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

ISCB2014_abstract_book

ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 57Wednesday, 27th August 2014 • 9:00-11:00 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust Wednesday, 27th August 2014 – 9:00-11:00 Invited session S1 STRATOS (Strengthening Thinking about Analyses of Observational Studies) initiative: first results & future steps Organizers: Willi Sauerbrei and Harbajan Chadha‐Boreham S1.1 Setting the stage with initial data analyses M Huebner1 , S Le Cessie2 , W Vach3 , M Blettner4 , D Bodicoat5 1 Michigan State University, East Lansing, United States, 2 Leiden University Medical Center, Leiden, The Netherlands, 3 Universitaet Freiburg, Freiburg, Germany, 4 Universitaet Mainz, Mainz, Germany, 5 University of Leicester, Leicester, United Kingdom   The importance of initial data analyses in observational studies has been recognized but is often neglected. Careful data preparation and descrip- tion is crucial before embarking on complex analyses to avoid spurious results. Obtaining high quality data starts far before data collection and includes a careful database design with variable definitions, within and between variable plausibility checks and date checks. Data should be cleaned systematically and carefully, especially when integrating multiple data sources. Changes to the data such as corrections, transformations, defini- tions of categories, or treatment of missing data need to be integrated in the programming code for reproducibility, rather than changes to the raw dataset. Data cleaning can take as much as 80% effort of the analysis, and the process may need to be automated for large data sets. The inclusion and exclusion criteria in the process of selecting the subset of data to be analyzed in the study should be described with an overview of missing measurements and follow-up data. Numerical and graphical descriptions include table summaries, illustrating correlations and confounding fac- tors, or examining distributions of variables and homogeneity of groups, keeping in mind the objectives of the study. It is especially important that the complete initial data analysis process is transparent and that researchers document all steps for reproducibility. STRATOS topic group 3 aims to provide guidance on this process based on an overview of existing literature with examples and feedback from experienced statisticians.   S1.2 Evaluation of incremental value of a marker: a historic perspective on the Net Reclassification Improvement EW Steyerberg1 , P Macaskill2 , AV Vickers3 1 Erasmus MC, Rotterdam, The Netherlands, 2 University of Sydney, Sydney, Australia, 3 Memorial Sloan Kettering Cancer Center, New York, United States   The net reclassification improvement (NRI) is an increasingly popular mea- sure for evaluating improvements in risk predictions. In a recent review, 67 publications were considered from high-impact general clinical journals that considered the NRI. Incomplete reporting of NRI methods, incorrect calculation, and common misinterpretations were found. To aid improved applications of the NRI, the article elaborated on aspects of the computa- tion and interpretation in various settings [1]. An accompanying Editorial emphasized conceptual problems (such as weighting reclassification inappropriately), and statistical problems (such as artificial inflation of NRI values and Type I error). It doubted that any- thing would be gained by reporting the NRI, either overall or in its compo- nents. The Editorial suggested that investigators should move away from statistical abstractions, such as the NRI, and illustrate the consequences of using a marker or model in straightforward clinical terms [2]. In this presentation, we will discuss the use of NRI and related perfor- mance measures from a historical perspective and suggest directions for improvement, particularly with respect to the use of decision analytic measures. References: 1. Leening MJ, Vedder MM, Witteman JC, Pencina MJ, Steyerberg EW. Net reclassification improvement: computation, interpretation, and contro- versies: a literature review and clinician’s guide. Ann Intern Med. 2014 Jan 21;160(2):122-31. 2. Vickers AJ, Pepe M. Does the net reclassification improvement help us evaluate models and markers? Ann Intern Med. 2014 Jan 21;160(2):136-7. S1.3 Review of methods used in recent observational epidemiological studies to select variables and their functional forms [STRATOS Task Group 2] M Abrahamowicz1 , RP Kyle1 1 McGill University, Montreal, Canada   The over-arching STRATOS goal is to improve statistical methodology used in real-life observational studies, focusing on selected ‘generic’ is- sues. Task Group 2 deals with selection of independent variables, and functional forms for continuous variables, in multivariable explanatory models. To convince end-users to adapt more sophisticated statistical methods, and demonstrate the weaknesses of‘conventional’, currently ap- plied methods, we reviewed methods in 50 papers published in 2013 in high-ranking epidemiology and clinical journals, which focused on some continuous variable(s). Whereas several studies selected independent variables a priori, on sub- stantive grounds, many used arbitrarily selected data-dependent criteria or procedures, and failed to account for, or even mention, their impact on the estimation and inference. As expected, most studies imposed a priori linearity of the effects of continuous variable(s), and did not test or evalu- ate this assumption. Review of additional papers revealed that linearity was often imposed even for associations consistently demonstrated to be non-linear, in previous flexible analyses (e.g. BMI versus mortality; age at diagnosis versus recurrence or death in various cancers). Studies that did consider possibly non-linear relationships employed different flexible methods (polynomials, fractional polynomials, and various spline-based approaches), and used different criteria to assess (non-)linearity. In conclusion, current applied research will benefit from evidence-based guidance, and a systematic comparison of methods, for selection of vari- ables and their functional forms.   S1.4 Causal questions and principled answers: a guide through the landscape for practicing statisticians E Goetghebeur1 , E Moodie2 , I Waernbaum3 , S Le Cessie4 1 Ghent University, Gent, Belgium, 2 Mc Gill University, Montreal, Canada, 3 Umea University, Umea, Sweden, 4 Leiden University Medical Center, Leiden, The Netherlands   Causal inference came a long way over the past decade. New method- ological approaches casting assumptions, models and results in terms of potential outcomes find their way into the clinical literature. This opens great potential for deepened understanding, but also for misunderstand- ings if subtle assumptions or interpretation of results are misunderstood. When choosing the causal analysis method or synthesizing evidence from different approaches one should be clear about the specific questions they aim to answer and about what can (not) be achieved from available

Pages Overview