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

ISCB2014_abstract_book

60 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsWednesday, 27th August 2014 • 9:00-10:48 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust We present a new implementation of Bayesian variable selection for sur- vival analysis under the Weibull regression model which is based on a Reversible Jump MCMC algorithm. In a realistic simulation study, we dem- onstrate superior power and specificity in comparison to an alternative LASSO based feature selection strategy. Subsequently we present a real data application in which 119 protein-based biomarkers are explored for association with breast cancer survival in a case cohort of 2,287 patients with ER-positive disease. Our method outperformed alternative strategies to provide evidence for three independent prognostic biomarkers of sur- vival, one of which is novel. C31.5 Approximate Bayesian model selection with the deviance statistic L Held1 , D Sabanés Bové2 1 University of Zurich, Zurich, Switzerland, 2 Roche, Basel, Switzerland   Bayesian model selection poses two main challenges: the specification of parameter priors for all models, and the computation of the resulting Bayes factors between models. There is now a large literature on automat- ic and objective parameter priors, which unburden the statistician from eliciting them manually in the absence of substantive prior information. One important class are g-priors, which were recently extended from lin- ear to generalized linear models. We show that the resulting Bayes factors can conveniently and accurately be approximated by test-based Bayes factors using the deviance statis- tic. For the estimation of the hyperparameter g, we show how empirical Bayes estimates correspond to shrinkage estimates from the literature, and propose a conjugate prior as a fully Bayes alternative. Considerable computational gains are obtained which enable an exhaustive evaluation of the model space in moderate size variable selection problems without the need to employ MCMC methods. We illustrate the methods with the development of a clinical prediction model for 30-day survival in the GUSTO-I trial, and with variable and func- tion selection in Cox regression for the survival times of primary biliary cirrhosis patients. This is joint work with Daniel Sabanès Bovè.   C31.6 A novel variable selection method for Monte Carlo logic regression M Malina1 , F Frommlet1 1 Medical University of Vienna, CeMSIIS, Vienna, Austria   Logic regression is a promising approach to detect epistatic effects in genetic association studies. A Bayesian version of logic regression called Monte Carlo Logic Regression (MCLR) is based on Markov Chain Monte Carlo (MCMC) search, where the relevance of a specific epistatic term is assessed by the frequency of appearance of this interaction among all models visited by the MCMC search. One problem with this approach is that the number of possible logic ex- pressions is growing rapidly with the number of observed genetic markers. Therefore the MCMC search might not visit relevant models often enough to provide reliable estimates of model posterior probabilities, and alterna- tive estimates are desirable. The aim of this talk is to compare MCLR with a novel variable selection approach based on model posterior estimates using Laplace approximation like in Schwarz BIC, but in combination with the more appropriate geometric prior distribution for the model size. Results of a systematic simulation study and real genetic data analysis are presented, which illustrate the properties and benefits of the newly pro- posed selection method to detect epistasis. C32 Longitudinal data analysis II C32.1 Mixed-effects location scale model for time to event data DS Courvoisier1,2 , TA Walls3 , D Hedeker4 1 University Hospitals of Geneva, Geneva, Switzerland, 2 Harvard School of Public Health, Boston, United States, 3 University of Rhode Island, Kingston, United States, 4 University of Illinois at Chicago, Chicago, United States   Compared to mixed-effects models with only a random intercept (i.e., mixed-effects location models), mixed-effects location scale (MELS) model have several advantages. First, these models allow researchers to include covariates not only on the random (between-subject) intercept, but also on the random within subject residual variance. Including covariates on the between-subject variance allows researchers to, for example, test whether drug A has a more similar action over all individuals (blood pres- sure decreases by, on average, 10mmHg +/- 5mmHg) than drug B (blood pressure decreases by, on average, 10mmHg +/- 20mmHg). Furthermore, including covariates on the within-subject variance allows researchers to test hypotheses about how well specific measurements can be estimated within each individual. A second advantage concerns the estimation of a random scale, which captures the unexplained variation in within-indi- vidual variability. MELS have been developed for outcomes that follow a normal or ordinal distribution. In this presentation, we extend this model to time to event data, allowing for censoring. We illustrate the model with data on time to first cigarette measured every day for 7 days in smokers randomized to a group asked to keep smoking or another group asked to stop.   C32.2 Bayesian growth mixture models to distinguish hemoglobin value trajectories in blood donors K Nasserinejad1 , JV Rosmalen1 , D Rizopoulos1 , WD Kort2 , M Baart2 , KV den Hurk2 , E Lesaffre1,3 1 Department of Biostatistics, Erasmus MC, Rotterdam, The Netherlands, 2 Sanquin Blood Supply, Nijmegen, The Netherlands, 3 L-Biostat, KU Leuven, Leuven, Belgium   Blood donation leads to a temporary reduction in the hemoglobin level necessitating a period after donation for the hemoglobin value to recover to its pre-donation level. A premature invitation may therefore result in a too low hemoglobin value, a deferral from donation and may demotivate the candidate donor. All in all this implies an inefficient planning of the donation process. The trajectory of the hemoglobin level after donation and the duration of the recovery period may differ between individuals. Here, we aim to clas- sify the longitudinal hemoglobin values measured in blood donors. More specifically, we wish to detect early which donors will run the risk of a too low hemoglobin value (8.4 mmol/l for men, 7.8 mmol/l for women). For this, we used a random sample of 2000 new-entrant whole-blood donors who have donated repeatedly during 2005-2012. This data set was col- lected by Sanquin Blood Supply (the Netherlands). To capture the unobserved heterogeneity of hemoglobin profiles, we implemented a Bayesian growth mixture model. This model assumes that each donor belongs to one of several latent classes. Within each class, the hemoglobin trajectory follows a linear mixed model. In addition we let the latent class membership depend on the age and hemoglobin value at first visit. Our fitted growth mixture model suggests different classes of hemoglobin trajectories. This model gives some insight in the donation process and is a start to better predict for which donors care needs to be exercised not to produce a too low hemoglobin level.

Pages Overview