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


50 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsTuesday, 26th August 2014 • 11:00-12:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust in making targeted therapies available for people with cancer in the UK, with the National Lung Matrix Trial forming the next major phase in the agenda. The trial consists of a series of parallel, multi-centre, single-arm phase II tri- als, each arm testing an experimental targeted drug in a population strati- fied by multiple pre-specified target biomarkers. There are currently 7 targeted drugs and 20 different drug-biomarker combinations. The aim of statistical analysis is to determine whether there is sufficient signal of ac- tivity in any drug-biomarker combination to warrant further investigation. Due to the complexity of the trial, we have chosen to use a Bayesian adaptive design that gives a more realistic approach to decision-making and flexibility to make conclusions without fixing the exact sample size. The design allows early stopping of recruitment to any drug-biomarker combinations that do not show sufficient promise at an interim analysis to warrant continuation. There have been many dilemmas over the de- sign and, for example, choices had to be made regarding whether to use a Bayesian approach, what priors to use and what outcome measure and criteria should be used for the decision to proceed. The paper will present the dilemmas and rationale for the final design. C25.5 A new framework using G-estimation for placebo-controlled randomized phase 3 trials with extensive crossovers for biomarker-driven molecularly targeted oncology agents S Nomura1,2 , T Shinozaki3 , C Hamada2 1 National Cancer Center, Chiba, Japan, 2 Tokyo University of Science, Tokyo, Japan, 3 University of Tokyo, Tokyo, Japan   Placebo-controlled randomized phase 3 trials for biomarker-driven mo- lecularly targeted agents (BMTAs) are playing increasingly important role in cancer drug development. In such phase 3 trials, crossover of treatment is often occurred because most of recent BMTAs are known to be over- whelmingly effective in trials of biomarker-enriched patient populations and therefore this evidence drives patients to hesitate the participation of the crossover-prohibited trials. While intention-to-treat (ITT) analysis is the most recommendable ap- proach in superiority trials, it is evident that the ITT analysis could not evaluate the causal survival benefit that would have been obtained had all patients complied protocol therapies in the presence of extensive non- random crossovers.To overcome this issue, we use a randomization-based G-estimation method with rank preserving structural accelerated failure time (RPSFT) models and incorporate it into phase 3 trial designs. In this proposed method, we first re-construct the potential overall survival (OS), defined as the time to death if the patients had received the assigned ther- apy throughout the study duration, using RPSFT models and then develop a G-test-like decision rule instead of the ITT log-rank test. Considering the three copula-type dependence structures between time to treatment switch and potential OS, we compared the performance between the pro- posed and ITT analysis with simulation studies. It is shown that the proposed method could increase the power with a slight inflation of type I error. We will show the performance of two-stage design which adaptively select the analysis strategies (proposed or ITT) from the first stage data.   C26 Network meta-analysis C26.1 A multi-state Markov model for network meta-analysis of studies with missing data O Efthimiou1 , S Leucht2 , G Salanti1 1 University of Ioannina, Ioannina, Greece, 2 Technische Universität München, Munich, Germany   Background: Missing data constitute a serious threat for the validity and precision of inferences from trials and their quantitative synthesis via meta-analysis. Various imputation methods have been proposed for the case that individual patient data (IPD) are available to systematic reviewers but their performance highly depends on the, typically unknown, missing mechanism. Methods: We propose a multi-state Markov model for network meta- analysis with a dichotomous outcome. The model combines IPD and ag- gregated study data on multiple competing treatments and incorporates possible differences in study duration while accounting for patients drop- ping out without making any imputation for their outcomes. Three states are included in the model: response to the treatment, non-response, and study discontinuation. Our model takes into consideration the exact time of each observation, thus allowing for time-dependent relative treatment effects. Results: We apply our model to compare the effectiveness of treatments for schizophrenia. Our model produces a joint estimation of the relative treatment effects and the dropout rates as a function of time and increases precision compared to popular imputation methods Conclusions: The suggested model constitutes a viable candidate for performing network meta-analysis in the presence of non-ignorable miss- ing data, with studies reporting on multiple time points and in different formats. C26.2 Investigating consistency of mixed treatment comparisons by approximating sub-networks J König1 , U Krahn1 , H Binder1 1 University Medical Center Mainz, IMBEI, Mainz, Germany   In network meta-analysis, evidence of different studies is pooled, each comparing only few treatments. The results are network-based effect es- timates for all pairs of treatments, taking indirect evidence into account. These mixed treatment comparisons are based on linear combinations of effect estimates, with coefficients that form a network of flows in a weight- ed directed acyclic graph. Consistency is crucial for the validity of these network based effect estimates. However, global assessment of consis- tency ignores which part of the network informs specific treatment com- parisons. We show up a way how to tailor investigation of consistency to a specific comparison.We therefor construct approximating sub-networks that capture the bulk of evidence relevant to the comparison at hand, and define the approximating evidence proportion and a residual evidence based pseudo effect for assessing goodness of fit. As sub-networks we consider independent path decompositions, obtained via shortest path algorithms, which are amenable to a forest plot display, and more gen- erally structured sub-networks obtained by selection based on evidence weights. The methods can be used both in fixed and random effects mod- els. When applying them to networks of antide­pressants and thrombolyt- ics, we achieve approximating evidence proportions of over 90% while discarding half of the designs. The corresponding visualizations highlight how evidence along a few short paths is pooled into a given mixed treat- ment comparison. In particular, we are able to identify mixed treatment comparisons that rely on evidence distributed evenly and consistently over different independent sources and thus might be more reassuring than others.

Pages Overview