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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 43Tuesday, 26th August 2014 • 9:00-10:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust cancer registry data from hepatocellular carcinoma patients. The results are contrasted to traditional hazard-based modeling, in par- ticular highlighting the difference in interpretation. In addition to more straightforward interpretation, identification of time-dependent covariate effect patterns on the cumulative incidence is seen to be feasible with a moderate number of observations, when using the proposed technique. C21.2 Imputing missing covariate values in presence of competing risks M Resche-Rigon1 , IR White2 , S Chevret1 1 INSERM UMR 1153, Université Paris 7, Paris, France, 2 MRC Biostatistics Unit, Institute of Public Health, Cambridge, United Kingdom Due to its flexibility, practicability and efficiency, multiple imputation by chained equations is widely used to impute missing data. To avoid bias in the substantive model, it is well known that the imputation model must include all the variables from the analysis model, including the outcome. In survival analysis, the outcome is defined by an event indicator D and the observed event or censoring time T. In 2009, White and Royston showed that when a Cox model is used for the analysis, the imputation model for each covariate should include the event indicator and the cumula- tive baseline hazard estimated by the Nelson-Aalen estimator (White IR, Royston P, Stat Med. 2009). In the competing risks setting, multiple imputation has been proposed only to impute missing information on the cause of failure, and has mostly been used in analyses of cumulative incidence functions. We extend the work of White and Royston to impute missing covariates in a compet- ing risks setting, where the substantive model is either a cause specific proportional hazards model or a sub-distribution proportional hazards model: we show that the event indicators and cumulative baseline haz- ards of all the competing events should be included in the imputation model. Consequently, even in a standard survival analysis framework, the cumulative baseline hazard of being censored should be included in the imputation model. These approaches will be evaluated by a simulation study, and then ap- plied to a sample of 278 adult patients with acute myeloid leukaemia. C21.3 Evaluation of a peritoneal dialysis program using semiparametric multi-state models in the presence of competing risks L Teixeira1 , C Cadarso-Suarez2 , A Rodrigues3,4 , D Mendonça1,5 1 ICBAS-UP, Porto, Portugal, 2 Department of Statistics and Operations Research, USC, Santiago de Compostela, Spain, 3 CHP-HGSA, Porto, Portugal, 4 UMIB/ICBAS-UP, Porto, Portugal, 5 ISPUP, Porto, Portugal Chronic kidney disease is becoming a major public health problem with a growing number of patients in need of replacement therapy, such as peri- toneal dialysis (PD). As the trajectory of PD patients is complex, character- ized by the presence of several transient and absorbing states, the evalu- ation of such programs may be addressed using a multi-state approach taking competing risks into account. The present study has as main objectives: (i) to discuss the use of flexible regression models like Structured Additive Regression (STAR) models in a multi-state competing risk framework, ex- pressing results of continuous covariates in terms of hazard ratio curves taking a specific covariate value as the reference; and (ii) to adapt the definition of time-dependent ROC curves to a multi-state competing risk framework to assess the predictive accuracy of the STAR model. The methodologies discussed were applied to explore the effects of ma- jor clinical covariates such as age, sex and diabetes in a PD data. These methods revealed to be very relevant for this type of real clinical data as the developed models were an informative tool for the evaluation of the patients and consequently for the medical decision process. The use of STAR models complemented with the use of temporal ROC curves in clini- cal context allowed to identify relevant factors associated with each one of the specific transitions. The identification of these factors, which could not have been obtained with standard survival models, contributes for a better knowledge of patient trajectories resulting in better management of treatment programs.   C21.4 Predicting optimal cumulative doses for breast cancer chemotherapy via competing risks regression models G Cortese1 1 University of Padua, Padua, Italy In breast cancer, the risk of cardiotoxicity due to chemotherapy increases with the cumulative dose of treatment over time.Therefore, it is of interest to estimate an optimal cumulative dosage over time that guarantees a low risk for cardiotoxicity, while controlling the competing risk for mortality by maximizing the antitumor effect. For this purpose, we consider a competing risks regression model with two events, cardiotoxicity and death. The aim is to predict optimal cumu- lative doses along a given treatment time that keep the cumulative risk for cardiotoxicity below a certain threshold (e.g. <5%). Data from breast cancer patients, treated with chemotherapy during follow-up, were analysed with different direct regression models for com- peting risks. The cumulative dose, predetermined according to given time schedules, was included as time-dependent covariate in addition to other risk factors. The cumulative incidence function for cardiotoxicity over a certain time window [s, t], Pc (s,t; X(s)), e.g. a one-year prediction t=s+1, was treated as a function of cumulative dose at time s, X(s). The direct regression models allow finding a one-to-one relationship between Pc (s,t; X(s)) and X(s). Then, the optimal cumulative doses at a se- quence of landmark time points, were found by inverting the one-year prediction of a cardiotoxicity risk equal to 5%. Confidence intervals for the doses were estimated by inverting pointwise confidence intervals for Pc (s,t; X(s)). To control also for increased risk of dying, we finally predict optimal cumu- lative doses by minimizing a combination of the two cumulative risks of cardiotoxicity and death. C21.5 The liability-threshold model for case-control family studies applied to censored time to event data L Cederkvist1,2 , KK Holst1 , T Scheike1 1 Copenhagen University, København K, Denmark, 2 Danish Cancer Society Research Center, København Ø, Denmark In case-control family studies, familial aggregation of a disease is investi- gated using families collected via case or control probands, who are cho- sen based on their disease status. By comparing the correlation for differ- ent family members, the presence and magnitude of familial aggregation of the disease can be assessed. A model which is often used is the liability-threshold model where the dis- ease status outcome is defined from a normal distributed latent variable, the so-called liability. The disease status outcome and the unobserved continuous liability are linked using the Probit function. A threshold on the liability scale determines whether an individual is affected by the dis- ease or not. If an individual´s liability exceeds the threshold he or she is af- fected. The variance of the latent liability can be decomposed into genetic and environmental components; a process that requires specification of

Pages Overview