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


52 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsTuesday, 26th August 2014 • 11:00-12:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust treatment from beginning, whereas the survival of patients under stan- dard treatment represents the counterfactual net survival where treat- ment switch is removed. Motivated by an international study on acute lymphoblastic leukemia with patients undergoing chemotherapy and with possible additional stem cell transplant, we aim to identify survival probabilities and suitable nonparametric estimators. In particular, we will compare the SM survival curve as well as the KM “clock back” curve, discussing the assumptions (e.g. Markov property) of both methods. Alternatives derived in the framework of multi-state models will be con- sidered.   C27.2 Using pseudo-values for comparing long-term survival after stem-cell transplantation (SCT) with long-term survival after chemotherapy U Pötschger1,2 , H Heinzl2 , MG Valsecchi3 , M Mittlböck2 1 Children’s Cancer Research Institute, Vienna, Austria, 2 Center for Medical Statistics, Medical University, Vienna, Austria, 3 Center of Biostatistics for Clinical Epidemiology, Monza, Italy   Aims: Allogeneic SCT is a therapeutic option in high-risk leukemia whose ability to improve long-term survival is still under discussion. From a meth- odological perspective, when investigating long-term survival rates, it is unclear and challenging how to adjust for waiting-time until donor identi- fication without relying on proportional hazards. Methods: With non-proportional hazards, the pseudo-value regression provides an attractive approach. An adaptation can be introduced that allows for a time-dependent covariate. Pseudo-values for survival expec- tations of patients with transplantation are generated which address the cumulative hazards before and after transplantation. These are compared in a generalised linear model to pseudo-values for baseline survival rates without SCT. Real data in leukaemia and a simulation study illustrate the practical value and the statistical properties of the novel approach. Results: The simulations show unbiasedness of the estimated parameters. In the real data example, SCT has worse early survival due to toxicities but better survival later on. Consequently Cox-regression is not suitable (HR=0.91, p=0.39) due to non-proportional hazards. However the pseudo- value approach shows that the cumulative hazards at 5-years under SCT is favorable to no SCT (HR=0.67, p=0.02). Additionally no dependence be- tween waiting time and resulting 5-years survival was observed. Conclusion: The proposed novel approach allows investigating the im- pact of a binary time-dependent covariate on long-term survival without relying on proportional hazards.   C27.3 Assessing effects of treatment change on survival when the measurement pattern of covariates and events are dependent I Schmidtmann1 , A Weinmann2 , A Schritz1,3 , D Zöller1 , H Binder1 1 IMBEI, Universitätsmedizin Mainz, Mainz, Germany, 2 I. Medizinische Klinik, Universitätsmedizin Mainz, Mainz, Germany, 3 Hochschule Koblenz, Rhein-Ahr-Campus, Remagen, Germany   Time-to-event data arising from clinical cancer registries often involve competing risks or other complex event patterns, such as time-depen- dent confounding, which may occur when treatment changes are driven by time-varying covariates that both influence prognosis and depend on previous treatment. Extensions of the Cox regression model allow for such complex event structures. A further complication arises when the pattern of measuring such time-varying covariates is irregular and updates are more frequent depending on patient health status or imminent treatment change. We investigate the effect of a treatment change in a hepatocellular carci- noma registry by considering conditional survival at several update time points after start of initial treatment. Covariates considered here include treatment (changed / not yet changed), baseline covariates, and covariate values prior to the update time points. The influence of the measurement pattern of the time-varying covariates is examined by fitting regression models using alternatively all available covariate values and covariate values from artificially coarsened measurement patterns. As illustrated for the Mainz hepatocellular carcinoma registry, the difference in regres- sion coefficients then indicates the potential extent of bias due to event- associated measurement patterns. Use of a propensity score approach is considered for reducing this bias while still retaining the information from all measurements.   C27.4 Clustering for treatment effect on recurrent events C Schramm1,2 , G Diao3 , S Katsahian1,3 1 INSERM UMRS 1138, Team 22, Paris, France, 2 INSERM U955, Team 1, Créteil, France, 3 Department of Statistics, George Mason University, Fairfax, United States   Recurrent events occur when a subject experiences repeated occurrences of the same event over follow-up time. Rate of events gives information about disease gravity. A specific treatment must reduce this rate and im- prove the quality of life of patients. However, treatment benefit could vary according to patients’ characteristics. Identifying a subpopulation of pa- tients that could most benefit from a treatment is interesting in order to target that treatment to those patients. The current work aims to define positive and negative treatment respond- ers according to the treatment effect on occurrence of recurrent events. The proposed method supposes that patients are followed-up before and after treatment initiation. We propose to use a Cox proportional hazards mixed model with the treatment as a time-dependent covariate. Gaussian frailty terms are associated to the intercept and the treatment effect. Then the random components of each patient are predicted and put in entry of non-supervised clustering algorithms in order to build clusters of patients. A simulation study is performed to examine the performance of the pro- posed methodology. Several non-supervised algorithms combined with several distances are envisaged. Methods are evaluated on the percentage of correctly classified patients and on the estimation of treatment effect in each cluster. Finding cluster of patients with a potential benefit from a treatment is es- sential in order to find biomarkers of a treatment efficacy.   C27.5 A total time approach for the simulation of recurrent event data when planning a clinical trial A Jahn-Eimermacher1 , K Ingel1 , A-K Ozga2 , H Binder1 1 University Medical Center, Mainz, Germany, 2 University of Applied Sciences, Koblenz, Germany   Simulation techniques are an important tool for planning clinical trials with complex time-to-event structure. We specifically consider simulation of recurrent events, which can follow two different general types of event- generating processes. In a gap time approach the hazard for experiencing events is defined on the time since last event, and simulation of the iden- tically distributed inter-event-times is straightforward. However, in many clinical settings the hazard depends on the time since study start, e.g. the time since beginning treatment of some progressive disease. This calls for a total time approach. Accordingly, we propose a method for simulating recurrent event data for an arbitrary hazard function defined on total time. We identify the distri- butions of inter-event-times conditional on the times of previous events

Pages Overview