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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 51Tuesday, 26th August 2014 • 11:00-12:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C26.3 Visualisation of networks in meta-analysis G Rücker1 1 Medical Center, University of Freiburg, Freiburg, Germany   In network meta-analysis, the network of treatments and comparisons should be visualised. In principle, there are innumerable ways to draw a complex network, all representations of the underlying graph in the plane. Several criteria exist for optimising this representation. Often a star- shaped presentation is chosen, where all vertices lie on a circle.This is easy to implement, but typically comes at the price of many crossings. Instead, it might be desired to avoid crossings as far as possible. For planar graphs, crossings can be completely avoided. Finally, a perspective view (i.e., a 2D projection of a 3D object) might be desired, particularly for regular geo- metrical objects such as a cube. Hall [Hall 1970] proposed to produce 2D representations of networks using as coordinates eigenvector coefficients of the Laplacian matrix, in order to select projections in which the distanc- es of neighboured edges are minimised. For regular planar networks, this often results in a nice perspective view when using the eigenvectors to the second and third smallest eigenvalues, whereas it results in star-shaped representations when using the eigenvectors to the largest eigenvalues. However, unfortunately, for irregular real networks this method does not work well. Alternative network visualisation algorithms aim at, e.g., keep- ing pre-specified `ideal´ distances between vertices, or minimising the number of crossings. We discuss the pros and cons of several criteria and approaches and give examples how to realise them using R. We argue that for sake of clarity a planar or perspective network representation is prefer- able to a star-shaped representation.   C26.4 Bayesian network meta-analysis for cluster-randomized trials L Uhlmann1 , K Jensen1 , M Kieser1 1 University of Heidelberg, Heidelberg, Germany   Cluster-randomized trials are used when randomization of single study participants is not possible. In the analysis of data from cluster-random- ized trials the correlation within clusters has to be taken into account, otherwise the type I error rate may be inflated. To combine the effects in a meta-analysis of cluster-randomized trials accordingly, either the corre- lation must be taken into account in the analysis of the single trials, or an adjustment of the effects or the variances must be performed. The variance inflation caused by the correlation can be taken into account in various ways. For classical pairwise meta-analysis, methods were pro- posed and their performance characteristics were compared. In our con- tribution we extend these approaches to network meta-analyses. In a first step, we illustrate how pairwise meta-analyses including cluster-random- ized trials can be conducted using a Bayesian approach. Furthermore, the models are extended such that they can be used to conduct multiple com- parisons in a meta-analysis. With these models, network meta-analyses of cluster-randomized trials can be performed. By use of simulation studies we evaluate the type I error rate to compare the derived methods. The results show that, in contrast to the unadjusted approach, our models do not lead to an inflation of the type I error rate. Finally, results of an investigation of the power characteristics are pre- sented.   C26.5 Precision of the estimates from a network meta-analysis model and their role in planning future studies A Nikolakopoulou1 , D Mavridis1,2 , G Salanti1 1 University of Ioannina, School of Medicine, Ioannina, Greece, 2 University of Ioannina, Department of Primary Education, Ioannina, Greece   When there are multiple competing interventions for a healthcare prob- lem the design of new studies could be based on the entire network of evidence as reflected in a network meta-analysis (NMA). There is a practi- cal need to answer how many (if any) studies are needed, of which design (the treatments being compared) and with what sample size to infer con- clusively about the relative treatment effects of all competing treatments and their relative ranking. We have previously addressed these questions based on the conditional power of NMA. Here we present methodology that approaches the same questions from a different angle. We consider the precision in the results obtained from NMA: the precision in the joint distribution of the estimated basic parameters of the model and the precision in the treatment rank- ing.We quantify the precision in the estimated effects by considering their variance-covariance matrix and estimate the precision in ranking by quan- tifying the dissimilarity of the density functions of summary estimates. Then, based on a target improvement in precision we calculate the re- quired sample size for each possible study design and number of study arms and we present visual tools that can help trialists select the optimal study design. We used a published network of interventions for the treat- ment of hepatocellular carcinoma to illustrate the suggested methodol- ogy. Results show that precision gain depends on the type of treatment com- parisons tested in new studies. The presented methodology can aid inves- tigators making informed and evidence based decisions about planning new studies.   C27 Survival analysis I C27.1 Survival probability with non-reversible time varying treatment indicator: theoretical quantities and nonparametric estimators L Antolini1 , DP Bernasconi1 , S Iacobelli2 , MG Valsecchi1 1 Università Milano Bicocca, Monza, Italy, 2 Università di Roma “Tor Vergata”, Roma, Italy   Inference on survival according to a non-reversible time varyingTreatment is often performed by applying the Cox model or the Mantel-Byar’s test while a reliable non-parametric description of the survival experience is still not fully established. Simon-Makuch (SM) curves were derived generalizing the Kaplan-Meier (KM) formula, initially classifying patients at a landmark time and dynami- cally updating the risk-sets for subsequent time-points. The curve for pa- tients switching treatment has been criticized, since it is unclear which quantity it estimates. The time-scale originates from start of standard treatment and the switch is considered as delayed entry into the alterna- tive treatment. SM estimates at time t from landmark are based on a mix- ture of individual hazards, heterogeneous with regard to time from switch. An alternative approach consists in using a clock-back scale where the survival is estimated at time from switch, i.e. considering homogeneous risks sets on this time-scale, but heterogeneous on the original one. Both curves estimate the counterfactual survival of patients under alternative

Pages Overview