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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Poster Presentations 111Tuesday, 26th August 2014 • 10:30-11:00 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust sensitivity analyses (n=252; IPW: OR=5.92, 95%CI: 1.80-19.42; MI: OR=3.12, 95%CI: 1.22-8.02). Conclusions: Anemia is a risk factor for poor cognitive outcome 3 months after ischemic stroke. Two different sensitivity analyses accounting for missing outcome data confirmed the negative impact of anemia.   P2.4 Penalized methods in high- and in low-dimensional regression analyses P2.4.23 Dimensional reduction in the flexible B-spline Cox model using functional principal components analysis MA Benadjaoud1,2,3 , H Cardot4 , F de Vathaire1,2,3 1 Radiation Epidemiology Group CESP - INSERM 1018, Villejuif, France, 2 Université Paris Sud, Le Kremlin-Bicêtre, France, 3 Institut Gustave Roussy, Villejuif, France, 4 Institut de Mathématiques de Bourgogne, Dijon, France   Radiation epidemiology is a rare events context, which restricts the num- ber of estimated parameters in flexible survival models. This requires the use of a low-dimensional spline regression constructed from a reduced number of interior knots or through Penalized splines (P-Splines) where the choosing of optimal degree of smoothness is crucial and still an issue even if various selection method were proposed such as Akaike informa- tion criterion (AIC), Bayesian information criterion (BIC) or generalized cross-validation (GCV). We propose a new dimension reduction technique based on a functional principal component analysis (FPCA) where the spline basis is replaced by a smaller number of score functions that summarize the effects of the initial interior knots sequence. We perform a simulation study in a flexible Cox regression context based on a dataset intended to reflect realistic radiation-epidemiological data. We vary the shape of the dose-response function, the number of interior knots, the number of principal components incorporated in the model. For assessing performance, we consider the integrated mean squared er- ror (IMSE). Finally, we apply the proposed method to investigate the expo- sure-response relation between the radiation dose to the thyroid and the radio-induced tumor risk. The FPCA estimator minimizes the IMSE better than the P-splines estima- tor and our rule of thumb is that 6 or 7 interior knots and 3 or 4 functional principal components are adequate for several practical situations. In conclusion, this study suggests the FPCA as a form of regularized es- timation which could represent an alternative to the classical penalized approach.   P2.4.56 Firth’s bias reduction method revisited: software implementation boosts application A Geroldinger1 1 Medical University of Vienna, CeMSIIS, Vienna, Austria   This presentation reviews David Firth’s bias reduction method for maxi- mum likelihood estimates of regression coefficients (Firth, Biometrika, 1993). Instead of correcting bias after estimation, Firth’s approach pre- vents bias by introducing a penalty to the likelihood function. This allows one to compute reliable finite regression coefficients even in the situation of separation. Separation, sometimes also termed‘monotone likelihood’, is frequently observed in models used in clinical biostatistics, including the logistic, the multinomial or the Cox model. Firth’s seminal publication was also followed by a series of papers discuss- ing different aspects of the penalization; for instance issues of inference, connections with Bayesian methods, interpretability of the estimates, or the application to high-dimensional predictor space. Some of these devel- opments are highlighted in our presentation. After the Firth correction was made available in add-on packages for stan- dard software, it has been widely used by many researchers to solve their small sample regression problems. Firth’s correction is now also available in the standard distribution of SAS, further enhancing its accessibility. We will provide an overview of implementations of Firth’s correction in statis- tical software, and will correlate the number of medical and non-medical citations of Firth’s method with the release dates of these software imple- mentations. This analysis reveals that besides methodological excellence, software availability is a very likely causative factor for getting many cita- tions. P2.5 Statistical methods for systems biology and genetics P2.5.40 Cancelled A two-stage approach to test for gene-gene interactions in family data based on within-family and between-family information L De Lobel1 , K Van Steen2 1 StatGent CRESCENDO - Ghent University, Ghent, Belgium, 2 Montefiore, Liege, Belgium   The search for susceptibility loci in gene-gene interactions imposes a methodological and computational challenge for statisticians due to the large dimensionality inherent to the modelling of gene-gene interactions or epistasis. In an era where genome-wide scans have become relatively common, new powerful methods are required to handle the huge amount of feasible gene-gene interactions and to weed out the false positives and negatives from these results. One solution to the dimensionality problem is to reduce the data by pre- liminary screening of markers to select the best candidates for further analysis. Ideally, this screening step is statistically independent of the testing phase. To obtain two independent steps to test for associations in family data, we can split up the genotypic information in a between- family and within-family component as is done in the QTDT. Those two components are orthogonal so that one of the components can be used for screening and the other can be used for testing. The QTDT proposes a definition of these components for one locus. In our research, we define analogous components for gene-gene interac- tions and investigate the proporties of this screening technique in differ- ent types of simulations.   P2.5.78 Topology-based pathway analysis of microarray and RNA-Seq data: an evaluation of existing methods I Ihnatova1 , E Budinska1,2 1 Institute of Biostatistics and Analyses, Brno, Czech Republic, 2 Masaryk Memorial Cancer Institute, Brno, Czech Republic   Pathway analysis methods for transcriptional microarray data analysis have reached their third generation, currently incorporating pathway topology information. However, high throughput parallel sequencing of transcriptome (RNA-Seq) has recently emerged as an appealing alterna- tive to microarrays and becomes widely available.

Pages Overview