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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 63Wednesday, 27th August 2014 • 9:00-10:48 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust The objectives of the present work are: 1) to adapt the Cortese method to parametric models; 2) to compare, using a simulation study, the performance of the method thus adapted with that of the Stare method; 3) to illustrate the use of these two methods on real data (none is currently used in practice). The performance criteria will be the ability to detect the non-proportion- ality of the effect of a covariate taking or not into account its linear or non- linear effect. Several scenarios were considered with changes in several modeling aspects, mainly the baseline distribution and the amplitudes of the parameters. Simulation results and an application to real cancer sur- vival data from French registries will be presented.   C33.6 Oblique decision trees for spatial clusters detection of net cancer survival rates N Grafféo1,2 , J Gaudart1,2,3 , K Ndiaye1,2 , R Giorgi1,2,3 1 INSERM, UMR912 (SESSTIM), Marseille, France, 2 Aix Marseille Université, UMR_S912, IRD, Marseille, France, 3 APHM, Hôpital Timone, BIOSTIC, Marseille, France   Net survival is the survival that would be observed, in a hypothetical world, if the disease under study was the only cause of death. Because of geographical variations in factors impacting on patients’ net cancer survival, spatial study is of particular interest. However, it relied on pre- specified administrative maps, which are not always appropriate in the case of epidemiological research. The goal of our work was to propose a method providing potential spatial clusters which could contain patients with similar net cancer survival rates at a given time without pre-specified boundaries. We extended to net survival analysis an oblique decision trees approach which had been developed for counted data. This non-parametric regres- sion model eliminates the need to define any specification of geographi- cal areas, shapes, or sizes of the clusters, provides potential aggregates with oblique partitions of the space, and allows adjusting on covariates. In this work, we used the Pohar-Perme estimator at time t, which yields consistent estimates of net survival. First, the algorithm splits the geo- graphic area into two adjacent partitions by maximizing the statistic of a Z-test, comparing net survival estimates between each potential split of angular sectors. Second, the algorithm goes on recursively until one of the proposed stopping rules is reached and the oblique decision tree is completed. Simulation studies will be used to investigate the performance of the pro- posed algorithm. Then, this method will be illustrated by a cancer popula- tion-based study. This approach could be useful to examine geographical variations in net survival rates. C34 Methodology C34.1 Constrained ordination analysis with an increased number of bell-shaped response functions with applications in metagenomics Y Zhang1 , O Thas1,2 1 Ghent University, Ghent, Belgium, 2 University of Wollongong, Wollongong, Australia   Ecologically meaningful bell-shaped responses of species to ecological gradients is a fundamental assumption of most current analytical meth- ods in community ecology. However, statistical methods often make no distinction between convex and concave response functions. The analy- sis output is therefore misleading and the conclusion are prone to errors. We identify this problem in classical model-based method, such as con- strained ordination analysis (COA), by means of several diagnostic graphi- cal tools. To solve the issue a penalty term similar to L1-penalization is proposed so as to penalize convexity in the likelihood ratio criterion. A fast method of determining tuning parameter is also introduced.   C34.2 Multivariate statistical process control for mixed-type data: an overview and a simulation study G Vidmar1 , N Majdič1 , R Blagus2 1 University Rehabilitation Institute, Ljubljana, Slovenia, 2 Institute for Biostatistics and Medical Informatics, Ljubljana, Slovenia   Multivariate statistical process control (MV SPC) based on mixed-type data (i.e., when some of the variables describing the process are numer- ic and some categorical) is a relatively new and undeveloped field. The usual approach to MV SPC addresses measurement data by construct- ing a Shewhart chart based on the Hotelling´s T2 statistic. We review the possibilities for MV SPC with mixed-type data and identify three main ap- proaches: multivariate outlier detection for mixed data; dimensionality reduction (via PCA, MDS or ICA) yielding numeric dimensions followed by T2 (or multivariate EMWA or multivariate CUSUM) control charts; and measuring distances between mixed-data points using Gower´s distance (i.e., Gower´s dissimilarity coefficient, Gower´s index or Gower´s general coefficient of similarity) and then constructing T2 charts, D2 charts (based on support vector data description, SVDD) or K2 charts (based on k-near- est neighbours data description, kNN). The control limits for the D2 and K2 charts are established via bootstrapping, whereby distances from the whole phase I sample (global) or just from the kNN (local) are considered. We present a simulation study comparing the Gower´s-distance-based ap- proaches and the T2 approach with categorical variables coded as binary indicator variables. The highly realistic simulations are based on a planned setup for health-care quality monitoring in the field of rehabilitation after lower-limb amputation. The results indicate that the Gower´s distance ap- proach improves as the number of categorical variable increases and that the local Gower´s-distance-based K2 chart outperforms the global one.  

Pages Overview