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

ISCB2014_abstract_book

42 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsTuesday, 26th August 2014 • 9:00-10:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C20.3 Missing data in individual patient data meta-analysis M Quartagno1 , J Carpenter2 1 London School of Hygiene and Tropical Medicine, London, United Kingdom, 2 London School of Hygiene & Tropical Medicine, London, United Kingdom Aim: Individual Patient Data (IPD) meta-analyses rely not only on results from different studies, but aim to directly combine individual data from each study, in order to better account for the statistical heterogeneity be- tween studies. Missing Data are common in clinical datasets and therefore also in meta- analyses of such datasets. This can take two forms: missing data within studies and important vari- ables not being available from some studies.We aim to define good strate- gies to handle issues raised by the presence of missing data in IPD meta- analyses. Methods: Multiple Imputation (MI) provides a natural approach to miss- ing data in this context. There are, broadly, two approaches to MI: Full Conditional Specification (FCS) and Joint Modelling (JM).We argue that JM provides a more natural approach for multilevel MI, and thus for IPD meta- analysis. However REALCOM, the only current software available for JM-MI, is too inefficient to handle the size and complexity of IPD Meta-Analysis datasets. We then present a new software that efficiently programmes JM and demonstrate its feasibility vis-a-vis FCS. Results: preliminary analyses show how our software is computationally feasible and competitive with FCS. We are going to investigate the main characteristics of JM imputation in IPD meta-analyses through simula- tions. We will then analyse real datasets, exploring advantages and issues raised by the use of JM in this setting. This includes comparing within-study im- putation with stratified imputation and finding if it is better to combine results from the analysis of different imputed datasets before the meta- analysis or vice versa.   C20.4 Multiple imputation of systematically missing predictors in an individual participant data meta-analysis: a generalized approach using MICE TPA Debray1 , S Jolani2 , H Koffijberg1 , S van Buuren3 , KGM Moons1 1 UMC Utrecht, Utrecht, The Netherlands, 2 Utrecht University, Utrecht, The Netherlands, 3 TNO Quality of Life, Leiden, The Netherlands Background: Individual participant data meta-analyses (IPD-MA) are in- creasingly used for developing and validating multivariable (diagnostic or prognostic) risk prediction models. Unfortunately, some predictors or even outcomes may not have been measured in each study and are thus systematically missing in the IPD-MA. As a consequence, it is no longer possible to evaluate between-study heterogeneity and to estimate study- specific predictor effects, which severely hampers the development and/ or validation of novel prediction models. Methods: Here, we describe a novel approach for imputing systemati- cally missing data and adopt a generalized linear mixed model to allow for between-study heterogeneity. This approach can be viewed as an exten- sion of Resche-Rigon´s method (Stat Med 2012), but relaxes assumptions regarding variance components and allows imputation of linear (e.g. con- tinuous) and non-linear (e.g. categorical) predictors. Results: We illustrate our approach in a case study with the IPD from 13 studies for predicting the presence of Deep Venous Thrombosis. We com- pare the results after applying various imputation methods, and make rec- ommendations about their implementation. Conclusions: Our approach improves the estimation of predictor effects and between-study heterogeneity, thereby facilitating the development and validation of novel prediction models from an IPD-MA. C20.5 Analysis of repeated ordinal measurements and trial planning in a rare neurological disorder L Tanadini1,2 , J Steeves3 , A Curt2 , T Hothorn1 1 Division of Biostatistics, IFSPM, University of Zurich, Zurich, Switzerland, 2 Spinal Cord Injury Center, Balgrist University Hospital, Zurich, Switzerland, 3 ICORD, University of British Columbia, Vancouver, Canada Despite three decades of methodological developments in the field of ordinal data analysis, clinical studies confronted with ordinal outcomes often resort to statistical methods that assume more refined measure- ment scales. The field of neurology, where ordinal scales are ubiquitous, provides concrete examples of this practice. In spinal cord injury, the neurological status of a patient is assessed through a sequential testing of multiple key muscles, each one being graded on a six-point ordinal scale. The current approach of analyzing the total sum of all key muscles scores is questionable on a number of levels. We provide a statistical framework for the analysis of this type of neuro- logical assessments that takes into account the ordinal nature of each muscle score. We show that a proportional odds model conditioning on the interaction between baseline level of lesion and distance from lesion provides an adequate description of the motor scores distribution at six months after injury. This holds true even when compared to a cumulative link mixed model explicitly incorporating the repeated measurements (multiple muscles) on the same patient. We further investigate the struc- turing of our initial, simplistic linear predictor via model-based boosting. Additionally, we simulate several clinical trials scenarios to provide bench- mark data for powering future trials. In addition to foster the use of correct methodology in neurology and re- lated disciplines, the proposed analysis framework is likely to be a missing, but essential element in the currently lagging translation of promising preclinical results into clinical therapies for humans.   C21 Survival analysis and competing risks C21.1 Stagewise pseudo-value regression for time-dependent effects on the cumulative incidence H Binder1 , D Zöller1 , I Schmidtmann1 , A Weinmann2 , T Gerds3 1 University Medical Center Mainz, IMBEI, Mainz, Germany, 2 University Medical Center Mainz, 1. Med, Mainz, Germany, 3 University of Copenhagen, Department of Biostatistics, Copenhagen, Denmark The cumulative incidence describes the absolute risk of an event as a func- tion of time in a competing risks setting. For regression analysis, one can either choose to model all competing events by separate cause-specific hazard models, or directly model the association between covariates and the cumulative incidence of one of the events.With a suitable link function, the direct regression models allow for a straightforward interpretation of covariate effects on the cumulative incidence. In practice where data can be right-censored, they are implemented using a pseudo-value approach. For a grid of time points the possibly unobserved binary event status is re- placed by a jackknife pseudo-value based on the Aalen-Johansen method. We combine a stagewise regression technique with the pseudo-value ap- proach to provide variable selection while allowing for time-dependent effects. This is implemented by coupling variable selection between the grid times, but determining estimates separately. The effect estimates are regularized to allow for model fitting also with a low to moderate number of observations. The technique is illustrated in an application to clinical

Pages Overview