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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 71Wednesday, 27th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C38.5 Random effect models for quality of life analysis in oncology A Barbieri1,2 , C Mollevi1 , C Lavergne2 1 Institut régional du Cancer de Montpellier (ICM), Montpellier, France, 2 Institut de Mathématiques et de Modélisation de Montpellier, Montpellier, France   In Oncology, the Health-related Quality of Life (QoL) has become an es- sential criterion in clinical trials. However, the longitudinal analysis of this criterion is complex and non-standardized. Indeed, the observations are obtained through self-questionnaires (Patient-Reported Outcomes) and there are both multiple responses, repeated and ordinal ones. From a statistical standpoint, QoL is not directly measurable and is considered as a latent trait which is accessible through responses to items. To evaluate QoL in most cancer clinical trials, the QLQ-C30 questionnaire has been used. Nowadays, the statistical analysis is done on a score from the EORTC recommendations, corresponding to the average of item responses. Longitudinal competing models are exploited such as a linear mixed mod- el (LMM) classically used for score modelling and generalized linear mixed models (GLMM) employed for ordinal categorical data. The latter model family builds on the Item Response Theory (IRT) and allows considering raw data (item responses). Regarding the longitudinal analysis, the IRT models are proposed as an alternative to LMM and extended to take into account the clinical covariates and data characteristics. These presented models were compared through the analysis of a dataset from a clinical trial and then a simulation study was performed. The IRT model for polytomous data is quite complex and fastidious to estimate the regression coefficients and to predict the random effects. Finally, a less complex approach of linearization advanced by Schall in 1991 is proposed to estimate these GLMM in order to complete the simulation study.   C39 Multistate models and competing risks I C39.1 Illness death models and their applications in cancer research M Yu1 1 University of Wisconsin-Madison, Madison, United States   Cancer studies frequently deal with non-terminal (T1) and terminal (T2) events. In many cases, T1 is disease progression and T2 is death. When T2 occurs first, it censors T1, but not vice versa. In practice, instead of con- sidering such bivariate outcomes, a composite outcome known as pro- gression free survival, defined as the minimum time to either of the two events, is frequently used. We illustrate problems with such approach by using the well-known illness-death models. We therefore advocate data analysis using illness- death models to account for possible dependent censoring of T1 by T2 and to improve prediction of T2 using T1. We propose flexible random effects to capture heterogeneous correlation structures that are usually present in real data. Our model also represents a generalization of the popular shared frailty models. We use Bayesian computation for analysis that can utilize existing software packages.The approach is demonstrated on both simulation and breast cancer data sets.   C39.2 Multi-state model for analysis of modified Rankin Scale in acute stroke trials: a new approach with a twist YY Palesch1 , W Zhao1 1 Medical University of South Carolina, Charleston, United States   Ordinal response outcomes in clinical trials are one of the more difficult types of outcome to analyze. In Phase III trials of acute therapy for stroke, modified Rankin Scale (mRS) score at 90 days from randomization is a stan- dard primary outcome measure assessing the subjects’ functionality. The 7-point ordinal scale of mRS ranges from 0 (normal function) to 6 (death). The most common method for analysis is to dichotomize the mRS (0-1 vs 2-6 or 0-2 vs 3-6 as good vs bad outcome) which yields clinically meaning- ful statistics (relative risk/benefit or odds ratio). Dichotomization has been criticized to be inefficient, leading to more recent approaches that use the full ordinal scale, such as proportional odds model and assumption-free Cochran-Mantel-Haenszel test. The former maintains the advantage of yielding a clinically meaningful interpretation using the common OR; however, it is sensitive to the pro- portionality of odds assumption. The latter has limitations in the number covariates and difficulty in clinical interpretation. We propose a multi-state modeling as an alternative approach to analyz- ing the mRS. The method utilizes the full ordinal scale, covariates can be accommodated, and transitional probabilities are clinically interpretable. One caveat is that mRS is unavailable at baseline, and we have explored a surrogate measure (NIH Stroke Scale) that yields good sensitivity and specificity against the mRS. Data from two recently completed large Phase III trials are re-analyzed and compared with the published results.   C39.3 The illness death model under left truncated and right censored data B Vakulenko-Lagun1 , M Mandel1 1 The Hebrew University of Jerusalem, Jerusalem, Israel   Left truncated data arise when a lifetime variable T and an independent truncation variable L are observed only if L

Pages Overview