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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 15Monday, 25th August 2014 • 9:00-10:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust We aim to assess and compare the properties of these strategies and iden- tify their strengths and weaknesses. First results of a simulation study for a continuous response variable in linear regression with two covariables with a spike at zero will be presented. Different aspects, such as how the distribution of zero and nonzero values influences the model fit, will be investigated. We measure the accuracy of the data fit calculating mean squared errors and R² separately in 4 categories of observations determined by the spike variables. References: Royston et al. Stat. Med. 2010; 29: 1219-27. Becher et al. Biom. Journal 2012;54: 4. 686-700.   C03.3 A new measure of association based on non-linear regression M Scharpenberg1 , W Brannath1 1 University of Bremen, Bremen, Germany Kernel smoothers and Spline methods are popular non-linear regression techniques that often lead to a better understanding of the regression dependencies than linear regression but are mainly used for descriptive purposes only. In order to derive some type of inferential conclusion from the non-linear regression curve they are often supplemented by confi- dence bands. However, there is no simple way to quantify the overall as- sociation between the variables from such type of non-linear regression analysis. In this talk, we present a new non-linear measure of association (called “mean impact”) which enables us to quantify the overall associa- tion between the target and independent variable when fitting a non- linear regression curve. The idea is to consider the maximum change in the population mean of the target variable when the distribution of the covariates is changed in a suitably standardized way. We show that linear and non-linear regressions provide conservative estimates for our new model-independent measure of association. Furthermore, we derive con- fidence intervals for the new association parameter based on normal ap- proximations as well as bootstrap based confidence intervals.The method is illustrated with examples and investigated in a simulation study. References: Brannath and Scharpenberg (2014). Interpretation of linear regression co- efficients under model-misspecification. In preparation. Martin Scharpenberg (2012). A population-based approach to analyse the influence of covariates. Diploma Thesis in Mathematics. University of Bremen.   C03.4 Weighted mean impact analysis S Burger1 , W Brannath1 1 University of Bremen, Bremen, Germany Linear regression analysis is a popular tool that is often applied in medi- cine and epidemiology. It relies on strong model assumptions that are rarely satisfied. To overcome this difficulty Brannath, Scharpenberg (2014) and Scharpenberg (2012) proposed a new population based interpreta- tion of linear regression coefficients. The idea is to quantify how much the unconditional mean of the dependent variable Y can be changed by changing the distribution of the independent variable X, the maximum change is called “mean impact”. They show that linear regression can be used to obtain conservative estimates of the mean impact and other population based association measures. This provides a clear interpreta- tion of linear regression coefficients also under miss-specifications of the mean structure. A disadvantage of the new association measures is their dependency on the distribution of the independent variables in the spe- cific study population. Hence, it may be difficult to compare the results between different studies with differing covariate distributions. To over- come this difficulty we develop a method to transfer the “mean impact” from one study population to another by reweighting the observations. Accordingly, we call the resulting estimates “weighted mean impact”. We derive the asymptotic distribution and develop bootstrap confidence in- tervals for the weighted mean impact, and we illustrate the utility of the new method by examples and results from a simulation study. References: Brannath, Scharpenberg (2014). Interpretation of linear regression coef- ficients under model-misspecification. In preparation. Martin Scharpenberg (2012). A population-based approach to analyse the influence of covariates. Diploma Thesis. University of Bremen. C03.5 Non-parametric self controlled case series method Y Ghebremichael-Weldeselassie1 , H Whitaker1 , P Farrington1 1 The Open University, Milton Keynes, United Kingdom The self-controlled case series (SCCS) method is an alternative to cohort and case control study designs used to investigate potential associations between vaccine or other drug exposures and adverse events (side ef- fects). It requires information only on cases, individuals who have experi- enced the adverse event at least once, and automatically controls all fixed confounding variables that could modify the true association between exposure and adverse event. Time-varying confounders (such as age, sea- son), on the other hand, are not automatically controlled. The SCCS method has been extended by modelling only the age effect or only the time-varying exposure effect using splines while representing the other by a piecewise constant step function. In these two extensions, there is a need to pre-specify exposure groups or age groups a priori. Mis- specification of these groups may lead to biased association between ex- posure and adverse event. In this talk, we propose a non-parametric SCCS method in which both age and exposure effects are represented as linear combinations of cubic M-splines at the same time. Cubic M-splines are piecewise polynomials of degree 3. To avoid a numerical integration of product of two spline func- tions in the likelihood function of the SCCS method we defined the first second and third integrals of I-splines based on the definition of integral of M-splines. Simulation studies showed that the new method performs well. This new method is applied to a data on paediatric vaccines. C04 Dose finding studies C04.1 Dose finding methods based on longitudinal ordinal data: Realistic prior hypotheses identified from 49 phase I studies X Paoletti1 , E Rizzo2 , L Collette2 1 Institut Curie, Paris, France, 2 EORTC Headquarters, Brussels, Belgium Context: Recently, dose finding designs have been extended to incorpo- rate ordinal graded toxicity with the proportional odds (PO) model, or re- peated toxicity measurements with mixed effect models. In a very large data warehouse of 63 studies of single targeted agents from the EORTC DLT-TARGETT initiative, we explored (i) the PO assumption for the relation between graded toxicity and both the dose and the cycle of treatment using residuals and score tests, (ii) the variability and the intra- patient serial correlation in longitudinal measurements using mixed-ef- fect models and markov chain modeling in cumulative logistic regression.

Pages Overview