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


80 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsWednesday, 27th August 2014 • 16:00-17:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust In particular, validation is based on the Bayesian testing of a hypothesis regarding possible values of AUC. Toward this aim, first, available informa- tion is translated into a prior distribution. Next, this prior information is used in a Bayesian design to estimate biomarker’s accuracy. This allows reducing the sample size as compared to the “classical”, frequentist ap- proach, in which the availability of any information on biomarker’s diag- nostic performance is ignored. Results: A simulation study is performed to evaluate the power of the proposed design. For each scenario, 200 studies of sample size 100, 400, 600, and 800 are simulated. The power of the Bayesian design to confirm a satisfactory diagnostic performance of a biomarker is compared to a cor- responding frequentist design. Conclusion: For each study size, the proposed Bayesian design leads to a significantly higher power than the frequentist design. In some of the con- sidered simulation settings, the Bayesian design required as little as ¼ of the frequentist‑trial sample size to reach approximately the same power.   C45 Multistate models and competing risks II C45.1 Regression models for expected length of stay MK Grand1 , H Putter1 1 Leiden University Medical Center, Leiden, The Netherlands   A multi-state model is a stochastic process with outcomes in a finite space that represents the states. The expected length of stay (ELoS) is defined as the time the process is expected to spend, in total, in a given state. ELoS is not a straightforward object to relate to covariates and the traditional ap- proach has been to construct regression models for the transition intensi- ties, and from these calculate ELoS. The disadvantage of this approach is that the effect of covariates on the intensities are not easily translated into the effect on the ELoS. Furthermore, it typically relies on the assumption that the process is Markov. We propose using pseudo-observations (Andersen et al., Biometrika 2003) to make regression models for ELoS, thereby allowing a direct interpre- tation of covariate effects and evading the Markov assumption. For this approach, all we need is a non-parametric (asymptotically) unbiased esti- mator for ELoS. For every subject (and for every state of interest) a pseudo- observation is constructed and they are then used as outcome variables in the regression model. We furthermore show how to construct longitudinal (pseudo-) data when combining the concept with landmarking. Covariates may then be time- varying and potential time-varying effects can be explored. The models can be fitted using generalised estimating equations and by applying the sandwich estimator. The method is illustrated using data from the US Health and Retirement Study to explore the impact of socio- economic factors on ELoS in health and disability. The efficiency of our approach is investigated through simulations.   C45.2 A multistate model to assess the impact of menstrual status in premenopausal breast cancer patients S Weber1 , M Schumacher1 1 Center for Medical Biometry and Medical Informatics, Freiburg, Germany   Adjuvant treatment in premenopausal breast cancer patients may affect the menstruation, a cessation of menses is possible. Depending on treat- ment, recovery of menses may occur afterwards. How does the menstrual status impact disease-free survival? The motivating question leads to a multistate model which allows the analysis of subsequent events concerning menstrual status and the re- spective transitions into the absorbing state defined by tumor recurrence and death. Problems like competing risk, right-censoring and left-trunca- tion have to be considered. The Zoladex Early Breast Cancer Research Association (ZEBRA, Jonat et al, Journal of Clinical Oncology 2002) study compares a hormone therapy with goserelin and a chemotherapy as adjuvant treatment in premeno- pausal patients with node-positive breast cancer. Since goserelin works via suppressing the ovarian estrogen production it induces cessation of menses but recovery is possible. We investigate the effects of time-dependent menstrual status within a multistate model thereby using several Cox models including time-de- pendent covariates. Furthermore, results of a similar clinical trial (trial VIII) of the International Breast Cancer Study Group are considered and compared with those de- rived from the ZEBRA study. C45.3 Variable selection in the illness-death model M Lauseker1 , J Schemenau2 , U Germing2 , VS Hoffmann1 1 Ludwig-Maximilians-Universität München, München, Germany, 2 Heinrich-Heine-Universität Düsseldorf, Düsseldorf, Germany   Preferably, variable selection should be done by content-related reflec- tions. But this is often not possible. The illness-death model is used more and more frequently, but recommendations on variable selection are rare. Data were simulated according to an illness-death model without recur- rent events. Using this simulated data a semi-Markov model was fitted. Simulated data sets included an interaction effect between the indepen- dent variables and a strong correlation between two variables. Backward selection based on AIC and BIC was used on the one hand di- rectly, on the other hand applying a bootstrap step. The performance of the selection procedures was measured via the inclusion fraction and the bias of the estimated coefficients. In the simulations, both selection criteria yielded reasonable models. As expected, BIC led to more parsimonious models than AIC, regardless if bootstrapping was used. BIC performed slightly better with regard to both inclusion fraction and bias of the coefficients. Bootstrapping did not generally improve the results. The results were illustrated with a real world data set on myelodysplastic syndromes (MDS) patients concerning an illness-death model with the states “MDS”, “acute myeloid leukaemia” and “death”. In this data a time- dependent effect was only detected by AIC. Model selection via AIC and BIC works well for illness-death models, even though the true model was only found infrequently. Especially when strong correlation between the covariates is present, bootstrapping can lead to difficulties. C45.4 Statistical models for improving prognosis of chronic cardiovascular diseases: hazard reconstruction and clustering of patients affected by heart failure F Ieva1 , AM Paganoni2 , T Pietrabissa2 1 Department of Mathematics, Università degli Studi di Milano, Milano, Italy, 2 Department of Mathematics, Politecnico di Milano, Milano, Italy   Heart Failure (HF) is nowadays among the leading causes of repeated hos- pitalisations in over 65 patients. The longitudinal dataset resulting from the discharge papers and its analysis are consequently becoming of a great interest for clinicians and statisticians worldwide in order to have insights of the burden of such an extensive disease.

Pages Overview