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

ISCB2014_abstract_book

72 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsWednesday, 27th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C39.4 Multi-state models for treatment success after stem cell transplantation LC de Wreede1,2 , J Schetelig2,3 , CJM Halkes4 , H Putter1 1 Medical Statistics, Leiden University Medical Center, Leiden, The Netherlands, 2 Clinical Trials Unit, DKMS, Dresden, Germany, 3 Medical Dept. I, University Hospital Carl Gustav Carus, Dresden, Germany, 4 Hematology, Leiden University Medical Center, Leiden, The Netherlands   The use of multi-state models to model complex disease histories has been advocated for over a decade; however, its use in clinical applications has been limited so far. This is especially striking in the field of hemato- poietic stem cell transplantations, since many examples in the statistical literature on this topic come from this field. We will show two examples where we tried to bridge the gap between statistical methodology and clinical questions. Our main outcome of interest is the probability of treatment succes over time. This outcome is both influenced by baseline characteristics and by intermediate events. Two related models will be used to analyse treatment success in 2 different transplantation settings. The first dataset describes a multi-center cohort of transplanted CLL (chronic lymphocytic leukemia) patients collected by the European Society for Blood and Marrow Transplantation. The second dataset gives detailed information about a cohort of acute leukemia pa- tients transplanted in Leiden UMC. In the first model, the impact of (base- line) covariates is considered. Dynamic prediction methods are applied to update the estimate of the probability of treatment success. All analyses will be performed by means of the‘survival’and‘mstate’pack- ages in R. Our examples show the potential of multi-state models in the analysis of clinically meaningful outcomes. The models are flexible and can easily be adjusted to different clinical questions. However, careful consideration of clinical aspects, data quality and limitations by small sample size are nec- essary to make applications successful. C39.5 Comparing multistate approaches for reducing the bias of relative risk estimates from cohort data with missing information due to death N Binder1,2 , M Schumacher1 1 Center for Medical Biometry and Medical Informatics, Freiburg, Germany, 2 Freiburg Center for Data Analysis and Modeling, Freiburg, Germany   In clinical and epidemiological studies information on the outcome of interest (e.g. disease status) is usually collected at a limited number of follow-up visits. The disease status can often only be retrieved retrospec- tively in individuals who are alive at follow-up, but will be missing for those who died before. Restricting the analysis to the survivors yields bi- ased hazard ratio estimates of a potential risk factor, and the bias can be in either direction. We focus on two approaches that use the same likelihood contributions derived from an illness-death multistate model for reducing this bias by including the death cases into the analysis: first, a penalized likelihood ap- proach by Leffondré et al. (Int J Epidemiol, 2013) and second, an imputa- tion based approach by Yu et al. (Biom J, 2010). We compare the two ap- proaches in simulation studies and evaluate them on completely recorded real data, where missing information due to death is artificially induced. For several scenarios, the bias is seen to be reduced compared to an ad- hoc analysis that right-censors the death cases at the last visit.   C40 Model performance evaluation C40.1 A new measure of predictive ability in a survival model: the total gain statistic B Choodari-Oskooei1 , P Royston1 , MKB Parmar1 1 MRC Clinical Trials Unit at UCL, London, United Kingdom   The results of prognostic factor studies are usually summarized in the form of statistics resulting from statistical significance testing, i.e. estimated parameters, confidence intervals, and p-values. These statistics do not inform us about whether prognostic factor information will lead to any substantial improvement in the prognostic assessment. Predictive ability measures can be used for this purpose since they provide important in- formation about the practical significance of prognostic factors. R2 -type indices are the most familiar forms of such measures in survival models, but they all have limitations and none is widely used. In this talk, we extend the total gain (TG) measure, proposed for a logistic regression model, to survival models and explore its properties using sim- ulations and real data. TG is based on the binary regression quantile plot, otherwise known as the predictiveness curve. Standardised TG ranges from 0 (no explanatory power) to 1 (`perfect´ explanatory power). The results of our simulations show that unlike most of the other R²-type predictive ability measures, TG is independent of random censoring. It increases as the effect of a covariate increases, and it remains largely unaffected by the categorisation of continuous prognostic factors. Furthermore, it can be applied to different types of survival models, in- cluding models with time-varying effects. Finally, we appliedTG to quanti- fy the predictive ability of prognostic models developed in several disease areas. On balance, TG performs well in our simulation studies and can be recommended as a measure to quantify the predictive ability in survival models. C40.2 A note on the time-profile of time-dependent area under the ROC curve for survival data J Lambert1 , R Porcher1,2,3 , S Chevret1,4,5 1 Inserm U1153, Paris, France, 2 Hôpital Hôtel Dieu, Paris, France, 3 Université Paris Descartes, Paris, France, 4 Hôpital Saint Louis, Paris, France, 5 Université Paris Diderot, Paris, France   In the setting of survival analysis, the time-dependent area under the re- ceiver operating characteristic curve (AUC) has been proposed as discrim- ination measure of interest. In contrast with the diagnostic setting, the definitions of time-dependent sensitivity and specificity are not unique and three time-dependent AUC are used in practice: cumulative/dynamic, incident/dynamic and incident /static. This work evaluates the time-de- pendent profile of these AUC(t). We show that, even when the effect of a binary biomarker on the haz- ard rate is constant, the value of AUC(t) varies over time according to the prevalence of the marker. The time- profile of continuous biomarker is il- lustrated with a simulation study, and data on several prognostic factors in AML are examined.  

Pages Overview