Please activate JavaScript!
Please install Adobe Flash Player, click here for download

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 13Monday, 25th August 2014 • 9:00-10:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C01.5 Long-term evaluation of different designs of a series of phase III clinical trial in rare cancers: a simulation study M-A Bayar1 , G Le Teuff1 , S Michiels1 , D Sargent2 , M-C Le Deley1 1 Institut Gustave Roussy, Department of Biostatistics and Epidemiology, Villejuif, France, 2 Mayo Clinic, Rochester, United States Background: In rare diseases, randomized trials designed with typical 5% α-level and 80%-power are unfeasible. A previous simulation study by Le Deley et al. suggested performing a series of small trials with relaxed α-levels; the treatment effect of each new treatment was characterized by the hazard ratio drawn from a certain distribution, yielding benefit accu- mulation trial after trial, which is a questionable hypothesis. We reviewed and extended this simulation framework. Methods: We simulated a series of two-treatment superiority trials over a 15-year research period. Trial parameters examined included the α-level and the number of trials run over the 15-year period (thus, the trial sample size). Each design was evaluated for different disease scenarios and accrual rates. In our simulation study, the treatment effect of each new treatment was defined by the hazard rate, avoiding benefit accumulation. Different assumptions of how treatments improve over time were considered. Results: Our simulation study shows that performing a series of small tri- als with relaxed α-levels leads on average to larger survival gains over a long research horizon than performing larger trials with typical 5% α-level while controlling the risk of selecting a worse treatment at the end of the research period. The survival gains were substantially lower than those shown in the previous work but the pattern is similar and recommenda- tions remain valid. Conclusion: Designs aiming to maximize the expected survival gain over a long research horizon across a series of trials are worth discussing in the context of rare diseases.   C02 Missing data C02.1 A comparison of multiple imputation methods for hierarchical data when there is whole cluster non-response K DiazOrdaz1 , MG Kenward1 , M Gomes1 , R Grieve1 1 LSHTM, London, United Kingdom Background: Missing data are common in cluster randomised trials and need to be handled appropriately. When using multiple imputation (MI), we must recognise the data struc- ture. Methods: We conducted a simulation study to compare complete-case analysis, multilevel MI, fixed-effects MI, and single-level MI, when the anal- ysis model is a linear mixed-model. Clustered data, consisting of normally-distributed outcomes with baseline individual and cluster-level covariates, were simulated with different levels of clustering, measured by the intra-cluster correlation coefficient (ICCs), number and size of clusters and probability of non-response. Missing data were introduced according to missing-at-random mecha- nisms driven by: (i) an individual-level covariate, (ii) cluster-level covariate, (iii) treatment, (iv) cluster size and included cluster non-response. For each missing data scenario, 1000 sets were simulated. Empirical bias of the treatment effect and confidence interval (CI) coverage were obtained for each method. Results: Complete-case analyses resulted in biased effect estimates when treatment was associated with missingness, while multilevel MI produced estimators with negligible bias across all the missingness mechanisms considered. Fixed-effects MI over-estimated the variance resulting in CI coverage in excess of nominal levels (up to 99.8%), especially for datasets with low ICC. Moreover, it resulted in biased estimates for scenarios where a non-neg- ligible number of clusters were missing and a cluster-level variable was associated with the missingness mechanism. Conclusion: The validity of inferences may depend on how clustering is accommodated in the imputation step. Multilevel MI performed well across all settings considered and is theoretically appropriate for studies that have a hierarchical design. C02.2 Pattern mixture models applied to clinical trials for chronic pain G Thoemmes1 , J Abellan-Andres1 , M Janssens2 1 Grunenthal GmbH, Aachen, Germany, 2 Hasselt University, Hasselt, Belgium Missing data are a common problem in chronic pain trials due to pre- mature termination of subjects experiencing side effects or having in- sufficient analgesia over the course of a long-term trial. Pattern mixture models have shown to provide a flexible and transparent framework for handling missing data. In practice they can be implemented using mul- tiple imputation. Over the last years the combination of pattern mixture models and multiple imputation has been recognized as one of the most promising approaches for analyzing trials with missing values. A number of pattern mixture models will be presented and some notes regarding implementation details will be given. A case study based on a pooled da- taset of several three-arm, parallel-group, active and placebo controlled chronic pain trials will presented.The different pattern mixture models will be applied to this dataset and compared with a direct likelihood analysis using a mixed model for repeated measures (MMRM), and also with well- known single imputation methods, e.g. LOCF and BOCF. C02.3 Conference Award for Scientists CD4+ counts in a 3-arm longitudinal clinical trial with substantial missing data: a sensitivity analysis AC Grobler1 , G Matthews2 , G Molenberghs3 1 CAPRISA, Durban, South Africa, 2 University of KwaZulu Natal, Durban, South Africa, 3 I-BioStat, Universiteit Hasselt and KU Leuven, Hasselt, Belgium The SAPiT trial was an open label, randomised controlled trial in HIV- tuberculosis co-infected patients. Patients were randomised to three arms; each initiating antiretroviral therapy at a different time. CD4 count was measured 6-monthly for 24 months. The assumption that missing data are missing completely at random (MCAR) was not supported by the observed data. We performed a range of sensitivity analyses under both missing at random (MAR) and missing not at random (MNAR) assumptions. Under MAR assumptions Bayesian analysis, multiple imputation and maxi- mum likelihood analyses (mixed model repeated measures with arm, time and the interaction between arm and time) were done. Under MNAR assumptions several pattern mixture models (PMMs) were fitted. These included analysis using the CCMV and NCMV identifying restrictions, PMMs using random effects mixed models and PMMs using multiple imputation. Selection models were fitted using Bayesian meth- ods. All these methods are based on different unobserved assumptions about the missing data processes; thus the importance of a sensitivity analysis. More than one third of participants were lost to follow-up. Results are given and contrasted from all these methods allowing conclusions taking the missing data into account. MAR analyses showed a larger difference between the treatment arms than the MNAR analyses. The conclusion is that mean CD4+ count increased more in the early and late integrated

Pages Overview