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


ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 67Wednesday, 27th August 2014 • 9:00-10:48 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C36.3 Graph based multiple testing strategies for confirmatory adaptive enrichment designs T Sugitani1 , M Posch1 , F Bretz2 , F König1 , F Klinglmueller1 1 Medical University of Vienna, CeMSIIS, Vienna, Austria, 2 Novartis Pharma AG, Basel, Switzerland   An important objective in the development of targeted therapies is the identification of a population where the treatment has a clinically relevant effect. Recently, adaptive enrichment designs have been proposed that allow to restrict enrolment to a subpopulation after an interim analysis. In addition to the selection of a subpopulation, the sample sizes in the subgroups may be adapted. In this work we apply the recently proposed adaptive graph based mul- tiple testing procedures to the analysis of adaptive enrichment designs. These procedures control the familywise error rate for hypothesis tests in multiple (sub-)populations and can be easily extended to cover also tests of multiple endpoints. The definition of the testing procedure by a graph allows to map the difference in importance and the logical structure of the tested hypothesis to the testing procedure. In addition, the graph provides a convenient tool to communicate the testing procedure to the clinical study team. In a simulation study we assess the operating charac- teristics of the graphical adaptive testing approaches and compare them to testing procedures based on group sequential tests. Furthermore, we investigate how the power of the adaptive testing procedures can be op- timized by stratification of the hypotheses tests to adjust for the hetero- geneity of treatment effects in subgroups. Finally, the application of the adaptive graph based testing strategy is illustrated with a case study for the development of a targeted therapy.   C36.4 Likelihood ratio tests for multiple nonlinear models G Gutjahr1 , B Bornkamp2 1 University of Bremen, Bremen, Germany, 2 Novartis Pharma AG, Basel, Switzerland   Consider a set of nonlinear models that predict a mean vector of normally distributed observations and the hypothesis that at least one of these models fits the data significantly better than a constant model. For a single ``sufficiently smooth´´ model, Hotelling showed that the likeli- hood ratio test statistic is a monotonous function of the correlation be- tween the observations and the maximum likelihood prediction from the model; using methods from differential geometry, the exact null distribu- tions of this statistic can be obtained. For multiple models, the best predic- tion from the multiple models is used in the likelihood ratio test statistic. The null distribution is determined by volumes of tubular neighborhoods on the unit sphere. We describe how such volumes can be approximated numerically. This approach can also be used to calculate the distribution under alterna- tive hypotheses and it does not required that the models are smooth. We compare the power of the likelihood ratio test with locally most powerful tests and with multiple-contrast tests and apply it to data from a dose- response clinical trial.   C36.5 Are multiple outcomes analysed appropriately in randomised controlled trials? A systematic review V Vickerstaff1 , G Ambler1 , R Omar1 1 University College London, London, United Kingdom   Many procedures for addressing multiplicity in clinical trials have been in- troduced in the literature; however, the techniques are rarely used in prac- tice. Investigators often analyse multiple primary outcomes using several independent tests with no adjustments. Reporting several unadjusted p- values can increase the probability of erroneously rejecting at least one true null hypothesis. We performed a review to quantify how many trials analysed multiple primary outcomes and how many analysed them appropriately. We re- viewed all randomised controlled trials published July 2011-June 2013 in top neurology and psychiatry journals: American Journal Psychiatry, JAMA Psychiatry, Psychotherapy and Psychosomatics, Lancet Neurology and Neurology. Typically in these areas, data on multiple correlated outcomes are collected. We focused on the results in the abstract, methods used for sample size calculation and statistical analysis. We identified 154 randomised controlled trials of which 70 analysed multi- ple primary outcomes. Among these, 55/70 did not adjust for the multiple comparisons. If multiplicity was addressed, the significance of the results and trial conclusions would have changed in several papers. Of the 15/70 papers which accounted for multiplicity; 5 performed MANOVA, 6 used Bonferroni’s correction and 4 used other correction methods. Nine trials provided a sample size calculation which considered multiplicity. Our review shows that multiple primary outcomes are commonly analysed in clinical trials and are often inadequately handled. Further methodologi- cal research is necessary to assess the appropriateness of existing meth- ods for addressing multiplicity in different scenarios, particularly when the outcomes are correlated and to provide guidance on their use in practice. C36.6 A multiple testing procedure for three primary endpoints R Ristl1 , F Frommlet1 , M Posch1 1 Medical University of Vienna, CeMSIIS, Vienna, Austria   When efficacy of a treatment is measured by co-primary endpoints, ef- ficacy is claimed only if for each endpoint an individual statistical test is significant at a local level α. While such a strategy controls the family-wise error rate (FWER) at level α, it may be strictly conservative and have low power. We improve the test of three co-primary endpoints to allow infer- ence also in settings where only two out of the three show a significant re- sult at the local level. While the test does not allow to reject an elementary null hypothesis in this case, it rejects an intersection hypothesis such that an effect in at least one of the endpoints can be inferred and the trial still serves as a proof of principle. We show under the assumption of multivariate normal test statistics with arbitrary correlation matrix that the procedure controls the FWER at level α in the strong sense. Besides the application to tests for co-primary end- points the result uniformly improves the Rüger test in the setting of tri- variate normal test statistics. The latter rejects if two out of three hypoth- eses are significant at level 2α/3 but controls the type 1 error rate at level α without the assumption of multivariate normality. We investigate the power of the improved test procedure and compare it to hierarchical and Bonferroni tests for co-primary endpoints. The test procedure is illustrated with a clinical trial for a rare disease. An application of the procedure in the assessment of diagnostic tools is discussed.  

Pages Overview