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

ISCB2014_abstract_book

74 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsWednesday, 27th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C41.2 Quantile regression and prediction intervals for survival data M Mayer1 , Q Li2,3 1 Consult AG Bern, Zurich, Switzerland, 2 Swiss Group for Clinical Cancer Research (SAKK), Bern, Switzerland, 3 University of Bern, IMSV, Bern, Switzerland   Cox models are by far the most traditional statistical modelling technique in survival data analysis, e.g. because the effects of predictor variables have a simple interpretation as hazard ratios. However, when used for pre- dicting survival times of individual patients (based on patient characteris- tics such as age, sex etc.), Cox models are unhandy and there is no simple way to quantify the precision of these individual predictions. A very powerful but still quite unknown alternative to the Cox model is quantile regression, originally introduced into survival data analysis by J. Powell in 1986. It allows modeling any quantile of the (log) survival time distribution, for instance the median and the two other quartiles, as a lin- ear function of the predictor variables. Quantile regression is almost as simple to use and to interprete as a multiple linear regression and is e.g. available in Roger Koenker’s“quantreg”library in R. We illustrate the method and its flavor using real survival data and show a trick how to use it to obtain “forecast” intervals for individual patients. Such an interval does not only quantify the precision of the corresponding point prediction but also answers the question“how long will I survive”in an honest and patient focused way. C41.3 A special case of the reduced rank model for modelling time varying effects in survival analysis A Perperoglou1 1 University of Essex, Colchester, United Kingdom   Consider the case of modelling time to event data, where the effect of some covariates on the hazard function might change with time. Starting from a proportional hazards model, one can introduce interactions of fixed covariates with time functions to model the dynamic behaviour of the effects. In Perperoglou et al (2006) Reduced Rank Hazard Regression was introduced as an approach to achieve parsimonious models with few parameters. The approach was further extended to include both fixed and time varying effects of the covariates (Perperoglou 2013). However, a seri- ous issue remain, which of the covariates in the model should be allowed to have time varying effects and which not. In this work we will present our findings on the suggestion of van Houwelingen and Putter, to fit a modified rank one model with all covari- ates having both time varying and time fixed effects. The special case of the rank=1 model can be written as: h(t|X)=h0 (t)exp(Xb1 +Xb2 γ´F´) where X is a matrix of covariates, b-s are the vectors of coefficients for the fixed effects and γ is a vector of coefficients for the interactions of fixed covariates with a matrix of time functions F. We will illustrate how to fit the model using an alternating least squares algorithm, the properties of this approach and results from a series of ap- plications in real and simulated data. C41.4 Estimating probability of non-response to treatment with survival data A Callegaro1 , B Spiessens1 1 GSK Vaccines, Rixensart, Belgium   The treatment effect reported from clinical trials represents the average of the individual benefit from treatment. Classical statistical approaches in cancer clinical trials evaluate treatment-effect heterogeneity by modeling the interaction between the treatment and some known covariates. However, often the underlying mechanism that causes variability is un- known and the relevant covariates are not observed. Further, if there is biological or empirical evidence that only a portion of treated patients respond to the treatment it is interesting to estimate the probability of patients to respond to the treatment. This probability could be used for personalized treatment selection i.e. to define a target population for a future comparative study. Different approaches to model the treatment effect heterogeneity for survival data (e.g. mixture models) will be pre- sented and compared by clinical trial simulations. Simulation results will give an idea of the amount of information (sample size, proportion of responders, treatment effect in responders) necessary to accurately estimate the probability of non-response to treatment on (oncology) trial data with survival outcome. Funding source: GlaxoSmithKline Biologicals SA   C41.5 An application of frailty modeling for family level clustering of infant mortality in Empowered Action Group states in India K Mani1 , RM Pandey1 1 All India Institute of Medical Sciences, New Delhi, India   Objectives: In India, the focused intervention policies led to a decline in mortality among children younger than five years, yet some of the states in India are having very high mortality rates.We explored the effects proxi- mate determinants on infant mortality by accounting for family level clus- tering using Cox frailty model in Empowered Action Group states (EAG) in India and compared the results with standard models. Methods: Analysis included 20,126 live births that occurred five years preceding the National Family Health Survey-3 (2005-06). The Cox frailty model was used to account for the family level clustering. Results: Of the 20126 live births, 1223 babies died before reaching their first birthday. The Cox frailty model showed that mother’s age at birth, composite variable of birth order and birth interval, size of the baby at birth and breastfeeding among proximate determinants were significant determinants of infant mortality after adjusting for familial effect. The fa- milial frailty effect was 2.52 in the EAG states. The inferences on the deter- minants for all the three models were similar except the death of a previ- ous child and mother’s age at birth in the Cox frailty model, which had the highest R2 and lowest log-likelihood. Public Health Implications: While planning for the child survival pro- gram in EAG states, parental competence which explains the unobserved familial effect needs to be considered along with significant proximate/ programmable determinants. The frailty model that provide statistically valid estimates of the covariate effects are recommended, when observa- tions are correlated.  

Pages Overview