer Institute for Parkinsons and Brain Research, Christchurch 8011, [email protected] [email protected]

Abstract. The diagnostic accuracy of the Montreal Cognitive Assessment (MoCA) has beenestablished recently for screening externally validated cognition in Parkinson's disease (PD)(Dalrymple-Alford et al, 2010). Patients were classied as having either normal cognition(PD-N), mild cognitive impairment (PD-MCI), or dementia (PD-D). ROC curve method-ology has been used to assess discrimination between two adjacent classes and the Youdenindex has been employed for cut-o point selection. ROC surface methodology has also beenused for the assessment of the simultaneous discrimination of the three classes. Recently,Nakas et al (2010) proposed a generalization of the Youden index for the assessment of accu-racy and cut-o point selection in simultaneous discrimination of three classes. In this work,we examine properties of the generalized Youden index and compare two- vs. three-classclassication accuracy approaches when screening for cognition status in PD.

Keywords

ROC analysis, Youden index, K-S statistic, Montreal cognitive assessment

References

DALRYMPLE-ALFORD, J.C., MACASKILL, M.R., NAKAS, C.T., et al (2010). TheMoCA: Well suited screen for cognitive impairment in Parkinson Disease. Neurology,75, 1717{1725.

NAKAS, C.T., ALONZO, T.A. and YIANNOUTSOS, C.T. (2010). Accuracy and cut-opoint selection in three-class classifcation problems using a generalization of the Youdenindex. Statistics in Medicine, 29, 2946{2955.

24

Formulation of Recommendation Domains forSugarcane Varieties: Using Modied StabilityAnalysis and Best Linear Unbiased Predictor

Njuho, P.M.1, Sewpersad, C.N.2, and Redshaw K A21 HIV/AIDS, STI and TB (HAST), Human Sciences Research Council, 750 Francois Road,Cato Manor, Durban 4001, South Africa. [email protected]

2 South African Sugarcane Research Institute, P.O. Box 700, Mount Edgecombe 4300,KwaZulu-Natal, South Africa [email protected]

Abstract. This paper discusses two alternative statistical procedures, the modied stabil-ity analysis (MSA) and the best linear unbiased predictor (BLUP), that are eective in theformulation of recommendation domains for sugarcane varieties grown across ten sites in twoSouthern African countries (South Africa and Swaziland). The procedures are illustrated us-ing yield data on fourteen sugarcane varieties not all grown in all the sites and in the sameyear. Furthermore, harvesting was done at dierent stages of ratoons, hence complicatingthe analysis. Modied stability analysis explicitly incorporates variation in eld manage-ment, soils and climatic conditions and in the process enables evaluation of the performanceof each variety relative to each site. The best linear unbiased predictor analysis takes intoconsideration factors that are xed and factors that are random. The two approaches helpscientists evaluate responses to treatments and partition sites into recommendation domains.The South African Sugar Association Research Institute (SASRI) and Swaziland Sugar As-sociation Technical Services (SSATS) are continually extending out- grower services to newgrowing areas. Evaluation of commercial varieties across a range of sites is conducted with aview to drawing recommendation domains to growers for dierent agro climatic conditionsand management practices. The current climatic conditions and socio-economic factors de-mand for development of high yielding varieties that utilise the available resources moreeciently. By formulating recommendation domains for these varieties, we respond to thisdemand. We demonstrate the process of the analysis, discuss the ndings and highlight thechallenges encountered. We conclude that no single analysis can handle this type of data andtherefore advocate analysis done on stages.

Keywords

xed and random eects, environmental index, inference space, prediction

25

Massively Parallel Nonparametrics in Neuroimaging

Philip T. Reiss1 and Lei Huang21 New York University, New York, NY, USA, and Nathan Kline Institute, Orangeburg,NY, USA [email protected]

2 New York University, New York, NY, USA [email protected]

Abstract. Given a sample of brain images, and associated clinical or demographic predic-tors, the standard modeling paradigm entails a collection of linear models, one at each oftens of thousands of brain locations (Friston et al., 1995). But in many applications, interestcenters on nonlinear dependence on predictors. For example, many studies have examinedhow image-derived quantities, such as functional connectivity or cortical thickness, vary withage at each point in the brain. The developmental trajectories are often not well described bylinear or other parametric models, so that nonparametric regression with penalized splinesis more suitable; but it has heretofore been impractical to perform spline smoothing in themassively parallel manner that these applications require. To surmount this diculty, we in-troduce new algebraic techniques that make it feasible to compute huge numbers of optimallysmoothed spline ts simultaneously. Our approach also enables testing a null hypothesis oflinear dependence against a smooth alternative at each point in the brain, using a restrictedlikelihood ratio test (Crainiceanu and Ruppert, 2004). We adapt our methods to a completelydierent application: a renement of higher-order diusion tensor imaging for mapping thebrain's white matter architecture.

Keywords

higher-order diusion tensor imaging, penalized splines, restricted likelihood ratiotest, smoothness selection

References

CRAINICEANU, C.M. and RUPPERT, D. (2004): Likelihood Ratio Tests in Linear MixedModels with One Variance Component. Journal of the Royal Statistical Society, SeriesB, 66(1), 165{185.

FRISTON, K.J., HOLMES, A.P., WORSLEY, K.J., POLINE, J.-B., FRITH, C.D., andFRACKOWIAK, R.S.J. (1995): Statistical Parametric Maps in Functional Imaging: AGeneral Linear Approach. Human Brain Mapping, 2, 189{210.

26

Statistics in a Translational Neuroscience Program

Allan R. SampsonUniversity of Pittsburgh

Abstract. This talk will consider several motivating problems in translational neurosciencethat lead to interesting statistical issues and is based upon the speaker's long involvementwith the Translational Neuroscience Program at the University of Pittsburgh. Focus will beon issues that have arisen from post-mortem brain tissue studies which are often employed fora number of psychiatric disorders to identify cellular biomarkers which distinguish subjectswith the disorder from normal controls. The studies we consider use matched subject-controldesigns for sampling and tissue processing, with subjects chosen from a Brain Bank main-tained at the University of Pittsburgh. The rst motivating problem that we consider in thiscontext is the ecient design of stereological tissue sampling schemes in order to test forpopulation level dierences. The other couple of issues we discuss arise from several projectswith diering purposes, but which integrate results across multiple such studies with a goalto better understand the neurobiology of schizophrenia. One issue deals with using mul-tiple studies to attempt to identify those neurobiological markers which best characterizeschizophrenia and the other issue uses multiple studies to try to nd clusters of subjects.There are specics of the available data for both these projects that require new approachesin terms of structured multivariate models, mixture modeling, missing data, and discrimi-nant techniques. The goal of our presentation is to provide a motivating overview of theseissues and outline briey some of our methodology.

(This is joint research with Josephine Asafu-Adjei (University of Pittsburgh) Qiang Wu(East Carolina University) and Wei Zhang (DVM, FDA).)

27

Generation Times in Epidemic Models

Gianpaolo Scalia TombaDept of Mathematics, University of Rome Tor Vergata, Italy [email protected]

Abstract. There has been recent interest in the so called generation time in epidemicmodels, i.e. the average time between the infection of a primary case and one of its secondarycases. It is related to the latent and infectious period distributions and is involved in auseful relationship between initial speed of growth and the basic reproductive number R0in SIR models. The natural framework for considerations about various times in epidemicmodels and analysis of their statistical properties is stochastic. Various facts about thegeneration time distribution will be presented and links to demography and statistics willalso be discussed.

Keywords

generation time, epidemic model

References

SCALIA TOMBA G., SVENSSON A., ASIKAINEN T. and GIESECKE J. (2010): Somemodel based considerations on observing generation times for communicable diseases.Mathematical Biosciences, 223, 24{31.

28

Real Time Prediction for Dengue Epidemic inHavana (2001) Using Model Averaging Methods

Ziv ShkedyI-Biostat, Universiteit Hasselt, Diepenbeek, Belgium

Abstract. Real time prediction of the nal size and the turning point of the dengue outbreakare of primary interest from a public health point of view. Typically (i.e., Hsieh and Ma,2009) a Richards model can be tted to weekly dengue notication numbers in order to todetect the turning point for the outbreak which enables us to study the impact of interventionmeasures relating to the turning point.

In this study we use t several non linear models for the outbreak data and use modelaveraging methodology to estimate both turning point and nal size of the outbreak. Theproposed methodology is applied to a single phase 2001 outbreak data from Havana city,Cuba.

29

A Measure of Explained Variation for EventHistory Data

Janez Stare1, Maja Pohar Perme2, and Robin Henderson31 University of Ljubljana, Slovenia [email protected] University of Ljubljana, Slovenia [email protected] Newcastle University, UK [email protected]

Abstract. There is no shortage of proposed measures of prognostic value of survival modelsin the statistical literature. They come under dierent names, including explained variation,correlation, explained randomness and information gain, but their goal is common: to denesomething analogous to the coecient of determination R2 in linear regression. None how-ever have been uniformly accepted, none have been extended to general event history data,including recurrent events, and many cannot incorporate time-varying eects or covariates.We present here a measure specically tailored for use with general dynamic event historyregression models. The measure is applicable and interpretable in discrete or continuoustime, with tied data or otherwise, with time-varying, time-xed or dynamic covariates, withtime-varying or time-constant eects, with single or multiple event times, with parametric orsemi-parametric models, and under general independent censoring/observation. For single-event survival data with neither censoring nor time-dependency it reduces to the concordanceindex. We give expressions for its population value and the variance of the estimator andexplore its use in simulations and applications. A web link to R software is provided.

Keywords

C-index, dynamic models, explained variation, rank correlation, recurrent events

References

STARE J, POHAR PERME M, HENDERSON R. A Measure of Explained Variation forEvent History Data. Biometrics 2011. In Press.

30

Modeling Reductions in Breast Cancer Mortality

Ester Vilaprinyo1 and Montserrat Rue21 Evaluation and Clinical Epidemiology, Hospital del Mar-IMIM, [email protected]

2 Basic Medical Sciences Department, Medical School, University of Lleida-Irblleida, [email protected]

Abstract. Reduction in breast cancer (BC) mortality in Western countries has been at-tributed to the use of screening mammography and to treatment improvement. We used astochastic model based on previous works [1, 2] to quantify the contribution of each inter-vention. We estimated standardized BC mortality rates for the age group 30-69 per 100 000women for calendar years 1975-2007 in four hypothetical scenarios: 1) Only screening, 2)Only treatment improvement, 3) Both, and 4) None.

Observed Catalan rates rose from 30.6 in 1977 to 37.5 in 1992, and afterwards contin-uously decreased to 23.1 in 2006. If none of the two interventions had been used, in 2006the estimated BC mortality would be 43. Mortality reduction due to treatment improve-ment was higher than reduction due to screening. Taking as reference the scenario withoutany intervention, the mortality reduction in 2006 for only screening was 20% and for onlytreatment improvement 37%. With both interventions the reduction was 51%, this value islower than the sum of the two individual contributions because there is a negative synergismbetween the two interventions (a dierence of 6%).

The agreement between observed data and estimations from the model seem to indicatethat the assumptions of the model and the inputs are correct.

Keywords

modelling, breast cancer, screening, mortality

References

Lee SJ and Zelen M. (2008): Mortality modeling of early detection programs. Biometrics64(2): 386{395.

Rue M, Vilaprinyo E, Lee S, Martinez-Alonso M, Carles MD, Marcos-Graguera R, Pla R,Espinas JA. (2009): Eectiveness of early detection on breast cancer mortality reductionin Catalonia (Spain). BMC Cancer 9: 326.

31

Accounting for Prediction Error in EnvironmentalExposure when Relating Public Health toEnvironmental Factors

Linda J. Young1, Kenneth K. Lopiano2, and Carol A. Gotway31 University of Florida, Gainesville FL USA, [email protected] University of Florida, Gainesville FL USA, [email protected] U.S. Centers for Disease Control and Prevention, Atlanta GA USA, [email protected]

Abstract. Publicly available data from disparate sources are increasingly combined forsubsequent statistical analyses. Because data from dierent sources are frequently measuredor associated with dierent geographic or spatial units, combining them for analysis usuallyrequires prediction of one or more of the variables of interest. Here it is assumed that healthoutcomes and related covariates are measured at residences (points), that environmentalexposure is measured at monitors (points), and that the two sets of points are mutuallyexclusive. To assess the association between the health outcome and environmental exposure,adjusting for covariates, the environmental exposure is predicted for the points for whichhealth outcomes are observed.

When exposure is predicted using a smoothing method, such as kriging, Berkson errorarises in the estimation of the parameter associated with environmental exposure in the re-gression of health outcomes on predicted environmental exposure, adjusting for covariates(Gryparis, et al. 2009, Szpiro, et al. 2011). As a consequence, the parameter is estimatedunbiasedly (unlike with classical measurement error), but the standard error is biased down-wards. Previously suggested methods for improving the estimated standard error will bebriey reviewed (Szpiro, et al. 2011; Lopiano, et al. 2011).

After aligning the health and environmental data sets using kriging, an iterativelyreweighted generalized least squares approach is suggested for relating health outcomes andenvironmental exposure, adjusting for covariates. The properties of the method are discussed,and simulation results illustrate the performance of the proposed approach. Using the pro-posed methodology, the association between birth weight and air quality is explored, andthe results contrasted to those obtained with other methods.

Keywordssensitivity, specicity, marker, censored survival dataReferencesGRYPARIS A., PACIOREK C.J., ZEKA A., SCHWARTZ J. and COULL B. (2009): Mea-

surement error caused by spatial misalignment in environmental epidemiology. Biostatis-tics, 10, 258-274; doi:10.1093/biostatistics/kxn033.

LOPIANO K. K.., GOTWAY C.A., and YOUNG L.J. (2010): A comparison of errors invariables methods for use in regression models with spatially misaligned data. StatisticalMethods in Medical Research, 20, 29-47. doi:10.1177/0962280210370266.

SZPIRO A.A., SHEPPARD L., and LUMLEY, T. (2011): Ecient measurement er-ror correction with spatially misaligned data. Biostatistics. Online early at doi:10.1093/biostatistics/kxq083z.

32

A Regularization/Extrapolation Corrected ScoreMethod for Nonlinear Regression Models withCovariate Error

David M. Zucker1, Malka Gorne2, Yi Li3, and Donna Spiegelman41 Hebrew University, Jerusalem, Israel, [email protected] Technion, Haifa, Israel, [email protected] Harvard School of Public Health, Boston MA, USA, [email protected] Harvard School of Public Health, Boston MA, USA, [email protected]

Abstract. Many regression analyses involve explanatory variables that are measured witherror, and failing to account for this error is well known to lead to biased estimates for the re-gression coecients. We present here a new general method for adjusting for covariate error.Our method consists of an approximate version of the Stefanski-Nakamura corrected scoreapproach, using the method of regularization for approximate solution of integral equations,along with an extrapolation device similar in spirit to that of the SIMEX method. Speci-cally, we compute estimates for various values of the regularization penalty parameter andextrapolate to a penalty parameter of zero. We develop the theory in the setting of classicallikelihood models, covering nonlinear regression, logistic regression, and Poisson regression.The method is extremely general in terms of the types of measurement error models covered,and is a functional method in the sense of not requiring information on the distribution of thetrue covariate. We present a simulation study in the logistic regression setting, and providean illustration on data from the Harvard Nurses' Health Study concerning the relationshipbetween physical activity and breast cancer death among patients with diagnosed breastcancer.

Keywords

errors in variables, nonlinear models, logistic regression

33

Contributed papers

34

Modelling the Mean and Covariance Structure forContinuous Bounded Longitudinal Data

Mouna Akacha1 and Jane L. Hutton21 Novartis Pharma AG, Basel, Switzerland, [email protected] University of Warwick, Coventry, United Kingdom, [email protected]

Abstract. In medical research it is common to measure patients' health status repeatedlyover time through questionnaires. Based on the answers, summary measures such as scorescan be derived. These scores will often have nite range, where one bound indicates 'nosymptoms' and the other bound 'extreme symptoms'.

For a continuous and bounded score, the classical approach is to transform the dataso that a linear regression model ts adequately. For some scores, however, a non-lineardependence of the transformed score on covariates persists. In addition, models based ontransformations cannot investigate the dependence of bounds on covariates as the boundsneed to be specied prior to the transformation.

In view of these limitations, we propose a non-linear mixed model for the mean scoreon the original scale as a function of covariates. The model is constructed for scores wherethe rate of recovery changes over time and has been motivated by the Collaborative AnkleSupport Trial, which is a randomized controlled trial comparing four treatments for acuteankle sprains.

Apart from modelling the mean score, we discuss models for the covariance structureof bounded longitudinal data. With repeated measurements, we expect higher correlationswhen the measurements are closer in time than when they are further apart. Additionally,with bounded data, correlations increase as measurements reach the bound regardless of thetime interval between measurements. Finally, the variances are rarely constant over time. Adata-driven regression approach introduced by Pourahmadi [1999] is adopted and extendedto allow for missing values.

Keywords

bounded data, non-linear mixed models, covariance models, missing data

References

Pourahmadi [1999]: Joint mean-covariance models with applications to longitudinal data:Unconstrained parametrisation. Biometrika, 86, 677{690.

35

Estimates of Clinically Useful Measures in SurvivalAnalysis

Federico Ambrogi, Elia Biganzoli, and Patrizia BoracchiSezione di Statistica Medica e Biometria "G.A. Maccacaro",Dipartimento di Medicina del Lavoro "Clinica del Lavoro L. Devoto"[email protected]

Abstract. In clinical studies where an event during patients follow-up is of interest, theanalysis of the hazard ratio is generally used to measure treatment or covariate eects. Inorder to support clinical decisions, the estimate of dierent clinical useful measures should bedirectly considered, such as relative risks, excess of risks, relative risk reduction and numberof patients needed to be treated.

The aim of this work is to provide a straightforward approach to obtain point andinterval estimates of the above measures, by using transformation models, through suitablelink functions. Modeling of the prognostic relationships in presence of variables measuredon continuous scale and of putative time dependent eects is a challenge in this context.In order to use standard software for model estimates, the proposal of Klein and Andersen,based on pseudo-values, was considered as starting point. This approach, which was originallyproposed to model competing risks and multi-state applications or as diagnostic tool forhazard regression, proved to be useful also in standard survival applications. The baselinerisk function was estimated resorting to regression spline on time. Time-varying eects ofcovariates were tested through interaction with baseline time functions. A large literaturedata set on breast cancer was used for illustration.

Keywords

clinically useful measures, treatment eect, pseudo-values, time varying eects

References

AMBROGI F., BIGANZOLI E., BORACCHI P. (2008): Estimates of clinically useful mea-sures in competing risks survival analysis. Stat Med, 27, 6407-25

KLEIN J.P. and ANDERSEN P.K. (2005): Regression modeling of competing risks databased on pseudovalues of the cumulative incidence function. Biometrics, 61, 223-9

36

Penalized Likelihood Methodology and Applications

E. Androulakis1, C. Koukouvinos2 and F. Vonta31 National Technical University of Athens [email protected] National Technical University of Athens [email protected] National Technical University of Athens [email protected]

Abstract. The penalized likelihood approach (Fan and Li, 2001), has been consistentlydemonstrated to be an attractive shrinkage and selection method. This procedure is dierentfrom traditional methods of variable selection in that it selects the signicant variables andestimates regression coecients simultaneously. As a result, the produced estimators are asecient as the oracle estimator.

This new methodology was extended further to the case where we have survival datawith censored observations and clusters, where the Cox proportional hazards model andthe Gamma frailty model (Fan and Li, 2002) are the two commonly used semi-parametricmodels. This prompted us to extend the penalized Gamma frailty model approach and topropose a generalized form of the full likelihood function designed for clusters, which allowsthe direct use of many dierent distributions for the frailty parameter.

However, the performance of the penalized likelihood estimators depends on the properchoice of the regularization parameter. To this end, we rstly propose new estimates of thenorm of the error in the generalized linear models framework, through the use of Kantorovichinequalities. Then these estimates are used in order to derive a tuning parameter selector inpenalized generalized linear models.

Keywords

penalized likelihood, penalized frailty model, generalized linear models, tuning pa-rameter estimation, error estimation

References

FAN, J. and LI, R. (2001): Variable selection via nonconcave penalized likelihood and itsoracle properties. J. Amer. Statist. Assos., 96, 1348{1360.

FAN, J. and LI, R. (2002): Variable selection for Cox's proportional hazards model andfrailty model. The Annals of Statistics, 30, 74{99.

37

Performance of Markers for Censored Failure TimeOutcome: Nonparametric Approach Based onProportions

Laura Antolini1Department of Clinical Medicine and Prevention, University of Milano-Bicocca, Monza(Mi), Italy. [email protected]

Abstract. Nonparametric analysis of classication probabilities of markers used to detectthe presence of disease at the time of marker measurement is based on inference on simpleproportions. When disease status is subject to verication bias, weighting/imputation tech-niques enable to work again on proportions calculated on a classication matrix with avail-able asymptotics. Inference becomes more complex because of censoring, when the markeraims at detecting the future development in time of disease. A fully nonparametric solutionbased on smoothed estimation of the bivariate survival function of time to disease and themarker was proposed, where bootstraap is recommended for asympotics. Semiparametric ap-proaches have also been developed. Here we consider censoring as a source of verication biason the time to development of disease. Full and partial imputation of the disease status byKaplan Meier estimation of survival functions conditional on the marker, and inverse prob-ability of censor