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


ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 83Wednesday, 27th August 2014 • 16:00-17:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C47.2 Weibull regression for a right-censored endpoint with one censored and an arbitrary number of non-censored covariates K Rufibach1 , S Hubeaux2 1 Biostatistics Oncology, F. Hoffmann-La Roche, Basel, Switzerland, 2 École polytechnique fédérale de Lausanne, Lausanne, Switzerland   Biomarker data is often subject to limits of quantification or limits of detection. Statistically, this corresponds to left- or interval-censoring. In applications, e.g. when a biomarker is a covariate in a regression model, such data is often imputed in some way, e.g. by considering the limit of detection an actual measurement. In order to be able to correctly account for the nature of the data, we have implemented maximum likelihood es- timation in Weibull regression for a right-censored endpoint, one interval- censored, and an arbitrary number of non-censored covariates. We discuss the assumptions made in the model and how to set up the like- lihood function and maximize it. Inference for estimated parameters can be received using standard maximum likelihood theory. We quantify the bias and mean-squared error for parameter estimates compared to com- monly used imputation methods. We illustrate the methodology by applying it to assess Prentice´s criteria for surrogacy in data simulated from a randomized clinical registration trial. The software is available on CRAN, as the package SurvRegCensCov.   C47.3 Accelerated failure time model with interval censored data and cure S Scolas1 , C Legrand1 , A El Ghouch1 1 Université Catholique de Louvain-La-Neuve, Louvain-la-Neuve, Belgium   Mild cognitive impairment (MCI) may be a precursor of Alzheimer disease or other dementia. Studying the time until conversion to MCI makes use of survival analysis theory. Generally, within this field, it is assumed that if the follow-up time is long enough, then the event of interest will be ob- served for each individual. In our case, not everybody will show signs of impairment. We then say that a proportion of the population is“cured”, or “long-term survivor”. Also, patients come to scheduled interviews and thus we can only detect MCI to have appeared between two visits. That is, the database contains interval censored data. Thus, we propose to extend the existing survival models to the case where interval censored data and cure may be present. In this paper, we present the method we want to use: to model event times (i.e. the latency part), we utilize an accelerated failure time (AFT) regres- sion model, adapted to interval censored data, together with an extended generalized gamma (EGG) distribution for the error term of the AFT. In ad- dition, modeling the cure proportion (i.e. the incidence part) is made by a logistic regression. Furthermore we show the good behavior of the method thanks to results of simulations. Then, we address some issues concerning variable selec- tion in such a model and finally, we apply this method to our Alzheimer disease database, which consist in 241 at-risk patients followed-up be- tween 1998 and 2008 with regular checks for the appearance of MCI. C47.4 Semiparametric Bayesian frailty model for clustered interval-censored data A Cetinyurek Yavuz1 , P Lambert1,2 1 Universite de Liege, Liege, Belgium, 2 Universite Catholique de Louvain, Louvain-la-Neuve, Belgium   Recently, there has been an increasing interest in statistical analysis of interval-censored time-to-event data. This type of data is quite usual for clinical trials or longitudinal studies especially in practical settings of AIDS and cancer research where the individuals have pre-scheduled visits but the event of interest occurs between the visits. Then, the event times are not known exactly but rather to lie in an interval of time. Moreover, in clinical trials, the units may be collected in clusters and they share some observed or unobserved characteristics, i.e. patients from mul- tiple centres, teeth of multiple subjects; and hence they tend to be corre- lated. Interval-censored data is a natural generalization of right censored time-to-event data for which a large number of statistical techniques are developed. However, less well developed procedures are available for ana- lysing interval-censored data. Here, we propose a semiparametric Bayesian frailty model for analyzing correlated interval censored data. We discuss parametric specifications for frailty distribution in the analysis of such data. Afterwards we call particu- lar attention to nonparametric specification of the frailty distribution. The results of the simulation study suggest that the proposed approach is robust to misspecification of the frailty distribution. Moreover, the perfor- mance of the proposed methodology is quite good in practical situations where the frailty distribution is multimodal or skewed.The approach is ap- plied to dental data arising from the Signal Tandmobiel Study.   C48 Drug development C48.1 Bayesian response-adaptive design development: practical experiences from the DexFEM trial CJ Weir1,2 , CH Hansen3,4 , P Warner1 , HOD Critchley1 1 University of Edinburgh, Edinburgh, United Kingdom, 2 Edinburgh Health Services Research Unit, Edinburgh, United Kingdom, 3 Mwanza Intervention Trials Unit, Mwanza, Tanzania, United Republic of, 4 London School of Hygiene and Tropical Medicine, London, United Kingdom   Background: Heavy menstrual bleeding (HMB) is common but non-sur- gical treatments are often judged ineffective by women. We developed DexFEM, a UK MRC-funded Bayesian response-adaptive parallel group trial to investigate whether oral dexamethasone reduces HMB and to iden- tify its optimal dose. Methods: We sought a design comparing placebo and several dexameth- asone doses, with randomisation probabilities adapting based on out- come data from patients already randomised to maximise learning about the dose-response. Bayesian Normal Dynamic Linear Modelling flexibly accommodated a range of potential shapes of dose-response curve. Design options considered were: number of doses; proportion assigned to placebo; adaptation criterion; number and timing of adaptations. We as- sessed design performance across plausible scenarios for: dose-response curve shape; treatment effect magnitude; outcome variance; recruitment rate; interaction and heteroscedasticity. A fractional factorial simulation study used SAS for data handling, generating scripts and executing analy- sis inWinBUGS. 200 trials were simulated for each of 150 scenarios. Normal linear modelling estimated the effect of each design option on empirical type I error and statistical power.

Pages Overview