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


90 ISCB 2014 Vienna, Austria • Abstracts - Poster PresentationsMonday, 25th August 2014 • 15:30-16:00 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust We extracted the Bayesian inference methods characteristics (type, aim, justification, etc.), involvement of a statistician, clinical domain, and the study quality according to STARD and SAMPL guidelines. Results: Among the 1666 identified articles, 281 were eligible, and 88 in- cluded. 44 (50%) papers were new methodological development articles, 29 (33%) clinical articles, 15 (17%) were intermediate between method- ological development and clinical studies, and mostly re-analysis of clini- cal datasets using Bayesian methods. Articles were classified based on an arborescence assessing the evolution of the methods and their use over time. Among the intermediate and clinical groups, more than 68% dealt with biological and microbiological concerns; and only 10 papers con- cerned practical clinical issues. Conclusion: Although there has been an increasing usage of Bayesian inference methods in the last few years with many methodological de- velopments, their use remains limited in practice. The spread of Bayesian methods for those studies should be fostered. P1.1.32 A Bayesian hybrid adaptive design for phase III survival trials M Moatti1 , WF Rosenberger2 , S Chevret1 1 INSERM UMR 1153, Université Paris 7, Paris, France, 2 Department of Statistics, George Mason University, Fairfax, United States   A few papers have described adaptive survival phase III randomised clini- cal trials in a frequentist context. We extend the method of Zhang and Rosenberger who proposed a response-adaptive randomization proce- dure that targets an optimal allocation for parametric survival data. However, while they explored only fixed sample size designs, we incor- porate an interim monitoring plan for estimating the log hazard ratio and propose stopping rules. Moreover, the main extension is in the Bayesian context, where we establish for the log hazard ratio a prior distribution based on either Spiegelhalter’s skeptical or enthusiastic normal priors, or a normal mixture derived from experts´ opinions. Combining the prior with the normal likelihood, the mean posterior estimate of the log hazard ratio allows deriving the optimal target allocation. We perform a simulation study to assess and compare the performances of this proposed Bayesian hybrid adaptive design to those of fixed, se- quential or adaptive frequentist designs.When using stopping rules, there was a gain in reducing the proportion of observed deaths in adaptive vs. non adaptive designs; this gain was maximal using a Bayes mixture prior. Such Bayesian hybrid adaptive survival trials may appear promising alter- natives to reduce the duration and the costs of survival trials, as well as optimizing the ethical concerns for patients enrolled in the trial.   P1.1.36 Copula functions in the presence of cure fraction EA Coelho-Barros1 , J Achcar2 , J Mazucheli3 1 UTFPR, Cornélio Procópio, Brazil, 2 USP, Ribeirão Preto, Brazil, 3 UEM, Maringá, Brazil   We introduce bivariate Weibull distributions derived from copula func- tions in presence of cure fraction, censored data and covariates. Two copula functions are explored: the FGM (Farlie - Gumbel - Morgenstern) copula and the Gumbel copula. Inferences for the proposed models are obtained under the Bayesian approach, using standard MCMC (Markov Chain Monte Carlo) methods. An illustration of the proposed methodol- ogy is given considering a medical data set. The use of copula functions could be a good alternative to analyse bivariate lifetime data in presence of censored data, cure fraction and covariates. Observe that in many applications of lifetime modelling we could have the presence of a cure fraction for individuals that are“long term survivors”or “cured individuals”.   P1.1.43 Reliable confidence intervals for fractional polynomials: a simulation study D Dunkler1 , M Gregorich1 1 Medical University of Vienna, CeMSIIS, Vienna, Austria Nonlinear associations of continuous risk factors with an outcome of in- terest can be modeled by fractional polynomials (FPs, Royston & Altman, 1994). Specifically, one or two elements of a polynomial transformation of the original variable are selected such that in subsequent regression analysis an optimal fit is obtained. This selection is usually ignored when constructing confidence intervals (CI) for the expected outcome at differ- ent values of the risk factor, or for contrasts between different risk factor values. This can lead to undercoverage if the selection is not stable, or if FPs are not flexible enough to capture a specific underlying shape of non- linearity. Reliable CI can be obtained by the bootstrap, but this requires a large number of models to be evaluated. Therefore, we consider some instant methods for CI estimation, such as model-based CIs with three degrees of freedom (MB3DF) to account for selection of two powers, and Bayesian model averaging (BMA) of several evaluated FP models. Using the setting of logistic regression, we compared these approaches with simple model-based CIs and bootstrapped CIs in a simulation study, assuming various nonlinear associations between a continuous risk factor and a binary outcome. Our simulations showed that with a true linear association BMA proves satisfactorily, and MB3DF may overcover. With nonlinear associations that can be modeled with FPs, both methods improve over simple model- based CIs without increasing the computational demand. In particular, we recommend the implementation of Bayesian model averaging CIs in software for FP estimation to stimulate their practical use.   P1.1.107 Determination of the minimum effective dose for correlated dose-response data using Bayesian variable selection (BVS) models LK Muchene1 , M Otava1 , Z Shkedy1 , T Jacobs2 1 Universiteit Hasselt, Center for Statistics, Diepenbeek, Belgium, 2 Janssen Pharmaceutical Companies of Johnson and Johnson, Beerse, Belgium In drug development, the determination of a minimum effective dose for a compound is of primary interest. Classically this involves testing the dif- ference in means of multiple doses against the mean in the first dose-level (typically, the control group) using analysis of variance with correction for multiple testing. The first dose level for which a significant difference is detected is defined as the minimum effective dose (MED). Alternatively, Bayesian variable selection (BVS) models can be used for se- lecting the most probable model for the dose -response relationship given a set of known candidate models. The model with the highest posterior model probability is selected and the MED is determined based on the selected model. Hence, the BVS model allows to estimate the MED taking into account model uncertainty. We apply Bayesian variable selection techniques to data from a differential reinforcement of low-rate 72 seconds (DRL-72) experiment.The DRL-72 ex- periment is commonly used in testing for clinically active anti-depressant compounds. In such an experiment, a rat pressed a lever and is expected to wait 72 seconds between two presses in order to receive a reward. The number of times the rat presses a lever is Poisson distributed and the number of rewards obtained is binomial distributed.The Bayesian variable selection model is applied to the data using joint binomial/Poisson mod- els while correcting for the design of the experiment, i.e. correlated mea- surements, cross-over drug administration design and over-dispersion in outcomes. Keywords: DRL-72, Bayesian variable selection models, hierarchical Bayesian models, Minimum effective dose.

Pages Overview